docs: add self-contained reproduction & usage guideline page

Add docs/guideline.html, a single self-contained documentation guide
(left-nav + content + on-this-page TOC) covering installation, data
preparation, training/eval, full configuration reference, framework
internals, and an API reference. Link it from the README with local,
htmlpreview, and GitHub Pages access instructions.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

.env.example

@@ -0,0 +1,29 @@
+# SkillOpt Environment Variables
+# Copy this file to .env and fill in your values.
+# Usage: set -a; source .env; set +a
+
+# ── Azure OpenAI (required for openai_chat backend) ──────────────────
+export AZURE_OPENAI_ENDPOINT=https://your-resource.openai.azure.com/
+export AZURE_OPENAI_API_VERSION=2024-12-01-preview
+# Authentication: choose one method
+#   Option 1: API Key
+export AZURE_OPENAI_API_KEY=
+#   Option 2: Azure CLI (no API key needed, recommended on Azure VMs)
+# export AZURE_OPENAI_AUTH_MODE=azure_cli
+#   Option 3: Managed Identity
+# export AZURE_OPENAI_AUTH_MODE=managed_identity
+# export AZURE_OPENAI_MANAGED_IDENTITY_CLIENT_ID=your-client-id
+
+# ── OpenAI-compatible endpoints ──────────────────────────────────────
+# Set AUTH_MODE to openai_compatible and reuse AZURE_OPENAI_ENDPOINT / _API_KEY.
+# The plain OpenAI client is used; no Azure auth, no api-version header.
+# export AZURE_OPENAI_ENDPOINT=https://api.openai.com/v1
+# export AZURE_OPENAI_API_KEY=sk-...
+# export AZURE_OPENAI_AUTH_MODE=openai_compatible
+
+# ── Anthropic / Claude (for claude_chat backend) ─────────────────────
+# export ANTHROPIC_API_KEY=sk-ant-...
+
+# ── Qwen Local Model (for qwen_chat backend) ────────────────────────
+# export QWEN_CHAT_BASE_URL=http://localhost:8000/v1
+# export QWEN_CHAT_MODEL=Qwen/Qwen3.5-4B

.gitignore

@@ -1,15 +1,42 @@
 __pycache__/
 *.pyc
+*.egg-info/
+build/
+dist/
+site/
+
 data/
 outputs/
 logs/
 external/
+
 /BabyVision/
 /MMRB/
 /SpreadsheetBench/
 /dl4ir-searchQA/
+
 configs/local/
 configs/**/*.local.yaml
 *.local.md
+*.secret.md
+*.bak
+
+.env
 .secrets/
 .codex_azure*/
+
+# Internal docs (not for open-source release)
+docs/ablation_plan.md
+docs/ablation_paper_tables.md
+docs/ablation_paper_tables.html
+docs/experiment_commands.md
+docs/slow_update_flowchart.md
+docs/session_memory.md
+docs/harness_fresh_machine_handoff.md
+docs/harness_monitoring_memory.md
+docs/harness_reproduction_secrets.secret.md
+docs/reflact_conda_env_export.yml
+docs/reflact_overview.html
+docs/render_ablation_paper_tables.py
+docs/让*
+.gradio/

CONTRIBUTING.md

@@ -0,0 +1,43 @@
+# Contributing to SkillOpt
+
+Thank you for your interest in contributing! SkillOpt welcomes contributions of all kinds.
+
+## Getting Started
+
+```bash
+git clone https://github.com/microsoft/SkillOpt.git
+cd SkillOpt
+pip install -e ".[dev]"
+```
+
+## How to Contribute
+
+### 🐛 Bug Reports
+Open a GitHub issue with reproduction steps, expected/actual behavior, and your config file (remove API keys).
+
+### 🔧 Add a Benchmark
+See the [guide](docs/guide/new-benchmark.md) and use the scaffold at `skillopt/envs/_template/`.
+
+### 🤖 Add a Model Backend
+See the [guide](docs/guide/new-backend.md).
+
+### 📝 Improve Documentation
+```bash
+pip install -e ".[docs]"
+mkdocs serve   # Preview at http://localhost:8000
+```
+
+## Pull Request Process
+
+1. Fork the repo and create a feature branch
+2. Make changes and test with an existing benchmark
+3. Submit a PR with a clear description
+4. Ensure CI passes
+
+## Code Style
+- Follow existing patterns in the codebase
+- Use type hints for function signatures
+- Keep docstrings concise
+
+## License
+By contributing, you agree your contributions are licensed under the [MIT License](LICENSE).

LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Microsoft Corporation
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

README.md

@@ -1,557 +1,288 @@
-# ReflACT: Reflective Agent Tuning
+# SkillOpt: Executive Strategy for Self-Evolving Agent Skills
 
-ReflACT is a framework for optimizing an external skill document through iterative rollout, reflection, editing, and gated validation.
+*Train agent skills like you train neural networks — with epochs, (mini-)batchsize, learning rates, and validation gates — but without touching model weights.*
 
-It does **not** fine-tune model weights. Instead, it treats the skill document as the optimization target:
+[![Project Page](https://img.shields.io/badge/Project%20Page-SkillOpt-8dbb3c)](https://microsoft.github.io/SkillOpt/) [![Paper](https://img.shields.io/badge/Paper-arXiv-b31b1b)](https://arxiv.org/abs/2605.23904) [![Project Video](https://img.shields.io/badge/Project%20Video-Watch%20Demo-ff0000)](https://youtu.be/JUBMDTCiM0M) [![Python 3.10+](https://img.shields.io/badge/Python-3.10%2B-blue.svg)](https://www.python.org/) [![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](LICENSE)
 
-- the **student** model executes tasks with the current skill
-- the **teacher** model analyzes trajectories and proposes edits
-- the framework merges, ranks, applies, and validates those edits
-- only validated skill updates are kept
+## 🎬 SkillOpt Demo Video
 
-This branch implements a full training loop with step-level skill optimization and optional epoch-level memory mechanisms (`slow_update`, `meta_skill`, `meta_reflect`).
+https://github.com/user-attachments/assets/eb12d3bc-371c-467f-904d-91b61f339ed7
 
-## Method Overview
+<p align="center">
+  <a href="https://youtu.be/JUBMDTCiM0M"><b>▶ Watch the full demo on YouTube</b></a>
+</p>
 
-### Optimization Target
+---
 
-Each run maintains a mutable markdown skill document. The framework repeatedly improves that document instead of changing model parameters.
+## Documentation
 
-This gives a training-style loop for prompt / policy optimization:
+A complete, self-contained **Documentation & Reproduction Guide** lives at
+[`docs/guideline.html`](docs/guideline.html). It covers installation, data
+preparation, training/eval commands, the full configuration reference, the
+framework internals (training loop, validation gate, slow update, meta skill),
+and an API/function reference — all in a single page with a left navigation
+sidebar.
 
-1. Roll out the current skill on a batch of tasks.
-2. Reflect on failures and successes.
-3. Merge patch proposals into a coherent candidate update.
-4. Rank and select a bounded number of edits.
-5. Apply those edits to produce a candidate skill.
-6. Validate the candidate skill on a held-out selection split.
-7. Keep the update only if the gate accepts it.
+Because GitHub shows raw source for `.html` files instead of rendering them,
+open the guide one of these ways:
 
-### Per-Step Pipeline
+- **Locally** — clone the repo and open `docs/guideline.html` in any browser
+  (no build step required).
+- **Rendered online (no setup)** — via the htmlpreview proxy:
+  [`htmlpreview.github.io/?…/docs/guideline.html`](https://htmlpreview.github.io/?https://github.com/microsoft/SkillOpt/blob/main/docs/guideline.html)
+- **GitHub Pages** — the repository's GitHub Pages site already serves the
+  project homepage from the repo root, so the guide is reachable alongside it at
+  `https://microsoft.github.io/SkillOpt/docs/guideline.html` (the homepage at
+  `https://microsoft.github.io/SkillOpt/` is unaffected).
 
-Every training step executes the following pipeline in `reflact/engine/trainer.py`:
+---
 
-1. **Rollout**
-   The student model runs a batch of tasks using the current skill.
+## Install
 
-2. **Reflect**
-   The teacher analyzes minibatches of trajectories and emits raw patches.
-   Failure-driven and success-driven patches are tracked separately.
-
-3. **Aggregate**
-   Raw patches are merged hierarchically. Metadata such as `support_count` and `source_type` is carried into the merged patch so later ranking can use it.
-
-4. **Select**
-   The teacher ranks the merged edit pool and keeps up to `edit_budget` edits.
-
-5. **Update**
-   The selected edits are applied to the skill document. The framework records an `edit_apply_report.json` so you can see which edits actually landed, which were skipped, and why.
-
-6. **Evaluate / Gate**
-   The candidate skill is evaluated on the selection split. Gate validation is mandatory in this branch. A candidate update is accepted only if it improves over the current selection score; a new global best is tracked separately.
-
-### Within-Epoch Memory
-
-Inside an epoch, the trainer maintains a step buffer containing:
-
-- compact failure-pattern summaries from previous steps
-- rejected edits and their score deltas
-
-That context is fed back into later reflection calls so the teacher can avoid repeating ineffective edits and can focus on unsolved error patterns.
-
-### Epoch-Level Mechanisms
-
-This branch supports three optional epoch-level mechanisms.
-
-#### Slow Update
-
-At the end of each epoch, `slow_update` compares the previous epoch’s terminal skill and current epoch’s terminal skill on a sampled train subset. It then writes longitudinal guidance into a protected slow-update region inside the skill document.
-
-Importantly, this guidance is **not** blindly written through. It is converted into a candidate skill and sent through the same selection gate as step-level updates.
-
-#### Meta Skill
-
-`meta_skill` is teacher-side cross-epoch memory. It does not directly edit the current skill. Instead, it writes a compact memory artifact describing longer-term patterns across adjacent epochs. That memory is loaded into later reflection / merge / ranking calls as extra context.
-
-#### Meta Reflect
-
-`meta_reflect` runs at epoch end over the step history of the current epoch. It looks at accepted and rejected directions from the whole epoch, proposes higher-level patch edits, applies them to a meta candidate, and then sends that candidate through the same selection gate.
-
-## What This Branch Guarantees
-
-The current implementation assumes the following as the mainline method contract:
-
-- gate validation is always on
-- the current skill, current score, best skill, and best score stay aligned
-- `slow_update` is gated before being committed
-- patch provenance (`source_type`, `support_count`) reaches selection
-- patch application is observable through per-edit reports
-- resume state is restored from `runtime_state.json` rather than inferred only from history
-- all benchmark model calls go through the unified backend router
-
-## Model Backends
-
-All model access now goes through the split teacher/student model layer in `reflact.model`.
-
-Supported teacher backends:
-
-- `openai_chat`
-- `claude_chat`
-
-Supported student backends:
-
-- `openai_chat`
-- `claude_chat`
-- `codex_exec`
-- `claude_code_exec`
-
-Recommended config shape:
-
-```yaml
-model:
-  teacher_backend: openai_chat
-  student_backend: codex_exec
-  teacher: gpt-5.4
-  student: gpt-5.4-codex
-  reasoning_effort: medium
-```
-
-Legacy `model.backend` and CLI flags like `--backend codex` still work. They are mapped onto the split backend model for backward compatibility.
-
-The same routing is used by:
-
-- training (`scripts/train.py`)
-- eval-only runs (`scripts/eval_only.py`)
-- SpreadsheetBench standalone prompt eval scripts
-- LiveMathematicianBench baseline eval script
-- benchmark rollout code inside the main framework
-
-### Azure OpenAI
-
-If you use `openai_chat`, configure either environment variables or config values:
+**Requirements:** Python 3.10+
 
 ```bash
-export AZURE_OPENAI_ENDPOINT="https://your-endpoint.openai.azure.com/"
-export AZURE_OPENAI_API_KEY="your-api-key"
-export AZURE_OPENAI_API_VERSION="2025-04-01-preview"
+git clone https://github.com/microsoft/SkillOpt.git
+cd SkillOpt
+pip install -e .
+
+# For ALFWorld benchmark (optional):
+pip install -e ".[alfworld]"
+alfworld-download
 ```
 
-The config supports both the old keys and the new explicit names:
+### Configure API Credentials
 
-```yaml
-model:
-  azure_openai_endpoint: "..."
-  azure_openai_api_version: "..."
-  azure_openai_api_key: ""
-  azure_openai_auth_mode: api_key
-  azure_openai_ad_scope: "https://cognitiveservices.azure.com/.default"
-  azure_openai_managed_identity_client_id: ""
+```bash
+cp .env.example .env
+# Edit .env with your API credentials, then:
+source .env
 ```
 
-`azure_openai_auth_mode` can be used for API-key auth or Azure AD / managed identity flows.
-
-### Exec Harness
-
-`codex_exec` and `claude_code_exec` run the student inside a workspace harness instead of a plain chat call. The harness writes task files, renders a dynamic `SKILL.md`, runs the student CLI, and saves raw execution artifacts such as:
-
-- `codex_raw.txt`
-- `codex_trace_summary.txt`
-- workspace-local task / skill files
-
-This branch keeps `meta_skill` and `apply_patch_with_report`, while upgrading the student path to the more realistic workspace-exec setup.
-
-### Trace-Aware Deep Reflect
-
-When `student_backend=codex_exec` and `gradient.use_deep_reflect=true`, deep reflection can probe a specific earlier Codex attempt:
-
-- the teacher sees a compact Codex trace summary
-- deep probe can target `probe_target_id`
-- the follow-up rollout can resume from `probe_after_step`
-
-This is wired for the dataset-backed environments in this branch.
-
-### Rewrite Mode
-
-Skill updates support two modes:
-
-- `optimizer.skill_update_mode=patch`
-- `optimizer.skill_update_mode=rewrite_from_suggestions`
-
-`patch` keeps the existing fine-grained edit application path and still records `edit_apply_report.json`.
-
-`rewrite_from_suggestions` asks the teacher to emit higher-level rewrite suggestions, then rewrites the whole skill in one pass. This is useful when patch edits become too fragmented.
-
-## Repository Layout
-
-```text
-reflact/
-  engine/
-    trainer.py                 main training loop
-  gradient/
-    reflect.py                 minibatch reflection
-    aggregate.py               hierarchical patch merge
-    deep_probe.py              diagnostic probing for deep reflect
-  optimizer/
-    clip.py                    edit ranking / selection
-    skill.py                   patch application + apply report
-    slow_update.py             epoch-level longitudinal guidance
-    meta_skill.py              teacher-side cross-epoch memory
-    meta_reflect.py            epoch-level macro editing
-  evaluation/
-    gate.py                    pure gate decision logic
-  model/
-    backend_config.py          teacher/student backend routing
-    azure_openai.py            Azure backend
-    codex_harness.py           workspace exec harness + Codex trace parsing
-    claude_backend.py          Claude backend
-  envs/
-    ...                        environment adapters and rollout logic
-scripts/
-  train.py                     unified training entry
-  eval_only.py                 evaluate one skill without training
-configs/
-  _base_/default.yaml          shared defaults
-  <env>/default.yaml           environment-specific configs
+**Azure OpenAI** (recommended):
+```bash
+export AZURE_OPENAI_ENDPOINT="https://your-resource.openai.azure.com/"
+# Option 1: API key auth
+export AZURE_OPENAI_API_KEY="your-key"
+# Option 2: Azure CLI auth (no API key needed)
+export AZURE_OPENAI_AUTH_MODE="azure_cli"
 ```
 
-## Configuration
+> **Note:** `AZURE_OPENAI_ENDPOINT` is required for all three modes (`api_key`, `azure_cli`,
+> `openai_compatible`). Without it, all LLM calls will fail.
 
-Configs use structured YAML with `_base_` inheritance.
-
-The base config is `configs/_base_/default.yaml`. Key defaults in this branch are:
-
-- `model.teacher_backend = openai_chat`
-- `model.student_backend = openai_chat`
-- `model.reasoning_effort = medium`
-- `optimizer.use_slow_update = true`
-- `optimizer.use_meta_skill = true`
-- `optimizer.use_meta_reflect = false`
-- `gradient.use_deep_reflect = false`
-- `optimizer.skill_update_mode = patch`
-
-Default setting snapshot:
-
-```yaml
-model:
-  backend: azure_openai
-  teacher: gpt-5.4
-  student: gpt-5.4
-  teacher_backend: openai_chat
-  student_backend: openai_chat
-  reasoning_effort: medium
-  rewrite_reasoning_effort: ""
-  rewrite_max_completion_tokens: 64000
-  codex_exec_path: codex
-  codex_exec_sandbox: workspace-write
-  codex_exec_profile: ""
-  codex_exec_full_auto: false
-  codex_exec_reasoning_effort: none
-  claude_code_exec_path: claude
-  claude_code_exec_profile: ""
-  codex_trace_to_teacher: true
-
-train:
-  num_epochs: 4
-  train_size: 0
-  batch_size: 80
-  accumulation: 1
-  seed: 42
-
-gradient:
-  minibatch_size: 16
-  merge_batch_size: 16
-  analyst_workers: 16
-  max_analyst_rounds: 3
-  failure_only: false
-  use_deep_reflect: false
-  deep_reflect_failures: 4
-  deep_reflect_successes: 2
-
-optimizer:
-  learning_rate: 8
-  min_learning_rate: 2
-  lr_scheduler: cosine
-  skill_update_mode: patch
-  use_meta_reflect: false
-  meta_learning_rate: 8
-  use_slow_update: true
-  slow_update_samples: 20
-  use_meta_skill: true
-
-evaluation:
-  use_gate: true
-  sel_env_num: 0
-  test_env_num: 0
-  eval_test: true
-
-env:
-  split_mode: ratio
-  split_ratio: "2:1:7"
-  split_seed: 42
+**OpenAI-compatible endpoints**:
+```bash
+export AZURE_OPENAI_ENDPOINT="https://api.openai.com/v1"
+export AZURE_OPENAI_API_KEY="sk-..."
+export AZURE_OPENAI_AUTH_MODE="openai_compatible"
 ```
 
-For the full source of truth, see [configs/_base_/default.yaml](/home/azureuser/workspace-yqh/skillopt_final/configs/_base_/default.yaml).
+This routes all calls through the plain OpenAI Python client (no Azure auth, no `api-version`
+header).
 
-Selected fields:
+> **Note:** SkillOpt reuses the `AZURE_OPENAI_*` env var names even in this mode — there is no
+> separate `OPENAI_API_KEY` knob.
 
-| Section | Key | Meaning |
+**Anthropic Claude**:
+```bash
+export ANTHROPIC_API_KEY="sk-ant-..."
+```
+
+**Qwen (local vLLM)**:
+```bash
+export QWEN_CHAT_BASE_URL="http://localhost:8000/v1"
+export QWEN_CHAT_MODEL="Qwen/Qwen3.5-4B"
+```
+
+---
+
+## Data Preparation
+
+SkillOpt expects data in a **split directory** with `train/`, `val/`, `test/` subdirectories, each containing a JSON file (e.g., `items.json`).
+
+```
+data/my_split/
+├── train/items.json
+├── val/items.json
+└── test/items.json
+```
+
+Each JSON file is an array of task items. The required fields depend on the benchmark. For example, SearchQA items look like:
+
+```json
+[
+  {
+    "id": "unique_item_id",
+    "question": "Who wrote the novel ...",
+    "context": "[DOC] relevant passage text ...",
+    "answers": ["expected answer"]
+  }
+]
+```
+
+See `skillopt/envs/<benchmark>/dataloader.py` for the exact format each benchmark expects.
+
+> **Note:** Benchmark datasets are not included in this repository. Prepare your own data following the format above.
+
+### Supported Benchmarks
+
+| Benchmark | Type | Config |
 |---|---|---|
-| `model` | `teacher_backend` | teacher backend: `openai_chat` or `claude_chat` |
-| `model` | `student_backend` | student backend: chat backend or exec backend |
-| `model` | `teacher` | teacher model / deployment |
-| `model` | `student` | student model / deployment |
-| `model` | `reasoning_effort` | reasoning budget passed to the backend when supported |
-| `model` | `codex_trace_to_teacher` | include Codex trace summaries in teacher reflection context |
-| `train` | `num_epochs` | number of epochs |
-| `train` | `train_size` | expected train split size, or `0` to infer |
-| `train` | `batch_size` | tasks per rollout batch |
-| `train` | `accumulation` | number of rollout/reflect minibatches per step |
-| `gradient` | `minibatch_size` | trajectories per analyst minibatch |
-| `gradient` | `merge_batch_size` | patches per aggregate batch |
-| `gradient` | `use_deep_reflect` | enable diagnostic probe rollouts |
-| `gradient` | `max_analyst_rounds` | teacher reflection retries / refinement budget |
-| `optimizer` | `learning_rate` | max edits kept after selection |
-| `optimizer` | `lr_scheduler` | edit-budget scheduler |
-| `optimizer` | `use_slow_update` | epoch-level longitudinal guidance |
-| `optimizer` | `use_meta_skill` | teacher-side epoch memory |
-| `optimizer` | `use_meta_reflect` | epoch-level macro editing |
-| `optimizer` | `skill_update_mode` | `patch` or `rewrite_from_suggestions` |
-| `evaluation` | `sel_env_num` | selection set size (`0` means full split) |
-| `evaluation` | `test_env_num` | test set size (`0` means full split) |
+| SearchQA | QA | `configs/searchqa/default.yaml` |
+| ALFWorld | Embodied agent | `configs/alfworld/default.yaml` |
+| DocVQA | Document QA | `configs/docvqa/default.yaml` |
+| LiveMathematicianBench | Math | `configs/livemathematicianbench/default.yaml` |
+| SpreadsheetBench | Code generation | `configs/spreadsheetbench/default.yaml` |
+| OfficeQA | Tool-augmented QA | `configs/officeqa/default.yaml` |
 
-### Important Branch Rule
+---
 
-`use_gate=false` is intentionally not supported in this branch. Gate validation is part of the method contract here.
+## Quick Start
 
-If an old config still contains `evaluation.use_gate: false`, the loader / trainer will raise instead of silently continuing.
-
-## Supported Environments
-
-The main training entry and eval-only entry now register 11 environments:
-
-| Env | Default rollout shape | Current default split / data setting | Branch alignment |
-|---|---|---|---|
-| `alfworld` | environment-backed episodic rollout | native ALFWorld train/eval splits | in `reflact_new_zzw` |
-| `babyvision` | single-round multimodal QA | `split_mode=ratio` from raw metadata/images, or prepared `split_dir` | in `reflact_new_zzw` |
-| `docvqa` | single-round multimodal QA | `split_dir: data/docvqa_split` | in `reflact_new_zzw` |
-| `livemathematicianbench` | single-round QA | `split_mode=ratio` or prepared `split_dir` | in `reflact_new_zzw` |
-| `mathverse` | single-round multimodal math QA | `data_root: data/MathVerse`, split files loaded from `split_dir` when provided | in `reflact_new_zzw` |
-| `mmrb` | single-round multimodal reasoning QA | `split_mode=ratio` or prepared `split_dir` | in `reflact_new_zzw` |
-| `officeqa` | multi-turn tool loop | `split_dir: data/officeqa_split` plus `data_dirs: [data/officeqa_docs_official]` | in `reflact_new_zzw` |
-| `sealqa` | multi-turn tool loop | `split_dir: data/sealqa_split` | in `reflact_new_zzw` |
-| `searchqa` | single-round QA (`max_turns=1`) | `split_dir: data/searchqa_split` | in `reflact_new_zzw` |
-| `spreadsheetbench` | codegen loop, default `mode=multi`, `max_turns=30` | `split_dir: data/spreadsheetbench_split`, `data_root: data/spreadsheetbench_verified_400` | in `reflact_new_zzw`, default adjusted here to multi-round |
-| `swebench` | mini-swe-agent multi-step bug-fixing rollout | `split_mode=ratio`, `dataset_name=lite`, repo-stratified `2:1:7` split materialized under `out_root/_generated_splits/...` unless `split_dir` is provided | added here, aligned to `swe-bench-old` |
-
-## Data Expectations
-
-The standard two-mode dataset entry path is:
-
-- `split_mode: ratio`
-  - load raw data from `env.data_path`
-  - build a deterministic `train/`, `val/`, `test/` split under `env.split_output_dir` (or under `out_root/_generated_splits/` if unset)
-  - default ratio is explicitly `2:1:7`
-- `split_mode: split_dir`
-  - load an existing `env.split_dir` with `train/`, `val/`, `test/` subdirectories
-
-This currently applies to:
-
-- `searchqa`
-- `spreadsheetbench`
-- `babyvision`
-- `livemathematicianbench`
-- `mmrb`
-- `swebench`
-
-`ALFWorld` is the exception: it is environment-backed rather than JSON split-backed.
-
-The following environments currently expect prepared split directories or extra rooted assets rather than the generic ratio-split path:
-
-- `docvqa`
-- `mathverse`
-- `officeqa`
-- `sealqa`
-
-At a high level:
-
-- `SearchQA`: raw QA json / jsonl or pre-split QA json files
-- `SpreadsheetBench`: raw task manifest json plus spreadsheet task directory, or a pre-split task manifest
-- `ALFWorld`: installed game environment and configured eval/train splits
-- `BabyVision`: raw `meta_data.jsonl` plus images, or a pre-split directory
-- `DocVQA`: pre-split CSV / JSON data under `split_dir`
-- `LiveMathematicianBench`: raw monthly QA json files, or a pre-split directory
-- `MathVerse`: split files plus `data_root` image assets
-- `MMRB`: raw extracted dataset json files, or a pre-split directory
-- `OfficeQA`: pre-split metadata plus resolved office document directories
-- `SealQA`: pre-split metadata for tool-augmented QA tasks
-- `SWEBench`: HuggingFace SWE-bench dataset alias (`lite` / `verified` / `full`) or a prepared split directory
-
-### Split References Across Branches
-
-The split-related defaults are not identical across `skillopt-final`, `reflact_new_zzw`, `gepa`, and `swe-bench-old`. The practical reference points are:
-
-| Source branch | Explicit split settings / dirs |
-|---|---|
-| `skillopt-final` | `searchqa -> data/searchqa_split`; `spreadsheetbench -> data/spreadsheetbench_split`; `docvqa -> data/docvqa_split`; `officeqa -> data/officeqa_split`; `sealqa -> data/sealqa_split`; `swebench -> ratio split 2:1:7 over the default lite dataset, materialized under out_root/_generated_splits/...` |
-| `reflact_new_zzw` | Same 10-benchmark env set as above except no `swebench`; explicit split dirs are `data/searchqa_split`, `data/spreadsheetbench_split`, `data/docvqa_split`, `data/officeqa_split`, `data/sealqa_split`; `spreadsheetbench` there defaults to `mode=single`; `officeqa` uses `max_tool_turns=24`; `sealqa` uses `max_tool_turns=12` |
-| `gepa` | `configs/spreadsheetbench.yaml` uses `data.splits_dir = data/spreadsheetbench/splits`, `eval.mode = react`, `eval.max_turns = 20`; `configs/swebench.yaml` uses `dataset = SWE-bench/SWE-bench_Verified` with `train_size = 100`, `val_size = 50`, `test_size = 350` |
-| `swe-bench-old` | Repo-stratified `2:1:7` split over `SWE-Bench_Lite`, persisted as `outputs/.../split/train.json`, `selection.json`, `test.json`; the example split in that branch is `train=60`, `selection=33`, `test=207` |
-
-For the 10 benches shared with `reflact_new_zzw`, the current branch is now aligned on env coverage. The main intentional delta is `spreadsheetbench`: this branch defaults to multi-round codegen, while `reflact_new_zzw` kept `mode=single` by default.
-
-## Running Training
-
-Example:
-
-```bash
-python scripts/train.py --config configs/searchqa/default.yaml
-```
-
-Explicit 2:1:7 split from raw data:
+### Training
 
 ```bash
+# Minimal example — train on SearchQA:
 python scripts/train.py \
-  --config configs/searchqa/default.yaml \
-  --split_mode ratio \
-  --data_path /path/to/searchqa_train_2000.json
-```
+    --config configs/searchqa/default.yaml \
+    --split_dir /path/to/your/searchqa_split \
+    --azure_openai_endpoint https://your-resource.openai.azure.com/ \
+    --optimizer_model gpt-5.5 \
+    --target_model gpt-5.5
 
-Directly consume a prepared split directory:
-
-```bash
+# Train on LiveMathematicianBench:
 python scripts/train.py \
-  --config configs/searchqa/default.yaml \
-  --split_mode split_dir \
-  --split_dir /path/to/searchqa_split
-```
+    --config configs/livemathematicianbench/default.yaml \
+    --split_dir /path/to/your/livemath_split \
+    --azure_openai_endpoint https://your-resource.openai.azure.com/ \
+    --optimizer_model gpt-5.5 \
+    --target_model gpt-5.5
 
-You can override structured config keys from the CLI:
-
-```bash
+# Train on ALFWorld:
 python scripts/train.py \
-  --config configs/spreadsheetbench/default.yaml \
-  --cfg-options model.teacher_backend=openai_chat model.student_backend=codex_exec train.batch_size=40 optimizer.learning_rate=4
+    --config configs/alfworld/default.yaml \
+    --split_dir /path/to/your/alfworld_split \
+    --azure_openai_endpoint https://your-resource.openai.azure.com/ \
+    --optimizer_model gpt-5.5 \
+    --target_model gpt-5.5
 ```
 
-Legacy flat overrides still work for common keys:
-
-```bash
-python scripts/train.py \
-  --config configs/searchqa/default.yaml \
-  --backend azure_openai \
-  --teacher_model gpt-5.4 \
-  --student_model gpt-5.4 \
-  --reasoning_effort medium
-```
-
-Exec harness example:
-
-```bash
-python scripts/train.py \
-  --config configs/searchqa/default.yaml \
-  --teacher_backend openai_chat \
-  --student_backend codex_exec \
-  --teacher_model gpt-5.4 \
-  --student_model gpt-5.4-codex \
-  --use_deep_reflect true \
-  --skill_update_mode rewrite_from_suggestions
-```
-
-SWEBench example:
-
-```bash
-python scripts/train.py \
-  --config configs/swebench/default.yaml \
-  --cfg-options env.dataset_name=lite env.split_ratio=2:1:7
-```
-
-## Eval-Only and Standalone Evaluation
-
-Evaluate a specific skill without training:
+Key CLI arguments:
+
+| Argument | Description | Example |
+|---|---|---|
+| `--config` | Benchmark config YAML | `configs/searchqa/default.yaml` |
+| `--split_dir` | Path to data split directory | `/path/to/split` |
+| `--azure_openai_endpoint` | Azure OpenAI endpoint URL | `https://your-resource.openai.azure.com/` |
+| `--optimizer_model` | Optimizer model deployment name | `gpt-5.5` |
+| `--target_model` | Target model deployment name | `gpt-5.5` |
+| `--num_epochs` | Number of training epochs | `4` |
+| `--batch_size` | Batch size per step | `40` |
+| `--workers` | Parallel rollout workers | `8` |
+| `--out_root` | Output directory | `outputs/my_run` |
+
+### Eval Only
+
+Evaluate a trained skill on specific data splits without training:
 
 ```bash
+# Evaluate the packaged GPT-5.5 SearchQA skill on the test split:
 python scripts/eval_only.py \
   --config configs/searchqa/default.yaml \
-  --skill reflact/envs/searchqa/skills/initial.md
+  --skill ckpt/searchqa/gpt5.5_skill.md \
+  --split valid_unseen \
+  --split_dir /path/to/searchqa_split \
+  --azure_openai_endpoint https://your-resource.openai.azure.com/
+
+# Evaluate on all splits (train + val + test):
+python scripts/eval_only.py \
+  --config configs/searchqa/default.yaml \
+  --skill ckpt/searchqa/gpt5.5_skill.md \
+  --split all \
+  --split_dir /path/to/searchqa_split \
+  --azure_openai_endpoint https://your-resource.openai.azure.com/
 ```
 
-The same dataset entry modes apply in eval-only runs:
+To evaluate a skill produced by a training run, replace `--skill` with that
+run's best-skill path, for example `outputs/my_run/best_skill.md`.
 
-- `--split_mode ratio --data_path ...`
-- `--split_mode split_dir --split_dir ...`
+| Split | Description |
+|---|---|
+| `valid_unseen` | Test set |
+| `valid_seen` | Validation set |
+| `train` | Training set |
+| `all` | All splits combined (default) |
 
-Standalone scripts also exist for benchmark-specific comparisons, including:
+### Output Structure
 
-- `scripts/eval_prompt_custom.py`
-- `scripts/eval_prompt_official.py`
-- `scripts/eval_livemathematicianbench_baseline.py`
+Each run writes to a structured output directory:
 
-These scripts now also support backend selection through the unified model layer.
+```
+outputs/<run_name>/
+├── config.json              # Flattened runtime config
+├── history.json             # Per-step training history
+├── runtime_state.json       # Resume checkpoint
+├── best_skill.md            # Best validated skill document
+├── skills/skill_vXXXX.md   # Skill snapshot per step
+├── steps/step_XXXX/        # Per-step artifacts (patches, evals)
+├── slow_update/epoch_XX/   # Slow update logs
+└── meta_skill/epoch_XX/    # Meta skill logs
+```
 
-## Output Structure
+Re-running the same command auto-resumes from the last completed step.
 
-Each run writes a structured output directory under `out_root`.
+---
 
-Important top-level artifacts:
+## Community-contributed configs
 
-- `config.json` — flattened runtime config
-- `history.json` — per-step history records
-- `runtime_state.json` — resume state for current/best skill tracking
-- `best_skill.md` — current best validated skill
-- `skills/skill_vXXXX.md` — persisted skill snapshot per step
+These are **not** default SkillOpt settings — they are reference configs
+contributed by users for specific scenarios. The paper-reported numbers
+were obtained with the default settings, not these.
 
-Per-step artifacts live under `steps/step_XXXX/`, including:
+- **`configs/examples/soft_gate.yaml`** *(PR #25, contributed by
+  [@lvbaocheng](https://github.com/lvbaocheng))* — switches the
+  validation gate from exact-match (`hard`) to soft / partial-credit
+  (`soft` or `mixed`). Useful when the held-out **selection split is
+  small** (e.g. ≤ ~10 items) and the **reward is continuous**, where the
+  discrete hard gate often rejects every candidate and training stalls.
+  See the comment at the top of the file for details and when not to use
+  it.
 
-- `merged_patch.json`
-- `ranked_edits.json`
-- `candidate_skill.md`
-- `edit_apply_report.json`
-- `rewrite_result.json` when rewrite mode is enabled
-- `selection_eval/`
-- `trajectory_digest.json`
-- rollout and patch subdirectories
+---
 
-Epoch-level artifacts live under:
+## WebUI
 
-- `slow_update/epoch_XX/`
-- `meta_skill/epoch_XX/`
-- `meta_reflect/epoch_XX/`
-
-## Resume Behavior
-
-The trainer resumes from `runtime_state.json` when present. That state tracks:
-
-- last completed step
-- current skill path
-- current score
-- best skill path
-- best score
-- origin tags for current and best skill
-
-This is important because skill state can change at both step level and epoch level; resuming only from `history.json` is not sufficient for this branch’s method logic.
-
-## Notes
-
-- This repository focuses on skill optimization logic; datasets are not included.
-- Patch application is intentionally observable. Inspect `edit_apply_report.json` when candidate skills do not behave as expected.
-- `SpreadsheetBench` now defaults to `mode=multi`. If you run an exec student backend there, override back to `env.mode=single` because exec backends are still only wired for SpreadsheetBench single-mode rollout.
-- `SWEBench` follows the older mini-swe-agent + `swebench.harness.run_evaluation` path, so it requires the SWE-bench / Docker toolchain rather than the generic chat-only stack.
-- `slow_update` writes into a protected skill region and normal edits are prevented from overwriting that region directly.
-- `meta_skill` is context memory, not a direct skill edit.
-- `meta_reflect` is a gated skill edit stage, not just logging.
-
-## Minimal Setup
+Launch the monitoring dashboard (optional):
 
 ```bash
-conda create -n reflact python=3.11
-conda activate reflact
-pip install openai pyyaml openpyxl
+pip install -e ".[webui]"
+python -m skillopt_webui.app
 ```
 
-Depending on the environment, you may also need:
+| Flag | Default | Description |
+|---|---|---|
+| `--port` | 7860 | Server port |
+| `--host` | `0.0.0.0` | Bind address |
+| `--share` | off | Create a public Gradio share link |
 
 ```bash
-pip install datasets gymnasium numpy ray regex
+# With public share link (useful for remote servers)
+python -m skillopt_webui.app --share
 ```
 
-For `SWEBench`, you also need a working Docker environment plus the SWE-bench / mini-swe-agent dependencies used in `swe-bench-old`.
+---
+
+## Citation
+
+```bibtex
+@misc{yang2026skilloptexecutivestrategyselfevolving,
+      title={SkillOpt: Executive Strategy for Self-Evolving Agent Skills}, 
+      author={Yifan Yang and Ziyang Gong and Weiquan Huang and Qihao Yang and Ziwei Zhou and Zisu Huang and Yan Li and Xuemei Gao and Qi Dai and Bei Liu and Kai Qiu and Yuqing Yang and Dongdong Chen and Xue Yang and Chong Luo},
+      year={2026},
+      eprint={2605.23904},
+      archivePrefix={arXiv},
+      primaryClass={cs.AI},
+      url={https://arxiv.org/abs/2605.23904}
+}
+```

ckpt/alfworld/gpt5.5_skill.md

@@ -0,0 +1,113 @@
+# ALFWorld Embodied Agent Skill
+
+## Overview
+This skill guides agents operating in the ALFWorld text-based embodied environment.
+The agent must complete household tasks by navigating rooms, interacting with objects,
+and using appliances. Actions must be chosen from the admissible action list provided
+at each step.
+
+**Output format**: Always output `<think>...</think>` for reasoning, then `<action>...</action>` for the chosen action.
+
+---
+
+## Task Types
+
+| Type | Goal | Key Steps |
+|------|------|-----------|
+| Pick & Place | Put object X in/on receptacle Y | Find X -> take X -> go to Y -> put X in/on Y |
+| Pick Two & Place | Put two instances of X in/on Y | Find X1 -> take -> place -> find X2 -> take -> place |
+
+### Pick Two Object Bookkeeping
+For `pick_two_obj_and_place`, choose one destination receptacle instance once it is opened/usable, and remember it as the target. Both object instances should be placed into that same remembered receptacle. After placing the first object, do not remove it again; if the second object was already seen, return directly to its remembered location rather than searching randomly. If the two objects are accidentally split across different receptacles, consolidate them into the chosen target receptacle.
+| Examine in Light | Examine object X under desklamp | Find X -> take X -> find desklamp -> use desklamp |
+
+| Examine in Light detail | Final interaction | While holding X where a desklamp is visible, use the desklamp; do not try to place X on the lamp first. |
+| Clean & Place | Clean object X and put in/on Y | Find X -> take X -> go to sink -> clean X -> go to Y -> put X |
+| Heat & Place | Heat object X and put in/on Y | Find X -> take X -> go to microwave -> heat X -> go to Y -> put X |
+| Cool & Place | Cool object X and put in/on Y | Find X -> take X -> go to fridge -> cool X -> go to Y -> put X |
+
+---
+
+## General Principles
+
+1. **Decompose the task**: Parse the goal into ordered sub-goals (locate, acquire, transform, deliver). Complete each before moving to the next.
+2. **Systematic exploration**: Search each surface and container exactly once before revisiting. Open closed containers (drawers, cabinets, fridge) before judging them empty.
+
+- Prioritize semantically likely locations first, then broaden systematically: food in fridges/on countertops or dining tables; dishes/utensils/cookware on countertops, dining tables, stoveburners, cabinets, or drawers; office/bedroom items on desks, shelves, dressers, sidetables, or in drawers; newspapers on coffeetables, sidetables, sofas, or tvstands; toiletries/cleaning items near sinks, bathroom counters, shelves, carts, or cabinets.
+
+- For portable kitchen targets such as bread, mugs, cups, plates, bowls, and utensils, check broad exposed surfaces early: after one or two empty countertops, try dining tables or other open surfaces before opening many cabinets/drawers. For small office/bedroom targets, alternate drawers with exposed desks, shelves, sidetables, and dressers rather than exhausting drawers first.
+
+- Keep a persistent **searched set** of receptacle instances, e.g. `drawer 1`, `shelf 3`, `countertop 2`. Once an observation shows no needed target object there, mark it searched and do not call it “unexplored” later.
+   - If all locations in the current preferred class are searched, **broaden to any unvisited admissible `go to ...` location** instead of restarting the same sequence. Search broadly across surfaces, furniture, containers, and appliances when relevant.
+   - If a visible object is itself an openable/container-like object, such as a box, and opening/examining it is admissible, inspect it before leaving the area.
+3. **Grab immediately**: When a required object is visible and reachable, take it right away before moving elsewhere.
+
+- Pick up only the exact requested object type. Similar or related objects, such as a cup when the task asks for a mug, a spoon when it asks for a knife, or a pot when it asks for a pan, are distractors; leave them in place and mark that location searched for the target.
+4. **Transform before placing**: If the task requires cleaning, heating, or cooling, perform the state change at the appropriate appliance before heading to the final destination.
+
+- Do not repeatedly revisit the sink, microwave, fridge, or final destination before holding the target object. If you find the appliance early, remember its location, then resume searching unvisited object locations until the target object is acquired.
+
+- Use direct admissible appliance/tool commands immediately when available, such as `clean X with sinkbasin`, `heat X with microwave`, `cool X with fridge`, or `use desklamp`. Do not waste steps opening, closing, toggling, or examining the appliance unless the needed action is unavailable or opening is required for searching/placing.
+5. **Direct delivery**: Once holding the transformed (or untransformed) goal object, navigate straight to the target receptacle and place it.
+
+- Remember known destination receptacles and return directly to the same instance after pickup/transformation. If the destination is also a semantically likely source location, check/open it early rather than only after exhaustive search: food may already be in the fridge, utensils may be on the diningtable, newspapers may be on/near the sofa, and a target drawer can be opened early for pick-two tasks. If the object starts at the destination but needs cleaning/heating/cooling, take it out, transform it, then return to that same instance and place it back.
+6. **Track progress**: Maintain an internal count of how many objects still need to be found and placed. Only stop searching when the count reaches zero.
+7. **Avoid loops**: Never repeat the same action more than twice in a row. If stuck, move to a different unexplored location.
+8. **Only choose admissible actions**: Always pick an action from the admissible action list. Do not invent actions.
+
+---
+
+## Common Mistakes to Avoid
+
+- **Revisiting searched locations**: Keep track of which surfaces/containers have been checked; do not re-examine them.
+- **Ignoring visible objects**: If the target object appears in the observation, pick it up immediately.
+- **Skipping state changes**: Do not place an object at the destination without first cleaning/heating/cooling it when required.
+- **Premature termination**: Do not stop the episode until all goal conditions are verified as met.
+- **Action loops**: Repeatedly toggling or examining the same object wastes steps. Move on to new locations instead.
+
+### Hard Search-Loop Recovery
+
+- **Exact-instance lockout before pickup**: once a receptacle/surface instance has been observed and does not contain the target object, do not go back to that exact instance while still searching for the object. A phase change, such as holding the object or needing final delivery, is the only reason to return.
+- **Fast broadening threshold**: after 3-4 misses in the same receptacle class, switch to a different likely class or any unvisited admissible location instead of continuing or restarting that class, unless the target has already been seen there.
+- **No search reset by recency**: do not say a location is "unsearched" merely because it was not in the last few observations. The searched set is global for the whole episode.
+- **Finite-class exhaustion**: if all visible instances of a small class have been checked once, such as all stoveburners, diningtables, countertops, or shelves, mark that class exhausted for object search and do not start a second pass. Remember a usable destination instance, then search different receptacle classes.
+- **Unvisited beats likely-but-searched**: after several misses, prefer any admissible unvisited `go to`, `open`, or `examine` target over revisiting a semantically likely but already-searched location.
+- **Destination surfaces before pickup**: if the destination receptacle is also a likely object location, inspect each instance at most once before pickup. If it lacks the object, remember it as the final destination but stop using it as a search target until the object has been transformed and is ready to place.
+- **Kitchen item fallback**: for cookware and dishware, after checking obvious burners/tables/counters once, broaden to unsearched cabinets, drawers, shelves, sinkbasins, and other kitchen storage/surfaces rather than cycling among the obvious locations.
+
+### Strict Search Ledger Action Filter
+
+Before every empty-handed search action, apply this hard filter:
+
+1. If a required target object is visible, take it immediately.
+2. Otherwise choose an exact receptacle/surface/container instance whose contents have not yet been observed in the current object-search phase.
+3. Reject any `go to`, `examine`, or `open` action for an exact instance already observed to lack the target, even if it is semantically likely, nearby, recently mentioned, or the final destination type.
+4. If all likely instances are rejected by the ledger, broaden to any unvisited admissible location/class instead of restarting from instance 1 of a searched class.
+
+The searched ledger survives inventory checks, appliance visits, placing the first object in a pick-two task, and putting down an irrelevant inspected object/container. These events are not permission to rescan shelves, drawers, cabinets, tables, counters, or destination receptacles from the beginning.
+
+### Destination-as-Source Lockout
+
+When the final receptacle type is also a plausible source location, inspect each visible destination instance at most once before pickup. After it lacks the target, remember a usable destination instance and lock that exact instance out of object search until you are holding the required object ready for delivery. Do not alternate between destination instances and other searched source instances while still empty-handed.
+
+### Pick-Two Phase Memory
+
+After placing the first object in a pick-two task, do not begin a fresh room/class search. If another required instance was previously seen, return directly to that remembered source location for the second pickup. If no second instance is remembered, continue from the existing unsearched-location ledger rather than revisiting locations already checked before the first placement.
+
+<!-- SLOW_UPDATE_START -->
+Preserve the successful pattern: when the exact requested object is visible, take it immediately; perform the required clean/heat/cool/use action as soon as the correct command is admissible; then deliver directly to the remembered destination.
+
+Treat tool locations as tools, not repeated search targets. If a sinkbasin, fridge, microwave, desklamp, or destination receptacle has already been checked and does not contain the target while you are empty-handed, remember it for later but do not revisit it until you are holding the required object or ready to place/use it.
+
+Use a next-unsearched-instance pointer for every numbered class. If you leave cabinets, drawers, shelves, countertops, or stoveburners and later return to that class, resume at the lowest exact instance not yet observed; never restart at instance 1 and never revisit an instance already observed to lack the target.
+
+For pan-to-stoveburner tasks, search in a step-efficient order: make one quick pass over stoveburners only to find a pan or remember an empty destination, then leave stoveburners until delivery. Next check countertops/islands and sinkbasins. Then prioritize cabinets in numeric order, opening each closed cabinet and observing its contents, before low-yield drawers. Do not abandon cabinet search to revisit searched stoveburners, countertops, or drawers.
+
+For kettle/teapot clean-and-place tasks, after checking obvious countertops/islands, check stoveburners and sinkbasins once, then cabinets in numeric order. If several cabinets are empty, continue to the next unsearched cabinet or broaden to unvisited shelves/carts/dining tables; do not return to already searched countertops. Remember one open/empty cabinet as the final destination, but do not keep using searched cabinets as search targets.
+
+For dishsponge clean-and-place tasks, check sinkbasin and nearby countertops once, then search unvisited cabinets, drawers, shelves, carts, and other storage/surfaces. Because the sink is needed for cleaning, remember it after the first visit; do not go back to the sink while empty-handed just because the sponge is likely near it. Because shelf is the destination, remember a usable shelf after inspecting it once; after a shelf lacks the sponge, search only unvisited shelves or other unvisited locations until the sponge is found.
+
+When the step budget is running and you are still empty-handed, prefer any unvisited admissible location over any searched likely location. A location being semantically likely, useful later, or recently mentioned is never a reason to rescan it before acquisition.
+
+Do not let the broadening threshold cause class restarts. Broadening means move to a different unvisited class or continue at the next unsearched instance of a promising storage class; it never means cycling back through exact instances already observed.
+<!-- SLOW_UPDATE_END -->

ckpt/docvqa/gpt5.5_skill.md

@@ -0,0 +1,26 @@
+# DocVQA Skill
+
+## Visual Evidence Discipline
+- Read the document carefully before answering.
+- Prefer the smallest exact text span that answers the question.
+
+- For questions asking for a value, count, page number, date, or graph reading, return only the requested value span; omit nearby labels, category names, units, or explanatory words unless the question explicitly asks for them.
+- When several nearby strings look similar, choose the one whose surrounding labels or layout best match the question.
+
+## Exact Answer Discipline
+- Copy names, numbers, and dates exactly from the document whenever possible.
+
+- Preserve the document's exact spelling and punctuation for names and quoted phrases; do not substitute similar letters or change straight/curly quotes, spacing, or parentheses when the visible text provides them.
+- Prefer direct extraction over paraphrase.
+- Before finalizing, compare the answer against nearby alternatives and keep the best-supported exact span.
+
+## Structured Layout Lookup
+- For tables, first find the row or entry named in the question, then read the value under the requested column, header, date, or category; answer with that cell only.
+- For forms, receipts, or labeled fields, locate the exact role, party, or field label mentioned in the question, then copy the filled-in value from the same line, box, block, or immediately adjacent field.
+- For table-of-contents, indexed, numbered, or bulleted lists, match the requested title, entry, or point number, then follow the same line or list item to the associated value; do not take a nearby value from another item.
+
+## Anchored Handwriting / Nearby Text
+- For handwritten or list/table questions with an anchor term, first locate the anchor, then inspect the immediately adjacent text in the same row, column, or nearby margin. If legible, provide the best-supported nearby span rather than leaving the answer blank.
+
+<!-- SLOW_UPDATE_START -->
+<!-- SLOW_UPDATE_END -->

ckpt/livemath/gpt5.5_skill.md

@@ -0,0 +1,35 @@
+# Live Mathematical MCQ Heuristics
+
+## Option Comparison
+
+### Meta-Options About Stronger Results
+- Treat options of the form “one of the remaining options is correct, but a stronger result can be proven” as serious candidates, especially when the question asks for the strongest statement.
+- If a concrete option is true but your theorem or derivation gives a strictly stronger conclusion not exactly listed, choose the meta-option rather than the weaker concrete statement.
+- When options are nested by strength, rank them explicitly before answering: e.g. finite-time blowup is stronger than merely “not globally bounded”; positive stable growth is stronger than ordinary unboundedness; sharper constants, rates, exceptional-set bounds, endpoint inclusion, or full equivalences are stronger than weaker asymptotic versions.
+- Compare all options before committing. The correct choice is often the strongest statement justified by the question, while nearby distractors are weaker, overstrong, or miss an equality case.
+- Track exact quantifiers such as "there exists", "for every", "if and only if", and "exactly when".
+
+## Theorem-Level Precision
+
+- Do not add converse, realization, or classification claims unless the theorem explicitly proves them. Phrases such as “conversely,” “every such parameter occurs,” “if and only if,” or “exactly all” add strength beyond a one-way implication.
+- Check whether an option weakens the conclusion by dropping a characterization, equality clause, or full equivalence.
+- Check whether an option overstates the theorem by upgrading regularity, removing scale restrictions, or changing an existential statement into a universal one.
+
+## Hypotheses
+
+### Exact Conditions and Thresholds
+- For biconditional/equivalence questions, reject conditions that are merely necessary or merely sufficient. A broader condition, such as congruence modulo a divisor instead of modulo the full modulus, is usually weaker and not equivalent unless the domain collapses the extra cases.
+- For threshold conditions, verify the exact sign and endpoint: distinguish \(\mu_0\) from \(-\mu_0\), \(<\) from \(\le\), and whether the equality case belongs to the positive, zero, or negative parameter regime.
+- When options differ by “for every” vs “for sufficiently large,” local vs global domains, strict vs non-strict inequalities, or dependence of constants, rank them by logical strength and match the sharpest justified version.
+- Verify the hypotheses and domain carefully. Distractors often keep the theorem shape but alter the required assumptions.
+- Pay close attention to equality cases, extremal conditions, and whether a result applies to the full family or only a restricted subfamily.
+
+## Final Answer
+- Output the final answer as the single option label only.
+
+## Exact Scope and Quantitative Wording
+- Distinguish global conclusions from localized or completed ones. Equivalence after localization, completion, or at each prime/scale is usually weaker than an unqualified equivalence.
+- In estimate-heavy options, compare every quantitative detail: exponent, derivative index range, constants and their parameter dependence, log factors, additive terms, one-sided vs two-sided notation, and pointwise vs uniform convergence.
+
+<!-- SLOW_UPDATE_START -->
+<!-- SLOW_UPDATE_END -->

ckpt/officeqa/gpt5.5_skill.md

@@ -0,0 +1,50 @@
+# OfficeQA Skill
+
+## Retrieval Discipline
+
+- When an external official time-series observation is needed, prefer the source's series/data-download/table page once identified. If exact-date or guessed-value searches return empty results, stop repeating them; broaden to the official series name/code plus `data` or `download` and use the table values.
+
+- Treat provided/oracle parsed pages as primary evidence: if they contain the relevant table and period, extract directly from them before searching elsewhere; search only for missing continuation pages, missing periods, or an official actual value not present.
+- Start by narrowing to the most likely candidate file before reading long passages.
+- Prefer targeted search terms that name the exact entity, period, measure, or table concept from the question.
+- After a promising match, read only a small surrounding span and verify it matches the requested year, basis, and unit.
+
+- If the requested date range extends beyond the provided/oracle page, first enumerate the required periods and verify that every period is present in evidence. Do not compute from a partial ledger or fill missing periods from memory; retrieve continuation pages, adjacent issues, or a later issue of the same table that contains the missing dates/revisions.
+
+## Evidence Discipline
+- Extract the exact value from the retrieved text before doing any arithmetic.
+- Keep track of each operand's period, unit, and semantic role so nearby proxy values are not mixed in.
+
+- For Treasury financing narratives, label each amount by transaction role before calculating: offered amount, tenders/subscriptions received, tenders accepted, competitive/noncompetitive accepted, foreign or Government-account exchange tenders, refunding, and **new cash** are not interchangeable.
+- When converting currencies or scales, make a direction ledger first: source table unit, source currency, exchange-rate orientation (foreign currency per U.S. dollar means divide by the rate; U.S. dollars per foreign unit means multiply), and requested final unit.
+
+- For tables, align values by row label and exact column header, not proximity alone; watch for continued or unlabeled columns, footnotes, adjacent amount-versus-percent columns, fiscal-year versus calendar-year sections, and repeated month rows under different year blocks.
+- If the question asks for a transformed or derived quantity, compute only after confirming every operand.
+
+- For derived comparisons, preserve the direction and sign implied by the wording: “change from A to B” means B minus A; “former than latter” means former minus latter; “share accounted for by X” means X divided by the stated total; paired “gap” questions require computing each within-row difference before ranking.
+- For statistical, regression, correlation, and growth-rate questions, write a formula ledger before calculating: confirm the exact series/endpoints, ordered vector, elapsed intervals, and requested convention such as continuously compounded rate, CAGR, Pearson correlation, or OLS index/year choice.
+- For multi-stage questions where one table determines the period/entity used in another lookup, freeze that derived key with evidence first, then retrieve the second measure only for that exact month/year/reporting date/entity.
+
+- For inclusive time-series ranges, make a period-by-period ledger covering every requested month/year exactly once, preserving calendar versus fiscal basis, end-of-month or end-of-fiscal-month status, source units, and any specified adjustments.
+
+- For statistical transforms over time-series windows, confirm endpoint inclusion/exclusion exactly as worded, use consecutive time indices for trend regressions when appropriate, sort values before medians, and for logarithmic growth use ln(final/initial) before converting to the requested percentage format.
+
+## Final Answer Discipline
+
+- Before finalizing, enforce the requested unit and format: convert thousands/millions/billions or full nominal dollars as needed, then apply no-comma, fixed-decimal, whole-number, or nearest-tenth/thousandth formatting exactly as asked.
+- Return the final answer only after one last consistency check against the retrieved evidence.
+- Copy the final answer from a checked value, not from an unverified intermediate guess.
+
+## Statistical and Time-Series Calculation Checks
+
+- Before computing any statistic, write the intended formula and denominator convention. If the prompt explicitly says **population standard deviation**, divide by `n`; if it says **sample**, divide by `n-1`; for a z-score comparing one observation against a small set of comparison months/periods and no population convention is stated, estimate dispersion with the **sample** standard deviation of the comparison set. Do not round intermediate operands, weighted averages, logs, exchange-rate conversions, or standard deviations before the final requested rounding.
+- For long inclusive ranges, first enumerate the expected count of observations and the first/last period, then verify the ledger has exactly that count. Exclude totals, cumulative-to-date columns, comparable-period columns, estimates, and extra latest-month columns outside the requested calendar or fiscal range.
+- When a page contains multiple nearby sections with similar labels, use only the section whose title and row label match the requested measure exactly; do not compute from the first visible table if the requested measure/table title is absent or only partially shown.
+- For Treasury security quotations, obey the table's quote basis. If the table states that price decimals are 32nds, convert quotes such as `99.27` as `99 + 27/32`, not as decimal `99.27`. If a task asks for smoothing, averaging, or forecasting in a target currency using period-specific exchange rates, convert each period's observation to the target currency first unless the prompt explicitly says to compute in the source currency and convert only the final result.
+
+## Stricter Final Formatting
+
+- Match any requested output template exactly. Unless the prompt explicitly asks for unit words or explanatory text, return only the numeric value or requested list; do not append words such as `million`, `dollars`, `percent`, or `percentage points`. Include symbols/commas only when the prompt requests currency-formatted output or the answer format clearly requires them.
+
+<!-- SLOW_UPDATE_START -->
+<!-- SLOW_UPDATE_END -->

ckpt/searchqa/gpt5.5_skill.md

@@ -0,0 +1,71 @@
+# Question Answering Skill
+
+(No learned rules yet. Rules will be added through the reflection process.)
+
+## Concise Answer Normalization
+- Prefer the shortest unambiguous answer that directly satisfies the question. Do not include generic descriptors, legal suffixes, or expanded formal names unless the question specifically asks for the full official name or the descriptor is necessary to identify the entity.
+
+- If the answer appears inside a longer descriptive phrase, strip words that merely repeat the clue's requested type or modifiers already stated in the clue. For short-answer trivia, return the distinctive core entity or headword rather than role titles, product flavor adjectives, or place/facility designators, even when those words are part of a fuller official phrase, unless the full official name is explicitly requested.
+- For place/name-etymology questions asking for “the name” or “the word” that means something, answer the distinctive name/word itself rather than a larger phrase with a generic type label.
+
+- For natural geographic features, preserve conventional feature designators such as “Lake,” “River,” “Bay,” “Gorge,” “Mount,” or “Island” when they are part of the proper name or match the requested feature type. Do not shorten “Lake Okeechobee,” “Tampa Bay,” or “Olduvai Gorge” to an ambiguous base name merely to be concise.
+- For companies, brands, and organizations, answer the common distinctive name when sufficient; omit additions such as “Company,” “Corporation,” “Inc.,” etc. unless explicitly required.
+
+- Preserve the answer surface form supported by the strongest evidence when exact variants differ: spelling, capitalization, punctuation, and word order can matter. Do not substitute an equivalent official/common variant such as an alternate spelling or inverted institution name if a direct title/snippet/answer field gives the expected form.
+
+- When copying titles or quoted names, preserve ordinary ASCII punctuation from the evidence, especially straight apostrophes (`'`). Do not replace them with typographic curly quotes/apostrophes unless that exact stylized form is explicitly shown as the supported answer.
+
+- For nicknames, epithets, saints, and quoted titles, copy the supported surface form exactly, including spacing, capitalization, and conventional abbreviations such as “St.” Do not normalize a stylized or quoted form into a lowercase dictionary word or an expanded spelling when the clue/evidence points to the stylized answer.
+
+- For person answers in trivia or crossword-style clues, prefer the conventional supported name. Use just a surname, first name, or saint/regnal name only when the clue/source clearly expects that short form; otherwise use the canonical full personal name from the strongest evidence or answer field, especially when a lone given name would be ambiguous.
+- Return the grammatical base form expected by the clue. Do not add a plural `s` merely because a crossword source pluralizes a shared name or category; if the clue lists people sharing a first name, answer the singular given name.
+
+- For common-noun category answers, default to the singular dictionary headword in trivia/crossword-style clues, even if the clue uses plural words like “these,” “those,” “places,” or “items” for grammar. Use a plural only when the term is inherently plural or an answer field/source clearly gives a plural phrase.
+
+- For common-noun clues about things being replaced, used in place of, or substituted by another system/item, answer the broad headword for the thing replaced unless a narrowing modifier is required by the clue or answer field. Do not add adjectives such as “letter,” “regular,” or “standard” merely because they appear in explanatory context.
+- For fill-in-the-blank or definitional clues using words like “this” or “that,” provide a standalone noun phrase. Avoid context-dependent pronouns or possessives from the source text; use a natural article such as “the” when needed (e.g., answer “the highest point,” not “its highest point”).
+
+## Context-Grounded Evidence Matching
+- Start by identifying the most distinctive terms in the question: proper names, dates, titles, quoted phrases, unusual words, roles, relationships, and category descriptors.
+- Prioritize passages or document titles where several distinctive clue terms occur together, especially if the wording directly repeats or closely paraphrases the question.
+- Treat document titles as useful evidence: the answer is often named in a title while the snippet confirms the clue facts.
+
+- Do not assume the document title itself is the answer. If the requested type differs from the title entity, use the title as context and extract the matching typed entity from the snippet or clue relationship.
+
+- For “known as,” “called,” “defined as,” or category/type clues, choose the canonical term explicitly used in the strongest matching title/snippet or scraped answer field rather than inventing a related derivative or near-synonym from the clue wording. When multiple plausible candidates appear, prefer the candidate whose evidence directly states the requested relationship and repeats the most distinctive clue facts.
+- Ignore noisy results that only match generic words; prefer evidence that directly connects the clue facts to one specific entity.
+
+## Clue Interpretation and Answer Type
+- For Jeopardy-style wording such as “this man,” “this group,” “this film,” “this country,” “this system,” “he,” or “his wife,” infer the expected answer type before choosing the answer.
+- Use that expected type to validate candidates: answer with the concise person, place, title, organization, object, term, or phrase requested by the clue.
+
+- Treat modifiers attached to the requested type as hard filters, not background flavor: constraints like dates, “largest,” “2-letter-named,” “1978 remake,” “hot dog brand,” “dual throne,” or “on this company’s board” must all fit the candidate before you answer.
+- For clues centered on creative works such as books, films, plays, songs, poems, or other media, first determine whether the clue asks for the work itself, its creator, a performer or cast member, a character, a quotation source, or a setting. Verbs such as “wrote,” “directed,” “stars,” “played,” and “set in,” plus pronouns like “he” or “her,” usually determine the target.
+
+- For fill-in-style clues with placeholders such as “this,” “these,” or “one of these,” substitute each candidate back into the clue and choose the concise answer that makes the full phrase, title, or fact read correctly.
+- For terse clues that are just examples or names separated by commas, slashes, or “or,” infer the shared category, class, or synonym that links them, then answer with that concise common term.
+
+- For crossword-style clues, treat parenthetical numbers or stated letter counts as hard constraints on the answer length, and omit generic labels that would violate them. In dual-definition clues using wording like “X, or what Y does,” choose the single word that satisfies both senses and preserve the required inflected form.
+- If the clue references an unavailable image or link with wording like “seen here,” “pictured,” or parenthetical visual hints, rely on the textual clues and context to infer the answer; do not treat the missing image as necessary evidence.
+- If multiple snippets support the same entity, use that corroboration to choose the canonical/common form of the answer.
+
+## Trivia / Jeopardy Snippet Formats
+- Retrieved trivia snippets may contain the clue and answer in scraped formats such as `CATEGORY | clue | answer`, `clue. ANSWER`, or labels like `right:`.
+- When the question text matches the clue in such a snippet, extract the answer field or adjacent answer name, not the category or the whole clue sentence.
+
+## Common Clue Traps
+- Watch for inverse relationships: if the clue says “His third wife was Jiang Qing,” the requested answer is the husband, not Jiang Qing.
+
+- More generally, preserve relation direction in clues: “A is evidence of this B,” “A is related to this language,” or “home to these characters” asks for the target of the relationship, not the entity already named in the clue.
+
+- When a clue says examples, models, breeds, members, or items “include,” “like,” or “such as” named entities, treat those names as evidence for the requested parent class or entity. Answer the encompassing brand, animal, category, place, or term requested by “this,” not one of the examples already given.
+- If the question gives the start of a quotation or phrase, answer with the exact missing continuation from the context.
+
+- For song, poem, nursery-rhyme, or quotation clues, first decide whether the question asks for a missing word or phrase from the quote or for the associated creator, performer, or work; use pronouns and answer-type signals to choose the right target.
+- When a clue asks for a constrained form such as a first name, abbreviation, acronym, or lyric word, return that exact form rather than the fuller person, title, or explanation; preserve conventional punctuation or spelling when it is part of the requested form.
+- If the clue contains wordplay, quotation marks, or puns, treat them as hints, but answer with the real entity supported by the evidence.
+
+- If a clue includes a quoted title, quoted narration or lyric, named event, slogan, or other distinctive phrase but asks for an associated “this” entity, treat the quote or name as evidence to identify the requested person, work, place, group, category, source, or term; do not return the quoted anchor unless the clue explicitly asks for it.
+
+<!-- SLOW_UPDATE_START -->
+<!-- SLOW_UPDATE_END -->

ckpt/spreadsheetbench/gpt5.5_skill.md

@@ -0,0 +1,133 @@
+# Spreadsheet Manipulation Skill (xlsx)
+
+## Overview
+This skill guides agents in manipulating Excel (.xlsx) spreadsheets using Python.
+
+**Primary libraries**: `openpyxl` (structure-preserving read/write), `pandas` (data transformation).
+Never use any other third-party libraries.
+
+---
+
+## Common Workflow
+
+1. **Explore** the input file: list sheets, inspect headers, check dimensions.
+
+- Inspect actual workbook data beyond the preview, including nearby rows/columns, sample outputs, formulas, labels, headers, and any reference/example sheets such as `Output`, `Manual Result`, or `Desired...` tabs.
+
+- Treat existing filled cells in the requested output area or adjacent example tables as semantic examples for edge cases and expected formats, but still recompute and write the complete requested target range.
+   - Scan the used range for complete header groups, not just row 1. Tables may start in later rows/columns, have title rows above them, or have multiple source/result tables on the same sheet; use nearby labels and the requested output range to distinguish sources from destinations.
+   - Locate tables, fields, and target ranges by header text, nearby labels, and surrounding nonblank structure rather than fixed coordinates. Build header maps from actual cells when useful, e.g. `{str(cell.value).strip(): cell.column}`.
+2. **Write `solution.py`** with `INPUT_PATH` and `OUTPUT_PATH` defined at the top.
+3. **Execute** `python solution.py` and verify the output file was created.
+4. **Confirm** the target cells/range contain the expected values.
+
+---
+
+## Library Selection
+
+| Use case | Library |
+|----------|---------|
+| Preserve formulas, formatting, named ranges | `openpyxl` |
+| Bulk data transformation, aggregation, sorting | `pandas` → write back with `openpyxl` |
+| Simple cell read/write | `openpyxl` |
+
+**Warning**: `pandas.to_excel()` silently destroys existing formulas and named ranges.
+When writing back to a spreadsheet that contains formulas, always use `openpyxl.save()`.
+
+**Formula evaluation caution**: `openpyxl` can write formulas but does **not** calculate them or update cached results. If the requested output will be checked as cell values, compute the result in Python and write literal values unless the user explicitly requires live formulas. When existing formulas are inputs to your logic, load a second workbook with `data_only=True` to read cached values while saving changes through the normal workbook:
+
+```python
+wb = openpyxl.load_workbook(INPUT_PATH)
+wb_values = openpyxl.load_workbook(INPUT_PATH, data_only=True)
+ws = wb["Sheet1"]
+ws_values = wb_values["Sheet1"]
+```
+
+Treat wording such as “write/fix a formula,” “SUMIFS/COUNTIFS,” “VBA,” or “macro” as a description of the spreadsheet logic unless the deliverable explicitly requires live formula text, an `.xlsm`, or a preserved VBA project. For normal `.xlsx` outputs, implement the equivalent logic in Python/openpyxl and write the computed final values to the requested cells so verification does not depend on Excel recalculation or macros.
+
+When the user provides an existing or broken formula, use it as a semantic specification: honor its referenced lookup ranges, criteria ranges, return ranges, aggregation intent, and error-handling behavior, then write the resulting values rather than guessing different source columns or leaving unevaluated formulas.
+
+---
+
+## solution.py Template
+
+```python
+import openpyxl
+import pandas as pd
+
+INPUT_PATH  = "..."   # set to the actual input path
+OUTPUT_PATH = "..."   # set to the actual output path
+
+wb = openpyxl.load_workbook(INPUT_PATH)
+ws = wb.active  # or wb["SheetName"]
+
+# --- perform manipulation ---
+
+wb.save(OUTPUT_PATH)
+```
+
+---
+
+## Output Requirements
+
+- Save the result to `OUTPUT_PATH`.
+- Do not hardcode row counts or column letters — iterate over actual rows in the workbook.
+- Preserve sheets and cells not mentioned in the instruction.
+
+## Matching and Target Range Hygiene
+
+- Choose the comparison operator from the instruction and examples: use `startswith` for “begins with”, substring search for “contains/search/occurrence”, and exact normalized equality only when a whole-cell match is implied.
+- Create small helper functions for comparisons and numeric parsing. Normalize text by trimming, collapsing repeated spaces/NBSPs, and casefolding; when names or labels have punctuation/spacing inconsistencies, consider punctuation-insensitive keys. Parse numeric text after removing commas/currency symbols while preserving signs and decimal points; skip `None`/blank and booleans for numeric tests, and handle placeholders such as `"-"`, `"$"`, `"$0"`, blanks, and numeric zero deliberately.
+- Normalize date keys deliberately: handle `datetime`/`date` objects, Excel serial numbers, and date-like strings, then compare at the granularity implied by the task, such as exact date, month, month/year, fiscal period, or year. For workday/date-window logic, compute the range in Python and exclude weekends/holidays as specified.
+
+- For monthly or period summary grids, canonicalize period labels from all sources: sheet names, title text, row/column headers, text months such as `March`, and actual date cells. Match summaries by normalized period plus the other stated criteria rather than by fixed month offsets or existing formulas.
+- For date ranges and rolling windows, infer endpoint inclusivity from wording and examples. Phrases like `X to Y`, `through`, and `up to`, or examples such as `2 to 5` meaning `4 days`, usually require inclusive boundary handling.
+- For time extraction or time-threshold logic, parse `datetime`, `time`, Excel serial/fractional times, and time-like strings into real Python `time`/`datetime` values. Write real time values with an Excel `number_format` such as `hh:mm:ss AM/PM`; do not write text substrings when the result should behave as a time.
+- For joins, deduplication, grouping, interval lookups, lookup grids, and ordered outputs, build explicit normalized keys, including composite keys when the task refers to multiple fields. Preserve original source order within each group unless sorting is explicitly requested.
+- For outputs that depend on other rows or lookup grids, make a first pass to build normalized dictionaries/groups/range structures, then a second pass to write results. Avoid nested full-sheet scans per row; split delimited tokens and ignore empty tokens, and treat error literals such as `#N/A` as meaningful sentinel values when the task refers to them.
+
+- For lookups, filters, joins, and label/header matching, normalize comparison keys consistently: trim whitespace, skip blanks explicitly, use case-insensitive text matching when appropriate, and treat numeric-looking IDs consistently (`330`, `330.0`, and `"330"`). Keep numeric outputs numeric; use `number_format` for display formatting instead of converting numbers to strings unless text is explicitly required.
+- When replacing a generated output area, clear only the instructed target range before writing new results so stale values/formulas do not remain. Preserve formatting, column widths, borders, formulas, and unrelated cells unless the instruction explicitly asks to change them.
+
+- If the instruction includes formatting changes, apply them exactly after writing values and only to the requested cells/range. Use `openpyxl` styles for fills, alignment, fonts, borders, and number formats; convert hex colors to ARGB when needed, for example `#FFC000` → `FFFFC000`. For “format as text,” set `number_format = '@'` and write string values when the expected cell values are text.
+
+- When the instruction names a destination range or columns, write derived results directly there. Do not insert rows/columns, relocate the source table, or sort/delete source records unless that structural change is explicitly requested.
+- For filtered lists, summaries, and aggregations, first collect all source records/results in memory, preserving the required order, then write from the first output row and clear leftover cells below the new results in the target columns. When adding rows, copy style/alignment/number format from an existing template row when appropriate; when deleting rows, delete from bottom to top to avoid row-index shifts.
+- Preserve intended blanks as empty cells (`None`) rather than placeholder text or `0` unless the task specifies otherwise.
+
+- For numeric aggregation, crosstab, SUMIFS-like, and INDEX/MATCH-style summary outputs, infer missing-match behavior from table semantics and examples: numeric summary grids usually require literal `0` for no matching records, while filtered lists or “show only once” outputs usually require blanks (`None`).
+- For blank-sensitive logic such as “if input is blank, output blank,” evaluate the driving input with `data_only=True` when it may itself be a formula, and write `None` for truly blank outputs rather than relying on a new formula returning `""`.
+
+## Robustness for Simple Fill Tasks
+
+- Prefer simple, auditable row/column loops over complex workbook XML parsing unless the task truly requires unsupported workbook internals. Before returning, run the script once to catch syntax/indentation errors and verify that representative target rows were actually written.
+
+<!-- SLOW_UPDATE_START -->
+When the user asks for a formula, macro, VBA code, or a fix to an Excel formula, still deliver the completed workbook state: compute the intended results in Python and write literal final values into the requested cells. Do not write formula strings unless the task explicitly says the output must contain live formulas.
+
+After writing, reload or inspect the saved workbook and verify that every requested/evaluated target cell contains a non-formula literal where a value is expected. If a target cell is still `None` unexpectedly, fix the script before finishing.
+
+Use existing formulas in the workbook as examples/specifications, not as output. If a cell contains a reference formula such as `=A25` or an INDEX/MATCH/SUMIFS pattern, parse what source cells/ranges/criteria it refers to, compute those results yourself, and overwrite the destination with the referenced or calculated value.
+
+For blank-sensitive formula tasks, compute the branch explicitly: if the driving source cell is truly blank, write `None`; otherwise write the actual result such as `0`, `1`, a category label, or a lookup value. Never rely on `IF(...,"",...)` formulas to be recalculated later.
+
+For lookup/category tasks, locate both the input rows and the lookup table by headers and nearby labels. Support exact keys, numeric-looking keys, and interval/range tables; then fill every destination row that has a driving input, not just the first visible example.
+
+For “every nth row” or OFFSET-style tasks, infer the source column, first source row, and step from the provided examples or formulas, then copy the actual source values into the requested output range as literals.
+
+For schedule/calendar fill tasks, build a cycle-day-to-periods mapping from the schedule/template area first, then fill the daily rows across all requested class columns based on each row’s cycle day. Preserve repeated/double periods exactly as shown by the template; do not leave formulas in the schedule cells.
+
+For INDEX/MATCH problems where the first row works but subsequent rows fail, treat row labels, column/year headers, region/type criteria, and expense/category labels as a multi-key lookup. Fill the whole result matrix with values from the source data table, using cached `data_only` values when source cells are formulas.
+
+For multi-step macro/VBA-style requests, implement every stated operation in the workbook, not just the first deletion/filtering step. Re-read the numbered requirements before saving and verify later computed columns, totals, and derived fields as well as the obvious filtered rows.
+
+When a target range includes special rows such as `Total`, `Grand Total`, `min`, `max`, constraints, headers, or blank separators, do not apply ordinary row logic blindly to those rows. Compute totals as aggregates when indicated, and leave constraint/header/blank cells untouched unless explicitly requested.
+
+For residual-balancing tasks, identify data rows separately from min/max constraint rows. Add positive residuals from unit 1 toward unit 5 without exceeding max values; subtract negative residuals from unit 5 toward unit 1 without going below min values; update only the unit cells in actual data rows.
+
+For time-threshold rows, decide per row whether it is a normal data row or a summary row. Normal rows use the before/after threshold rule; summary rows should aggregate the computed normal-row results if the workbook labels or examples indicate a total.
+
+Keep scripts simple enough to run cleanly. Avoid unnecessary dynamic code generation and fragile f-strings with regex expressions inside them. Always execute the final `solution.py`; fix any syntax, indentation, or runtime error, then verify representative target cells.
+
+If workbook cells contain arbitrary sample text that could be sensitive or trigger content filters, do not quote large raw cell contents in your response. Process them locally in Python with neutral variable names and output only the completed script/workbook changes.
+<!-- SLOW_UPDATE_END -->

configs/_base_/default.yaml

@@ -1,12 +1,12 @@
-# ReflACT default configuration — base for all environments.
+# SkillOpt default configuration — base for all environments.
 # Environment configs should inherit via: _base_: default.yaml
 
 model:
   backend: azure_openai
-  teacher: gpt-5.5
-  student: gpt-5.5
-  teacher_backend: openai_chat
-  student_backend: openai_chat
+  optimizer: gpt-5.5
+  target: gpt-5.5
+  optimizer_backend: openai_chat
+  target_backend: openai_chat
   reasoning_effort: medium
   rewrite_reasoning_effort: ""
   rewrite_max_completion_tokens: 64000
@@ -24,25 +24,25 @@ model:
   claude_code_exec_use_sdk: auto
   claude_code_exec_effort: medium
   claude_code_exec_max_thinking_tokens: 16384
-  codex_trace_to_teacher: true
-  azure_openai_endpoint: "https://t2vgoaigpt4o3.openai.azure.com/"
+  codex_trace_to_optimizer: true
+  azure_openai_endpoint: ""           # e.g. "https://your-resource.openai.azure.com/"
   azure_openai_api_version: "2024-12-01-preview"
   azure_openai_api_key: ""       # Fill locally if you do not export AZURE_OPENAI_API_KEY
-  azure_openai_auth_mode: azure_cli
+  azure_openai_auth_mode: ""           # empty → fall back to AZURE_OPENAI_AUTH_MODE env (default "azure_cli")
   azure_openai_ad_scope: "https://cognitiveservices.azure.com/.default"
   azure_openai_managed_identity_client_id: ""
-  teacher_azure_openai_endpoint: "https://t2vgoaigpt4o3.openai.azure.com/"
-  teacher_azure_openai_api_version: "2024-12-01-preview"
-  teacher_azure_openai_api_key: ""
-  teacher_azure_openai_auth_mode: azure_cli
-  teacher_azure_openai_ad_scope: "https://cognitiveservices.azure.com/.default"
-  teacher_azure_openai_managed_identity_client_id: ""
-  student_azure_openai_endpoint: "https://t2vgoaigpt4o3.openai.azure.com/"
-  student_azure_openai_api_version: "2024-12-01-preview"
-  student_azure_openai_api_key: ""
-  student_azure_openai_auth_mode: azure_cli
-  student_azure_openai_ad_scope: "https://cognitiveservices.azure.com/.default"
-  student_azure_openai_managed_identity_client_id: ""
+  optimizer_azure_openai_endpoint: ""   # e.g. "https://your-resource.openai.azure.com/"
+  optimizer_azure_openai_api_version: "2024-12-01-preview"
+  optimizer_azure_openai_api_key: ""
+  optimizer_azure_openai_auth_mode: ""           # empty → fall back to OPTIMIZER_AZURE_OPENAI_AUTH_MODE env, then shared
+  optimizer_azure_openai_ad_scope: "https://cognitiveservices.azure.com/.default"
+  optimizer_azure_openai_managed_identity_client_id: ""
+  target_azure_openai_endpoint: ""   # e.g. "https://your-resource.openai.azure.com/"
+  target_azure_openai_api_version: "2024-12-01-preview"
+  target_azure_openai_api_key: ""
+  target_azure_openai_auth_mode: ""           # empty → fall back to TARGET_AZURE_OPENAI_AUTH_MODE env, then shared
+  target_azure_openai_ad_scope: "https://cognitiveservices.azure.com/.default"
+  target_azure_openai_managed_identity_client_id: ""
 
 train:
   num_epochs: 4
@@ -57,9 +57,6 @@ gradient:
   analyst_workers: 16
   max_analyst_rounds: 3
   failure_only: false
-  use_deep_reflect: false
-  deep_reflect_failures: 4
-  deep_reflect_successes: 2
 
 optimizer:
   learning_rate: 4          # max edits per step (edit_budget)
@@ -67,10 +64,9 @@ optimizer:
   lr_scheduler: cosine      # constant / linear / cosine / autonomous
   lr_control_mode: fixed    # fixed / autonomous / none
   skill_update_mode: patch  # patch / rewrite_from_suggestions / full_rewrite_minibatch
-  use_meta_reflect: false
-  meta_learning_rate: 4     # max edits per epoch-level meta-reflect
   use_slow_update: true
   slow_update_samples: 20
+  slow_update_gate_with_selection: false
   longitudinal_pair_policy: mixed  # mixed / changed / unchanged
   use_meta_skill: true
 
@@ -84,10 +80,9 @@ env:
   name: ""
   skill_init: ""
   split_mode: ratio       # ratio = build deterministic split from data_path; split_dir = use pre-split train/val/test
-  split_ratio: "2:1:7"    # explicit default for dataset-backed benchmarks: train:val:test
   split_seed: 42
   split_dir: ""
   data_path: ""
   split_output_dir: ""
-  exec_timeout: 120      # per student model/code-agent call timeout in seconds
+  exec_timeout: 120      # per target model/code-agent call timeout in seconds
   out_root: ""

configs/ablation_study/README.md

@@ -1,305 +0,0 @@
-# Ablation Study Configuration Manifest
-
-This folder records the final, reproducible settings for the ablation runs used
-in `docs/ablation_paper_tables.md`.
-
-It is intentionally separate from the benchmark default configs. The benchmark
-configs under `configs/<benchmark>/default.yaml` remain the source task configs;
-this folder records the exact matrix-level overrides, run roots, launch commands,
-and validation rules used for the paper ablations.
-
-## Files
-
-- `matrix.yaml`: canonical ablation matrix, common overrides, benchmark splits,
-  token/output caps, and invalid-run rules.
-- `launch_commands.sh`: exact launcher commands for the valid run roots.
-- `validation.md`: monitoring, result extraction, and invalidation checklist.
-
-## Source Of Truth
-
-Use the matrix launcher:
-
-```bash
-/home/azureuser/workspace-gzy/miniconda3/envs/reflact/bin/python scripts/run_ablation_matrix.py
-```
-
-The launcher builds runs from the same defaults and values recorded in
-`matrix.yaml`. It skips completed runs by checking `summary.json` and skips
-active runs by checking `env.out_root` in active `scripts/train.py` processes.
-
-Do not manually rerun a completed run into the same `env.out_root`. If a run is
-invalid, archive or remove its output directory first, then let the launcher
-start it cleanly.
-
-## Current Correct Run Roots
-
-- SearchQA / SpreadsheetBench original ablations:
-  `outputs/ablation_20260502_040604_unique48`
-- SearchQA / SpreadsheetBench batch-size ablations:
-  `outputs/ablation_batch_searchqa_spreadsheet_20260503_153902_run`
-- LiveMathBench / ALFWorld clean ablations:
-  `outputs/ablation_livemath_alfworld_clean_20260503_155155_run`
-- DocVQA ablations:
-  `outputs/ablation_docvqa_20260503_160225_run`
-
-Archived, superseded, misaligned, dry-run, or pre-fix directories must not be
-used for paper tables.
-
-## End-To-End Runbook
-
-### Environment
-
-Run from the repository root:
-
-```bash
-cd /home/azureuser/workspace-gzy/SkillReflection
-```
-
-Always use:
-
-```bash
-PY=/home/azureuser/workspace-gzy/miniconda3/envs/reflact/bin/python
-export ALFWORLD_DATA=/home/azureuser/.cache/alfworld
-```
-
-Default model/auth settings are generated by `scripts/run_ablation_matrix.py`:
-
-```text
-teacher=gpt-5.5
-student=gpt-5.5
-teacher_backend=openai_chat
-student_backend=openai_chat
-reasoning_effort=medium
-teacher/student endpoint=https://t2vgoaigpt4o3.openai.azure.com/
-teacher/student api_version=2024-12-01-preview
-teacher/student auth_mode=azure_cli
-```
-
-Core training settings:
-
-```text
-train.num_epochs=4
-train.train_size=0
-train.batch_size=40
-train.accumulation=1
-train.seed=42
-gradient.minibatch_size=8
-gradient.merge_batch_size=8
-gradient.analyst_workers=16
-gradient.use_deep_reflect=false
-optimizer.learning_rate=4
-optimizer.min_learning_rate=2
-optimizer.lr_scheduler=cosine
-optimizer.lr_control_mode=fixed
-optimizer.use_slow_update=true
-optimizer.slow_update_samples=20
-optimizer.use_meta_skill=true
-optimizer.use_meta_reflect=false
-optimizer.longitudinal_pair_policy=mixed
-evaluation.use_gate=true
-evaluation.eval_test=true
-env.split_mode=split_dir
-```
-
-`train.train_size=0` is intentional. The dataloader derives the train size from
-the fixed split. Batch-size ablations rely on the default `ceil(train_size /
-batch_size)` behavior; the last batch can be smaller than `train.batch_size`.
-
-### Fixed Splits
-
-Default split directories:
-
-```text
-searchqa: data/ablation_splits/searchqa/2-1-7_seed42
-spreadsheetbench: data/ablation_splits/spreadsheetbench/2-1-7_seed42
-livemathematicianbench: data/ablation_splits/livemathematicianbench/2-1-7_seed42
-alfworld: data/ablation_splits/alfworld/2-1-7_seed42
-docvqa: /home/azureuser/zisu/SkillReflection/data/docvqa/splits
-```
-
-Default train/val/test sizes:
-
-| Benchmark | Train | Val | Test |
-| --- | ---: | ---: | ---: |
-| SearchQA | 400 | 200 | 1400 |
-| SpreadsheetBench | 80 | 40 | 280 |
-| LiveMathBench | 35 | 18 | 124 |
-| ALFWorld | 39 | 18 | 134 |
-| DocVQA | 1070 | 535 | 3744 |
-
-DocVQA images are not copied. The valid setup uses:
-
-```text
-data/docvqa_images -> /home/azureuser/zisu/SkillReflection/data/docvqa_images
-```
-
-2026-05-05 DocVQA data correction: all DocVQA final reruns should use the zisu
-10% split above and a fresh output root such as
-`outputs/ablation_docvqa_zisu10pct_20260505_run`. The older local
-`data/ablation_splits/docvqa/2-1-7_seed42` contains the same 5349 questionId
-pool but a different train/val/test assignment, so its completed summaries are
-historical only.
-
-### Matrix Groups
-
-Use these group names with `scripts/run_ablation_matrix.py`:
-
-```text
-default split batch mbs lr sched slown mod smodel longpair lrctrl
-```
-
-`longpair` is the slow-update/meta-skill comparison-example ablation. It keeps
-all prompts and training settings unchanged and only overrides:
-
-```text
-optimizer.longitudinal_pair_policy=changed
-optimizer.longitudinal_pair_policy=unchanged
-```
-
-The default paper setting remains `mixed`.
-
-`lrctrl` contains the two learning-rate-control baselines:
-
-```text
-optimizer.lr_control_mode=autonomous
-optimizer.lr_control_mode=none + optimizer.skill_update_mode=full_rewrite_minibatch
-```
-
-The autonomous run logs the chosen integer per step in `lr_decision.json` and
-`lr_history.jsonl`. The full-rewrite run removes the LR/edit-selection concept:
-each minibatch analyst produces a complete skill candidate, and aggregate/merge
-produces the candidate skill directly.
-
-Batch-size values are:
-
-```text
-8 / 24 / 40 / 56 / full
-```
-
-`40` is the default point. `full` expands to the benchmark train size.
-
-### Launch Commands Used In This Session
-
-The exact commands are recorded in `launch_commands.sh` and in
-`docs/ablation_plan.md`. The important current policy is:
-
-- SearchQA / SpreadsheetBench batch-only matrix can run at `--max-parallel 8`.
-- DocVQA matrix can run with its launcher at `--max-parallel 8`; later top-up used `--max-parallel 16` only because completed runs were skipped and active roots were checked.
-- LiveMathBench is safe as API-only benchmark after the token cap fix.
-- ALFWorld must not be mixed into a 24-way run on this shared machine. Use `--bench alfworld --max-parallel 1` only after memory is available.
-
-### Token And Timeout Fixes
-
-LiveMathBench must use a large student completion cap:
-
-```text
-max_completion_tokens=16384
-timeout=300
-```
-
-The old 768/512 cap produced many empty visible responses because hidden
-reasoning consumed the budget.
-
-ALFWorld must use:
-
-```text
-max_completion_tokens=2048
-empty response fallback -> <action>look</action>
-missing action fallback -> <action>look</action>
-```
-
-### Invalid Runs
-
-Never fill paper tables from these archive directories:
-
-```text
-outputs/ablation_livemath_alfworld_clean_20260503_155155_run/archive_livemath_token768_20260504_022258/
-outputs/ablation_livemath_alfworld_clean_20260503_155155_run/archive_alfworld_token512_20260504_021417/
-outputs/ablation_livemath_alfworld_clean_20260503_155155_run/archive_alfworld_empty_action_20260504_025311/
-outputs/ablation_livemath_alfworld_clean_20260503_155155_run/archive_alfworld_prefallback_20260504_025402/
-outputs/ablation_livemath_alfworld_clean_20260503_155155_run/archive_alfworld_oom_partial_20260504_050517/
-outputs/ablation_livemath_alfworld_clean_20260503_155155_run/archive_alfworld_serial_lowmem_20260504_1300/
-```
-
-### Current Resource Notes
-
-ALFWorld model calls are API calls. The ablation branch now creates local
-ALFWorld/TextWorld environments through multiprocessing workers ported from
-`skillopt_final_zzw`, not through Ray actors. Old Ray-based archived runs are
-not valid for table fill. The observed historical failure mode in this session
-was system RAM pressure and Ray OOM prevention, not model GPU memory.
-
-GPU memory currently shown by `nvidia-smi` came from unrelated Ray Serve visual
-models under:
-
-```text
-/home/azureuser/workspace-gzy/zyf/gca-skill
-```
-
-Those processes are `GroundingDINOModel` / `DA3Model`, not the
-SkillReflection ablation ALFWorld run.
-
-There are also unrelated ALFWorld jobs under:
-
-```text
-/home/azureuser/zisu/skill_distill
-```
-
-Do not confuse those with this repository's ablation outputs.
-
-### Monitoring
-
-Active run and duplicate output-root check:
-
-```bash
-$PY - <<'PY'
-import subprocess, re, collections, time
-try:
-    raw = subprocess.check_output(["pgrep", "-af", "scripts/train.py"], text=True)
-except subprocess.CalledProcessError:
-    raw = ""
-roots = []
-for line in raw.splitlines():
-    m = re.search(r"env\.out_root=([^\s]+)", line)
-    if m:
-        roots.append(m.group(1))
-ctr = collections.Counter(roots)
-print("time", time.strftime("%F %T"))
-print("active_count", len(roots))
-print("duplicates", [r.rsplit("/", 1)[-1] for r, c in ctr.items() if c > 1])
-for root in sorted(roots):
-    print(root.rsplit("/", 1)[-1])
-PY
-```
-
-Error scan:
-
-```bash
-rg -n "Traceback|ERROR|Error code|AuthenticationError|BadRequest|RateLimit|content_filter|Killed|OutOfMemory|CUDA out of memory|\\[FAIL\\]|LLM call failed" \
-  outputs/ablation_docvqa_20260503_160225_run/logs \
-  outputs/ablation_livemath_alfworld_clean_20260503_155155_run/logs \
-  outputs/ablation_batch_searchqa_spreadsheet_20260503_153902_run/logs \
-  -g '*.log' | tail -100 || true
-```
-
-Resource checks:
-
-```bash
-free -h | sed -n '1,3p'
-df -h /tmp
-du -sh /tmp/ray 2>/dev/null || true
-nvidia-smi
-```
-
-### Filling Tables
-
-Only use top-level `summary.json` from valid run roots. Fill
-`docs/ablation_paper_tables.md` from:
-
-```text
-best_selection_hard
-baseline_test_hard
-test_hard
-test_delta_hard
-token_summary._total.total_tokens
-```

configs/ablation_study/launch_commands.sh

@@ -1,81 +0,0 @@
-#!/usr/bin/env bash
-set -euo pipefail
-
-cd /home/azureuser/workspace-gzy/SkillReflection
-
-PY=/home/azureuser/workspace-gzy/miniconda3/envs/reflact/bin/python
-export ALFWORLD_DATA=/home/azureuser/.cache/alfworld
-
-# Original SearchQA / SpreadsheetBench full matrix reproduction command.
-# Do not run this into the existing root unless intentionally reproducing from
-# scratch; the current valid root is already populated:
-# outputs/ablation_20260502_040604_unique48
-#
-# setsid "$PY" scripts/run_ablation_matrix.py \
-#   --groups default split mbs lr sched slown mod smodel \
-#   --bench searchqa spreadsheetbench \
-#   --run-root /home/azureuser/workspace-gzy/SkillReflection/outputs/ablation_20260502_040604_unique48 \
-#   --max-parallel 24 \
-#   --execute \
-#   > /home/azureuser/workspace-gzy/SkillReflection/outputs/ablation_20260502_040604_unique48/launcher_reproduce_full_matrix.log 2>&1 < /dev/null &
-#
-# SearchQA / SpreadsheetBench batch-size ablations only.
-# Original non-batch SearchQA/SpreadsheetBench ablations live in:
-# outputs/ablation_20260502_040604_unique48
-setsid "$PY" scripts/run_ablation_matrix.py \
-  --groups batch \
-  --bench searchqa spreadsheetbench \
-  --run-root /home/azureuser/workspace-gzy/SkillReflection/outputs/ablation_batch_searchqa_spreadsheet_20260503_153902_run \
-  --max-parallel 8 \
-  --execute \
-  > /home/azureuser/workspace-gzy/SkillReflection/outputs/ablation_batch_searchqa_spreadsheet_20260503_153902_run/launcher_parallel8.log 2>&1 < /dev/null &
-
-# DocVQA full matrix.
-setsid "$PY" scripts/run_ablation_matrix.py \
-  --groups default split batch mbs lr sched slown mod smodel \
-  --bench docvqa \
-  --run-root /home/azureuser/workspace-gzy/SkillReflection/outputs/ablation_docvqa_20260503_160225_run \
-  --max-parallel 8 \
-  --execute \
-  > /home/azureuser/workspace-gzy/SkillReflection/outputs/ablation_docvqa_20260503_160225_run/launcher_parallel8.log 2>&1 < /dev/null &
-
-# LiveMathBench clean matrix. ALFWorld should be launched separately at lower
-# concurrency because Ray OOM occurred when many ALFWorld runs were mixed into a
-# 24-way run.
-setsid "$PY" scripts/run_ablation_matrix.py \
-  --groups default split batch mbs lr sched slown mod smodel \
-  --bench livemathematicianbench \
-  --run-root /home/azureuser/workspace-gzy/SkillReflection/outputs/ablation_livemath_alfworld_clean_20260503_155155_run \
-  --max-parallel 8 \
-  --execute \
-  > /home/azureuser/workspace-gzy/SkillReflection/outputs/ablation_livemath_alfworld_clean_20260503_155155_run/launcher_livemath_parallel8.log 2>&1 < /dev/null &
-
-# ALFWorld clean matrix. Increase to 2 only after checking memory, /tmp/ray,
-# and that no other ALFWorld run is active. Do not use 8/16/24 for ALFWorld on
-# the current shared machine unless resources are explicitly reserved.
-setsid "$PY" scripts/run_ablation_matrix.py \
-  --groups default split batch mbs lr sched slown mod smodel \
-  --bench alfworld \
-  --run-root /home/azureuser/workspace-gzy/SkillReflection/outputs/ablation_livemath_alfworld_clean_20260503_155155_run \
-  --max-parallel 1 \
-  --execute \
-  > /home/azureuser/workspace-gzy/SkillReflection/outputs/ablation_livemath_alfworld_clean_20260503_155155_run/launcher_alfworld_parallel1.log 2>&1 < /dev/null &
-
-# Longitudinal comparison-example policy ablations. This intentionally excludes
-# ALFWorld. The only varied setting is optimizer.longitudinal_pair_policy.
-setsid "$PY" scripts/run_ablation_matrix.py \
-  --groups longpair \
-  --bench searchqa spreadsheetbench livemathematicianbench docvqa \
-  --run-root /home/azureuser/workspace-gzy/SkillReflection/outputs/ablation_longpair_20260504_run \
-  --max-parallel 8 \
-  --execute \
-  > /home/azureuser/workspace-gzy/SkillReflection/outputs/ablation_longpair_20260504_run/launcher_longpair_parallel8.log 2>&1 < /dev/null &
-
-# Learning-rate-control baselines. This intentionally excludes ALFWorld.
-setsid "$PY" scripts/run_ablation_matrix.py \
-  --groups lrctrl \
-  --bench searchqa spreadsheetbench livemathematicianbench docvqa \
-  --run-root /home/azureuser/workspace-gzy/SkillReflection/outputs/ablation_lrctrl_20260504_run \
-  --max-parallel 8 \
-  --execute \
-  > /home/azureuser/workspace-gzy/SkillReflection/outputs/ablation_lrctrl_20260504_run/launcher_lrctrl_parallel8.log 2>&1 < /dev/null &

configs/ablation_study/matrix.yaml

@@ -1,257 +0,0 @@
-version: 2026-05-04
-purpose: "Canonical paper ablation settings matching the valid current runs."
-
-launcher:
-  script: scripts/run_ablation_matrix.py
-  python: /home/azureuser/workspace-gzy/miniconda3/envs/reflact/bin/python
-  skip_completed_by: summary.json
-  skip_active_by: "active scripts/train.py env.out_root"
-
-environment:
-  working_directory: /home/azureuser/workspace-gzy/SkillReflection
-  required_env:
-    ALFWORLD_DATA: /home/azureuser/.cache/alfworld
-  docvqa_images_symlink: "data/docvqa_images -> /home/azureuser/zisu/SkillReflection/data/docvqa_images"
-
-common_overrides:
-  model.teacher_backend: openai_chat
-  model.student_backend: openai_chat
-  model.teacher: gpt-5.5
-  model.student: gpt-5.5
-  model.teacher_azure_openai_endpoint: https://t2vgoaigpt4o3.openai.azure.com/
-  model.teacher_azure_openai_api_version: 2024-12-01-preview
-  model.teacher_azure_openai_auth_mode: azure_cli
-  model.student_azure_openai_endpoint: https://t2vgoaigpt4o3.openai.azure.com/
-  model.student_azure_openai_api_version: 2024-12-01-preview
-  model.student_azure_openai_auth_mode: azure_cli
-  model.reasoning_effort: medium
-  train.num_epochs: 4
-  train.train_size: 0
-  train.batch_size: 40
-  train.accumulation: 1
-  train.seed: 42
-  gradient.minibatch_size: 8
-  gradient.merge_batch_size: 8
-  gradient.analyst_workers: 16
-  gradient.use_deep_reflect: false
-  optimizer.learning_rate: 4
-  optimizer.min_learning_rate: 2
-  optimizer.lr_scheduler: cosine
-  optimizer.skill_update_mode: patch
-  optimizer.use_slow_update: true
-  optimizer.slow_update_samples: 20
-  optimizer.use_meta_skill: true
-  optimizer.use_meta_reflect: false
-  evaluation.use_gate: true
-  evaluation.eval_test: true
-  env.split_mode: split_dir
-
-benchmarks:
-  searchqa:
-    config: configs/searchqa/default.yaml
-    run_roots:
-      original_matrix: outputs/ablation_20260502_040604_unique48
-      batch_matrix: outputs/ablation_batch_searchqa_spreadsheet_20260503_153902_run
-    default_split: data/ablation_splits/searchqa/2-1-7_seed42
-    train: 400
-    val: 200
-    test: 1400
-    student_rollout:
-      function: reflact/envs/searchqa/rollout.py::chat_student
-      max_completion_tokens:
-        first_turn: 512
-        refinement: 512
-      rationale: "Short-answer QA; sampled empties are low and not LiveMath-like."
-
-  spreadsheetbench:
-    config: configs/spreadsheetbench/default.yaml
-    run_roots:
-      original_matrix: outputs/ablation_20260502_040604_unique48
-      batch_matrix: outputs/ablation_batch_searchqa_spreadsheet_20260503_153902_run
-    default_split: data/ablation_splits/spreadsheetbench/2-1-7_seed42
-    train: 80
-    val: 40
-    test: 280
-    student_rollout:
-      function: reflact/envs/spreadsheetbench/codegen_agent.py::run_multi
-      max_output_tokens: 16384
-      result_note: "results.jsonl stores execution fields, not a response field."
-
-  livemathematicianbench:
-    config: configs/livemathematicianbench/default.yaml
-    run_roots:
-      clean_matrix: outputs/ablation_livemath_alfworld_clean_20260503_155155_run
-    default_split: data/ablation_splits/livemathematicianbench/2-1-7_seed42
-    train: 35
-    val: 18
-    test: 124
-    student_rollout:
-      function: reflact/envs/livemathematicianbench/rollout.py::chat_student
-      max_completion_tokens:
-        first_turn: 16384
-        refinement: 16384
-      timeout_seconds: 300
-      invalid_old_caps:
-        first_turn: 768
-        refinement: 512
-      invalid_archive: outputs/ablation_livemath_alfworld_clean_20260503_155155_run/archive_livemath_token768_20260504_022258
-      rationale: "GPT-5 reasoning consumed small budgets and produced many empty visible responses."
-
-  alfworld:
-    config: configs/alfworld/default.yaml
-    run_roots:
-      clean_matrix: outputs/ablation_livemath_alfworld_clean_20260503_155155_run
-    default_split: data/ablation_splits/alfworld/2-1-7_seed42
-    train: 39
-    val: 18
-    test: 134
-    student_rollout:
-      function: reflact/envs/alfworld/rollout.py::chat_student
-      max_completion_tokens: 2048
-      timeout_seconds: 120
-      max_steps: 50
-      fallback_action: look
-      invalid_old_cap: 512
-      invalid_archives:
-        - outputs/ablation_livemath_alfworld_clean_20260503_155155_run/archive_alfworld_token512_20260504_021417
-        - outputs/ablation_livemath_alfworld_clean_20260503_155155_run/archive_alfworld_empty_action_20260504_025311
-        - outputs/ablation_livemath_alfworld_clean_20260503_155155_run/archive_alfworld_prefallback_20260504_025402
-        - outputs/ablation_livemath_alfworld_clean_20260503_155155_run/archive_alfworld_oom_partial_20260504_050517
-      concurrency_note: "Do not mix many ALFWorld runs into 24-way total concurrency; Ray OOM occurred. Prefer 1-2 ALFWorld runs at a time unless resources are clearly free."
-
-  docvqa:
-    config: configs/docvqa/default.yaml
-    run_roots:
-      matrix: outputs/ablation_docvqa_zisu10pct_20260505_run
-    default_split: /home/azureuser/zisu/SkillReflection/data/docvqa/splits
-    train: 1070
-    val: 535
-    test: 3744
-    data_note: "2026-05-05: use zisu-provided 5349-item DocVQA split directly; previous local 2-1-7_seed42 used the same item pool but a different train/val/test assignment and must not be used for final DocVQA reruns."
-    student_rollout:
-      function: reflact/envs/docvqa/rollout.py::chat_student_messages
-      max_completion_tokens:
-        first_turn: 768
-        refinement: 512
-      rationale: "Short-answer VQA output; preserve current setting for alignment unless explicitly rerunning all affected DocVQA."
-
-splits:
-  tags:
-    1shot:
-      extra_overrides:
-        optimizer.slow_update_samples: 1
-    1-1-8: {}
-    2-1-7:
-      default: true
-    4-1-5: {}
-  paths:
-    searchqa:
-      1shot: data/ablation_splits/searchqa/1shot_seed42
-      1-1-8: data/ablation_splits/searchqa/1-1-8_seed42
-      2-1-7: data/ablation_splits/searchqa/2-1-7_seed42
-      4-1-5: data/ablation_splits/searchqa/4-1-5_seed42
-    spreadsheetbench:
-      1shot: data/ablation_splits/spreadsheetbench/1shot_seed42
-      1-1-8: data/ablation_splits/spreadsheetbench/1-1-8_seed42
-      2-1-7: data/ablation_splits/spreadsheetbench/2-1-7_seed42
-      4-1-5: data/ablation_splits/spreadsheetbench/4-1-5_seed42
-    livemathematicianbench:
-      1shot: data/ablation_splits/livemathematicianbench/1shot_seed42
-      1-1-8: data/ablation_splits/livemathematicianbench/1-1-8_seed42
-      2-1-7: data/ablation_splits/livemathematicianbench/2-1-7_seed42
-      4-1-5: data/ablation_splits/livemathematicianbench/4-1-5_seed42
-    alfworld:
-      1shot: data/ablation_splits/alfworld/1shot_seed42
-      1-1-8: data/ablation_splits/alfworld/1-1-8_seed42
-      2-1-7: data/ablation_splits/alfworld/2-1-7_seed42
-      4-1-5: data/ablation_splits/alfworld/4-1-5_seed42
-    docvqa:
-      1shot: data/ablation_splits/docvqa/1shot_seed42
-      1-1-8: data/ablation_splits/docvqa/1-1-8_seed42
-      2-1-7: /home/azureuser/zisu/SkillReflection/data/docvqa/splits
-      4-1-5: data/ablation_splits/docvqa/4-1-5_seed42
-
-groups:
-  default:
-    run_id: "DEFAULT-{benchmark}-5.5"
-    overrides: {}
-  split:
-    values: [1shot, 1-1-8, 4-1-5]
-    skip_default_2_1_7: true
-    override_template: "env.split_dir={split_path}"
-  batch:
-    values: [8, 24, 56, full]
-    default_value_reused: 40
-    full_values:
-      searchqa: 400
-      spreadsheetbench: 80
-      livemathematicianbench: 35
-      alfworld: 39
-      docvqa: 1070
-    fixed_overrides:
-      gradient.minibatch_size: 8
-  mbs:
-    values: [1, 2, 4, 16, 32]
-    default_value_reused: 8
-    override_template: "gradient.minibatch_size={value}"
-  lr:
-    values: [1, 2, 4, 8, 16]
-    fixed_overrides:
-      optimizer.lr_scheduler: constant
-      optimizer.min_learning_rate: 1
-    override_template: "optimizer.learning_rate={value}"
-  sched:
-    values: [constant, linear]
-    default_value_reused: cosine
-    override_template: "optimizer.lr_scheduler={value}"
-  slown:
-    values: [5, 10, 40]
-    default_value_reused: 20
-    override_template: "optimizer.slow_update_samples={value}"
-  mod:
-    values:
-      slow-only:
-        optimizer.use_slow_update: true
-        optimizer.use_meta_skill: false
-      meta-only:
-        optimizer.use_slow_update: false
-        optimizer.use_meta_skill: true
-      none:
-        optimizer.use_slow_update: false
-        optimizer.use_meta_skill: false
-    default_value_reused: slow-meta
-  longpair:
-    values: [changed, unchanged]
-    default_value_reused: mixed
-    override_template: "optimizer.longitudinal_pair_policy={value}"
-    note: "Only changes slow-update/meta-skill comparison examples; prompts and other settings remain unchanged."
-  lrctrl:
-    values:
-      autonomous:
-        optimizer.lr_control_mode: autonomous
-      full-rewrite:
-        optimizer.lr_control_mode: none
-        optimizer.skill_update_mode: full_rewrite_minibatch
-    default_value_reused: "fixed patch learning_rate=4"
-    note: "autonomous records lr_decision.json/lr_history.jsonl; full-rewrite removes LR/select/apply-edit and uses full skill candidates."
-  smodel:
-    values:
-      "5.4":
-        model.student: gpt-5.4-pro
-        model.student_azure_openai_endpoint: https://t2vgoaigpt4o3.openai.azure.com/
-        model.student_azure_openai_api_version: 2025-03-01-preview
-        model.student_azure_openai_auth_mode: azure_cli
-      "5.4-mini":
-        model.student: gpt-5.4-mini
-        model.student_azure_openai_endpoint: https://searchagent5.cognitiveservices.azure.com/
-        model.student_azure_openai_api_version: 2024-12-01-preview
-        model.student_azure_openai_auth_mode: azure_cli
-    default_value_reused: "5.5"
-
-validity_rules:
-  use_for_tables:
-    - "Only runs with summary.json in valid run roots."
-    - "Do not use archive, archived, MISALIGNED, SUPERSEDED, dryrun, smoke, or debug directories."
-    - "Do not use ALFWorld runs started before empty/missing-action fallback."
-    - "Do not use old LiveMath runs with 768/512 token caps."
-  rerun_rule: "Archive or remove invalid out_root before relaunch; never write a rerun into a polluted output directory."

configs/ablation_study/validation.md

@@ -1,141 +0,0 @@
-# Ablation Validation Checklist
-
-Use this checklist before launch, during monitoring, and before filling
-`docs/ablation_paper_tables.md`.
-
-## Before Launch
-
-Run from repo root:
-
-```bash
-cd /home/azureuser/workspace-gzy/SkillReflection
-export ALFWORLD_DATA=/home/azureuser/.cache/alfworld
-```
-
-Verify syntax for edited files:
-
-```bash
-/home/azureuser/workspace-gzy/miniconda3/envs/reflact/bin/python -m py_compile \
-  scripts/run_ablation_matrix.py \
-  scripts/train.py \
-  reflact/model/azure_openai.py \
-  reflact/envs/searchqa/rollout.py \
-  reflact/envs/spreadsheetbench/rollout.py \
-  reflact/envs/livemathematicianbench/rollout.py \
-  reflact/envs/alfworld/rollout.py \
-  reflact/envs/docvqa/rollout.py
-```
-
-Check active runs and duplicate `env.out_root` before starting more:
-
-```bash
-/home/azureuser/workspace-gzy/miniconda3/envs/reflact/bin/python - <<'PY'
-import subprocess, re, collections
-try:
-    raw = subprocess.check_output(["pgrep", "-af", "scripts/train.py"], text=True)
-except subprocess.CalledProcessError:
-    raw = ""
-roots = []
-for line in raw.splitlines():
-    m = re.search(r"env\.out_root=([^\s]+)", line)
-    if m:
-        roots.append(m.group(1))
-ctr = collections.Counter(roots)
-print("train_count", len(roots))
-print("duplicate_roots", [r.rsplit("/", 1)[-1] for r, c in ctr.items() if c > 1])
-for root in sorted(roots):
-    print(root.rsplit("/", 1)[-1])
-PY
-```
-
-## During Monitoring
-
-Check launchers:
-
-```bash
-pgrep -af 'scripts/run_ablation_matrix.py' || true
-tail -80 outputs/ablation_docvqa_20260503_160225_run/launcher_parallel8.log 2>/dev/null || true
-tail -80 outputs/ablation_livemath_alfworld_clean_20260503_155155_run/launcher_livemath_parallel8.log 2>/dev/null || true
-tail -80 outputs/ablation_livemath_alfworld_clean_20260503_155155_run/launcher_alfworld_parallel1.log 2>/dev/null || true
-```
-
-Scan current logs for new hard failures:
-
-```bash
-rg -n "Traceback|ERROR|Error code|AuthenticationError|BadRequest|RateLimit|content_filter|Killed|OutOfMemory|\\[FAIL\\]|\\[RETRY\\]" \
-  outputs/ablation_docvqa_20260503_160225_run/logs \
-  outputs/ablation_livemath_alfworld_clean_20260503_155155_run/logs \
-  outputs/ablation_batch_searchqa_spreadsheet_20260503_153902_run/logs \
-  -g '*.log' | tail -160 || true
-```
-
-Check resource pressure:
-
-```bash
-df -h /tmp
-du -sh /tmp/ray 2>/dev/null || true
-free -h | sed -n '1,3p'
-```
-
-## Quality Checks
-
-LiveMathBench current valid runs should not look like old 768/512 runs:
-
-```bash
-/home/azureuser/workspace-gzy/miniconda3/envs/reflact/bin/python - <<'PY'
-import json, pathlib
-root = pathlib.Path("outputs/ablation_livemath_alfworld_clean_20260503_155155_run")
-for run in sorted(root.glob("*livemathematicianbench*")):
-    if not run.is_dir() or "archive" in str(run):
-        continue
-    for rel in ["test_eval_baseline/results.jsonl", "test_eval/results.jsonl"]:
-        p = run / rel
-        if not p.exists():
-            continue
-        rows = [json.loads(l) for l in p.open(errors="ignore") if l.strip()]
-        empty = sum(1 for r in rows if not str(r.get("response", "")).strip())
-        answer = sum(1 for r in rows if "<answer>" in str(r.get("response", "")).lower())
-        if empty:
-            print(run.name, rel, "empty", empty, "answer", answer, "n", len(rows))
-PY
-```
-
-ALFWorld valid runs must not contain empty action or missing action:
-
-```bash
-/home/azureuser/workspace-gzy/miniconda3/envs/reflact/bin/python - <<'PY'
-import json, pathlib
-root = pathlib.Path("outputs/ablation_livemath_alfworld_clean_20260503_155155_run")
-for run in sorted(root.glob("*alfworld*")):
-    if not run.is_dir() or "archive" in str(run):
-        continue
-    bad = []
-    fallback = 0
-    for c in run.glob("**/conversation.json"):
-        data = json.load(c.open(errors="ignore"))
-        for step in data:
-            if step.get("step") is None:
-                continue
-            if not step.get("action"):
-                bad.append(str(c.relative_to(run)))
-                break
-            mr = str(step.get("model_response", ""))
-            if "empty model response" in mr or "missing action tag" in mr:
-                fallback += 1
-    print(run.name, "bad_action_files", len(bad), "fallback", fallback)
-PY
-```
-
-## Filling Tables
-
-Use only `summary.json` fields:
-
-- `best_selection_hard` -> Best Sel
-- `baseline_test_hard` -> Base Test
-- `test_hard` -> Best Test
-- `test_delta_hard` -> Delta
-- `total_accepts` -> Accept
-- `total_rejects` -> Reject
-- `token_summary._total.total_tokens` -> Tokens
-
-Do not fill table rows from logs alone.

configs/alfworld/default.yaml

@@ -10,7 +10,6 @@ gradient:
 
 optimizer:
   learning_rate: 4
-  use_meta_reflect: false
 
 evaluation:
   sel_env_num: 0
@@ -18,13 +17,13 @@ evaluation:
 
 env:
   name: alfworld
-  skill_init: reflact/envs/alfworld/skills/initial.md
+  skill_init: skillopt/envs/alfworld/skills/initial.md
   split_mode: split_dir
-  split_ratio: "2:1:7"
   split_dir: data/ablation_splits/alfworld/2-1-7_seed42
   data_path: ""
   split_output_dir: ""
   max_steps: 50
+  max_completion_tokens: 16384
   workers: 8
   max_api_workers: 8
   limit: 0

configs/alfworld/meta_reflect.yaml

@@ -1,4 +0,0 @@
-_base_: default.yaml
-
-optimizer:
-  use_meta_reflect: true

configs/babyvision/default.yaml

@@ -1,21 +0,0 @@
-_base_: ../_base_/default.yaml
-
-train:
-  batch_size: 64
-  accumulation: 1
-
-env:
-  name: babyvision
-  skill_init: reflact/envs/babyvision/skills/initial.md
-  split_mode: ratio
-  split_ratio: "2:1:7"
-  split_dir: ""
-  data_path: ""
-  split_output_dir: ""
-  max_turns: 1
-  workers: 16
-  limit: 0
-  image_detail: auto
-  judge_model: gpt-5.4
-  judge_max_completion_tokens: 256
-  judge_retries: 5

configs/docvqa/default.yaml

@@ -16,13 +16,13 @@ optimizer:
 
 env:
   name: docvqa
-  skill_init: reflact/envs/docvqa/skills/initial.md
+  skill_init: skillopt/envs/docvqa/skills/initial.md
   split_mode: split_dir
-  split_ratio: "2:1:7"
-  split_dir: /home/azureuser/zisu/SkillReflection/data/docvqa/splits
+  split_dir: data/docvqa/splits
   data_path: ""
   split_output_dir: ""
   max_turns: 1
+  max_completion_tokens: 16384
   workers: 16
   image_detail: auto
   limit: 0

configs/examples/soft_gate.yaml

@@ -0,0 +1,47 @@
+# ─────────────────────────────────────────────────────────────────────────────
+# Example: soft / mixed validation-gate metric (community-contributed, PR #25)
+# ─────────────────────────────────────────────────────────────────────────────
+#
+# This is NOT a default SkillOpt setting and was NOT used to produce the
+# numbers reported in the paper. It is provided as a reference for users
+# who encounter a specific scenario where the default `hard` gate is too
+# coarse to drive training.
+#
+# When to consider this:
+#   - You are running on a custom environment.
+#   - Your held-out *selection* split has very few items (e.g. ≤ ~10).
+#   - Your reward function is continuous / partial-credit (e.g. F1, BLEU,
+#     soft match) rather than purely binary 0/1.
+#
+# Symptom this addresses:
+#   With a small selection split + continuous rewards, candidate skills
+#   often improve per-item soft scores (e.g. 0.06 → 0.26 on one item) but
+#   never flip the discrete hard outcome. The default `hard` gate then
+#   rejects every candidate and training stalls. Switching the gate to
+#   `soft` or `mixed` lets these partial improvements be accepted.
+#
+# When NOT to use this:
+#   - When reproducing the paper. The paper-reported numbers were obtained
+#     under the default `hard` gate.
+#   - When your selection split is large (dozens+ items) and / or your
+#     reward is already binary — `hard` is the more conservative choice
+#     and matches the design described in the paper.
+#
+# To use: inherit your env config from this file, e.g.
+#   _base_: ../examples/soft_gate.yaml
+# or copy the `evaluation:` block below into your config.
+# ─────────────────────────────────────────────────────────────────────────────
+
+_base_: ../_base_/default.yaml
+
+evaluation:
+  # Three options:
+  #   'hard'  — default; exact-match accuracy. Use this to reproduce the paper.
+  #   'soft'  — per-item soft / partial-credit score (recommended for the
+  #             small-split + continuous-reward scenario described above).
+  #   'mixed' — weighted average: (1 - w) * hard + w * soft, with `w` set by
+  #             `gate_mixed_weight` below.
+  gate_metric: soft
+
+  # Only used when gate_metric == 'mixed'. Ignored otherwise.
+  gate_mixed_weight: 0.5

configs/livemathematicianbench/default.yaml

@@ -7,13 +7,13 @@ train:
 
 env:
   name: livemathematicianbench
-  skill_init: reflact/envs/livemathematicianbench/skills/initial.md
+  skill_init: skillopt/envs/livemathematicianbench/skills/initial.md
   split_mode: split_dir
-  split_ratio: "2:1:7"
   split_dir: data/ablation_splits/livemathematicianbench/2-1-7_seed42
   data_path: ""
   split_output_dir: ""
   max_turns: 1
+  max_completion_tokens: 16384
   exec_timeout: 300
   workers: 64
   limit: 0

configs/mathverse/default.yaml

@@ -1,23 +0,0 @@
-_base_: ../_base_/default.yaml
-
-model:
-  codex_exec_sandbox: danger-full-access
-
-train:
-  batch_size: 64
-  accumulation: 1
-
-env:
-  name: mathverse
-  skill_init: reflact/envs/mathverse/skills/initial.md
-  split_dir: ""
-  data_root: data/MathVerse
-  problem_version: Text Lite
-  use_text_dominant_reference: false
-  max_turns: 1
-  workers: 16
-  limit: 0
-  image_detail: auto
-  judge_model: gpt-5.4
-  judge_max_completion_tokens: 256
-  judge_retries: 5

configs/mmrb/default.yaml

@@ -1,18 +0,0 @@
-_base_: ../_base_/default.yaml
-
-train:
-  batch_size: 128
-  accumulation: 1
-
-env:
-  name: mmrb
-  skill_init: reflact/envs/mmrb/skills/initial.md
-  split_mode: ratio
-  split_ratio: "2:1:7"
-  split_dir: ""
-  data_path: ""
-  split_output_dir: ""
-  max_turns: 1
-  workers: 16
-  limit: 0
-  image_detail: auto

configs/officeqa/default.yaml

@@ -16,10 +16,19 @@ optimizer:
 
 env:
   name: officeqa
-  skill_init: reflact/envs/officeqa/skills/initial.md
+  skill_init: skillopt/envs/officeqa/skills/initial.md
+  split_mode: split_dir
   split_dir: data/officeqa_split
   data_dirs:
     - data/officeqa_docs_official
   workers: 4
   max_tool_turns: 24
+  max_completion_tokens: 16384
+  search_mode: offline
+  max_queries_per_turn: 4
+  search_api_url: http://apisix.westus2.cloudapp.azure.com/search_tool/search
+  search_auth_env: OFFICEQA_CUSTOM_SEARCH_AUTH
+  search_provider: duckduckgo
+  search_max_num_results: 4
+  search_timeout_seconds: 20
   limit: 0

configs/sealqa/default.yaml

@@ -1,23 +0,0 @@
-_base_: ../_base_/default.yaml
-
-model:
-  reasoning_effort: medium
-
-train:
-  batch_size: 10
-  accumulation: 1
-
-gradient:
-  minibatch_size: 8
-  merge_batch_size: 8
-
-optimizer:
-  learning_rate: 4
-
-env:
-  name: sealqa
-  skill_init: reflact/envs/sealqa/skills/initial.md
-  split_dir: data/sealqa_split
-  workers: 4
-  max_tool_turns: 12
-  limit: 0

configs/searchqa/default.yaml

@@ -21,12 +21,12 @@ evaluation:
 
 env:
   name: searchqa
-  skill_init: reflact/envs/searchqa/skills/initial.md
+  skill_init: skillopt/envs/searchqa/skills/initial.md
   split_mode: split_dir
-  split_ratio: "2:1:7"
   split_dir: data/searchqa_split
   data_path: ""
   split_output_dir: ""
   max_turns: 1
+  max_completion_tokens: 16384
   workers: 24
   limit: 0

configs/spreadsheetbench/default.yaml

@@ -21,14 +21,14 @@ evaluation:
 
 env:
   name: spreadsheetbench
-  skill_init: reflact/envs/spreadsheetbench/skills/initial.md
+  skill_init: skillopt/envs/spreadsheetbench/skills/initial.md
   split_mode: split_dir
-  split_ratio: "2:1:7"
   split_dir: data/spreadsheetbench_split
   data_path: ""
   split_output_dir: ""
   data_root: data/spreadsheetbench_verified_400
   mode: multi
   max_turns: 30
+  max_completion_tokens: 16384
   exec_timeout: 600
   workers: 24

configs/swebench/default.yaml

@@ -1,36 +0,0 @@
-_base_: ../_base_/default.yaml
-
-model:
-  reasoning_effort: medium
-
-train:
-  batch_size: 20
-  accumulation: 1
-
-gradient:
-  minibatch_size: 4
-  merge_batch_size: 8
-
-optimizer:
-  learning_rate: 4
-
-evaluation:
-  sel_env_num: 0
-  test_env_num: 0
-
-env:
-  name: swebench
-  skill_init: reflact/envs/swebench/skills/initial.md
-  split_mode: ratio
-  split_ratio: "2:1:7"
-  split_dir: ""
-  data_path: ""
-  split_output_dir: ""
-  dataset_name: lite
-  hf_split: test
-  workers: 8
-  eval_workers: 8
-  step_limit: 50
-  cost_limit: 3.0
-  timeout_per_instance: 600
-  limit: 0

data/searchqa_id_split/val/sel.json

@@ -0,0 +1,602 @@
+[
+  {
+    "id": "1758dc50625e46ee814e44a6061f091d"
+  },
+  {
+    "id": "52967b20d9564d5b94756bc0b3d113d6"
+  },
+  {
+    "id": "e77dbfc0d45e476b9ed5b2204b6d9fc8"
+  },
+  {
+    "id": "675f01b5343e46599626104968c0f45a"
+  },
+  {
+    "id": "dccd0f04cd3b41bc8c4182e41a413b3e"
+  },
+  {
+    "id": "efde5820a29949488feaa18ce7e429f9"
+  },
+  {
+    "id": "98773558971e490d8de41088e41a3207"
+  },
+  {
+    "id": "70e3e1e1d41f4795ba45678d5f4b6524"
+  },
+  {
+    "id": "206a9ffb8ed24686a24db1492c9f10dc"
+  },
+  {
+    "id": "e91f1225cdb84ee7a2d773b5177945c1"
+  },
+  {
+    "id": "b68a14a7b43842a0a50a3965d3bbffbb"
+  },
+  {
+    "id": "e2e679989b544e94bf8dbc3466e49003"
+  },
+  {
+    "id": "c51c26c07b384bbe865976b81a86736d"
+  },
+  {
+    "id": "83928d21f3504843aa6294ce9d9c8da1"
+  },
+  {
+    "id": "f822cde36d864c508b426f3e2b5245b2"
+  },
+  {
+    "id": "f03757766e034c539bf2595428c03626"
+  },
+  {
+    "id": "349be67be2e141438287d21cb6a94131"
+  },
+  {
+    "id": "7f3e18272c6d43dd8c6385c4758b685e"
+  },
+  {
+    "id": "66dc8aee931142b1834da63363511491"
+  },
+  {
+    "id": "52ac78645e144b4598958d67920617dc"
+  },
+  {
+    "id": "32a66c67a8d7490b8572387d8f1d762a"
+  },
+  {
+    "id": "b028bea99d2641268532cd2a32d4ad2d"
+  },
+  {
+    "id": "c9f3a971abe64853aca3e0bb1fea6962"
+  },
+  {
+    "id": "5324041c25d04316a800ccba2d79a2c7"
+  },
+  {
+    "id": "9960a3f1ac954282a7ea1d7f7c8a86f9"
+  },
+  {
+    "id": "03bd8fc86a9d465e9370dcdaedb3c422"
+  },
+  {
+    "id": "a1d0d03fcac54483a1e88542b767afd5"
+  },
+  {
+    "id": "c21a41745ab9402da8b1deea67055a8d"
+  },
+  {
+    "id": "d5c48bd600824c8a9328df80a8baf964"
+  },
+  {
+    "id": "746806ac0b1c48cb80d7ab5ad0a63517"
+  },
+  {
+    "id": "e3b96fc1b99247ebada09a84fa5fd205"
+  },
+  {
+    "id": "bbeeec20466a4dab9e4d73d86c964cba"
+  },
+  {
+    "id": "2b85b097d130404daeb26a3349bdc297"
+  },
+  {
+    "id": "fbaae95260474ed58ccbb05ac482089b"
+  },
+  {
+    "id": "28b837ac304845bd961a46cedc9246af"
+  },
+  {
+    "id": "691e3a24e5994c749ff11c6bc43b2d1c"
+  },
+  {
+    "id": "12d5b51c793a43b1aa3b7e1bf5f4bc73"
+  },
+  {
+    "id": "a3a3892296c74fc4b23b2475156901b1"
+  },
+  {
+    "id": "8065f2e8639848f5b2f16a4f2877b7c2"
+  },
+  {
+    "id": "b06e0d2be0a5465ea12ff2bf2b27d387"
+  },
+  {
+    "id": "e50328dc70634863ac39580ef6f8dd5a"
+  },
+  {
+    "id": "17b441db9b444435951d859e6e081f3b"
+  },
+  {
+    "id": "0fc37a4bd8424ecf8eede56e956edb3a"
+  },
+  {
+    "id": "09c769c7262e43eb95a52e874042c02a"
+  },
+  {
+    "id": "2417ce75564b4b8eab1ddd755b10d962"
+  },
+  {
+    "id": "8f4beaa8a7704eba93f753a0e7040165"
+  },
+  {
+    "id": "6d8e99531264435badb67bf6ac209075"
+  },
+  {
+    "id": "940ddc06648f4c9bac6b3bf5df402e5a"
+  },
+  {
+    "id": "fa0fcf36e99644d0b51be95869d87adc"
+  },
+  {
+    "id": "1882f20b777146608474ac0056d54a5e"
+  },
+  {
+    "id": "439e908aa7c7422daf5632af7e0440ec"
+  },
+  {
+    "id": "2d90c36ab1dd4724b1c108b2507316e4"
+  },
+  {
+    "id": "efa5cc2f750c421a807bc4315db0dc1e"
+  },
+  {
+    "id": "3b1c596e1720499ebe790d14f34a67af"
+  },
+  {
+    "id": "c87234a3a1074b5ea5500114ad9aa16d"
+  },
+  {
+    "id": "9d08ada791844519b7a4bb247d50c89c"
+  },
+  {
+    "id": "74c45e35e5a740798fbd528ff08a70e3"
+  },
+  {
+    "id": "650b683d39dd45c99ad7435fc7d33fe3"
+  },
+  {
+    "id": "191ce59c97e143f7b4c33400913d3570"
+  },
+  {
+    "id": "74441a0f7dae42ca9ee19d8be8c9b318"
+  },
+  {
+    "id": "baef3cef84f14ed98655ac73293a56db"
+  },
+  {
+    "id": "099a1ae148184bd5a93c9bc9889306e7"
+  },
+  {
+    "id": "4fb4332a7a0e44e4b5655de1d7c2e5ac"
+  },
+  {
+    "id": "cdf86d1982f64cb49d6595680b1c53b8"
+  },
+  {
+    "id": "fa4cbba4ffed4f4e8ed4c148711cbc78"
+  },
+  {
+    "id": "41c691d3b7e94ffd993647c7e531e94e"
+  },
+  {
+    "id": "df43ea868a6d409ea67eb632fc9f285c"
+  },
+  {
+    "id": "746448dcf441410db997a59f7137a2c8"
+  },
+  {
+    "id": "3d0cad58c3c145f3b98d51c06a65a0c9"
+  },
+  {
+    "id": "9cb4317831024d589622e2a3cf544360"
+  },
+  {
+    "id": "802015dbc9f5406d8df3b468497f6553"
+  },
+  {
+    "id": "5463f1d3fc124cadb2ad016076a699f5"
+  },
+  {
+    "id": "8c77419b4e864778b965753f411bb469"
+  },
+  {
+    "id": "2e112a7c9d97482fa0deb079c6c09172"
+  },
+  {
+    "id": "f0c26f4f2e9f4d1d84b4f13cb201cbeb"
+  },
+  {
+    "id": "231d7d9955d54f60a1f1ff4f0bf1c5b6"
+  },
+  {
+    "id": "9b56a920d67a48e8b33932802d17d374"
+  },
+  {
+    "id": "092a8aab433c4184bde9c8aa08789464"
+  },
+  {
+    "id": "266a4337fa5742c8bbcc5aa79e571e85"
+  },
+  {
+    "id": "d67cecf38f954800a7c651cca70bdc66"
+  },
+  {
+    "id": "e634fa14bd4941ef96e1d21e796a9cd3"
+  },
+  {
+    "id": "dbc29aef324e4c4f9f19082dcb1b5cb1"
+  },
+  {
+    "id": "e688c5081a84488694b111a11a50ef7e"
+  },
+  {
+    "id": "80b114d03dbb4598b5da648bbe174d0f"
+  },
+  {
+    "id": "ce42d55ff5874c6380e018896b16e128"
+  },
+  {
+    "id": "ef2b37212c98453390c9dca445d69de1"
+  },
+  {
+    "id": "a66639dbb5ed4dfb9e65fe359ed8dc03"
+  },
+  {
+    "id": "10eacaf72a4e4ad49cc43d3a710ee24a"
+  },
+  {
+    "id": "b7e3d45f77bd4a4db768505fc4726592"
+  },
+  {
+    "id": "0d1592f640f04ecb9c7add7164f87c5a"
+  },
+  {
+    "id": "296a678ada814b09bdd0a53629f7f3f9"
+  },
+  {
+    "id": "6e6fd6d62ce54031a4e36f88c2d5b63d"
+  },
+  {
+    "id": "8d0dd71f94cc4abba255f3cdd6c2d940"
+  },
+  {
+    "id": "7c26258ff99c4adeb4b9b35001005453"
+  },
+  {
+    "id": "d13c24c99c704e31bbb3aaa585884d56"
+  },
+  {
+    "id": "dbfb81f100dc474f8267e6f9ab0ae51c"
+  },
+  {
+    "id": "3b4a5a2f617b43c8aa37b0890f576749"
+  },
+  {
+    "id": "c49a8bf2d0ad4afd9b0bb0d03c1b9d96"
+  },
+  {
+    "id": "fbfb4851dd494e44b9a21475d7e6dbaa"
+  },
+  {
+    "id": "e0ba941f8e2c4fa6a0dea2b562b0e4be"
+  },
+  {
+    "id": "7c9ec2b1cd8d4a8eaa6e1154e6fb40c4"
+  },
+  {
+    "id": "4ae645d2721c4751b1e71c83870bbe16"
+  },
+  {
+    "id": "c7bd32aee95f485d85d0b6749d542c51"
+  },
+  {
+    "id": "20e140b44e0f4a6484e9cf776b09679b"
+  },
+  {
+    "id": "5deb6bfcdd8c486d9f5f630309968763"
+  },
+  {
+    "id": "18d1fcc3ed394b23ada799aad3beca90"
+  },
+  {
+    "id": "74a655bcaead4857b92a49ec5f93ee23"
+  },
+  {
+    "id": "bfec094ae9da4aaebd91713b39d3008e"
+  },
+  {
+    "id": "80ebe13e36eb413a9acaace84648b36d"
+  },
+  {
+    "id": "982e414b371248788c72732975949957"
+  },
+  {
+    "id": "963ad06f3e314d3b99e0b6b9eaeaf2d8"
+  },
+  {
+    "id": "90b6478bd7ae4145976c474dc42da80f"
+  },
+  {
+    "id": "ea475dcd08d54da6896f56e330fa2a1c"
+  },
+  {
+    "id": "64899cac91e5423fa3617bf576e6df4d"
+  },
+  {
+    "id": "6231d57d040143e3bc56cafaabaa0975"
+  },
+  {
+    "id": "3fdb66ab2b0a49318b053edda6633237"
+  },
+  {
+    "id": "7dff080db1b8438e9ed737df1d7c32e3"
+  },
+  {
+    "id": "878108ec7e9742a090175bd33390e2c5"
+  },
+  {
+    "id": "3b2acfd62bdf4361aa4e51895d12df8b"
+  },
+  {
+    "id": "0335847f4a44451db5a1b775cc019765"
+  },
+  {
+    "id": "b6bfa34339c649398ae3b4540ed95fcf"
+  },
+  {
+    "id": "82fd2c61b9a64b5f8971b28a1cfd8c1b"
+  },
+  {
+    "id": "a9385b35a3a5492283aa461b93fdb361"
+  },
+  {
+    "id": "847f117d4638487d92539d52eb07a999"
+  },
+  {
+    "id": "de103586389245e7858b26b9b9a1ba57"
+  },
+  {
+    "id": "82d95f6b3df64bb48668f7e98ddf5b9c"
+  },
+  {
+    "id": "b1f6907b8f0f41ec82dc2eae7f84666b"
+  },
+  {
+    "id": "1f232b1de1f34793b6557c8688bf4d9c"
+  },
+  {
+    "id": "c16f718bd6554e998a95fe1d07623a13"
+  },
+  {
+    "id": "dc2add9004fd40e69fcbd2eea80c513c"
+  },
+  {
+    "id": "464e204e258340b69eca2b90451f06b8"
+  },
+  {
+    "id": "c12fd75ea3db450bbfcabeb4f62f7e8f"
+  },
+  {
+    "id": "5dbcadd728fa4956a4f739c1168fbf18"
+  },
+  {
+    "id": "1730220302104d54ba2bb781ca219e52"
+  },
+  {
+    "id": "ffcf25fe8cce4198bd483f2faaa17889"
+  },
+  {
+    "id": "98ed5d627d0746e0aecb42eae39c14dc"
+  },
+  {
+    "id": "30718db44c7f44a5b4b0c68842b1053b"
+  },
+  {
+    "id": "4e90a93f4548458bb5d35ac9c767233b"
+  },
+  {
+    "id": "dabccc32402a43a7ab9c82408f3603a1"
+  },
+  {
+    "id": "80a8a1ea275944228e71fd0dfc74fb74"
+  },
+  {
+    "id": "687434d75186430896ea636fec156e6d"
+  },
+  {
+    "id": "28e5233352ba44ea9f7a8d4cf7f6b86a"
+  },
+  {
+    "id": "b11499ee2a1f4a2b8fa8c9f561136430"
+  },
+  {
+    "id": "6efcba5974464dcbaf5d9c5e9322bd9c"
+  },
+  {
+    "id": "a1d460152c3840d98ed3bc45bf80d5c1"
+  },
+  {
+    "id": "0d482f0e297544dcb2ce1e87e4c3b4b4"
+  },
+  {
+    "id": "1a5baeb7a51d453d892ba70a9de2146d"
+  },
+  {
+    "id": "ab84169c66b2415180563269295f53be"
+  },
+  {
+    "id": "08c0e9678f1548aebeada27ab7b954e1"
+  },
+  {
+    "id": "efd2d367aeec4cc4ae72cd3eaa040a0f"
+  },
+  {
+    "id": "ae54626131ed42eaa55755f10ee64a76"
+  },
+  {
+    "id": "1787cbe3e160427ca4e61931daf66e51"
+  },
+  {
+    "id": "24143f01e7dd455e86c34a4a6f4ed3b9"
+  },
+  {
+    "id": "5e92b869dd0b4ecbaf8694041891eb75"
+  },
+  {
+    "id": "8d130f744b214055ad0c4cdce459a4b9"
+  },
+  {
+    "id": "87913f6be2004e938ba60cda550bc675"
+  },
+  {
+    "id": "68c7f892cbd94fc59ebe4ef1e1ad0c89"
+  },
+  {
+    "id": "f7bf36628c7540da907dda84b48aac82"
+  },
+  {
+    "id": "357af030b55046fc856e03fdac36032a"
+  },
+  {
+    "id": "3a90f09909344dfdbb1a73f80c983a87"
+  },
+  {
+    "id": "44fd76ce5f864cc2aef99c75cfd4904e"
+  },
+  {
+    "id": "95e54b962ce1420d82c716ff428d84fe"
+  },
+  {
+    "id": "8e22ba3b6fc34886a86d60dd1a1db025"
+  },
+  {
+    "id": "162917ea12b24d0e874540feecf9b445"
+  },
+  {
+    "id": "c8e87f5da41c47c7a3116bd1ed0cc34d"
+  },
+  {
+    "id": "a3458bd73564491585df515468475a66"
+  },
+  {
+    "id": "e27a4c1fb25c48809feb3cc438b5eb3c"
+  },
+  {
+    "id": "f36338c39e2846959491ea48cbce07f6"
+  },
+  {
+    "id": "eae01f064523415c95e57ff1e8dadc43"
+  },
+  {
+    "id": "8b845d50f998467a898367107377c64b"
+  },
+  {
+    "id": "a50ae9e7cb6a4f1e8dad07a9bbace810"
+  },
+  {
+    "id": "5c2009e4f1384b27b4b4fe9422f048ed"
+  },
+  {
+    "id": "0baaa0fe4fde4bae99bc6d2be230d525"
+  },
+  {
+    "id": "d2563a0ca702457497995a6959ff1c61"
+  },
+  {
+    "id": "b07871d2991d47e899cfa68a91f16b52"
+  },
+  {
+    "id": "b745738e5b3c482aa2ae94ad5688246b"
+  },
+  {
+    "id": "d66f1f0158054f638ff1b8acf3e18a8a"
+  },
+  {
+    "id": "ede0705d606d4435a300e06beab2322f"
+  },
+  {
+    "id": "e9d415f928644b59b5ed19b7011f627d"
+  },
+  {
+    "id": "1a78a8cf476c4a0dbf551558da1d879b"
+  },
+  {
+    "id": "ea7c66d5dae241eda155763541749cd1"
+  },
+  {
+    "id": "78ae271549144680a42139693d1117e9"
+  },
+  {
+    "id": "7eb9a69a23b8469b9f20f1915abff3be"
+  },
+  {
+    "id": "db73ca2786684b9e92deb1fcc796effd"
+  },
+  {
+    "id": "a876fb1accf64f4eb1b20b4a0ee089db"
+  },
+  {
+    "id": "71d6dabd553c471b98debe941341a4c0"
+  },
+  {
+    "id": "1c6add1cbd464f34b6e90485780f1c3e"
+  },
+  {
+    "id": "5e9b0b44825e44d89ba0501c3b819685"
+  },
+  {
+    "id": "4cfc841cb81e4c90822cfb7c0ec9f86e"
+  },
+  {
+    "id": "2933df89dd6146c791eac1a85f0d02c8"
+  },
+  {
+    "id": "5af06dca120a46889964109e94fff0a0"
+  },
+  {
+    "id": "77970683cb6444388e5cfdd743784db6"
+  },
+  {
+    "id": "3e40d5d91615427d976a56932be3c955"
+  },
+  {
+    "id": "70bc65e0c17041729a4eabe86ff36d53"
+  },
+  {
+    "id": "b913e55670474b69a70736d314822155"
+  },
+  {
+    "id": "ada9cbe2dd8e48deb55171645e47a24a"
+  },
+  {
+    "id": "fba813b8837249cbb39a1996dc657ef0"
+  },
+  {
+    "id": "f7ecc32a212e4c3b8471ce26c32f3135"
+  },
+  {
+    "id": "015641bf588b43aebba1aee07a8afde0"
+  },
+  {
+    "id": "bcecb67a14f34fc6896bc5d65c58405a"
+  }
+]

docs/contributing.md

@@ -0,0 +1,69 @@
+# Contributing to SkillOpt
+
+Thank you for your interest in contributing to SkillOpt! This guide covers how to get started.
+
+## Development Setup
+
+```bash
+git clone https://github.com/microsoft/SkillOpt.git
+cd SkillOpt
+pip install -e ".[dev]"
+```
+
+## Ways to Contribute
+
+### 🐛 Bug Reports
+
+Open an issue with:
+- Steps to reproduce
+- Expected vs actual behavior
+- Config file used (sanitize API keys)
+- Python version and OS
+
+### 🔧 New Benchmark
+
+See [Add a New Benchmark](guide/new-benchmark.md) for the implementation guide.
+
+**Checklist:**
+- [ ] Data loader in `skillopt/envs/<benchmark>/loader.py`
+- [ ] Environment adapter in `skillopt/envs/<benchmark>/env.py`
+- [ ] Config file in `configs/<benchmark>/default.yaml`
+- [ ] Registration in `skillopt/envs/__init__.py`
+- [ ] Documentation page in `docs/`
+
+### 🤖 New Model Backend
+
+See [Add a New Model Backend](guide/new-backend.md) for the implementation guide.
+
+**Checklist:**
+- [ ] Backend in `skillopt/model/<backend>.py`
+- [ ] Registration in `skillopt/model/__init__.py`
+- [ ] API key entry in `.env.example`
+- [ ] Documentation update
+
+### 📝 Documentation
+
+Documentation is built with MkDocs Material:
+
+```bash
+pip install -e ".[docs]"
+mkdocs serve  # Preview at http://localhost:8000
+```
+
+## Code Style
+
+- Follow existing patterns in the codebase
+- Use type hints for function signatures
+- Keep docstrings concise
+
+## Pull Request Process
+
+1. Fork the repository
+2. Create a feature branch: `git checkout -b feature/my-benchmark`
+3. Make your changes
+4. Test with an existing benchmark config
+5. Submit a PR with a clear description
+
+## License
+
+By contributing, you agree that your contributions will be licensed under the MIT License.

docs/guide/configuration.md

@@ -0,0 +1,109 @@
+# Configuration Guide
+
+SkillOpt uses YAML configuration files with a hierarchical override system.
+
+## Config Structure
+
+```
+configs/
+├── _base_/
+│   └── default.yaml          # Global defaults
+├── searchqa/
+│   └── default.yaml          # SearchQA overrides
+├── docvqa/
+│   └── default.yaml          # DocVQA overrides
+└── alfworld/
+    └── default.yaml          # ALFWorld overrides
+```
+
+Benchmark configs inherit from `_base_/default.yaml` and override specific values.
+
+## Key Parameters
+
+### Model
+
+```yaml
+model:
+  backend: azure_openai          # azure_openai | openai_chat | claude_code_exec | qwen
+  optimizer: gpt-5.5               # Optimizer model (for reflection)
+  target: gpt-5.5               # Target model (for rollout)
+```
+
+### Training
+
+```yaml
+train:
+  num_epochs: 4                  # Number of training epochs
+  batch_size: 40                 # Tasks per step (batch size)
+  accumulation: 1                # Gradient accumulation
+  seed: 42
+```
+
+### Gradient (Reflection)
+
+```yaml
+gradient:
+  minibatch_size: 8              # Reflect minibatch size
+  analyst_workers: 16            # Parallel reflection workers
+  max_analyst_rounds: 3          # Max rounds of analyst reflection
+  failure_only: false            # Only reflect on failures
+```
+
+### Optimizer
+
+```yaml
+optimizer:
+  learning_rate: 4               # Max edits per step (edit budget)
+  min_learning_rate: 2           # Min edits for decay schedulers
+  lr_scheduler: cosine           # constant | linear | cosine | autonomous
+  use_slow_update: true          # Momentum-like blending at epoch boundary
+  slow_update_samples: 20        # Samples for slow update evaluation
+  use_meta_skill: true           # Cross-epoch strategy memory
+```
+
+### Evaluation
+
+```yaml
+evaluation:
+  use_gate: true                 # Validation gating (accept/reject updates)
+  eval_test: true                # Run test evaluation after training
+```
+
+### Environment (Data)
+
+```yaml
+env:
+  name: searchqa                 # Benchmark name
+  split_mode: ratio              # ratio | split_dir
+  split_ratio: "2:1:7"           # train:val:test ratio
+  data_path: ""                  # Path to dataset
+  exec_timeout: 120              # Per-task timeout (seconds)
+```
+
+## CLI Overrides
+
+Override any config value from the command line:
+
+```bash
+python scripts/train.py \
+  --config configs/searchqa/default.yaml \
+  optimizer.learning_rate=16 \
+  optimizer.lr_scheduler=linear \
+  gradient.analyst_workers=8
+```
+
+## Environment Variables
+
+Model credentials are loaded from environment variables:
+
+| Variable | Backend | Description |
+|---|---|---|
+| `AZURE_OPENAI_ENDPOINT` | azure_openai | Azure resource endpoint |
+| `AZURE_OPENAI_API_KEY` | azure_openai | Azure API key |
+| `OPENAI_API_KEY` | openai | OpenAI API key |
+| `ANTHROPIC_API_KEY` | claude | Anthropic API key |
+| `QWEN_API_BASE` | qwen | Local Qwen vLLM endpoint |
+
+## Full Reference
+
+See [Configuration Reference](../reference/config.md) for the complete parameter list.

docs/guide/dl-analogy.md

@@ -0,0 +1,51 @@
+# Deep Learning ↔ SkillOpt Analogy
+
+SkillOpt is designed around a core insight: **optimizing natural-language prompts follows the same structure as training neural networks**. This page maps every DL concept to its SkillOpt counterpart.
+
+## Complete Mapping
+
+| Deep Learning | SkillOpt | Description |
+|---|---|---|
+| **Model weights** | Skill document (Markdown) | The thing being optimized |
+| **Forward pass** | Rollout | Target executes tasks using current skill |
+| **Loss function** | Task evaluator | Scores task execution quality |
+| **Backpropagation** | Reflect | Optimizer analyzes failures → edit patches |
+| **Gradients** | Edit patches | Proposed changes to the skill |
+| **Gradient aggregation** | Patch aggregation | Merge similar edits |
+| **Gradient clipping** | Edit selection | Cap max edits per step |
+| **Learning rate** | `learning_rate` | Max number of edits applied per step |
+| **LR scheduler** | `lr_scheduler` | Decay schedule: cosine, linear, constant |
+| **SGD step** | Skill update | Apply selected patches to document |
+| **Validation set** | Selection split | Gate checks improvement before accepting |
+| **Early stopping** | Gate patience | Reject updates that don't improve |
+| **Training step** | Step | One rollout → reflect → update cycle |
+| **Epoch** | Epoch | Full pass with slow update + meta memory |
+| **Momentum** | Slow update | Longitudinal comparison at epoch boundary |
+| **Meta-learning** | Meta skill | Cross-epoch optimizer strategy memory |
+| **Batch size** | `batch_size` | Tasks sampled per rollout |
+| **Data parallelism** | `analyst_workers` | Parallel reflection workers |
+| **Training set** | Train split | Items used for rollout |
+| **Test set** | Test split | Held-out final evaluation |
+| **Warm-up** | (implicit) | High LR early steps explore broadly |
+| **Checkpointing** | Skill snapshots | Saved after each accepted step |
+| **Transfer learning** | Seed skill / cross-benchmark init | Start from pre-trained skill |
+
+## Why This Analogy Matters
+
+1. **Familiar mental model**: ML practitioners immediately understand how to tune SkillOpt
+2. **Principled hyperparameter search**: Grid search over `learning_rate` × `lr_scheduler` works just like in DL
+3. **Proven mechanisms**: Gating ≈ validation-based selection, patience ≈ early stopping, slow update ≈ momentum — all with strong theoretical motivation
+
+## Hyperparameter Transfer Rules
+
+From our experiments, these DL intuitions transfer well:
+
+!!! success "What transfers"
+    - **Cosine schedule > constant** — same as in DL, cosine annealing helps convergence
+    - **Moderate LR (4-16) > very high/low** — too few edits = slow learning, too many = noisy
+    - **Slow update helps** — longitudinal comparison prevents catastrophic forgetting across epochs
+    - **Meta skill memory improves reflection** — optimizer benefits from cross-epoch strategy notes
+
+!!! warning "What doesn't transfer"
+    - **Batch size ≠ better** — larger rollout batches have diminishing returns due to API costs
+    - **More epochs ≠ better** — skills converge faster than neural networks (2-4 epochs usually enough)

docs/guide/first-experiment.md

@@ -0,0 +1,110 @@
+# Your First Experiment
+
+This guide walks through running a complete SkillOpt training on SearchQA.
+
+## 1. Choose a Benchmark
+
+SkillOpt includes ready-to-use configs for several benchmarks:
+
+| Benchmark | Difficulty | Typical Runtime |
+|---|---|---|
+| SearchQA | ⭐ Easy | ~30 min |
+| DocVQA | ⭐⭐ Medium | ~2 hours |
+| ALFWorld | ⭐⭐⭐ Hard | ~3 hours |
+
+We'll use **SearchQA** as it's the fastest to complete.
+
+## 2. Configure
+
+Review the config file:
+
+```bash
+cat configs/searchqa/default.yaml
+```
+
+Key parameters (deep learning analogy in parentheses):
+
+```yaml
+train:
+  num_epochs: 4           # (epochs)
+  batch_size: 40          # (batch size)
+
+optimizer:
+  learning_rate: 4        # (max edits per step)
+  lr_scheduler: cosine    # (learning rate schedule)
+  use_slow_update: true   # (momentum at epoch boundary)
+  use_meta_skill: true    # (cross-epoch optimizer memory)
+
+gradient:
+  analyst_workers: 16     # (parallel reflection workers)
+
+evaluation:
+  use_gate: true          # (validation gating)
+```
+
+## 3. Train
+
+```bash
+python scripts/train.py --config configs/searchqa/default.yaml
+```
+
+You'll see output like:
+
+```
+[Step 1/8] Rollout: 20 items, 4 workers...
+[Step 1/8] Score: 0.65 → Reflect...
+[Step 1/8] 6 edit patches generated
+[Step 1/8] Selected 4 edits (lr=8, cosine → 7.7)
+[Step 1/8] Gate: val score 0.68 > 0.65 ✓ ACCEPT
+[Step 2/8] ...
+```
+
+## 4. Monitor
+
+Training outputs are saved to `outputs/<benchmark>/<run_id>/`:
+
+```
+outputs/searchqa/2024-01-15_10-30-00/
+├── steps/
+│   ├── step_0001/
+│   │   ├── candidate_skill.md
+│   │   ├── step_record.json
+│   │   └── trajectory_digest.json
+│   └── step_0002/
+├── slow_update/
+│   └── epoch_02/
+├── meta_skill/
+│   └── epoch_02/
+├── skills/
+│   └── step_0001.md
+├── best_skill.md
+├── history.json
+└── config.yaml
+```
+
+## 5. Evaluate
+
+Evaluate the best skill on the test split:
+
+```bash
+python scripts/eval_only.py \
+  --config configs/searchqa/default.yaml \
+  --skill outputs/searchqa/<run_id>/skills/best_skill.md
+```
+
+## WebUI
+
+Prefer a graphical interface? Launch the WebUI:
+
+```bash
+pip install -e ".[webui]"
+python -m skillopt_webui.app
+```
+
+Then open `http://localhost:7860` in your browser to configure parameters and launch training.
+
+## Next Steps
+
+- [Understand the training loop](training-loop.md)
+- [Configuration reference](../reference/config.md)
+- [Add a new benchmark](new-benchmark.md)

docs/guide/installation.md

@@ -0,0 +1,89 @@
+# Installation
+
+## Requirements
+
+- Python ≥ 3.10
+- At least one model API key (Azure OpenAI, OpenAI, Anthropic, or local Qwen)
+
+## Quick Install
+
+```bash
+git clone https://github.com/microsoft/SkillOpt.git
+cd SkillOpt
+pip install -e .
+```
+
+## Optional Dependencies
+
+Install extras for specific benchmarks or backends:
+
+=== "ALFWorld"
+
+    ```bash
+    pip install -e ".[alfworld]"
+    ```
+
+=== "Claude Backend"
+
+    ```bash
+    pip install -e ".[claude]"
+    ```
+
+=== "Qwen (Local)"
+
+    ```bash
+    pip install -e ".[qwen]"
+    ```
+
+=== "WebUI"
+
+    ```bash
+    pip install -e ".[webui]"
+    ```
+
+=== "Development"
+
+    ```bash
+    pip install -e ".[dev]"
+    ```
+
+=== "All"
+
+    ```bash
+    pip install -e ".[alfworld,claude,qwen,webui,dev]"
+    ```
+
+## Environment Variables
+
+Copy the example `.env` file and fill in your credentials:
+
+```bash
+cp .env.example .env
+```
+
+Edit `.env` with your API keys:
+
+```ini
+# Azure OpenAI (default backend)
+AZURE_OPENAI_ENDPOINT=https://your-resource.openai.azure.com/
+AZURE_OPENAI_API_KEY=your-key
+
+# Or use OpenAI directly
+OPENAI_API_KEY=sk-...
+
+# Or Anthropic Claude
+ANTHROPIC_API_KEY=sk-ant-...
+```
+
+!!! tip
+    You only need credentials for the backend you plan to use. Azure OpenAI is the default.
+
+## Verify Installation
+
+```bash
+python -c "import skillopt; print('SkillOpt ready!')"
+```
+
+## Next Steps
+
+→ [Run your first experiment](first-experiment.md)

docs/guide/local-env-smoke.md

@@ -0,0 +1,143 @@
+# Local Environment Smoke Tests
+
+This guide describes a lightweight pattern for testing a custom SkillOpt environment before connecting it to expensive model calls or a full benchmark dataset.
+
+The goal is to validate the training loop plumbing first:
+
+- config loading
+- adapter construction
+- dataloader splits
+- rollout output shape
+- reflection patch shape
+- merge/rank/update control flow
+- artifact creation under `out_root`
+
+Once those are stable, you can switch the same environment to real model calls and larger evaluation splits.
+
+## 1. Add a tiny fixture split
+
+Start with a handful of deterministic examples that cover the expected pass/fail cases for your environment. Keep them small enough that a single training step can run locally.
+
+A minimal fixture item usually needs:
+
+```json
+{
+  "id": "example-1",
+  "split": "train",
+  "question": "...",
+  "expected": "..."
+}
+```
+
+Use the split names your adapter maps to SkillOpt phases:
+
+- `train` for optimization rollouts
+- `val` or `valid_seen` for selection/gating
+- `test` or `valid_unseen` for final evaluation
+
+## 2. Support an offline mock mode
+
+Add a configuration flag such as `mock: true` to your adapter. In mock mode, `rollout()` should return deterministic responses without calling external model APIs.
+
+This lets you verify the SkillOpt loop with a fast command such as:
+
+```bash
+python scripts/train.py \
+  --config configs/myenv/tiny_mock.yaml
+```
+
+Mock mode should still write the same artifacts as a real run, for example:
+
+- `responses.json`
+- `rollout_results.json`
+- `ranked_edits.json`
+- `candidate_skill.md`
+- `summary.json`
+
+## 3. Keep the smoke config tiny
+
+A CI-friendly smoke config should run a single small step:
+
+```yaml
+train:
+  num_epochs: 1
+  train_size: 3
+  batch_size: 3
+
+gradient:
+  minibatch_size: 1
+  merge_batch_size: 2
+  analyst_workers: 1
+  max_analyst_rounds: 1
+
+optimizer:
+  learning_rate: 1
+  min_learning_rate: 1
+  lr_scheduler: constant
+  skill_update_mode: patch
+  use_slow_update: false
+
+evaluation:
+  use_gate: true
+  sel_env_num: 2
+  test_env_num: 2
+  eval_test: false
+
+env:
+  name: myenv
+  out_root: outputs/myenv_tiny_mock
+  mock: true
+```
+
+Prefer a mock config that runs without credentials. That makes it useful for contributors and CI.
+
+## 4. Validate optimizer JSON before returning it
+
+If your environment or extension asks an LLM to merge or rank skill edits, validate the returned JSON before passing it back into SkillOpt. This avoids silent fallbacks from empty, malformed, or out-of-range responses.
+
+Useful checks for edit payloads:
+
+- response is a JSON object
+- `edits` is a non-empty list
+- every edit is an object
+- every edit has an allowed operation
+- required fields such as `content` or `target` are present for that operation
+
+Useful checks for ranking payloads:
+
+- `selected_indices` exists
+- indices are integers
+- indices are unique
+- indices are within the candidate edit range
+- selected count does not exceed the edit budget
+
+On failure, retry with a compact prompt that includes the schema error. If retries fail, raise an explicit error instead of silently accepting malformed output.
+
+## 5. Run progressively stronger checks
+
+A good development sequence is:
+
+```bash
+python -m py_compile scripts/train.py skillopt/envs/myenv/adapter.py
+python scripts/train.py --config configs/myenv/tiny_mock.yaml
+python scripts/train.py --config configs/myenv/tiny.yaml
+```
+
+For the real tiny run, verify that:
+
+- the run completes
+- `summary.json` is written
+- `ranked_edits.json` contains the expected ranking metadata
+- any optimizer bridge log marks the response schema as valid
+- no generated files are written outside `out_root`
+
+## 6. Keep custom environments isolated
+
+When adding a custom environment to the registry, avoid side effects for existing benchmarks:
+
+- lazy-import optional dependencies
+- install environment-specific hooks only when `cfg["env"]` matches your environment
+- keep mock behavior behind an explicit config flag
+- write generated artifacts only under `out_root`
+
+This makes it easier to review and test a custom integration without affecting the built-in benchmarks.

docs/guide/new-backend.md

@@ -0,0 +1,130 @@
+# Add a New Model Backend
+
+SkillOpt supports multiple LLM backends. This guide shows how to add your own.
+
+## Backend Architecture
+
+```
+skillopt/model/
+├── base.py           # Abstract base class
+├── azure_openai.py   # Azure OpenAI backend
+├── openai_model.py   # Direct OpenAI backend
+├── claude.py         # Anthropic Claude backend
+├── qwen.py           # Local Qwen (vLLM) backend
+└── your_backend.py   # Your new backend
+```
+
+## Step 1: Create the Backend
+
+Create `skillopt/model/your_backend.py`:
+
+```python
+from skillopt.model.base import ModelBackend, ModelResponse
+
+class YourBackend(ModelBackend):
+    """Your custom model backend."""
+    
+    def __init__(self, cfg: dict):
+        super().__init__(cfg)
+        self.model_name = cfg.get('model_name', 'your-default-model')
+        self.api_key = os.environ.get('YOUR_API_KEY', '')
+        self.client = self._init_client()
+    
+    def _init_client(self):
+        """Initialize API client."""
+        # TODO: Set up your API client
+        pass
+    
+    async def generate(
+        self,
+        messages: list[dict],
+        temperature: float = 0.7,
+        max_tokens: int = 4096,
+        **kwargs
+    ) -> ModelResponse:
+        """
+        Generate a completion.
+        
+        Args:
+            messages: Chat messages [{"role": "...", "content": "..."}]
+            temperature: Sampling temperature
+            max_tokens: Maximum tokens in response
+            
+        Returns:
+            ModelResponse with content, usage, and metadata
+        """
+        response = await self.client.chat(
+            model=self.model_name,
+            messages=messages,
+            temperature=temperature,
+            max_tokens=max_tokens,
+        )
+        
+        return ModelResponse(
+            content=response.text,
+            usage={
+                'prompt_tokens': response.usage.input,
+                'completion_tokens': response.usage.output,
+            },
+            model=self.model_name,
+        )
+    
+    async def generate_with_tools(
+        self,
+        messages: list[dict],
+        tools: list[dict],
+        **kwargs
+    ) -> ModelResponse:
+        """Generate with tool/function calling support."""
+        # Optional: implement if your model supports tool use
+        raise NotImplementedError("Tool use not supported")
+```
+
+## Step 2: Register the Backend
+
+Add to `skillopt/model/__init__.py`:
+
+```python
+from .your_backend import YourBackend
+
+BACKEND_REGISTRY = {
+    # ... existing backends ...
+    'your_backend': YourBackend,
+}
+```
+
+## Step 3: Configure
+
+Use your backend in any config:
+
+```yaml
+model:
+  backend: your_backend
+  model_name: your-model-id
+  temperature: 0.7
+  max_tokens: 4096
+```
+
+Set credentials via environment variable:
+
+```bash
+export YOUR_API_KEY="your-key"
+```
+
+## Required Interface
+
+Your backend must implement these methods:
+
+| Method | Required | Description |
+|---|---|---|
+| `generate()` | ✅ | Basic text generation |
+| `generate_with_tools()` | Optional | Tool/function calling |
+| `count_tokens()` | Optional | Token counting for context management |
+
+## Tips
+
+!!! tip
+    - Test your backend with `python -c "from skillopt.model.your_backend import YourBackend"` first
+    - Use `async` methods for all API calls — SkillOpt uses asyncio throughout
+    - Implement retry logic with exponential backoff for production use
+    - Add your API key to `.env.example` when submitting a PR

docs/guide/new-benchmark.md

@@ -0,0 +1,181 @@
+# Add a New Benchmark
+
+Extend SkillOpt with your own benchmark in ~100 lines of code.
+
+## Overview
+
+To add a benchmark, you need:
+
+1. **Data Loader** — Loads and splits your dataset
+2. **Environment Adapter** — Executes tasks and returns scores
+3. **Config** — YAML configuration file
+
+## Step 1: Create the Benchmark Package
+
+```bash
+mkdir -p skillopt/envs/my_benchmark
+touch skillopt/envs/my_benchmark/__init__.py
+```
+
+## Step 2: Implement the Data Loader
+
+Create `skillopt/envs/my_benchmark/loader.py`:
+
+```python
+from skillopt.data.base import DataLoader, DataItem
+
+class MyBenchmarkDataLoader(DataLoader):
+    """Load and split your benchmark data."""
+    
+    def __init__(self, data_dir: str, **kwargs):
+        super().__init__(**kwargs)
+        self.data_dir = data_dir
+    
+    def setup(self, cfg: dict):
+        """Initialize splits based on config."""
+        self.split_mode = cfg.get('split_mode', 'ratio')
+        # Load your data here
+        self.items = self._load_items()
+        self._create_splits(cfg)
+    
+    def _load_items(self) -> list[DataItem]:
+        """Load raw data into DataItem objects."""
+        items = []
+        # TODO: Load your data
+        for entry in your_data:
+            items.append(DataItem(
+                id=entry['id'],
+                input=entry['question'],
+                ground_truth=entry['answer'],
+                metadata=entry.get('metadata', {})
+            ))
+        return items
+    
+    def get_split_items(self, split: str) -> list[DataItem]:
+        """Return items for a given split (train/valid/test)."""
+        return self.splits[split]
+```
+
+## Step 3: Implement the Environment Adapter
+
+Create `skillopt/envs/my_benchmark/env.py`:
+
+```python
+from skillopt.envs.base import EnvAdapter, TaskResult
+
+class MyBenchmarkEnv(EnvAdapter):
+    """Execute tasks and evaluate results."""
+    
+    def __init__(self, cfg: dict):
+        super().__init__(cfg)
+    
+    async def execute(self, item: DataItem, skill: str, model) -> TaskResult:
+        """
+        Execute a single task.
+        
+        Args:
+            item: The data item to process
+            skill: Current skill document content
+            model: The target model instance
+            
+        Returns:
+            TaskResult with prediction, score, and trajectory
+        """
+        # Build prompt with skill document
+        prompt = self.build_prompt(item, skill)
+        
+        # Get model response
+        response = await model.generate(prompt)
+        
+        # Extract prediction
+        prediction = self.parse_response(response)
+        
+        # Score against ground truth
+        score = self.evaluate(prediction, item.ground_truth)
+        
+        return TaskResult(
+            item_id=item.id,
+            prediction=prediction,
+            score=score,
+            trajectory=[
+                {"role": "system", "content": skill},
+                {"role": "user", "content": item.input},
+                {"role": "assistant", "content": response}
+            ]
+        )
+    
+    def evaluate(self, prediction: str, ground_truth: str) -> float:
+        """
+        Score a prediction against ground truth.
+        
+        Returns:
+            Float between 0.0 and 1.0
+        """
+        # TODO: Implement your scoring logic
+        # Examples: exact match, F1, ANLS, etc.
+        return float(prediction.strip() == ground_truth.strip())
+    
+    def build_prompt(self, item, skill: str) -> str:
+        """Combine skill document with task input."""
+        return f"{skill}\n\n---\n\nQuestion: {item.input}"
+    
+    def parse_response(self, response: str) -> str:
+        """Extract the answer from model response."""
+        return response.strip()
+```
+
+## Step 4: Register the Benchmark
+
+Add to `skillopt/envs/__init__.py`:
+
+```python
+from .my_benchmark.env import MyBenchmarkEnv
+from .my_benchmark.loader import MyBenchmarkDataLoader
+
+BENCHMARK_REGISTRY = {
+    # ... existing benchmarks ...
+    'my_benchmark': {
+        'env': MyBenchmarkEnv,
+        'loader': MyBenchmarkDataLoader,
+    },
+}
+```
+
+## Step 5: Create Config
+
+Create `configs/my_benchmark/default.yaml`:
+
+```yaml
+_base_: ['../_base_/default.yaml']
+
+env:
+  name: my_benchmark
+  data_path: data/my_benchmark
+  split_mode: ratio
+  split_ratio: "2:1:7"
+
+train:
+  num_epochs: 4
+  batch_size: 40
+
+optimizer:
+  learning_rate: 4
+  lr_scheduler: cosine
+  use_slow_update: true
+  use_meta_skill: true
+
+gradient:
+  analyst_workers: 16
+```
+
+## Step 6: Run
+
+```bash
+python scripts/train.py --config configs/my_benchmark/default.yaml
+```
+
+## Tips
+
+!!! tip
+    - Use a small `batch_size` (10-20) for initial testing
+    - The `evaluate()` method is critical — a noisy metric will confuse the optimizer

docs/guide/skill-document.md

@@ -0,0 +1,78 @@
+# Skill Document
+
+A **skill document** is a Markdown file that serves as the "prompt weights" of your agent. SkillOpt trains this document through iterative optimization.
+
+## What is a Skill Document?
+
+A skill document is a structured set of instructions that tells a language model **how** to approach a specific type of task. It's analogous to learned weights in a neural network — encoding task-specific knowledge in natural language rather than floating-point parameters.
+
+## Structure
+
+A typical skill document contains:
+
+```markdown
+# Task Strategy
+
+## General Approach
+- Break complex problems into sub-steps
+- Always verify intermediate results
+
+## Common Patterns
+- When you see X, try approach Y
+- Avoid Z because it leads to errors
+
+## Edge Cases
+- If the input contains A, handle it specially by...
+- Watch out for B — it requires C
+
+## Output Format
+- Always include reasoning before the answer
+- Format numbers with proper units
+```
+
+## How It Evolves
+
+During training, the skill document is modified by **edit patches**:
+
+1. **Additions**: New rules or strategies discovered from failed trajectories
+2. **Modifications**: Refining existing rules that are partially correct
+3. **Deletions**: Removing rules that consistently lead to errors
+
+Each edit is validated through the **gate** mechanism before being permanently accepted.
+
+## Initial Skill
+
+You can start training with:
+
+- **Empty skill**: The system learns everything from scratch
+- **Seed skill**: Provide initial instructions to bootstrap training
+- **Pre-trained skill**: Transfer a skill from a related benchmark
+
+Configure the initial skill in your YAML:
+
+```yaml
+train:
+  init_skill: "path/to/initial_skill.md"  # or omit for empty
+```
+
+## Skill Quality Metrics
+
+Track your skill's evolution through:
+
+- **Validation score**: Primary metric on the selection split
+- **Test score**: Final metric on held-out test data
+- **Skill length**: Total tokens in the document
+- **Edit acceptance rate**: Fraction of proposed edits that pass gating
+
+## Best Practices
+
+!!! tip "Tips for better skills"
+    1. **Start with a seed skill** (`env.skill_init`) if you have domain knowledge — it converges faster
+    2. **Use cosine LR schedule** — aggressive early exploration + careful late refinement
+    3. **Enable slow update** (`use_slow_update: true`) to prevent forgetting across epochs
+    4. **Enable meta skill** (`use_meta_skill: true`) so the optimizer accumulates strategy memory
+
+## Next Steps
+
+- [Deep Learning Analogy](dl-analogy.md)
+- [Configuration Reference](../reference/config.md)

docs/guide/training-loop.md

@@ -0,0 +1,92 @@
+# The Training Loop
+
+SkillOpt's core insight: **optimizing natural-language skill documents follows the same structure as training neural networks**.
+
+## Overview
+
+```
+┌─────────────────────────────────────────────────────────┐
+│                    Training Loop                         │
+│                                                         │
+│  for epoch in epochs:                                   │
+│    for step in steps:                                   │
+│      1. Rollout   — Target executes tasks              │
+│      2. Reflect   — Optimizer analyzes trajectories       │
+│      3. Aggregate — Hierarchical merge of patches       │
+│      4. Select    — Rank & clip edits (learning rate)   │
+│      5. Update    — Apply patches to skill doc          │
+│      6. Gate      — Validate & accept/reject            │
+│                                                         │
+│    Epoch Boundary:                                       │
+│      • Slow Update (longitudinal comparison & guidance) │
+│      • Meta Skill  (cross-epoch strategy memory)        │
+└─────────────────────────────────────────────────────────┘
+```
+
+## Stage Details
+
+### 1. Rollout (Forward Pass)
+
+The **target** model executes tasks using the current skill document as its prompt. Each task produces a trajectory and a score.
+
+```python
+# Analogy: forward pass through the network
+predictions = model(input, skill_document)
+scores = evaluate(predictions, ground_truth)
+```
+
+### 2. Reflect (Backward Pass)
+
+The **optimizer** model analyzes failed trajectories and produces **edit patches** — structured suggestions for improving the skill document.
+
+Two modes:
+
+- **Shallow**: Analyze each trajectory independently
+- **Deep**: Cross-reference multiple failures to find systemic issues
+
+```python
+# Analogy: computing gradients
+gradients = loss.backward()  # → edit patches
+```
+
+### 3. Aggregate
+
+Semantically similar edit patches are merged to avoid redundant edits.
+
+### 4. Select (Gradient Clipping)
+
+Edits are ranked by relevance score. The `learning_rate` parameter caps how many edits are applied per step — just like gradient clipping prevents overshooting.
+
+```python
+# Analogy: gradient clipping + optimizer step size
+selected = top_k(edits, k=learning_rate)
+```
+
+The `lr_scheduler` adjusts this over training:
+
+- **cosine**: Start aggressive, taper smoothly
+- **linear**: Linear decay
+- **constant**: Fixed rate
+
+### 5. Update (Parameter Update)
+
+Selected edits are applied to the skill document, producing a new version.
+
+### 6. Gate (Validation)
+
+The updated skill is evaluated on a **selection split** (analogous to a validation set). The update is only accepted if performance improves.
+
+## Epoch Boundary Mechanisms
+
+### Slow Update
+
+At the end of each epoch (starting from epoch 2), the system performs a **longitudinal comparison**: it rolls out both the previous epoch's skill and the current skill on the same samples, categorizes items as improved/regressed/persistent_fail/stable_success, then generates high-level **guidance** that is injected into the skill document. This prevents catastrophic forgetting of earlier improvements.
+
+### Meta Skill
+
+A **meta-skill memory** accumulates high-level strategy notes across the entire training run. At the end of each epoch, the optimizer reflects on what changed between epochs and produces a compact memory that is provided as additional context during future reflection steps.
+
+## Next Steps
+
+- [Understand Skill Documents](skill-document.md)
+- [DL ↔ SkillOpt analogy table](dl-analogy.md)

docs/guideline.html

@@ -0,0 +1,911 @@
+<!DOCTYPE html>
+<html lang="en">
+<head>
+<meta charset="UTF-8">
+<meta name="viewport" content="width=device-width, initial-scale=1.0">
+<title>SkillOpt — Documentation &amp; Reproduction Guide</title>
+<meta name="description" content="Complete documentation and reproduction guide for SkillOpt: installation, data preparation, training, configuration reference, framework internals, and API reference.">
+<link rel="icon" type="image/svg+xml" href="data:image/svg+xml,%3Csvg xmlns='http://www.w3.org/2000/svg' viewBox='0 0 23 23'%3E%3Crect width='10' height='10' fill='%23F25022'/%3E%3Crect x='13' width='10' height='10' fill='%237FBA00'/%3E%3Crect y='13' width='10' height='10' fill='%2300A4EF'/%3E%3Crect x='13' y='13' width='10' height='10' fill='%23FFB900'/%3E%3C/svg%3E">
+<style>
+  :root {
+    --bg: #ffffff;
+    --bg-soft: #f7f8fb;
+    --sidebar-bg: #fbfcfe;
+    --ink: #1f2733;
+    --muted: #5b6675;
+    --quiet: #8a94a3;
+    --line: #e6e9ef;
+    --line-strong: #d3d9e3;
+    --brand: #4f46e5;
+    --brand-soft: #eef0fe;
+    --accent: #0ea5e9;
+    --green: #16a34a;
+    --amber: #d97706;
+    --red: #dc2626;
+    --code-bg: #0f172a;
+    --code-ink: #e2e8f0;
+    --inline-code-bg: #eef1f6;
+    --inline-code-ink: #b3146b;
+    --sidebar-w: 300px;
+    --toc-w: 220px;
+    --mono: "SFMono-Regular", "JetBrains Mono", Consolas, "Liberation Mono", monospace;
+    --sans: "Inter", -apple-system, BlinkMacSystemFont, "Segoe UI", Roboto, Helvetica, Arial, sans-serif;
+  }
+  * { box-sizing: border-box; }
+  html { scroll-behavior: smooth; }
+  body {
+    margin: 0;
+    font-family: var(--sans);
+    color: var(--ink);
+    background: var(--bg);
+    font-size: 15px;
+    line-height: 1.65;
+    -webkit-font-smoothing: antialiased;
+  }
+
+  /* ── Top bar ─────────────────────────────────────────── */
+  header.topbar {
+    position: sticky; top: 0; z-index: 40;
+    height: 56px;
+    display: flex; align-items: center; gap: 14px;
+    padding: 0 20px;
+    background: rgba(255,255,255,0.92);
+    backdrop-filter: blur(8px);
+    border-bottom: 1px solid var(--line);
+  }
+  .topbar .logo { width: 22px; height: 22px; flex: none; }
+  .topbar .brand { font-weight: 700; font-size: 16px; letter-spacing: -0.01em; }
+  .topbar .brand span { color: var(--brand); }
+  .topbar .tag { color: var(--quiet); font-size: 13px; border-left: 1px solid var(--line-strong); padding-left: 14px; }
+  .topbar .spacer { flex: 1; }
+  .topbar a.gh {
+    display: inline-flex; align-items: center; gap: 6px;
+    font-size: 13px; font-weight: 600; color: var(--muted);
+    text-decoration: none; padding: 6px 12px; border: 1px solid var(--line-strong);
+    border-radius: 8px;
+  }
+  .topbar a.gh:hover { color: var(--brand); border-color: var(--brand); }
+  #menuBtn {
+    display: none; background: none; border: 1px solid var(--line-strong);
+    border-radius: 8px; width: 38px; height: 34px; cursor: pointer; font-size: 18px; color: var(--muted);
+  }
+
+  /* ── Layout ──────────────────────────────────────────── */
+  .layout { display: flex; align-items: flex-start; }
+
+  /* ── Sidebar (left nav) ──────────────────────────────── */
+  nav.sidebar {
+    position: sticky; top: 56px;
+    width: var(--sidebar-w); flex: none;
+    height: calc(100vh - 56px);
+    overflow-y: auto;
+    background: var(--sidebar-bg);
+    border-right: 1px solid var(--line);
+    padding: 22px 14px 60px 20px;
+  }
+  nav.sidebar .group { margin-bottom: 22px; }
+  nav.sidebar .group > .glabel {
+    display: flex; align-items: center; gap: 8px;
+    font-size: 11.5px; font-weight: 700; text-transform: uppercase;
+    letter-spacing: 0.07em; color: var(--quiet);
+    margin: 0 0 8px 2px;
+  }
+  nav.sidebar .group > .glabel .num {
+    display: inline-flex; align-items: center; justify-content: center;
+    width: 18px; height: 18px; border-radius: 5px;
+    background: var(--brand-soft); color: var(--brand);
+    font-size: 11px; font-weight: 700;
+  }
+  nav.sidebar a {
+    display: block; text-decoration: none;
+    color: var(--muted); font-size: 13.5px;
+    padding: 5px 10px; border-radius: 7px; margin: 1px 0;
+    border-left: 2px solid transparent;
+  }
+  nav.sidebar a:hover { background: #eef1f6; color: var(--ink); }
+  nav.sidebar a.active {
+    color: var(--brand); background: var(--brand-soft);
+    border-left-color: var(--brand); font-weight: 600;
+  }
+
+  /* ── Content ─────────────────────────────────────────── */
+  main.content {
+    flex: 1; min-width: 0;
+    padding: 38px 46px 120px;
+    max-width: 900px;
+  }
+  main.content section { scroll-margin-top: 72px; }
+  main h1 { font-size: 30px; line-height: 1.2; letter-spacing: -0.02em; margin: 0 0 8px; }
+  main h2 {
+    font-size: 23px; letter-spacing: -0.015em; margin: 52px 0 14px;
+    padding-bottom: 10px; border-bottom: 1px solid var(--line);
+  }
+  main section:first-of-type h2 { margin-top: 8px; }
+  main h3 { font-size: 17.5px; margin: 30px 0 10px; letter-spacing: -0.01em; }
+  main h4 { font-size: 15px; margin: 22px 0 8px; color: var(--ink); }
+  main p { margin: 12px 0; color: #2c3645; }
+  main ul, main ol { margin: 12px 0; padding-left: 22px; }
+  main li { margin: 5px 0; }
+  main a { color: var(--brand); text-decoration: none; }
+  main a:hover { text-decoration: underline; }
+  .lead { font-size: 16.5px; color: var(--muted); margin: 6px 0 4px; }
+  .eyebrow { color: var(--brand); font-weight: 700; font-size: 12.5px; letter-spacing: 0.08em; text-transform: uppercase; }
+
+  /* code */
+  code {
+    font-family: var(--mono); font-size: 0.86em;
+    background: var(--inline-code-bg); color: var(--inline-code-ink);
+    padding: 2px 6px; border-radius: 5px;
+  }
+  pre {
+    background: var(--code-bg); color: var(--code-ink);
+    border-radius: 12px; padding: 16px 18px; overflow-x: auto;
+    font-family: var(--mono); font-size: 13px; line-height: 1.6;
+    margin: 14px 0; border: 1px solid #1e293b;
+  }
+  pre code { background: none; color: inherit; padding: 0; font-size: inherit; }
+  .tok-c { color: #7c8aa5; }   /* comment */
+  .tok-k { color: #c4b5fd; }   /* keyword */
+  .tok-s { color: #86efac; }   /* string */
+  .tok-f { color: #93c5fd; }   /* flag/path */
+  .tok-n { color: #fca5a5; }   /* number/value */
+
+  /* tables */
+  .table-wrap { overflow-x: auto; margin: 16px 0; border: 1px solid var(--line); border-radius: 12px; }
+  table { border-collapse: collapse; width: 100%; font-size: 13.5px; }
+  th, td { text-align: left; padding: 9px 13px; border-bottom: 1px solid var(--line); vertical-align: top; }
+  thead th { background: var(--bg-soft); font-weight: 700; color: var(--ink); white-space: nowrap; }
+  tbody tr:last-child td { border-bottom: none; }
+  td code { white-space: nowrap; }
+  td.def { color: var(--muted); font-family: var(--mono); font-size: 12px; }
+
+  /* callouts */
+  .note { border-radius: 10px; padding: 12px 16px; margin: 16px 0; border: 1px solid; font-size: 14px; }
+  .note p { margin: 4px 0; }
+  .note .nh { font-weight: 700; display: block; margin-bottom: 2px; }
+  .note.info  { background: #eff6ff; border-color: #bfdbfe; }
+  .note.info .nh { color: #1d4ed8; }
+  .note.tip   { background: #ecfdf5; border-color: #a7f3d0; }
+  .note.tip .nh { color: #047857; }
+  .note.warn  { background: #fffbeb; border-color: #fde68a; }
+  .note.warn .nh { color: #b45309; }
+
+  .pill { display:inline-block; font-size: 11px; font-weight:700; padding: 1px 8px; border-radius: 999px; vertical-align: middle; }
+  .pill.def { background:#eef2ff; color:#4338ca; }
+  .pill.opt { background:#f1f5f9; color:#475569; }
+
+  /* card grid */
+  .cards { display: grid; grid-template-columns: repeat(auto-fit, minmax(230px,1fr)); gap: 14px; margin: 18px 0; }
+  .card { border: 1px solid var(--line); border-radius: 12px; padding: 16px; background: var(--bg-soft); }
+  .card h4 { margin: 0 0 6px; font-size: 14.5px; }
+  .card p { margin: 0; font-size: 13px; color: var(--muted); }
+
+  /* anchor link on hover */
+  .anchor { color: var(--quiet); text-decoration: none; font-weight: 400; opacity: 0; margin-left: 8px; font-size: 0.8em; }
+  h2:hover .anchor, h3:hover .anchor { opacity: 1; }
+
+  /* ── Right TOC ───────────────────────────────────────── */
+  aside.toc {
+    position: sticky; top: 56px;
+    width: var(--toc-w); flex: none;
+    height: calc(100vh - 56px); overflow-y: auto;
+    padding: 38px 18px; border-left: 1px solid var(--line);
+  }
+  aside.toc .tl { font-size: 11.5px; font-weight: 700; text-transform: uppercase; letter-spacing: 0.07em; color: var(--quiet); margin-bottom: 10px; }
+  aside.toc a { display: block; color: var(--muted); text-decoration: none; font-size: 12.5px; padding: 4px 8px; border-left: 2px solid var(--line); line-height: 1.45; }
+  aside.toc a:hover { color: var(--ink); }
+  aside.toc a.active { color: var(--brand); border-left-color: var(--brand); font-weight: 600; }
+
+  .footer-note { margin-top: 60px; padding-top: 20px; border-top: 1px solid var(--line); color: var(--quiet); font-size: 13px; }
+
+  /* responsive */
+  @media (max-width: 1180px) { aside.toc { display: none; } }
+  @media (max-width: 860px) {
+    #menuBtn { display: inline-block; }
+    nav.sidebar {
+      position: fixed; left: 0; top: 56px; z-index: 35;
+      transform: translateX(-100%); transition: transform 0.22s ease;
+      box-shadow: 0 16px 40px rgba(15,23,42,0.18);
+    }
+    nav.sidebar.open { transform: translateX(0); }
+    main.content { padding: 28px 20px 100px; }
+    .topbar .tag { display: none; }
+  }
+</style>
+</head>
+<body>
+
+<header class="topbar">
+  <button id="menuBtn" aria-label="Toggle navigation">&#9776;</button>
+  <svg class="logo" viewBox="0 0 23 23"><rect width="10" height="10" fill="#F25022"/><rect x="13" width="10" height="10" fill="#7FBA00"/><rect y="13" width="10" height="10" fill="#00A4EF"/><rect x="13" y="13" width="10" height="10" fill="#FFB900"/></svg>
+  <span class="brand">Skill<span>Opt</span></span>
+  <span class="tag">Documentation &amp; Reproduction Guide</span>
+  <span class="spacer"></span>
+  <a class="gh" href="https://github.com/microsoft/SkillOpt" target="_blank" rel="noopener">GitHub ↗</a>
+  <a class="gh" href="https://arxiv.org/abs/2605.23904" target="_blank" rel="noopener">Paper ↗</a>
+</header>
+
+<div class="layout">
+
+  <!-- ───────────── LEFT NAV ───────────── -->
+  <nav class="sidebar" id="sidebar">
+    <div class="group">
+      <div class="glabel"><span class="num">1</span> Overview</div>
+      <a href="#what-is">What is SkillOpt</a>
+      <a href="#analogy">DL ↔ SkillOpt analogy</a>
+      <a href="#features">Key features</a>
+      <a href="#layout">Repository layout</a>
+    </div>
+    <div class="group">
+      <div class="glabel"><span class="num">2</span> Installation</div>
+      <a href="#requirements">Requirements</a>
+      <a href="#install">Install the package</a>
+      <a href="#credentials">Configure credentials</a>
+      <a href="#verify">Verify installation</a>
+    </div>
+    <div class="group">
+      <div class="glabel"><span class="num">3</span> Data Preparation</div>
+      <a href="#split-dir">Split directory format</a>
+      <a href="#item-schema">Item JSON schema</a>
+      <a href="#split-modes">Split modes</a>
+    </div>
+    <div class="group">
+      <div class="glabel"><span class="num">4</span> Quick Start</div>
+      <a href="#train">Train a skill</a>
+      <a href="#eval">Evaluate a skill</a>
+      <a href="#outputs">Output structure</a>
+      <a href="#resume">Auto-resume</a>
+    </div>
+    <div class="group">
+      <div class="glabel"><span class="num">5</span> How It Works</div>
+      <a href="#loop">The training loop</a>
+      <a href="#stages">The six per-step stages</a>
+      <a href="#gate">Validation gate</a>
+      <a href="#slow-update">Slow update (momentum)</a>
+      <a href="#meta-skill">Meta skill (memory)</a>
+      <a href="#skill-doc">Skill document anatomy</a>
+    </div>
+    <div class="group">
+      <div class="glabel"><span class="num">6</span> Configuration</div>
+      <a href="#config-system">Config system</a>
+      <a href="#cfg-model">model.*</a>
+      <a href="#cfg-train">train.*</a>
+      <a href="#cfg-gradient">gradient.*</a>
+      <a href="#cfg-optimizer">optimizer.*</a>
+      <a href="#cfg-evaluation">evaluation.*</a>
+      <a href="#cfg-env">env.*</a>
+    </div>
+    <div class="group">
+      <div class="glabel"><span class="num">7</span> Benchmarks</div>
+      <a href="#bench-list">Supported benchmarks</a>
+      <a href="#bench-new">Add a new benchmark</a>
+    </div>
+    <div class="group">
+      <div class="glabel"><span class="num">8</span> API Reference</div>
+      <a href="#module-map">Module map</a>
+      <a href="#functions">Core functions</a>
+      <a href="#cli">CLI scripts</a>
+      <a href="#webui">WebUI</a>
+    </div>
+  </nav>
+
+  <!-- ───────────── MAIN CONTENT ───────────── -->
+  <main class="content">
+
+    <span class="eyebrow">Microsoft Research</span>
+    <h1>SkillOpt Documentation &amp; Reproduction Guide</h1>
+    <p class="lead">Train agent skills like you train neural networks — with epochs, (mini-)batch size, learning rates, and validation gates — but without touching any model weights.</p>
+    <p>This guide walks you from a clean checkout to a reproduced result and a full reference for every configuration knob and core function. It is generated from, and kept consistent with, the current state of the codebase.</p>
+
+    <!-- ===================== 1. OVERVIEW ===================== -->
+    <section id="what-is">
+      <h2>1.1 What is SkillOpt <a class="anchor" href="#what-is">#</a></h2>
+      <p><strong>SkillOpt</strong> is a text-space optimizer that improves a <em>frozen</em> language agent by iteratively editing a natural-language <strong>skill document</strong> — never the model weights. The skill document is a Markdown file that conditions a target model as it executes tasks. SkillOpt treats this document as the "weights" and runs a training loop that mirrors deep-learning training: rollout (forward pass), reflect (backward pass / gradients), select &amp; apply edits (optimizer step), and a validation gate (accept/reject).</p>
+      <p>Two roles split every model call:</p>
+      <ul>
+        <li><strong>Target</strong> — executes tasks using the current skill document (the agent being improved).</li>
+        <li><strong>Optimizer</strong> — analyzes the target's trajectories and proposes edits to the skill document.</li>
+      </ul>
+      <p>The same loop drives six benchmarks out of the box (QA, document QA, embodied agents, math, spreadsheet code generation, and tool-augmented QA).</p>
+    </section>
+
+    <section id="analogy">
+      <h2>1.2 Deep-Learning ↔ SkillOpt Analogy <a class="anchor" href="#analogy">#</a></h2>
+      <p>Every concept below maps to a concrete code construct, so deep-learning intuitions transfer directly to hyperparameter tuning.</p>
+      <div class="table-wrap">
+      <table>
+        <thead><tr><th>Deep learning</th><th>SkillOpt</th><th>Where it lives</th></tr></thead>
+        <tbody>
+          <tr><td>Model weights</td><td>Skill document (Markdown)</td><td><code>skillopt/optimizer/skill.py</code></td></tr>
+          <tr><td>Forward pass</td><td>Rollout — target runs tasks</td><td><code>envs/&lt;bench&gt;/rollout.py</code></td></tr>
+          <tr><td>Loss / score</td><td>Task evaluator</td><td><code>envs/&lt;bench&gt;/evaluator.py</code></td></tr>
+          <tr><td>Backprop / gradients</td><td>Reflect → edit patches</td><td><code>gradient/reflect.py</code></td></tr>
+          <tr><td>Gradient aggregation</td><td>Hierarchical patch merge</td><td><code>gradient/aggregate.py</code></td></tr>
+          <tr><td>Gradient clipping</td><td>Rank &amp; select top-k edits</td><td><code>optimizer/clip.py</code></td></tr>
+          <tr><td>Learning rate</td><td><code>optimizer.learning_rate</code> (edits/step)</td><td><code>optimizer/scheduler.py</code></td></tr>
+          <tr><td>LR scheduler</td><td><code>lr_scheduler</code> (cosine/linear/…)</td><td><code>optimizer/scheduler.py</code></td></tr>
+          <tr><td>Optimizer step</td><td>Apply patches to the document</td><td><code>optimizer/skill.py</code></td></tr>
+          <tr><td>Validation set</td><td>Selection split (<code>valid_seen</code>)</td><td><code>evaluation/gate.py</code></td></tr>
+          <tr><td>Early stopping / accept</td><td>Validation gate</td><td><code>evaluation/gate.py</code></td></tr>
+          <tr><td>Momentum</td><td>Slow update (epoch boundary)</td><td><code>optimizer/slow_update.py</code></td></tr>
+          <tr><td>Meta-learning</td><td>Meta skill (cross-epoch memory)</td><td><code>optimizer/meta_skill.py</code></td></tr>
+          <tr><td>Batch / minibatch</td><td><code>batch_size</code> / <code>minibatch_size</code></td><td><code>engine/trainer.py</code></td></tr>
+          <tr><td>Epoch</td><td>Epoch (+ slow update &amp; meta skill)</td><td><code>engine/trainer.py</code></td></tr>
+        </tbody>
+      </table>
+      </div>
+      <div class="note tip"><span class="nh">What transfers from DL</span>
+        <p>Cosine schedule tends to beat constant; moderate learning rates (≈4–16 edits/step) beat very high/low; slow update curbs cross-epoch forgetting; meta-skill memory improves reflection quality. Conversely, bigger rollout batches and many epochs show diminishing returns — skills converge in ~2–4 epochs.</p>
+      </div>
+    </section>
+
+    <section id="features">
+      <h2>1.3 Key Features <a class="anchor" href="#features">#</a></h2>
+      <div class="cards">
+        <div class="card"><h4>Validation gating</h4><p>Every candidate skill is scored on a held-out selection split and only accepted if it beats the current/best skill.</p></div>
+        <div class="card"><h4>Slow update</h4><p>Epoch-boundary longitudinal comparison writes guidance into a protected region — momentum against forgetting. Force-injected or selection-gated.</p></div>
+        <div class="card"><h4>Meta skill</h4><p>Optimizer-side memory that reflects on what worked across epochs and feeds back into reflection.</p></div>
+        <div class="card"><h4>Pluggable backends</h4><p>OpenAI / Azure OpenAI, Anthropic Claude, local Qwen (vLLM), plus Codex/Claude-Code exec backends for the target.</p></div>
+        <div class="card"><h4>Six benchmarks</h4><p>SearchQA, DocVQA, ALFWorld, LiveMathematicianBench, SpreadsheetBench, OfficeQA — each a self-contained env module.</p></div>
+        <div class="card"><h4>Auto-resume</h4><p>Every run is checkpointed step-by-step; re-running the same command continues from the last completed step.</p></div>
+      </div>
+    </section>
+
+    <section id="layout">
+      <h2>1.4 Repository Layout <a class="anchor" href="#layout">#</a></h2>
+<pre><code><span class="tok-c"># top level</span>
+configs/            <span class="tok-c"># YAML configs (_base_ + per-benchmark)</span>
+scripts/            <span class="tok-c"># train.py, eval_only.py CLIs</span>
+ckpt/               <span class="tok-c"># packaged reference skills (e.g. gpt5.5_skill.md)</span>
+docs/               <span class="tok-c"># this guide + mkdocs sources</span>
+skillopt/           <span class="tok-c"># the package</span>
+ ├─ config.py        <span class="tok-c"># YAML loading, _base_ inheritance, flatten</span>
+ ├─ engine/trainer.py<span class="tok-c"># the training loop (ReflACTTrainer)</span>
+ ├─ gradient/        <span class="tok-c"># reflect.py (analyst), aggregate.py (merge)</span>
+ ├─ optimizer/       <span class="tok-c"># skill edits, scheduler, clip, slow_update, meta_skill</span>
+ ├─ evaluation/      <span class="tok-c"># gate.py (accept/reject logic)</span>
+ ├─ model/           <span class="tok-c"># backend clients + routing</span>
+ └─ envs/&lt;benchmark&gt;/ <span class="tok-c"># adapter, dataloader, rollout, evaluator, reflect</span></code></pre>
+    </section>
+
+    <!-- ===================== 2. INSTALLATION ===================== -->
+    <section id="requirements">
+      <h2>2.1 Requirements <a class="anchor" href="#requirements">#</a></h2>
+      <ul>
+        <li>Python ≥ 3.10</li>
+        <li>Credentials for at least one model backend (Azure OpenAI, OpenAI-compatible, Anthropic, or a local Qwen server)</li>
+        <li>Benchmark datasets are <strong>not</strong> bundled — prepare your own splits (see §3)</li>
+      </ul>
+    </section>
+
+    <section id="install">
+      <h2>2.2 Install the Package <a class="anchor" href="#install">#</a></h2>
+<pre><code><span class="tok-k">git</span> clone https://github.com/microsoft/SkillOpt.git
+<span class="tok-k">cd</span> SkillOpt
+<span class="tok-k">pip</span> install -e .
+
+<span class="tok-c"># Optional extras (install only what you need):</span>
+<span class="tok-k">pip</span> install -e <span class="tok-s">".[alfworld]"</span>   <span class="tok-c"># ALFWorld benchmark</span>
+<span class="tok-k">pip</span> install -e <span class="tok-s">".[claude]"</span>     <span class="tok-c"># Anthropic Claude backend</span>
+<span class="tok-k">pip</span> install -e <span class="tok-s">".[qwen]"</span>       <span class="tok-c"># local Qwen backend</span>
+<span class="tok-k">pip</span> install -e <span class="tok-s">".[webui]"</span>      <span class="tok-c"># monitoring dashboard</span>
+
+<span class="tok-c"># ALFWorld also needs its data assets:</span>
+<span class="tok-k">alfworld-download</span></code></pre>
+    </section>
+
+    <section id="credentials">
+      <h2>2.3 Configure Credentials <a class="anchor" href="#credentials">#</a></h2>
+      <p>Copy the template and fill in whichever backend you will use:</p>
+<pre><code><span class="tok-k">cp</span> .env.example .env
+<span class="tok-c"># edit .env, then:</span>
+<span class="tok-k">set</span> -a; <span class="tok-k">source</span> .env; <span class="tok-k">set</span> +a</code></pre>
+      <div class="note info"><span class="nh">One env-var family for all OpenAI modes</span>
+        <p>SkillOpt reuses the <code>AZURE_OPENAI_*</code> variable names even for plain OpenAI — there is no separate <code>OPENAI_API_KEY</code> knob. <code>AZURE_OPENAI_ENDPOINT</code> is required for every OpenAI auth mode.</p>
+      </div>
+      <h4>Azure OpenAI (default)</h4>
+<pre><code><span class="tok-k">export</span> AZURE_OPENAI_ENDPOINT=<span class="tok-s">"https://your-resource.openai.azure.com/"</span>
+<span class="tok-k">export</span> AZURE_OPENAI_API_VERSION=<span class="tok-s">"2024-12-01-preview"</span>
+<span class="tok-c"># Auth option 1 — API key:</span>
+<span class="tok-k">export</span> AZURE_OPENAI_API_KEY=<span class="tok-s">"your-key"</span>
+<span class="tok-c"># Auth option 2 — Azure CLI (no key; recommended on Azure VMs):</span>
+<span class="tok-k">export</span> AZURE_OPENAI_AUTH_MODE=azure_cli
+<span class="tok-c"># Auth option 3 — Managed Identity:</span>
+<span class="tok-k">export</span> AZURE_OPENAI_AUTH_MODE=managed_identity
+<span class="tok-k">export</span> AZURE_OPENAI_MANAGED_IDENTITY_CLIENT_ID=<span class="tok-s">"your-client-id"</span></code></pre>
+      <h4>OpenAI-compatible endpoint</h4>
+<pre><code><span class="tok-k">export</span> AZURE_OPENAI_ENDPOINT=<span class="tok-s">"https://api.openai.com/v1"</span>
+<span class="tok-k">export</span> AZURE_OPENAI_API_KEY=<span class="tok-s">"sk-..."</span>
+<span class="tok-k">export</span> AZURE_OPENAI_AUTH_MODE=openai_compatible</code></pre>
+      <h4>Anthropic Claude / local Qwen</h4>
+<pre><code><span class="tok-k">export</span> ANTHROPIC_API_KEY=<span class="tok-s">"sk-ant-..."</span>          <span class="tok-c"># claude_chat backend</span>
+
+<span class="tok-k">export</span> QWEN_CHAT_BASE_URL=<span class="tok-s">"http://localhost:8000/v1"</span> <span class="tok-c"># local vLLM</span>
+<span class="tok-k">export</span> QWEN_CHAT_MODEL=<span class="tok-s">"Qwen/Qwen3.5-4B"</span></code></pre>
+    </section>
+
+    <section id="verify">
+      <h2>2.4 Verify Installation <a class="anchor" href="#verify">#</a></h2>
+<pre><code><span class="tok-k">python</span> -c <span class="tok-s">"import skillopt; print('SkillOpt ready!')"</span></code></pre>
+    </section>
+
+    <!-- ===================== 3. DATA ===================== -->
+    <section id="split-dir">
+      <h2>3.1 Split Directory Format <a class="anchor" href="#split-dir">#</a></h2>
+      <p>With <code>env.split_mode: split_dir</code> (the recommended, deterministic mode), SkillOpt reads a directory containing <code>train/</code>, <code>val/</code>, and <code>test/</code> subfolders, each holding a JSON array of task items:</p>
+<pre><code>data/my_split/
+ ├─ train/items.json   <span class="tok-c"># used for rollout (the "train split")</span>
+ ├─ val/items.json     <span class="tok-c"># selection split → validation gate (valid_seen)</span>
+ └─ test/items.json    <span class="tok-c"># held-out final eval (valid_unseen)</span></code></pre>
+      <div class="note info"><span class="nh">Split naming</span>
+        <p>Internally the splits are referred to as <code>train</code>, <code>valid_seen</code> (validation/selection), and <code>valid_unseen</code> (test). The <code>--split</code> flag of <code>eval_only.py</code> uses these names.</p>
+      </div>
+    </section>
+
+    <section id="item-schema">
+      <h2>3.2 Item JSON Schema <a class="anchor" href="#item-schema">#</a></h2>
+      <p>Required fields depend on the benchmark; consult <code>skillopt/envs/&lt;benchmark&gt;/dataloader.py</code> for the exact contract. A SearchQA item, for example:</p>
+<pre><code>[
+  {
+    <span class="tok-f">"id"</span>:       <span class="tok-s">"unique_item_id"</span>,
+    <span class="tok-f">"question"</span>: <span class="tok-s">"Who wrote the novel ..."</span>,
+    <span class="tok-f">"context"</span>:  <span class="tok-s">"[DOC] relevant passage text ..."</span>,
+    <span class="tok-f">"answers"</span>:  [<span class="tok-s">"expected answer"</span>]
+  }
+]</code></pre>
+      <div class="note warn"><span class="nh">Datasets not included</span>
+        <p>This repository ships no benchmark data. Prepare your own splits in the format above before training.</p>
+      </div>
+    </section>
+
+    <section id="split-modes">
+      <h2>3.3 Split Modes <a class="anchor" href="#split-modes">#</a></h2>
+      <div class="table-wrap"><table>
+        <thead><tr><th><code>env.split_mode</code></th><th>Behavior</th></tr></thead>
+        <tbody>
+          <tr><td><code>split_dir</code></td><td>Use a pre-built directory with explicit <code>train/val/test</code> folders (set <code>env.split_dir</code>). Deterministic and reproducible.</td></tr>
+          <tr><td><code>ratio</code></td><td>Build a deterministic split on the fly from a single <code>env.data_path</code>, using <code>split_seed</code> (and a train:val:test ratio). Convenient for quick experiments.</td></tr>
+        </tbody>
+      </table></div>
+    </section>
+
+    <!-- ===================== 4. QUICK START ===================== -->
+    <section id="train">
+      <h2>4.1 Train a Skill <a class="anchor" href="#train">#</a></h2>
+<pre><code><span class="tok-c"># Minimal SearchQA run</span>
+<span class="tok-k">python</span> scripts/train.py \
+    <span class="tok-f">--config</span> configs/searchqa/default.yaml \
+    <span class="tok-f">--split_dir</span> /path/to/your/searchqa_split \
+    <span class="tok-f">--azure_openai_endpoint</span> https://your-resource.openai.azure.com/ \
+    <span class="tok-f">--optimizer_model</span> gpt-5.5 \
+    <span class="tok-f">--target_model</span> gpt-5.5</code></pre>
+      <p>Swap the config for another benchmark (e.g. <code>configs/livemathematicianbench/default.yaml</code>, <code>configs/alfworld/default.yaml</code>). Common CLI arguments:</p>
+      <div class="table-wrap"><table>
+        <thead><tr><th>Argument</th><th>Description</th></tr></thead>
+        <tbody>
+          <tr><td><code>--config</code></td><td>Benchmark config YAML (required)</td></tr>
+          <tr><td><code>--split_dir</code></td><td>Path to the data split directory</td></tr>
+          <tr><td><code>--azure_openai_endpoint</code></td><td>Azure OpenAI endpoint URL</td></tr>
+          <tr><td><code>--optimizer_model</code> / <code>--target_model</code></td><td>Deployment names for optimizer / target</td></tr>
+          <tr><td><code>--num_epochs</code> / <code>--batch_size</code></td><td>Epochs and rollout batch size</td></tr>
+          <tr><td><code>--out_root</code></td><td>Output directory</td></tr>
+          <tr><td><code>--cfg-options k=v ...</code></td><td>Override any config key (see §6.1)</td></tr>
+        </tbody>
+      </table></div>
+    </section>
+
+    <section id="eval">
+      <h2>4.2 Evaluate a Skill <a class="anchor" href="#eval">#</a></h2>
+      <p>Evaluate any skill document (a packaged reference skill, or a trained run's <code>best_skill.md</code>) without training:</p>
+<pre><code><span class="tok-c"># Evaluate the packaged GPT-5.5 SearchQA skill on the test split</span>
+<span class="tok-k">python</span> scripts/eval_only.py \
+  <span class="tok-f">--config</span> configs/searchqa/default.yaml \
+  <span class="tok-f">--skill</span> ckpt/searchqa/gpt5.5_skill.md \
+  <span class="tok-f">--split</span> valid_unseen \
+  <span class="tok-f">--split_dir</span> /path/to/searchqa_split \
+  <span class="tok-f">--azure_openai_endpoint</span> https://your-resource.openai.azure.com/</code></pre>
+      <div class="table-wrap"><table>
+        <thead><tr><th><code>--split</code></th><th>Meaning</th></tr></thead>
+        <tbody>
+          <tr><td><code>valid_unseen</code></td><td>Test set (held-out)</td></tr>
+          <tr><td><code>valid_seen</code></td><td>Validation / selection set</td></tr>
+          <tr><td><code>train</code></td><td>Training set</td></tr>
+          <tr><td><code>all</code></td><td>All splits combined (default)</td></tr>
+        </tbody>
+      </table></div>
+    </section>
+
+    <section id="outputs">
+      <h2>4.3 Output Structure <a class="anchor" href="#outputs">#</a></h2>
+<pre><code>outputs/&lt;run_name&gt;/
+ ├─ config.json          <span class="tok-c"># flattened runtime config</span>
+ ├─ history.json         <span class="tok-c"># per-step training history</span>
+ ├─ runtime_state.json   <span class="tok-c"># resume checkpoint</span>
+ ├─ best_skill.md        <span class="tok-c"># best validated skill document</span>
+ ├─ skills/skill_vXXXX.md<span class="tok-c"># skill snapshot per step</span>
+ ├─ steps/step_XXXX/     <span class="tok-c"># per-step artifacts (patches, evals)</span>
+ ├─ slow_update/epoch_XX/<span class="tok-c"># slow-update logs &amp; rollouts</span>
+ └─ meta_skill/epoch_XX/ <span class="tok-c"># meta-skill logs</span></code></pre>
+    </section>
+
+    <section id="resume">
+      <h2>4.4 Auto-Resume <a class="anchor" href="#resume">#</a></h2>
+      <p>Each completed step persists its state to <code>runtime_state.json</code> and a <code>steps/step_XXXX/</code> directory. Re-running the <em>same command</em> against the same <code>out_root</code> detects finished work and continues from the last completed step — including epoch-boundary slow-update and meta-skill stages.</p>
+    </section>
+
+    <!-- ===================== 5. HOW IT WORKS ===================== -->
+    <section id="loop">
+      <h2>5.1 The Training Loop <a class="anchor" href="#loop">#</a></h2>
+      <p>The loop lives in <code>ReflACTTrainer</code> (<code>skillopt/engine/trainer.py</code>). Each epoch runs a series of optimization steps over rollout batches, then performs two epoch-boundary stages.</p>
+<pre><code><span class="tok-k">for</span> epoch <span class="tok-k">in</span> epochs:
+    <span class="tok-k">for</span> step <span class="tok-k">in</span> steps:
+        1. Rollout    <span class="tok-c"># target executes a batch of tasks</span>
+        2. Reflect    <span class="tok-c"># optimizer analyzes trajectories → edit patches</span>
+        3. Aggregate  <span class="tok-c"># hierarchically merge similar patches</span>
+        4. Select     <span class="tok-c"># rank &amp; clip edits to the learning rate</span>
+        5. Update     <span class="tok-c"># apply patches → candidate skill</span>
+        6. Gate       <span class="tok-c"># score on selection split → accept / reject</span>
+
+    <span class="tok-c"># epoch boundary (from epoch 2 onward)</span>
+    Slow update   <span class="tok-c"># longitudinal comparison → protected guidance</span>
+    Meta skill    <span class="tok-c"># cross-epoch optimizer memory</span></code></pre>
+    </section>
+
+    <section id="stages">
+      <h2>5.2 The Six Per-Step Stages <a class="anchor" href="#stages">#</a></h2>
+      <div class="table-wrap"><table>
+        <thead><tr><th>Stage</th><th>What happens</th><th>Source</th></tr></thead>
+        <tbody>
+          <tr><td><strong>1. Rollout</strong></td><td>The target model runs each task in the batch with the current skill as context, producing trajectories and scores.</td><td><code>envs/&lt;b&gt;/rollout.py</code></td></tr>
+          <tr><td><strong>2. Reflect</strong></td><td>The optimizer runs an error analyst (and optional success analyst) over minibatches of trajectories, emitting structured edit patches. Runs in parallel across <code>analyst_workers</code>.</td><td><code>gradient/reflect.py</code></td></tr>
+          <tr><td><strong>3. Aggregate</strong></td><td>Semantically similar patches are merged hierarchically to remove redundancy.</td><td><code>gradient/aggregate.py</code> → <code>merge_patches</code></td></tr>
+          <tr><td><strong>4. Select</strong></td><td>Patches are ranked and clipped to the current learning rate (max edits this step), set by the scheduler.</td><td><code>optimizer/clip.py</code> → <code>rank_and_select</code></td></tr>
+          <tr><td><strong>5. Update</strong></td><td>Selected edits are applied to the skill document, producing a candidate skill (patch / rewrite modes).</td><td><code>optimizer/skill.py</code>, <code>update_modes.py</code></td></tr>
+          <tr><td><strong>6. Gate</strong></td><td>The candidate is scored on the selection split and accepted only if it improves (see §5.3).</td><td><code>evaluation/gate.py</code> → <code>evaluate_gate</code></td></tr>
+        </tbody>
+      </table></div>
+    </section>
+
+    <section id="gate">
+      <h2>5.3 Validation Gate <a class="anchor" href="#gate">#</a></h2>
+      <p><code>evaluate_gate</code> is a pure decision function. It compares the candidate's selection-set score against the <em>current</em> and <em>best</em> skills:</p>
+      <ul>
+        <li><strong>accept_new_best</strong> — candidate &gt; current <em>and</em> candidate &gt; best → becomes both current and best.</li>
+        <li><strong>accept</strong> — candidate &gt; current but ≤ best → becomes current only.</li>
+        <li><strong>reject</strong> — candidate ≤ current → discarded; current/best unchanged.</li>
+      </ul>
+      <p>The comparison metric is configurable via <code>evaluation.gate_metric</code>:</p>
+      <div class="table-wrap"><table>
+        <thead><tr><th>Metric</th><th>Score used</th></tr></thead>
+        <tbody>
+          <tr><td><code>hard</code> <span class="pill def">default</span></td><td>Exact-match / discrete score</td></tr>
+          <tr><td><code>soft</code></td><td>Partial-credit / continuous score</td></tr>
+          <tr><td><code>mixed</code></td><td>Weighted blend, controlled by <code>gate_mixed_weight</code></td></tr>
+        </tbody>
+      </table></div>
+      <div class="note info"><span class="nh">When to use soft/mixed</span>
+        <p>The <code>soft</code>/<code>mixed</code> metrics (contributed config <code>configs/examples/soft_gate.yaml</code>) help when the selection split is small and rewards are continuous, where a discrete hard gate may reject every candidate and stall training. Paper numbers use the default <code>hard</code> gate.</p>
+      </div>
+    </section>
+
+    <section id="slow-update">
+      <h2>5.4 Slow Update (Momentum) <a class="anchor" href="#slow-update">#</a></h2>
+      <p>At each epoch boundary (from epoch 2), the slow update rolls out both the <em>previous</em> epoch's skill and the <em>current</em> skill on the same sampled tasks, categorizes items (improved / regressed / persistent-fail / stable-success), and asks the optimizer to write a free-form <strong>guidance</strong> block. This guidance lands in a <strong>protected region</strong> of the skill that step-level edits cannot touch — only the slow update overwrites it. It is SkillOpt's analogue of momentum, countering cross-epoch forgetting.</p>
+      <p>Acceptance has two modes, selected by <code>optimizer.slow_update_gate_with_selection</code>:</p>
+      <div class="table-wrap"><table>
+        <thead><tr><th>Mode</th><th>Behavior</th></tr></thead>
+        <tbody>
+          <tr><td><code>false</code> <span class="pill def">default</span> — force-injected</td><td>Guidance is injected into both current and best skills unconditionally. The longitudinal guidance always persists; it is not gated by step-level selection scores.</td></tr>
+          <tr><td><code>true</code> — gated</td><td>The slow-update candidate is scored on the selection split and accepted/rejected through the same validation gate as step-level updates.</td></tr>
+        </tbody>
+      </table></div>
+    </section>
+
+    <section id="meta-skill">
+      <h2>5.5 Meta Skill (Optimizer Memory) <a class="anchor" href="#meta-skill">#</a></h2>
+      <p>The meta skill is <strong>optimizer-side memory</strong> — it never modifies the target skill document. At the end of each epoch (skipped for epoch 1), the optimizer compares the previous and current epoch's last-step skills on the same sampled tasks and writes a compact, evidence-based reflection on what kind of edits helped or hurt. That memory is then injected as extra context into the next epoch's reflect / merge / learning-rate / ranking stages, so the optimizer accumulates strategy across the run.</p>
+    </section>
+
+    <section id="skill-doc">
+      <h2>5.6 Skill Document Anatomy <a class="anchor" href="#skill-doc">#</a></h2>
+      <p>A skill document is plain Markdown. Initial skills can be empty (learn from scratch) or seeded with domain knowledge via <code>env.skill_init</code>. During training the document accrues rules, patterns, and edge-case handling through accepted edit patches. A dedicated protected region holds the slow-update guidance, delimited by HTML-comment markers:</p>
+<pre><code><span class="tok-c"># Question Answering Skill</span>
+
+<span class="tok-c">## Learned rules ...</span>
+- When the context contains multiple candidates, prefer ...
+
+<span class="tok-c">&lt;!-- SLOW_UPDATE_START --&gt;</span>
+<span class="tok-c"># (epoch-level longitudinal guidance — only the slow update writes here)</span>
+<span class="tok-c">&lt;!-- SLOW_UPDATE_END --&gt;</span></code></pre>
+      <p>Helpers in <code>optimizer/slow_update.py</code> manage this region: <code>inject_empty_slow_update_field</code> (placeholder at epoch 1), <code>extract_slow_update_field</code> (read), and <code>replace_slow_update_field</code> (overwrite). Step-level edits are blocked from modifying anything inside the markers.</p>
+    </section>
+
+    <!-- ===================== 6. CONFIGURATION ===================== -->
+    <section id="config-system">
+      <h2>6.1 Configuration System <a class="anchor" href="#config-system">#</a></h2>
+      <p>Configs are <strong>structured YAML</strong> with section blocks (<code>model</code>, <code>train</code>, <code>gradient</code>, <code>optimizer</code>, <code>evaluation</code>, <code>env</code>) and <code>_base_</code> inheritance. A benchmark config inherits the shared defaults and overrides only what differs:</p>
+<pre><code><span class="tok-c"># configs/searchqa/default.yaml</span>
+<span class="tok-f">_base_</span>: ../_base_/default.yaml
+<span class="tok-f">train</span>:
+  <span class="tok-f">train_size</span>: <span class="tok-n">400</span>
+  <span class="tok-f">batch_size</span>: <span class="tok-n">40</span>
+<span class="tok-f">optimizer</span>:
+  <span class="tok-f">learning_rate</span>: <span class="tok-n">4</span>
+<span class="tok-f">env</span>:
+  <span class="tok-f">name</span>: searchqa
+  <span class="tok-f">split_dir</span>: data/searchqa_split</code></pre>
+      <p>Override any key at the command line without editing files:</p>
+<pre><code><span class="tok-k">python</span> scripts/train.py --config configs/searchqa/default.yaml \
+  <span class="tok-f">--cfg-options</span> optimizer.learning_rate=<span class="tok-n">16</span> optimizer.lr_scheduler=linear</code></pre>
+      <div class="note info"><span class="nh">Reading the tables below</span>
+        <p>Each section lists the key (relative to its YAML block), type, default (from <code>configs/_base_/default.yaml</code>), allowed values, and meaning. Defaults shown are the shipped base defaults.</p>
+      </div>
+    </section>
+
+    <section id="cfg-model">
+      <h2>6.2 <code>model.*</code> <a class="anchor" href="#cfg-model">#</a></h2>
+      <div class="table-wrap"><table>
+        <thead><tr><th>Key</th><th>Type</th><th>Default</th><th>Description / options</th></tr></thead>
+        <tbody>
+          <tr><td><code>backend</code></td><td>str</td><td class="def">azure_openai</td><td>High-level backend label for the run.</td></tr>
+          <tr><td><code>optimizer</code></td><td>str</td><td class="def">gpt-5.5</td><td>Optimizer model deployment (writes skill edits).</td></tr>
+          <tr><td><code>target</code></td><td>str</td><td class="def">gpt-5.5</td><td>Target model deployment (executes tasks).</td></tr>
+          <tr><td><code>optimizer_backend</code></td><td>str</td><td class="def">openai_chat</td><td>Client path for the optimizer: <code>openai_chat</code> or <code>claude_chat</code>.</td></tr>
+          <tr><td><code>target_backend</code></td><td>str</td><td class="def">openai_chat</td><td>Client path for the target: <code>openai_chat</code> / <code>claude_chat</code> / <code>qwen_chat</code> / <code>codex_exec</code> / <code>claude_code_exec</code>.</td></tr>
+          <tr><td><code>reasoning_effort</code></td><td>str</td><td class="def">medium</td><td><code>low</code> / <code>medium</code> / <code>high</code> / <code>xhigh</code> / <code>max</code> (or empty).</td></tr>
+          <tr><td><code>rewrite_reasoning_effort</code></td><td>str</td><td class="def">""</td><td>Override effort for full-rewrite calls (empty = inherit).</td></tr>
+          <tr><td><code>rewrite_max_completion_tokens</code></td><td>int</td><td class="def">64000</td><td>Token cap for full-rewrite optimizer calls.</td></tr>
+          <tr><td><code>azure_openai_endpoint</code></td><td>str</td><td class="def">""</td><td>Azure resource URL (or via <code>AZURE_OPENAI_ENDPOINT</code>).</td></tr>
+          <tr><td><code>azure_openai_api_version</code></td><td>str</td><td class="def">2024-12-01-preview</td><td>Azure API version header.</td></tr>
+          <tr><td><code>azure_openai_auth_mode</code></td><td>str</td><td class="def">""</td><td><code>api_key</code> / <code>azure_cli</code> / <code>managed_identity</code> / <code>openai_compatible</code> (empty → env default).</td></tr>
+        </tbody>
+      </table></div>
+      <div class="note info"><span class="nh">Separate optimizer / target endpoints</span>
+        <p>Every <code>azure_openai_*</code> key also has <code>optimizer_azure_openai_*</code> and <code>target_azure_openai_*</code> variants, letting you point the optimizer and target at different Azure resources. Exec backends (<code>codex_exec</code>, <code>claude_code_exec</code>) add their own <code>codex_exec_*</code> / <code>claude_code_exec_*</code> knobs (sandbox, reasoning effort, SDK mode, etc.).</p>
+      </div>
+    </section>
+
+    <section id="cfg-train">
+      <h2>6.3 <code>train.*</code> <a class="anchor" href="#cfg-train">#</a></h2>
+      <div class="table-wrap"><table>
+        <thead><tr><th>Key</th><th>Type</th><th>Default</th><th>DL analogy</th><th>Description</th></tr></thead>
+        <tbody>
+          <tr><td><code>num_epochs</code></td><td>int</td><td class="def">4</td><td>Epochs</td><td>Number of training epochs.</td></tr>
+          <tr><td><code>train_size</code></td><td>int</td><td class="def">0</td><td>Train-set size</td><td>0 = derive from the dataset split. (Fixed by split size when using <code>split_dir</code>.)</td></tr>
+          <tr><td><code>batch_size</code></td><td>int</td><td class="def">40</td><td>Batch size</td><td>Tasks rolled out per optimization step.</td></tr>
+          <tr><td><code>accumulation</code></td><td>int</td><td class="def">1</td><td>Grad accumulation</td><td>Accumulation rounds per step.</td></tr>
+          <tr><td><code>seed</code></td><td>int</td><td class="def">42</td><td>Random seed</td><td>Reproducibility seed.</td></tr>
+        </tbody>
+      </table></div>
+    </section>
+
+    <section id="cfg-gradient">
+      <h2>6.4 <code>gradient.*</code> <a class="anchor" href="#cfg-gradient">#</a></h2>
+      <div class="table-wrap"><table>
+        <thead><tr><th>Key</th><th>Type</th><th>Default</th><th>Description</th></tr></thead>
+        <tbody>
+          <tr><td><code>minibatch_size</code></td><td>int</td><td class="def">8</td><td>Trajectories per reflect minibatch.</td></tr>
+          <tr><td><code>merge_batch_size</code></td><td>int</td><td class="def">8</td><td>Patches per merge batch during aggregation.</td></tr>
+          <tr><td><code>analyst_workers</code></td><td>int</td><td class="def">16</td><td>Parallel reflection workers (data parallelism).</td></tr>
+          <tr><td><code>max_analyst_rounds</code></td><td>int</td><td class="def">3</td><td>Max rounds of analyst reflection per step.</td></tr>
+          <tr><td><code>failure_only</code></td><td>bool</td><td class="def">false</td><td>Reflect only on failed trajectories when true.</td></tr>
+        </tbody>
+      </table></div>
+    </section>
+
+    <section id="cfg-optimizer">
+      <h2>6.5 <code>optimizer.*</code> <a class="anchor" href="#cfg-optimizer">#</a></h2>
+      <div class="table-wrap"><table>
+        <thead><tr><th>Key</th><th>Type</th><th>Default</th><th>DL analogy</th><th>Description / options</th></tr></thead>
+        <tbody>
+          <tr><td><code>learning_rate</code></td><td>int</td><td class="def">4</td><td>Learning rate</td><td>Max edit patches applied per step (the "edit budget").</td></tr>
+          <tr><td><code>min_learning_rate</code></td><td>int</td><td class="def">2</td><td>Min LR</td><td>Floor edit budget for decaying schedulers.</td></tr>
+          <tr><td><code>lr_scheduler</code></td><td>str</td><td class="def">cosine</td><td>LR schedule</td><td><code>constant</code> / <code>linear</code> / <code>cosine</code> / <code>autonomous</code>.</td></tr>
+          <tr><td><code>lr_control_mode</code></td><td>str</td><td class="def">fixed</td><td>—</td><td><code>fixed</code> / <code>autonomous</code> / <code>none</code>.</td></tr>
+          <tr><td><code>skill_update_mode</code></td><td>str</td><td class="def">patch</td><td>—</td><td><code>patch</code> / <code>rewrite_from_suggestions</code> / <code>full_rewrite_minibatch</code>.</td></tr>
+          <tr><td><code>use_slow_update</code></td><td>bool</td><td class="def">true</td><td>Momentum</td><td>Enable epoch-boundary slow update.</td></tr>
+          <tr><td><code>slow_update_samples</code></td><td>int</td><td class="def">20</td><td>—</td><td>Tasks sampled for the longitudinal comparison.</td></tr>
+          <tr><td><code>slow_update_gate_with_selection</code></td><td>bool</td><td class="def">false</td><td>—</td><td><code>false</code> = force-inject guidance; <code>true</code> = gate it on the selection split (see §5.4).</td></tr>
+          <tr><td><code>longitudinal_pair_policy</code></td><td>str</td><td class="def">mixed</td><td>—</td><td><code>mixed</code> / <code>changed</code> / <code>unchanged</code> — which comparison pairs to keep.</td></tr>
+          <tr><td><code>use_meta_skill</code></td><td>bool</td><td class="def">true</td><td>Meta-learning</td><td>Enable cross-epoch optimizer memory.</td></tr>
+        </tbody>
+      </table></div>
+    </section>
+
+    <section id="cfg-evaluation">
+      <h2>6.6 <code>evaluation.*</code> <a class="anchor" href="#cfg-evaluation">#</a></h2>
+      <div class="table-wrap"><table>
+        <thead><tr><th>Key</th><th>Type</th><th>Default</th><th>Description / options</th></tr></thead>
+        <tbody>
+          <tr><td><code>use_gate</code></td><td>bool</td><td class="def">true</td><td>Validation gating is mandatory in this branch (must remain <code>true</code>).</td></tr>
+          <tr><td><code>gate_metric</code></td><td>str</td><td class="def">hard</td><td><code>hard</code> / <code>soft</code> / <code>mixed</code> — score used by the gate (see §5.3).</td></tr>
+          <tr><td><code>gate_mixed_weight</code></td><td>float</td><td class="def">0.5</td><td>Weight on the soft score when <code>gate_metric = mixed</code>.</td></tr>
+          <tr><td><code>sel_env_num</code></td><td>int</td><td class="def">0</td><td>Selection-split eval size (0 = use full split).</td></tr>
+          <tr><td><code>test_env_num</code></td><td>int</td><td class="def">0</td><td>Test-split eval size (0 = use full split).</td></tr>
+          <tr><td><code>eval_test</code></td><td>bool</td><td class="def">true</td><td>Run a final test evaluation after training.</td></tr>
+        </tbody>
+      </table></div>
+      <div class="note warn"><span class="nh">Gate is required</span>
+        <p>Setting <code>evaluation.use_gate: false</code> raises an error — validation gating cannot be disabled in this branch.</p>
+      </div>
+    </section>
+
+    <section id="cfg-env">
+      <h2>6.7 <code>env.*</code> <a class="anchor" href="#cfg-env">#</a></h2>
+      <div class="table-wrap"><table>
+        <thead><tr><th>Key</th><th>Type</th><th>Default</th><th>Description</th></tr></thead>
+        <tbody>
+          <tr><td><code>name</code></td><td>str</td><td class="def">""</td><td>Benchmark name (<code>searchqa</code>, <code>docvqa</code>, <code>alfworld</code>, …). Selects the env module.</td></tr>
+          <tr><td><code>skill_init</code></td><td>str</td><td class="def">""</td><td>Path to a seed skill (empty = start from scratch).</td></tr>
+          <tr><td><code>split_mode</code></td><td>str</td><td class="def">ratio</td><td><code>ratio</code> or <code>split_dir</code> (see §3.3).</td></tr>
+          <tr><td><code>split_dir</code></td><td>str</td><td class="def">""</td><td>Pre-split directory (when <code>split_mode = split_dir</code>).</td></tr>
+          <tr><td><code>data_path</code></td><td>str</td><td class="def">""</td><td>Single dataset path (when <code>split_mode = ratio</code>).</td></tr>
+          <tr><td><code>split_seed</code></td><td>int</td><td class="def">42</td><td>Seed for deterministic ratio splitting.</td></tr>
+          <tr><td><code>exec_timeout</code></td><td>int</td><td class="def">120</td><td>Per-task target/code-agent timeout (seconds).</td></tr>
+          <tr><td><code>out_root</code></td><td>str</td><td class="def">""</td><td>Output directory for the run.</td></tr>
+        </tbody>
+      </table></div>
+      <div class="note info"><span class="nh">Benchmark-specific env keys</span>
+        <p>Env blocks may carry extra benchmark-specific keys (e.g. <code>max_turns</code>, <code>workers</code>, <code>max_completion_tokens</code>, <code>limit</code>). Unmapped env keys are passed straight through to the benchmark adapter — check the relevant <code>configs/&lt;benchmark&gt;/default.yaml</code>.</p>
+      </div>
+    </section>
+
+    <!-- ===================== 7. BENCHMARKS ===================== -->
+    <section id="bench-list">
+      <h2>7.1 Supported Benchmarks <a class="anchor" href="#bench-list">#</a></h2>
+      <div class="table-wrap"><table>
+        <thead><tr><th>Benchmark</th><th>Type</th><th>Config</th></tr></thead>
+        <tbody>
+          <tr><td>SearchQA</td><td>Question answering</td><td><code>configs/searchqa/default.yaml</code></td></tr>
+          <tr><td>DocVQA</td><td>Document QA</td><td><code>configs/docvqa/default.yaml</code></td></tr>
+          <tr><td>ALFWorld</td><td>Embodied agent</td><td><code>configs/alfworld/default.yaml</code></td></tr>
+          <tr><td>LiveMathematicianBench</td><td>Math reasoning</td><td><code>configs/livemathematicianbench/default.yaml</code></td></tr>
+          <tr><td>SpreadsheetBench</td><td>Spreadsheet code generation</td><td><code>configs/spreadsheetbench/default.yaml</code></td></tr>
+          <tr><td>OfficeQA</td><td>Tool-augmented QA</td><td><code>configs/officeqa/default.yaml</code></td></tr>
+        </tbody>
+      </table></div>
+      <p>Each benchmark is a self-contained module under <code>skillopt/envs/&lt;benchmark&gt;/</code> with an <code>adapter.py</code>, <code>dataloader.py</code>, <code>rollout.py</code>, and <code>evaluator.py</code> (some add a custom <code>reflect.py</code>). Packaged reference skills live in <code>ckpt/&lt;benchmark&gt;/</code>.</p>
+    </section>
+
+    <section id="bench-new">
+      <h2>7.2 Add a New Benchmark <a class="anchor" href="#bench-new">#</a></h2>
+      <p>Use <code>skillopt/envs/_template/</code> as a starting point. At minimum, implement:</p>
+      <ol>
+        <li><strong>Dataloader</strong> — read your item JSON into the framework's item dicts (<code>dataloader.py</code>).</li>
+        <li><strong>Rollout</strong> — run the target on one item with the current skill and return a trajectory + score (<code>rollout.py</code>).</li>
+        <li><strong>Evaluator</strong> — score predictions against ground truth (<code>evaluator.py</code>).</li>
+        <li><strong>Adapter</strong> — wire the above into the trainer's expected interface and register the env name (<code>adapter.py</code>).</li>
+      </ol>
+      <p>Then add a <code>configs/&lt;name&gt;/default.yaml</code> inheriting <code>_base_/default.yaml</code> and set <code>env.name</code> to your new benchmark.</p>
+    </section>
+
+    <!-- ===================== 8. API REFERENCE ===================== -->
+    <section id="module-map">
+      <h2>8.1 Module Map <a class="anchor" href="#module-map">#</a></h2>
+      <div class="table-wrap"><table>
+        <thead><tr><th>Module</th><th>Responsibility</th></tr></thead>
+        <tbody>
+          <tr><td><code>skillopt/config.py</code></td><td>Load structured YAML, resolve <code>_base_</code> inheritance, flatten to the trainer's flat dict, apply CLI overrides.</td></tr>
+          <tr><td><code>skillopt/engine/trainer.py</code></td><td><code>ReflACTTrainer</code> — orchestrates the whole loop, gating, slow update, meta skill, resume, and artifact writing.</td></tr>
+          <tr><td><code>skillopt/gradient/</code></td><td>Reflection ("backward pass"): <code>reflect.py</code> analysts, <code>aggregate.py</code> patch merging.</td></tr>
+          <tr><td><code>skillopt/optimizer/</code></td><td>The "optimizer": edit application, learning-rate scheduling, edit selection, slow update, meta skill, rewrite modes.</td></tr>
+          <tr><td><code>skillopt/evaluation/gate.py</code></td><td>Pure accept/reject decision and metric selection.</td></tr>
+          <tr><td><code>skillopt/model/</code></td><td>Backend clients (OpenAI/Azure, Claude, Qwen, Codex/Claude-Code exec) and routing.</td></tr>
+          <tr><td><code>skillopt/envs/&lt;b&gt;/</code></td><td>Per-benchmark dataloader, rollout, evaluator, adapter.</td></tr>
+        </tbody>
+      </table></div>
+    </section>
+
+    <section id="functions">
+      <h2>8.2 Core Functions <a class="anchor" href="#functions">#</a></h2>
+      <div class="table-wrap"><table>
+        <thead><tr><th>Function</th><th>File</th><th>Purpose</th></tr></thead>
+        <tbody>
+          <tr><td><code>load_config</code> / <code>flatten_config</code> / <code>apply_overrides</code></td><td><code>config.py</code></td><td>Load YAML with inheritance; flatten sections; apply <code>key=value</code> overrides.</td></tr>
+          <tr><td><code>run_minibatch_reflect</code></td><td><code>gradient/reflect.py</code></td><td>Run error/success analysts over trajectory minibatches → edit patches.</td></tr>
+          <tr><td><code>merge_patches</code></td><td><code>gradient/aggregate.py</code></td><td>Hierarchically merge semantically similar patches.</td></tr>
+          <tr><td><code>rank_and_select</code></td><td><code>optimizer/clip.py</code></td><td>Rank edits and clip to the learning-rate budget.</td></tr>
+          <tr><td><code>build_scheduler</code></td><td><code>optimizer/scheduler.py</code></td><td>Construct the LR (edit-budget) scheduler: constant/linear/cosine/autonomous.</td></tr>
+          <tr><td><code>decide_autonomous_learning_rate</code></td><td><code>optimizer/lr_autonomous.py</code></td><td>Let the optimizer pick the next learning rate (autonomous mode).</td></tr>
+          <tr><td><code>apply_patch</code> / <code>apply_edit</code></td><td><code>optimizer/skill.py</code></td><td>Apply edits to the skill document (respecting the protected region).</td></tr>
+          <tr><td><code>rewrite_skill_from_suggestions</code></td><td><code>optimizer/rewrite.py</code></td><td>Full-rewrite update mode from accumulated suggestions.</td></tr>
+          <tr><td><code>evaluate_gate</code> / <code>select_gate_score</code></td><td><code>evaluation/gate.py</code></td><td>Accept/reject decision; compute hard/soft/mixed score.</td></tr>
+          <tr><td><code>run_slow_update</code></td><td><code>optimizer/slow_update.py</code></td><td>Produce epoch-boundary longitudinal guidance.</td></tr>
+          <tr><td><code>replace_slow_update_field</code> / <code>extract_slow_update_field</code></td><td><code>optimizer/slow_update.py</code></td><td>Read/overwrite the protected guidance region.</td></tr>
+          <tr><td><code>run_meta_skill</code> / <code>format_meta_skill_context</code></td><td><code>optimizer/meta_skill.py</code></td><td>Generate cross-epoch optimizer memory and render it into reflection context.</td></tr>
+        </tbody>
+      </table></div>
+    </section>
+
+    <section id="cli">
+      <h2>8.3 CLI Scripts <a class="anchor" href="#cli">#</a></h2>
+      <h4>scripts/train.py</h4>
+      <p>Runs a full training loop. Required: <code>--config</code>. Override config via <code>--cfg-options section.key=value …</code> or legacy flat flags (<code>--num_epochs</code>, <code>--batch_size</code>, <code>--optimizer_model</code>, <code>--target_model</code>, <code>--lr_scheduler</code>, <code>--edit_budget</code>, <code>--split_dir</code>, …).</p>
+      <h4>scripts/eval_only.py</h4>
+      <p>Evaluates a skill document without training. Required: <code>--config</code> and <code>--skill</code>. Use <code>--split</code> to choose <code>train</code> / <code>valid_seen</code> / <code>valid_unseen</code> / <code>all</code>.</p>
+<pre><code><span class="tok-k">python</span> scripts/eval_only.py \
+  --config configs/searchqa/default.yaml \
+  --skill outputs/my_run/best_skill.md \
+  --split valid_unseen</code></pre>
+    </section>
+
+    <section id="webui">
+      <h2>8.4 WebUI <a class="anchor" href="#webui">#</a></h2>
+      <p>An optional Gradio dashboard to configure parameters and monitor runs:</p>
+<pre><code><span class="tok-k">pip</span> install -e <span class="tok-s">".[webui]"</span>
+<span class="tok-k">python</span> -m skillopt_webui.app          <span class="tok-c"># http://localhost:7860</span>
+<span class="tok-k">python</span> -m skillopt_webui.app --share  <span class="tok-c"># public share link</span></code></pre>
+      <div class="table-wrap"><table>
+        <thead><tr><th>Flag</th><th>Default</th><th>Description</th></tr></thead>
+        <tbody>
+          <tr><td><code>--port</code></td><td class="def">7860</td><td>Server port.</td></tr>
+          <tr><td><code>--host</code></td><td class="def">0.0.0.0</td><td>Bind address.</td></tr>
+          <tr><td><code>--share</code></td><td class="def">off</td><td>Create a public Gradio share link.</td></tr>
+        </tbody>
+      </table></div>
+
+      <div class="footer-note">
+        SkillOpt — Executive Strategy for Self-Evolving Agent Skills ·
+        <a href="https://github.com/microsoft/SkillOpt">github.com/microsoft/SkillOpt</a> ·
+        <a href="https://arxiv.org/abs/2605.23904">arXiv:2605.23904</a><br>
+        This guide reflects the current configuration defaults in <code>configs/_base_/default.yaml</code>. When in doubt, the code is the source of truth.
+      </div>
+    </section>
+
+  </main>
+
+  <!-- ───────────── RIGHT TOC ───────────── -->
+  <aside class="toc" id="toc">
+    <div class="tl">On this page</div>
+    <div id="tocLinks"></div>
+  </aside>
+
+</div>
+
+<script>
+(function () {
+  // Build right-hand "On this page" from <h2> elements
+  var sections = Array.prototype.slice.call(document.querySelectorAll('main.content section[id]'));
+  var tocLinks = document.getElementById('tocLinks');
+  var h2s = Array.prototype.slice.call(document.querySelectorAll('main.content h2'));
+  h2s.forEach(function (h) {
+    var sec = h.closest('section');
+    if (!sec || !sec.id) return;
+    var a = document.createElement('a');
+    a.href = '#' + sec.id;
+    a.textContent = h.textContent.replace(/#$/, '').trim();
+    a.dataset.target = sec.id;
+    tocLinks.appendChild(a);
+  });
+
+  var sideLinks = Array.prototype.slice.call(document.querySelectorAll('nav.sidebar a'));
+  var tocAnchors = Array.prototype.slice.call(tocLinks.querySelectorAll('a'));
+
+  function setActive(id) {
+    sideLinks.forEach(function (a) {
+      a.classList.toggle('active', a.getAttribute('href') === '#' + id);
+    });
+    tocAnchors.forEach(function (a) {
+      a.classList.toggle('active', a.dataset.target === id);
+    });
+  }
+
+  // Scroll spy
+  var observer = new IntersectionObserver(function (entries) {
+    entries.forEach(function (e) {
+      if (e.isIntersecting) setActive(e.target.id);
+    });
+  }, { rootMargin: '-64px 0px -75% 0px', threshold: 0 });
+  sections.forEach(function (s) { observer.observe(s); });
+
+  // Mobile sidebar toggle
+  var btn = document.getElementById('menuBtn');
+  var sidebar = document.getElementById('sidebar');
+  btn.addEventListener('click', function () { sidebar.classList.toggle('open'); });
+  sideLinks.forEach(function (a) {
+    a.addEventListener('click', function () { sidebar.classList.remove('open'); });
+  });
+})();
+</script>
+</body>
+</html>

docs/index.md

@@ -0,0 +1,170 @@
+---
+hide:
+  - navigation
+---
+
+<div class="hero" markdown>
+
+# SkillOpt
+
+### Train Agent Skills Like Neural Networks
+
+*Optimize natural-language skill documents through iterative rollout, reflection, and gated validation — with epochs, learning rates, and validation gates — without touching model weights.*
+
+[Get Started :material-rocket-launch:](guide/installation.md){ .md-button .md-button--primary }
+[View on GitHub :material-github:](https://github.com/microsoft/SkillOpt){ .md-button }
+
+</div>
+
+---
+
+## How It Works
+
+<div class="pipeline-container" markdown>
+<div class="pipeline-wrapper">
+
+<div class="pipeline-stage" id="stage-rollout">
+<div class="stage-icon">🎯</div>
+<div class="stage-label">Rollout</div>
+<div class="stage-desc">Target executes tasks</div>
+</div>
+
+<div class="pipeline-arrow"><div class="flow-line"></div></div>
+
+<div class="pipeline-stage" id="stage-reflect">
+<div class="stage-icon">🔍</div>
+<div class="stage-label">Reflect</div>
+<div class="stage-desc">Optimizer analyzes trajectories</div>
+</div>
+
+<div class="pipeline-arrow"><div class="flow-line"></div></div>
+
+<div class="pipeline-stage" id="stage-aggregate">
+<div class="stage-icon">🔗</div>
+<div class="stage-label">Aggregate</div>
+<div class="stage-desc">Merge edit patches</div>
+</div>
+
+<div class="pipeline-arrow"><div class="flow-line"></div></div>
+
+<div class="pipeline-stage" id="stage-select">
+<div class="stage-icon">✂️</div>
+<div class="stage-label">Select</div>
+<div class="stage-desc">Rank & clip edits</div>
+</div>
+
+<div class="pipeline-arrow"><div class="flow-line"></div></div>
+
+<div class="pipeline-stage" id="stage-update">
+<div class="stage-icon">📝</div>
+<div class="stage-label">Update</div>
+<div class="stage-desc">Apply to skill doc</div>
+</div>
+
+<div class="pipeline-arrow"><div class="flow-line"></div></div>
+
+<div class="pipeline-stage" id="stage-gate">
+<div class="stage-icon">🚦</div>
+<div class="stage-label">Gate</div>
+<div class="stage-desc">Validate & accept</div>
+</div>
+
+</div>
+
+<div class="pipeline-epoch-bar">
+<div class="epoch-mechanism">🔄 Slow Update</div>
+<div class="epoch-mechanism">🧠 Meta Skill</div>
+<div class="epoch-label">Epoch Boundary</div>
+</div>
+
+</div>
+
+---
+
+## Deep Learning Analogy
+
+SkillOpt brings the familiar deep-learning training paradigm to agentic prompt optimization:
+
+| Deep Learning | SkillOpt | 
+|---|---|
+| Model weights | Skill document (Markdown) |
+| Forward pass | Rollout (target executes tasks) |
+| Loss / gradient | Reflect (optimizer produces edit patches) |
+| Gradient clipping | Edit selection (`learning_rate` = max edits) |
+| SGD step | Patch application to skill |
+| Validation set | Gated evaluation on selection split |
+| LR schedule | `lr_scheduler`: cosine, linear, constant |
+| Epochs | Multi-epoch with slow update & meta skill memory |
+
+---
+
+## Supported Benchmarks
+
+| Benchmark | Type | Config |
+|---|---|---|
+| **DocVQA** | Document QA | `configs/docvqa/` |
+| **ALFWorld** | Embodied AI | `configs/alfworld/` |
+| **OfficeQA** | Enterprise QA | `configs/officeqa/` |
+| **SearchQA** | Open-domain QA | `configs/searchqa/` |
+| **LiveMathBench** | Math reasoning | `configs/livemathematicianbench/` |
+| **SWEBench** | Software Engineering | `configs/swebench/` |
+| + 5 more | Various | See [docs](guide/first-experiment.md) |
+
+---
+
+## Quick Example
+
+```bash
+# Install
+pip install -e .
+
+# Configure credentials
+export AZURE_OPENAI_ENDPOINT="https://your-resource.openai.azure.com/"
+export AZURE_OPENAI_API_KEY="your-key"
+
+# Train on SearchQA
+python scripts/train.py --config configs/searchqa/default.yaml
+
+# Evaluate best skill
+python scripts/eval_only.py \
+  --config configs/searchqa/default.yaml \
+  --skill outputs/best_skill.md
+```
+
+---
+
+<div class="grid cards" markdown>
+
+-   :material-book-open-variant:{ .lg .middle } **Getting Started**
+
+    ---
+
+    Install SkillOpt, configure your API keys, and run your first experiment in 5 minutes.
+
+    [:octicons-arrow-right-24: Installation](guide/installation.md)
+
+-   :material-puzzle:{ .lg .middle } **Add a Benchmark**
+
+    ---
+
+    Extend SkillOpt with your own benchmark in ~100 lines of code.
+
+    [:octicons-arrow-right-24: Extension Guide](guide/new-benchmark.md)
+
+-   :material-cog:{ .lg .middle } **Configuration**
+
+    ---
+
+    Full reference for all hyperparameters with deep learning analogies.
+
+    [:octicons-arrow-right-24: Config Reference](reference/config.md)
+
+-   :material-monitor-dashboard:{ .lg .middle } **WebUI**
+
+    ---
+
+    Configure, launch, and monitor training from your browser.
+
+    [:octicons-arrow-right-24: WebUI Guide](guide/first-experiment.md#webui)
+
+</div>

docs/reference/api.md

@@ -0,0 +1,81 @@
+# API Reference
+
+## Core Classes
+
+### `EnvAdapter`
+
+Abstract base class for benchmark environments.
+
+```python
+class EnvAdapter(ABC):
+    async def execute(self, item, skill, model) -> TaskResult
+    def evaluate(self, prediction, ground_truth) -> float
+    def build_prompt(self, item, skill) -> str
+```
+
+### `DataLoader`
+
+Abstract base class for data loading and splitting.
+
+```python
+class DataLoader(ABC):
+    def setup(self, cfg: dict) -> None
+    def get_split_items(self, split: str) -> list[DataItem]
+```
+
+### `ModelBackend`
+
+Abstract base class for LLM backends.
+
+```python
+class ModelBackend(ABC):
+    async def generate(self, messages, **kwargs) -> ModelResponse
+    async def generate_with_tools(self, messages, tools, **kwargs) -> ModelResponse
+```
+
+### `Trainer`
+
+Main training loop orchestrator.
+
+```python
+class Trainer:
+    def __init__(self, cfg: dict)
+    async def train(self) -> TrainResult
+    async def evaluate(self, skill: str, split: str) -> EvalResult
+```
+
+## Data Classes
+
+### `DataItem`
+
+```python
+@dataclass
+class DataItem:
+    id: str
+    input: str
+    ground_truth: str
+    metadata: dict = field(default_factory=dict)
+```
+
+### `TaskResult`
+
+```python
+@dataclass
+class TaskResult:
+    item_id: str
+    prediction: str
+    score: float
+    trajectory: list[dict]
+```
+
+### `ModelResponse`
+
+```python
+@dataclass
+class ModelResponse:
+    content: str
+    usage: dict
+    model: str
+```
+
+For detailed source code, see the [`skillopt/`](https://github.com/microsoft/SkillOpt/tree/main/skillopt) directory.

docs/reference/cli.md

@@ -0,0 +1,71 @@
+# CLI Reference
+
+## Training
+
+```bash
+python scripts/train.py --config <config.yaml> [overrides...]
+```
+
+### Arguments
+
+| Argument | Description |
+|---|---|
+| `--config` | Path to YAML config file (required) |
+| `key=value` | Override any config parameter |
+
+### Examples
+
+```bash
+# Basic training
+python scripts/train.py --config configs/searchqa/default.yaml
+
+# With overrides
+python scripts/train.py \
+  --config configs/searchqa/default.yaml \
+  --cfg-options optimizer.learning_rate=16 optimizer.lr_scheduler=linear
+
+# With custom initial skill
+python scripts/train.py \
+  --config configs/searchqa/default.yaml \
+  --cfg-options env.skill_init=skills/my_seed.md
+```
+
+## Evaluation
+
+```bash
+python scripts/eval_only.py --config <config.yaml> --skill <skill.md>
+```
+
+### Arguments
+
+| Argument | Description |
+|---|---|
+| `--config` | Path to YAML config file (required) |
+| `--skill` | Path to skill document to evaluate (required) |
+| `--split` | Evaluation split: `test` (default), `valid`, `train` |
+
+### Examples
+
+```bash
+# Evaluate best skill on test set
+python scripts/eval_only.py \
+  --config configs/searchqa/default.yaml \
+  --skill outputs/searchqa/run_001/skills/best_skill.md
+
+# Evaluate on validation set
+python scripts/eval_only.py \
+  --config configs/searchqa/default.yaml \
+  --skill outputs/searchqa/run_001/skills/best_skill.md \
+  --split valid
+```
+
+## WebUI
+
+```bash
+python -m skillopt_webui.app [--port PORT] [--share]
+```
+
+| Argument | Default | Description |
+|---|---|---|
+| `--port` | 7860 | Port number |
+| `--share` | false | Create public Gradio link |

docs/reference/config.md

@@ -0,0 +1,72 @@
+# Configuration Reference
+
+Complete reference for all SkillOpt configuration parameters.
+
+## Model
+
+| Parameter | Type | Default | Description |
+|---|---|---|---|
+| `model.backend` | str | `azure_openai` | Backend: `azure_openai` / `openai_chat` / `claude_code_exec` / `qwen` |
+| `model.optimizer` | str | `gpt-5.5` | Optimizer model (for reflection & slow update) |
+| `model.target` | str | `gpt-5.5` | Target model (for rollout execution) |
+| `model.reasoning_effort` | str | `medium` | Reasoning effort level |
+
+## Training (`train`)
+
+| Parameter | Type | Default | DL Analogy | Description |
+|---|---|---|---|---|
+| `train.num_epochs` | int | 4 | Epochs | Number of training epochs |
+| `train.batch_size` | int | 40 | Batch size | Tasks sampled per step |
+| `train.accumulation` | int | 1 | Gradient accumulation | Accumulation rounds per step |
+| `train.seed` | int | 42 | Random seed | Reproducibility seed |
+
+## Gradient / Reflection (`gradient`)
+
+| Parameter | Type | Default | Description |
+|---|---|---|---|
+| `gradient.minibatch_size` | int | 8 | Reflect minibatch size |
+| `gradient.merge_batch_size` | int | 8 | Patch merge batch size |
+| `gradient.analyst_workers` | int | 16 | Parallel reflection workers |
+| `gradient.max_analyst_rounds` | int | 3 | Max rounds of analyst reflection |
+| `gradient.failure_only` | bool | `false` | Only reflect on failures |
+
+## Optimizer (`optimizer`)
+
+| Parameter | Type | Default | DL Analogy | Description |
+|---|---|---|---|---|
+| `optimizer.learning_rate` | int | 4 | Learning rate | Max edit patches per step (edit budget) |
+| `optimizer.min_learning_rate` | int | 2 | Min LR | Min edits for decay schedulers |
+| `optimizer.lr_scheduler` | str | `cosine` | LR schedule | `constant` / `linear` / `cosine` / `autonomous` |
+| `optimizer.skill_update_mode` | str | `patch` | — | `patch` / `rewrite_from_suggestions` / `full_rewrite_minibatch` |
+| `optimizer.use_slow_update` | bool | `true` | Momentum | Epoch-boundary longitudinal comparison & guidance |
+| `optimizer.slow_update_samples` | int | 20 | — | Samples for slow update evaluation |
+| `optimizer.use_meta_skill` | bool | `true` | Meta-learning | Cross-epoch optimizer-side strategy memory |
+| `optimizer.longitudinal_pair_policy` | str | `mixed` | — | `mixed` / `changed` / `unchanged` |
+
+## Evaluation (`evaluation`)
+
+| Parameter | Type | Default | Description |
+|---|---|---|---|
+| `evaluation.use_gate` | bool | `true` | Enable validation gating (accept/reject updates) |
+| `evaluation.eval_test` | bool | `true` | Run test evaluation after training |
+
+## Environment (`env`)
+
+| Parameter | Type | Default | Description |
+|---|---|---|---|
+| `env.name` | str | — | Benchmark name (e.g., `searchqa`, `docvqa`) |
+| `env.data_path` | str | — | Path to dataset |
+| `env.skill_init` | str | — | Path to initial seed skill (optional) |
+| `env.split_mode` | str | `ratio` | `ratio` or `split_dir` |
+| `env.split_ratio` | str | `2:1:7` | Train:val:test ratio |
+| `env.exec_timeout` | int | 120 | Per-task timeout in seconds |
+| `env.out_root` | str | — | Output directory |
+
+## Azure OpenAI Credentials
+
+| Variable | Description |
+|---|---|
+| `AZURE_OPENAI_ENDPOINT` / `model.azure_openai_endpoint` | Azure resource endpoint |
+| `AZURE_OPENAI_API_KEY` / `model.azure_openai_api_key` | Azure API key |
+| `OPENAI_API_KEY` | OpenAI API key (for `openai_chat` backend) |
+| `ANTHROPIC_API_KEY` | Anthropic API key (for `claude_code_exec` backend) |

mkdocs.yml

@@ -0,0 +1,78 @@
+site_name: SkillOpt Documentation
+site_url: https://microsoft.github.io/SkillOpt
+site_description: "SkillOpt: Agentic Skill Optimization via Reflective Training Loops"
+repo_url: https://github.com/microsoft/SkillOpt
+repo_name: microsoft/SkillOpt
+
+theme:
+  name: material
+  palette:
+    - scheme: default
+      primary: indigo
+      accent: deep purple
+      toggle:
+        icon: material/brightness-7
+        name: Switch to dark mode
+    - scheme: slate
+      primary: indigo
+      accent: deep purple
+      toggle:
+        icon: material/brightness-4
+        name: Switch to light mode
+  features:
+    - navigation.instant
+    - navigation.tracking
+    - navigation.sections
+    - navigation.expand
+    - navigation.top
+    - content.code.copy
+    - content.tabs.link
+    - search.suggest
+    - search.highlight
+  icon:
+    repo: fontawesome/brands/github
+  font:
+    text: Inter
+    code: JetBrains Mono
+
+
+
+nav:
+  - Home: index.md
+  - Getting Started:
+    - Installation: guide/installation.md
+    - First Experiment: guide/first-experiment.md
+    - Configuration: guide/configuration.md
+  - Core Concepts:
+    - Training Loop: guide/training-loop.md
+    - Skill Document: guide/skill-document.md
+    - Deep Learning Analogy: guide/dl-analogy.md
+  - Extension Guides:
+    - Add a New Benchmark: guide/new-benchmark.md
+    - Local Environment Smoke Tests: guide/local-env-smoke.md
+    - Add a New Model Backend: guide/new-backend.md
+  - Reference:
+    - Configuration Reference: reference/config.md
+    - CLI Reference: reference/cli.md
+    - API Reference: reference/api.md
+  - Contributing: contributing.md
+
+markdown_extensions:
+  - admonition
+  - pymdownx.details
+  - pymdownx.superfences
+  - pymdownx.tabbed:
+      alternate_style: true
+  - pymdownx.highlight:
+      anchor_linenums: true
+  - pymdownx.inlinehilite
+  - pymdownx.emoji:
+      emoji_index: !!python/name:material.extensions.emoji.twemoji
+      emoji_generator: !!python/name:material.extensions.emoji.to_svg
+  - attr_list
+  - md_in_html
+  - toc:
+      permalink: true
+
+plugins:
+  - search

pyproject.toml

@@ -0,0 +1,75 @@
+[build-system]
+requires = ["setuptools>=68.0", "wheel"]
+build-backend = "setuptools.build_meta"
+
+[project]
+name = "skillopt"
+version = "0.1.0"
+description = "SkillOpt: Agentic Skill Optimization via Reflective Training Loops"
+readme = "README.md"
+license = {text = "MIT"}
+requires-python = ">=3.10"
+authors = [
+    {name = "SkillOpt Team"},
+]
+keywords = ["agent", "prompt-optimization", "skill-learning", "LLM", "agentic"]
+classifiers = [
+    "Development Status :: 3 - Alpha",
+    "Intended Audience :: Science/Research",
+    "License :: OSI Approved :: MIT License",
+    "Programming Language :: Python :: 3",
+    "Programming Language :: Python :: 3.10",
+    "Programming Language :: Python :: 3.11",
+    "Programming Language :: Python :: 3.12",
+    "Topic :: Scientific/Engineering :: Artificial Intelligence",
+]
+dependencies = [
+    "openai>=1.30.0",
+    "pyyaml>=6.0",
+    "numpy>=1.24.0",
+    "openpyxl>=3.1.0",
+    "azure-identity>=1.15.0",
+    "azure-core>=1.30.0",
+    "httpx>=0.27.0",
+]
+
+[project.optional-dependencies]
+# Benchmark-specific dependencies
+alfworld = ["alfworld>=0.4.0", "gymnasium>=0.29.0"]
+# Claude model backend
+claude = ["claude-agent-sdk>=0.1.0"]
+# Qwen local model backend (via vLLM)
+qwen = ["vllm>=0.4.0"]
+# Documentation site
+docs = ["mkdocs-material>=9.5.0", "mkdocstrings[python]>=0.24.0"]
+# WebUI dashboard
+webui = ["gradio>=4.0.0"]
+# Development tools
+dev = ["ruff>=0.4.0", "pytest>=8.0.0"]
+# All optional dependencies (except docs/dev/webui)
+all = [
+    "alfworld>=0.4.0",
+    "gymnasium>=0.29.0",
+    "claude-agent-sdk>=0.1.0",
+]
+
+[project.scripts]
+skillopt-train = "scripts.train:main"
+skillopt-eval = "scripts.eval_only:main"
+
+[project.urls]
+Homepage = "https://github.com/microsoft/SkillOpt"
+Documentation = "https://microsoft.github.io/SkillOpt"
+Repository = "https://github.com/microsoft/SkillOpt"
+Issues = "https://github.com/microsoft/SkillOpt/issues"
+
+[tool.setuptools.packages.find]
+include = ["skillopt*", "scripts*"]
+
+[tool.ruff]
+line-length = 120
+target-version = "py310"
+
+[tool.ruff.lint]
+select = ["E", "F", "I", "W"]
+ignore = ["E501"]

reflact/envs/alfworld/prompts/deep_probe.md

@@ -1,35 +0,0 @@
-You are an expert diagnostic-probe designer for ALFWorld embodied tasks.
-
-You will design one short diagnostic instruction to append to the student's prompt
-for a handful of representative ALFWorld trajectories.
-
-The goal is to expose whether the student has the right intermediate subgoal,
-object/receptacle state, and next-step intention without substantially changing
-the current scaffold.
-
-## Hard Constraints
-1. Do NOT substantially change the student's existing action-selection scaffold.
-2. Do NOT prescribe a brand-new planner or long multi-step policy.
-3. Do NOT ask for exhaustive search over all objects or all admissible actions.
-4. Keep the diagnostic readout brief and place it inside the existing <think>...</think> block.
-5. The student must still output exactly one admissible action inside <action>...</action>.
-6. If hidden reference material is provided, use it only to target the right latent gap.
-7. Never copy hidden reference content into the student-facing probe.
-
-## Good Probe Targets
-- current subgoal
-- target object / target receptacle / target state
-- decisive missing precondition
-- why one candidate action is better than a tempting alternative
-- whether the current step should explore, transform an object, or place it
-
-## Bad Probe Targets
-- a full optimal plan from start to finish
-- exhaustive object inventories
-- a new theorem-like or planner-like protocol
-
-Respond ONLY with a valid JSON object:
-{
-  "reasoning": "<why this probe reveals the latent skill gap>",
-  "probe_instruction": "<the exact instruction text to append to the student prompt>"
-}

reflact/envs/babyvision/__init__.py

@@ -1 +0,0 @@
-"""BabyVision environment package for ReflACT."""

reflact/envs/babyvision/adapter.py

@@ -1,267 +0,0 @@
-"""BabyVision environment adapter for ReflACT."""
-from __future__ import annotations
-
-import json
-import os
-
-from reflact.gradient.deep_probe import generate_deep_probe_instruction
-from reflact.datasets.base import BatchSpec
-from reflact.gradient.reflect import run_minibatch_reflect
-from reflact.envs.base import EnvAdapter
-from reflact.envs.babyvision.dataloader import BabyVisionDataLoader
-from reflact.envs.babyvision.rollout import run_batch
-from reflact.model import get_student_backend
-
-
-class BabyVisionAdapter(EnvAdapter):
-    """BabyVision adapter."""
-
-    def build_reference_text(self, item: dict) -> str:
-        cot = str(item.get("cot") or "").strip()
-        if not cot:
-            return ""
-        return f"## Reference CoT\n{cot}"
-
-    def get_reference_metadata(self, item: dict) -> dict:
-        cot = str(item.get("cot") or "").strip()
-        if not cot:
-            return {"fields": [], "preview": ""}
-        return {
-            "fields": ["cot"],
-            "preview": cot[:400],
-        }
-
-    def __init__(
-        self,
-        split_dir: str = "",
-        data_path: str = "",
-        split_mode: str = "ratio",
-        split_ratio: str = "2:1:7",
-        split_seed: int = 42,
-        split_output_dir: str = "",
-        max_turns: int = 1,
-        workers: int = 32,
-        analyst_workers: int = 16,
-        failure_only: bool = False,
-        minibatch_size: int = 8,
-        edit_budget: int = 4,
-        seed: int = 42,
-        limit: int = 0,
-        image_detail: str = "auto",
-        judge_model: str = "gpt-5.4",
-        judge_max_completion_tokens: int = 256,
-        judge_retries: int = 5,
-        use_deep_reflect: bool = False,
-        deep_reflect_failures: int = 4,
-        deep_reflect_successes: int = 2,
-    ) -> None:
-        self.max_turns = max_turns
-        self.workers = workers
-        self.analyst_workers = analyst_workers
-        self.failure_only = failure_only
-        self.minibatch_size = minibatch_size
-        self.edit_budget = edit_budget
-        self.image_detail = image_detail
-        self.judge_model = judge_model
-        self.judge_max_completion_tokens = judge_max_completion_tokens
-        self.judge_retries = judge_retries
-        self.use_deep_reflect = use_deep_reflect
-        self.deep_reflect_failures = deep_reflect_failures
-        self.deep_reflect_successes = deep_reflect_successes
-        self.dataloader = BabyVisionDataLoader(
-            split_dir=split_dir,
-            data_path=data_path,
-            split_mode=split_mode,
-            split_ratio=split_ratio,
-            split_seed=split_seed,
-            split_output_dir=split_output_dir,
-            seed=seed,
-            limit=limit,
-        )
-
-    def setup(self, cfg: dict) -> None:
-        super().setup(cfg)
-        self.dataloader.setup(cfg)
-
-    def get_dataloader(self):
-        return self.dataloader
-
-    def build_env_from_batch(self, batch: BatchSpec, **kwargs):
-        return list(batch.payload or [])
-
-    def build_train_env(self, batch_size: int, seed: int, **kwargs):
-        batch = self.dataloader.build_train_batch(batch_size=batch_size, seed=seed, **kwargs)
-        return self.build_env_from_batch(batch, **kwargs)
-
-    def build_eval_env(self, env_num: int, split: str, seed: int, **kwargs):
-        batch = self.dataloader.build_eval_batch(env_num=env_num, split=split, seed=seed, **kwargs)
-        return self.build_env_from_batch(batch, **kwargs)
-
-    def rollout(
-        self,
-        env_manager,
-        skill_content: str,
-        out_dir: str,
-        **kwargs,
-    ) -> list[dict]:
-        items: list[dict] = env_manager
-        return run_batch(
-            items=items,
-            out_root=out_dir,
-            skill_content=skill_content,
-            max_turns=self.max_turns,
-            workers=self.workers,
-            image_detail=self.image_detail,
-            judge_model=self.judge_model,
-            judge_max_completion_tokens=self.judge_max_completion_tokens,
-            judge_retries=self.judge_retries,
-            diagnostic_mode=kwargs.get("diagnostic_mode", False),
-            diagnostic_instruction=kwargs.get("diagnostic_instruction", ""),
-            diagnostic_trace_context_by_id=kwargs.get("diagnostic_trace_context_by_id"),
-        )
-
-    def reflect(
-        self,
-        results: list[dict],
-        skill_content: str,
-        out_dir: str,
-        **kwargs,
-    ) -> list[dict | None]:
-        prediction_dir = kwargs.get("prediction_dir", os.path.join(out_dir, "predictions"))
-        patches_dir = kwargs.get("patches_dir", os.path.join(out_dir, "patches"))
-        random_seed = kwargs.get("random_seed")
-        step_buffer_context = kwargs.get("step_buffer_context", "")
-        meta_skill_context = kwargs.get("meta_skill_context", "")
-
-        return run_minibatch_reflect(
-            results=results,
-            skill_content=skill_content,
-            prediction_dir=prediction_dir,
-            patches_dir=patches_dir,
-            workers=self.analyst_workers,
-            failure_only=self.failure_only,
-            minibatch_size=self.minibatch_size,
-            edit_budget=self.edit_budget,
-            random_seed=random_seed,
-            error_system=self.get_error_minibatch_prompt(),
-            success_system=self.get_success_minibatch_prompt(),
-            step_buffer_context=step_buffer_context,
-            meta_skill_context=meta_skill_context,
-            update_mode=getattr(self, "_cfg", {}).get("skill_update_mode", "patch"),
-        )
-
-    def deep_reflect(
-        self,
-        results: list[dict],
-        skill_content: str,
-        out_dir: str,
-        **kwargs,
-    ) -> list[dict | None]:
-        if not self.use_deep_reflect:
-            return []
-
-        env_manager = kwargs.get("env_manager")
-        prediction_dir = kwargs.get("prediction_dir", os.path.join(out_dir, "predictions"))
-        random_seed = kwargs.get("random_seed")
-        step_buffer_context = kwargs.get("step_buffer_context", "")
-        meta_skill_context = kwargs.get("meta_skill_context", "")
-        codex_backend = get_student_backend() == "codex_exec"
-        selected_items = self.select_representative_items(
-            results,
-            env_manager if isinstance(env_manager, list) else None,
-            n_failures=self.deep_reflect_failures,
-            n_successes=self.deep_reflect_successes,
-            seed=random_seed,
-        )
-        if not selected_items:
-            return []
-        selected_ids = {str(item["id"]) for item in selected_items}
-        selected_results = [row for row in results if str(row.get("id")) in selected_ids]
-        selected_examples = self.attach_reference_context(selected_results, selected_items)
-        if codex_backend:
-            selected_examples = self.attach_codex_probe_context(selected_examples, prediction_dir)
-        selected_metadata = []
-        cot_count = 0
-        for item in selected_items:
-            meta = self.get_reference_metadata(item)
-            if meta["fields"]:
-                cot_count += 1
-            selected_metadata.append({
-                "id": str(item["id"]),
-                "task_type": str(item.get("subtype") or item.get("task_type") or "babyvision"),
-                "reference_fields": meta["fields"],
-                "reference_preview": meta["preview"],
-            })
-
-        deep_dir = os.path.join(out_dir, "deep_reflect")
-        rollout_dir = os.path.join(deep_dir, "rollout")
-        patches_dir = os.path.join(deep_dir, "patches")
-        os.makedirs(deep_dir, exist_ok=True)
-        print(
-            f"    [2b/6 DEEP REFLECT setup] selected={len(selected_items)} "
-            f"reference_fields=cot({cot_count}/{len(selected_items)})"
-        )
-        probe = generate_deep_probe_instruction(
-            skill_content=skill_content,
-            items=selected_examples,
-            prediction_dir=prediction_dir,
-            system_prompt=self.get_codex_deep_probe_prompt() if codex_backend else self.get_deep_probe_prompt(),
-            step_buffer_context=step_buffer_context,
-            meta_skill_context=meta_skill_context,
-        )
-        if not probe:
-            return []
-        diagnostic_trace_context_by_id = None
-        if codex_backend:
-            selected_items, diagnostic_trace_context_by_id, probe = self.resolve_codex_probe_target(
-                selected_items=selected_items,
-                selected_examples=selected_examples,
-                prediction_dir=prediction_dir,
-                probe=probe,
-            )
-        probe_record = {
-            **probe,
-            "reference_summary": {
-                "selected_count": len(selected_items),
-                "field_counts": {
-                    "cot": cot_count,
-                },
-            },
-            "selected_examples": selected_metadata,
-        }
-        with open(os.path.join(deep_dir, "probe.json"), "w", encoding="utf-8") as f:
-            json.dump(probe_record, f, ensure_ascii=False, indent=2)
-        deep_results = run_batch(
-            items=selected_items,
-            out_root=rollout_dir,
-            skill_content=skill_content,
-            max_turns=self.max_turns,
-            workers=min(self.workers, max(len(selected_items), 1)),
-            image_detail=self.image_detail,
-            judge_model=self.judge_model,
-            judge_max_completion_tokens=self.judge_max_completion_tokens,
-            judge_retries=self.judge_retries,
-            diagnostic_mode=True,
-            diagnostic_instruction=probe["probe_instruction"],
-            diagnostic_trace_context_by_id=diagnostic_trace_context_by_id,
-        )
-        deep_results = self.attach_reference_context(deep_results, selected_items)
-        return run_minibatch_reflect(
-            results=deep_results,
-            skill_content=skill_content,
-            prediction_dir=os.path.join(rollout_dir, "predictions"),
-            patches_dir=patches_dir,
-            workers=self.analyst_workers,
-            failure_only=self.failure_only,
-            minibatch_size=self.minibatch_size,
-            edit_budget=self.edit_budget,
-            random_seed=random_seed,
-            error_system=self.get_error_minibatch_prompt(),
-            success_system=self.get_success_minibatch_prompt(),
-            step_buffer_context=step_buffer_context,
-            meta_skill_context=meta_skill_context,
-            update_mode=getattr(self, "_cfg", {}).get("skill_update_mode", "patch"),
-        )
-
-    def get_task_types(self) -> list[str]:
-        return self.dataloader.get_task_types()

reflact/envs/babyvision/dataloader.py

@@ -1,214 +0,0 @@
-"""BabyVision task dataloader."""
-from __future__ import annotations
-
-import json
-import os
-from typing import Any
-
-from reflact.datasets.base import SplitDataLoader
-
-
-# ── Raw data loading utilities (for preprocessing / standalone eval) ─────
-
-_CHOICE_LABELS = ["A", "B", "C", "D", "E", "F", "G"]
-
-
-def _iter_jsonl(path: str) -> list[dict]:
-    items: list[dict] = []
-    with open(path, encoding="utf-8") as f:
-        for line in f:
-            line = line.strip()
-            if not line:
-                continue
-            items.append(json.loads(line))
-    return items
-
-
-def _normalize_ans_type(raw: Any, options: list[dict], choice_answer: Any) -> str:
-    text = str(raw or "").strip().lower()
-    if text in {"choice", "multiple_choice", "mcq", "option"}:
-        return "choice"
-    if text in {"blank", "open", "open_ended", "fill_blank", "short_answer"}:
-        return "blank"
-    if options or choice_answer not in (None, "", []):
-        return "choice"
-    return "blank"
-
-
-def _coerce_options(raw: Any) -> list[dict]:
-    options: list[dict] = []
-    if isinstance(raw, list):
-        for idx, item in enumerate(raw):
-            if isinstance(item, dict):
-                text = str(item.get("text") or item.get("content") or item.get("option") or "").strip()
-                label = str(item.get("label") or _CHOICE_LABELS[idx]).strip()
-            else:
-                text = str(item).strip()
-                label = _CHOICE_LABELS[idx]
-            if text:
-                options.append({"label": label, "text": text})
-    elif isinstance(raw, dict):
-        for idx, (key, value) in enumerate(raw.items()):
-            text = str(value).strip()
-            if text:
-                options.append({"label": str(key).strip() or _CHOICE_LABELS[idx], "text": text})
-    return options
-
-
-def _normalize_choice_answer(choice_answer: Any, options: list[dict]) -> dict[str, str]:
-    if not options:
-        return {"label": "", "text": ""}
-
-    if isinstance(choice_answer, dict):
-        label = str(choice_answer.get("label") or "").strip().upper()
-        text = str(choice_answer.get("text") or "").strip()
-        for option in options:
-            if label and option["label"].strip().upper() == label:
-                return {"label": option["label"], "text": option["text"]}
-            if text and option["text"] == text:
-                return {"label": option["label"], "text": option["text"]}
-
-    if isinstance(choice_answer, int):
-        idx = choice_answer
-        if 0 <= idx < len(options):
-            return dict(options[idx])
-        if 1 <= idx <= len(options):
-            return dict(options[idx - 1])
-
-    text = str(choice_answer or "").strip()
-    label = text.upper().rstrip(".):")
-    for option in options:
-        if option["label"].strip().upper() == label:
-            return dict(option)
-        if option["text"] == text:
-            return dict(option)
-
-    return {"label": "", "text": ""}
-
-
-def _coerce_blank_answers(raw: Any) -> list[str]:
-    if isinstance(raw, list):
-        return [str(item).strip() for item in raw if str(item).strip()]
-    if raw is None:
-        return []
-    text = str(raw).strip()
-    return [text] if text else []
-
-
-def load_items(data_path: str) -> list[dict]:
-    """Load and normalise BabyVision items from a directory or JSONL file."""
-    if not data_path:
-        raise ValueError("BabyVision requires data_path pointing to a local dataset directory or meta_data.jsonl.")
-
-    if os.path.isdir(data_path):
-        meta_path = os.path.join(data_path, "meta_data.jsonl")
-        image_root = os.path.join(data_path, "images")
-    else:
-        meta_path = data_path
-        image_root = os.path.join(os.path.dirname(data_path), "images")
-
-    if not os.path.exists(meta_path):
-        raise ValueError(
-            "BabyVision expected a meta_data.jsonl file. "
-            f"Could not find: {meta_path}"
-        )
-
-    raw_items = _iter_jsonl(meta_path)
-    items: list[dict] = []
-    for idx, raw in enumerate(raw_items):
-        options = _coerce_options(raw.get("options") or raw.get("choices") or raw.get("choiceOptions"))
-        ans_type = _normalize_ans_type(raw.get("ansType"), options, raw.get("choiceAns"))
-        correct_choice = _normalize_choice_answer(raw.get("choiceAns"), options)
-        blank_answers = _coerce_blank_answers(raw.get("blankAns"))
-
-        image_name = str(
-            raw.get("image")
-            or raw.get("image_path")
-            or raw.get("image_file")
-            or raw.get("img")
-            or ""
-        ).strip()
-        if not image_name:
-            continue
-        image_path = image_name if os.path.isabs(image_name) else os.path.join(image_root, image_name)
-        if not os.path.exists(image_path):
-            alt = os.path.join(os.path.dirname(meta_path), image_name)
-            if os.path.exists(alt):
-                image_path = alt
-            else:
-                continue
-
-        task_id = str(raw.get("taskId") or raw.get("id") or idx + 1)
-        task_type = str(raw.get("type") or raw.get("taskType") or "unknown").strip() or "unknown"
-        subtype = str(raw.get("subtype") or raw.get("subType") or task_type).strip() or task_type
-        question = str(raw.get("question") or raw.get("query") or "").strip()
-        if not question:
-            continue
-
-        if ans_type == "choice" and not correct_choice["label"]:
-            continue
-        if ans_type != "choice" and not blank_answers:
-            continue
-
-        items.append({
-            "id": task_id,
-            "task_type": task_type,
-            "subtype": subtype,
-            "question": question,
-            "image_path": os.path.abspath(image_path),
-            "ans_type": ans_type,
-            "choices": options,
-            "correct_choice": correct_choice,
-            "blank_answers": blank_answers,
-            "cot": str(raw.get("coT") or raw.get("cot") or "").strip(),
-            "source_path": os.path.abspath(meta_path),
-        })
-
-    if not items:
-        raise ValueError(f"No valid BabyVision items loaded from {data_path}")
-    return items
-
-
-# ── Dataloader ───────────────────────────────────────────────────────────
-
-class BabyVisionDataLoader(SplitDataLoader):
-    """BabyVision dataloader."""
-
-    def __init__(
-        self,
-        split_dir: str = "",
-        data_path: str = "",
-        split_mode: str = "ratio",
-        split_ratio: str = "2:1:7",
-        split_seed: int = 42,
-        split_output_dir: str = "",
-        seed: int = 42,
-        limit: int = 0,
-        **kwargs,
-    ) -> None:
-        super().__init__(
-            split_dir=split_dir,
-            data_path=data_path,
-            split_mode=split_mode,
-            split_ratio=split_ratio,
-            split_seed=split_seed,
-            split_output_dir=split_output_dir,
-            seed=seed,
-            limit=limit,
-        )
-        self._task_types: list[str] = []
-
-    def load_raw_items(self, data_path: str) -> list[dict]:
-        return load_items(data_path)
-
-    def setup(self, cfg: dict) -> None:
-        super().setup(cfg)
-        all_items = self.train_items + self.val_items + self.test_items
-        task_types = {
-            item.get("subtype") or item.get("task_type") or "unknown"
-            for item in all_items
-        }
-        self._task_types = sorted(task_types)
-
-    def get_task_types(self) -> list[str]:
-        return list(self._task_types)

reflact/envs/babyvision/evaluator.py

@@ -1,160 +0,0 @@
-"""BabyVision evaluation helpers using the official-style LLM judge."""
-from __future__ import annotations
-
-import re
-import string
-
-import regex
-
-from reflact.model import chat_with_deployment
-from reflact.prompts import load_prompt
-
-_EVAL_MODE = "babyvision_judge_v2_official_style"
-
-def normalize_text(text: str) -> str:
-    text = str(text).strip().lower()
-    text = "".join(ch for ch in text if ch not in string.punctuation)
-    return " ".join(text.split())
-
-
-def extract_boxed_answer(text: str | None) -> str | None:
-    """Extract the final answer using the official BabyVision rule."""
-    if text is None:
-        return None
-
-    pattern = r'\\boxed\{((?:[^{}]|{(?:[^{}]|{.*})*})*)\}'
-    matches = regex.findall(pattern, text)
-    if matches:
-        return matches[-1]
-
-    pattern_alt = r'<\|begin_of_box\|>(.*?)<\|end_of_box\|>'
-    matches_alt = regex.findall(pattern_alt, text)
-    if matches_alt:
-        return matches_alt[-1].strip()
-
-    return None
-
-
-def _token_f1(prediction: str, gold: str) -> float:
-    pred_tokens = normalize_text(prediction).split()
-    gold_tokens = normalize_text(gold).split()
-    if not pred_tokens and not gold_tokens:
-        return 1.0
-    if not pred_tokens or not gold_tokens:
-        return 0.0
-    pred_set = {}
-    gold_set = {}
-    for tok in pred_tokens:
-        pred_set[tok] = pred_set.get(tok, 0) + 1
-    for tok in gold_tokens:
-        gold_set[tok] = gold_set.get(tok, 0) + 1
-    common = 0
-    for tok, count in pred_set.items():
-        common += min(count, gold_set.get(tok, 0))
-    if common == 0:
-        return 0.0
-    precision = common / len(pred_tokens)
-    recall = common / len(gold_tokens)
-    return 2 * precision * recall / (precision + recall)
-
-
-def _format_choices(choices: list[dict]) -> str:
-    return "\n".join(f"{choice['label']}. {choice['text']}" for choice in choices)
-
-
-def _judge_answer(
-    *,
-    item: dict,
-    prediction_text: str,
-    extracted_answer: str,
-    judge_model: str,
-    max_completion_tokens: int,
-    retries: int,
-) -> dict:
-    if item["ans_type"] == "choice":
-        ground_truth = str(item["correct_choice"]["label"])
-    else:
-        if len(item["blank_answers"]) == 1:
-            ground_truth = item["blank_answers"][0]
-        else:
-            ground_truth = " | ".join(item["blank_answers"])
-
-    question = str(item["question"])
-    if item["ans_type"] == "choice" and item.get("choices"):
-        question = f"{question}\nChoices:\n{_format_choices(item['choices'])}"
-
-    raw, _ = chat_with_deployment(
-        deployment=judge_model,
-        system="You are a careful and strict evaluator.",
-        user=load_prompt("judge", env="babyvision").format(
-            question=question,
-            groundtruth=ground_truth,
-            modeloutput=extracted_answer,
-        ),
-        max_completion_tokens=max_completion_tokens,
-        retries=retries,
-        stage="babyvision_judge",
-    )
-    judge_response_clean = str(raw).strip().lower()
-    if "true" in judge_response_clean:
-        correct = True
-    elif "false" in judge_response_clean:
-        correct = False
-    else:
-        correct = False
-    return {
-        "raw": raw,
-        "correct": correct,
-        "reason": judge_response_clean,
-        "matched_gold": ground_truth if correct else "",
-    }
-
-
-def evaluate_item(
-    *,
-    item: dict,
-    prediction_text: str,
-    judge_model: str,
-    max_completion_tokens: int = 256,
-    retries: int = 5,
-) -> dict:
-    answer = extract_boxed_answer(prediction_text)
-    judge = _judge_answer(
-        item=item,
-        prediction_text=prediction_text,
-        extracted_answer=answer,
-        judge_model=judge_model,
-        max_completion_tokens=max_completion_tokens,
-        retries=retries,
-    )
-    hard = 1.0 if judge["correct"] else 0.0
-
-    result = {
-        "evaluation_mode": _EVAL_MODE,
-        "predicted_answer": answer,
-        "em": hard,
-        "f1": hard,
-        "sub_em": hard,
-        "judge_model": judge_model,
-        "judge_raw": judge["raw"],
-        "judge_reason": judge["reason"],
-        "matched_gold": judge["matched_gold"],
-    }
-
-    if item["ans_type"] == "choice":
-        result["predicted_label"] = str(answer or "").strip().upper().rstrip(".):")
-        result["predicted_text"] = ""
-        result["correct_label"] = str(item["correct_choice"].get("label") or "")
-        result["correct_text"] = str(item["correct_choice"].get("text") or "")
-    else:
-        result["gold_answers"] = list(item["blank_answers"])
-        best_f1 = 0.0
-        for gold in item["blank_answers"]:
-            best_f1 = max(best_f1, _token_f1(str(answer or ""), gold))
-        result["string_f1"] = best_f1
-
-    return result
-
-
-def evaluation_mode() -> str:
-    return _EVAL_MODE

reflact/envs/babyvision/prompts/analyst_error.md

@@ -1,36 +0,0 @@
-You are an expert failure-analysis agent for child-level visual reasoning tasks.
-
-You will be given MULTIPLE failed BabyVision trajectories from a minibatch and the current skill document.
-Each trajectory includes the text prompt, the model answer, and the evaluation result.
-You do not have direct access to raw pixel content during reflection, so focus on general reasoning,
-option-selection, and visual-question-answering behaviors that can be improved through prompting.
-
-## Failure Type Categories
-- **visual_detail_miss**: the agent likely overlooked a salient visual attribute, relation, count, or object state
-- **option_mismatch**: the agent selected the wrong option despite relevant evidence likely being present
-- **instruction_slip**: the agent ignored output format or answered too vaguely
-- **answer_granularity**: the agent gave an answer that was too broad, too narrow, or mismatched the expected specificity
-- **other**: none of the above
-
-## Rules
-1. Focus on patterns recurring across the minibatch.
-2. Prefer reusable behaviors for inspecting images and grounding answers in visible evidence.
-3. Do not memorize dataset-specific answers.
-4. Only patch gaps not already covered by the current skill.
-
-Respond ONLY with a valid JSON object:
-{
-  "batch_size": <number>,
-  "failure_summary": [
-    {"failure_type": "<type>", "count": <int>, "description": "<one-line>"}
-  ],
-  "patch": {
-    "reasoning": "<why these edits address the common failures>",
-    "edits": [
-      {"op": "append",       "content": "<markdown>"},
-      {"op": "insert_after", "target": "<heading/text>", "content": "<markdown>"},
-      {"op": "replace",      "target": "<old text>",     "content": "<new text>"},
-      {"op": "delete",       "target": "<exact text to remove>"}
-    ]
-  }
-}

reflact/envs/babyvision/prompts/analyst_success.md

@@ -1,25 +0,0 @@
-You are an expert success-pattern analyst for child-level visual reasoning tasks.
-
-You will be given MULTIPLE successful BabyVision trajectories from a minibatch and the current skill document.
-Identify generalizable behavior patterns that help the agent inspect the image carefully and answer at the right level of specificity.
-
-## Rules
-- Focus on broadly useful visual QA behaviors.
-- Prefer patterns about systematic image inspection, comparing options, and concise grounded answers.
-- Do not add dataset-specific facts.
-- "edits" may be empty if the skill already captures the useful patterns.
-
-Respond ONLY with a valid JSON object:
-{
-  "batch_size": <number>,
-  "success_patterns": ["<pattern 1>", "<pattern 2>"],
-  "patch": {
-    "reasoning": "<why these patterns matter>",
-    "edits": [
-      {"op": "append",       "content": "<markdown>"},
-      {"op": "insert_after", "target": "<heading/text>", "content": "<markdown>"},
-      {"op": "replace",      "target": "<old text>",     "content": "<new text>"},
-      {"op": "delete",       "target": "<exact text to remove>"}
-    ]
-  }
-}

reflact/envs/babyvision/prompts/deep_probe.md

@@ -1,25 +0,0 @@
-You are an expert diagnostic-probe designer for BabyVision-style visual reasoning tasks.
-
-You will be shown representative trajectories, the current student skill, and the student's original prompt context.
-Design one SMALL diagnostic instruction that exposes the student's intermediate visual judgment without materially changing the original scaffold.
-
-## Hard Constraints
-1. Do NOT substantially change the original scaffold.
-2. Do NOT prescribe a new step-by-step solving method.
-3. You MAY ask for a short structured list of a few intermediate conclusions, candidate cues, or counted units, as long as it stays close to the original scaffold.
-4. Do NOT ask for exhaustive listing of all cells, all objects, or a full chain-of-thought.
-5. Ask only for a short readout that reveals the student's current latent state.
-6. Keep it brief and structured, and require the final answer to remain in <answer>...</answer>.
-
-## Good Probe Targets
-- top answer and runner-up
-- decisive visual cue
-- suspicious region or compared objects
-- counting unit or formatting interpretation
-- 2-4 short intermediate conclusions that directly support the final answer
-
-Respond ONLY with a valid JSON object:
-{
-  "reasoning": "<why this probe is informative>",
-  "probe_instruction": "<the exact instruction text to append to the student prompt>"
-}

reflact/envs/babyvision/prompts/judge.md

@@ -1,35 +0,0 @@
-You are a careful and strict evaluator. You will be given:
-
-1. **Question**
-2. **Ground Truth Answer** (correct answer)
-3. **Model Output** (answer from another model)
-
-**Your goal:** Determine if the Model Output **accurately matches** the Ground Truth Answer in meaning.
-
-* Matching means: the facts, entities, and key details are equivalent, even if phrasing differs.
-* Not matching means: the Model Output is wrong, incomplete, contains extra incorrect facts, or changes the meaning.
-
-**Process (internal reasoning):**
-
-1. Read and understand the Question, Ground Truth Answer, and Model Output.
-2. Ignore small wording differences, formatting, or synonyms.
-3. If all factual content matches, conclude `1`. Otherwise, conclude `0`.
-
-**Important:**
-
-* Think through your decision step-by-step **internally** before responding.
-* In your final output, return **only** True or False, with no extra text or explanation.
-
-**Output format:**
-
-True
-
-or
-
-False
-
-**Input:**
-
-Question: {question},
-Ground Truth Answer: {groundtruth},
-Model Output: {modeloutput}

reflact/envs/babyvision/prompts/rollout_system.md

@@ -1,13 +0,0 @@
-You are an expert visual reasoning agent solving child-level image understanding tasks.
-
-{skill_section}## Task Format
-You will receive one image and one question about it.
-Inspect the image carefully before answering. Ground the answer in visible evidence.
-
-## Answer Format
-Think step by step, then provide your final answer in \boxed{{Answer}} format.
-- For multiple-choice questions, output only the single choice label, such as \boxed{{A}}.
-- For open questions, output only a short final answer inside \boxed{{...}}.
-
-Example:
-\boxed{{B}}

reflact/envs/babyvision/reflect.py

@@ -1,4 +0,0 @@
-"""BabyVision Reflect stage.
-
-Prompts are now loaded from .md files by the base adapter.
-"""

reflact/envs/babyvision/rollout.py

@@ -1,467 +0,0 @@
-"""BabyVision rollout — multimodal visual QA with image input."""
-from __future__ import annotations
-
-import base64
-import json
-import mimetypes
-import os
-from concurrent.futures import ThreadPoolExecutor, as_completed
-
-from reflact.envs.babyvision.evaluator import evaluate_item, evaluation_mode, extract_boxed_answer
-from reflact.model import chat_student_messages, get_student_backend, is_student_exec_backend
-from reflact.model.codex_harness import prepare_workspace, render_skill_md, run_student_exec
-from reflact.prompts import load_prompt
-
-def _build_system(skill_content: str) -> str:
-    if skill_content.strip():
-        skill_section = f"## Skill\n{skill_content.strip()}\n\n"
-    else:
-        skill_section = ""
-    return load_prompt("rollout_system", env="babyvision").format(skill_section=skill_section)
-
-
-def _format_choices(choices: list[dict]) -> str:
-    return "\n".join(f"{choice['label']}. {choice['text']}" for choice in choices)
-
-
-def _build_user_text(
-    item: dict,
-    *,
-    diagnostic_mode: bool = False,
-    diagnostic_instruction: str = "",
-    diagnostic_trace_context: str = "",
-) -> str:
-    parts = []
-    if diagnostic_trace_context.strip():
-        parts.append(
-            "## Previous Codex Trace Snapshot\n"
-            "This is a partial transcript from an earlier attempt. Use it as your current reasoning context.\n\n"
-            f"{diagnostic_trace_context.strip()}"
-        )
-    parts.append(f"## Question\n{item['question']}")
-    if item["ans_type"] == "choice":
-        parts.append(f"## Choices\n{_format_choices(item['choices'])}")
-        parts.append("Answer using the single correct option label in \\boxed{...}.")
-    else:
-        parts.append("Answer with a short phrase in \\boxed{...}.")
-    if diagnostic_mode and diagnostic_instruction.strip():
-        parts.append(f"## Training Readout\n{diagnostic_instruction.strip()}")
-    return "\n\n".join(parts)
-
-
-def _image_to_data_uri(path: str) -> str:
-    mime = mimetypes.guess_type(path)[0] or "image/png"
-    with open(path, "rb") as f:
-        encoded = base64.b64encode(f.read()).decode("ascii")
-    return f"data:{mime};base64,{encoded}"
-
-
-def _build_messages(
-    item: dict,
-    skill_content: str,
-    image_detail: str,
-    *,
-    diagnostic_mode: bool = False,
-    diagnostic_instruction: str = "",
-    diagnostic_trace_context: str = "",
-) -> tuple[list[dict], str, str]:
-    system = _build_system(skill_content)
-    user_text = _build_user_text(
-        item,
-        diagnostic_mode=diagnostic_mode,
-        diagnostic_instruction=diagnostic_instruction,
-        diagnostic_trace_context=diagnostic_trace_context,
-    )
-    image_url = {
-        "url": _image_to_data_uri(item["image_path"]),
-    }
-    if image_detail and image_detail != "auto":
-        image_url["detail"] = image_detail
-    messages = [
-        {"role": "system", "content": system},
-        {
-            "role": "user",
-            "content": [
-                {"type": "text", "text": user_text},
-                {"type": "image_url", "image_url": image_url},
-            ],
-        },
-    ]
-    return messages, system, user_text
-
-
-def _build_codex_skill(skill_content: str) -> str:
-    return render_skill_md(
-        skill_content,
-        description="Dynamic ReflACT skill for solving the current BabyVision visual reasoning question.",
-        preamble=(
-            "Use this skill when answering the current visual reasoning question.\n"
-            "Inspect the attached image carefully and return the final answer in \\boxed{...}."
-        ),
-    )
-
-
-def _run_codex_once(
-    *,
-    pred_dir: str,
-    item: dict,
-    skill_content: str,
-    model: str,
-    timeout: int,
-    image_detail: str,
-    diagnostic_mode: bool = False,
-    diagnostic_instruction: str = "",
-    diagnostic_trace_context: str = "",
-    previous_response: str = "",
-) -> tuple[str, str, str, str]:
-    user_text = _build_user_text(
-        item,
-        diagnostic_mode=diagnostic_mode,
-        diagnostic_instruction=diagnostic_instruction,
-        diagnostic_trace_context=diagnostic_trace_context,
-    )
-    task_parts = [user_text]
-    if previous_response:
-        task_parts.append(
-            "## Previous Attempt\n"
-            f"{previous_response}\n\n"
-            "Review the same image and question carefully. If needed, correct the answer."
-        )
-    task_text = "\n\n".join(task_parts)
-    skill_md = _build_codex_skill(skill_content)
-    work_dir = os.path.join(pred_dir, "codex_exec")
-    prepare_workspace(
-        work_dir=work_dir,
-        skill_md=skill_md,
-        task_text=task_text,
-        images=[item["image_path"]],
-    )
-    prompt = (
-        "Use the `reflact-student` skill available in this workspace.\n"
-        "Read `task.md`, inspect the attached image, and answer the question.\n"
-        "Return the final answer in \\boxed{...}."
-    )
-    final_message, raw = run_student_exec(
-        work_dir=work_dir,
-        prompt=prompt,
-        model=model,
-        timeout=timeout,
-        images=[item["image_path"]],
-    )
-    return final_message or raw, raw, skill_md, task_text
-
-
-def process_one(
-    item: dict,
-    out_root: str,
-    skill_content: str,
-    *,
-    max_turns: int = 1,
-    image_detail: str = "auto",
-    judge_model: str = "gpt-5.4",
-    judge_max_completion_tokens: int = 256,
-    judge_retries: int = 5,
-    diagnostic_mode: bool = False,
-    diagnostic_instruction: str = "",
-    diagnostic_trace_context: str = "",
-) -> dict:
-    item_id = str(item["id"])
-    result = {
-        "id": item_id,
-        "question": item["question"],
-        "task_type": item.get("subtype") or item.get("task_type") or "babyvision",
-        "task_description": item["question"],
-        "hard": 0,
-        "soft": 0.0,
-        "predicted_answer": "",
-        "predicted_label": "",
-        "predicted_text": "",
-        "response": "",
-        "fail_reason": "",
-        "agent_ok": False,
-        "n_turns": 0,
-        "image_path": item["image_path"],
-        "ans_type": item["ans_type"],
-        "evaluation_mode": evaluation_mode(),
-        "judge_model": judge_model,
-    }
-    if item["ans_type"] == "choice":
-        result["correct_label"] = item["correct_choice"]["label"]
-        result["correct_text"] = item["correct_choice"]["text"]
-    else:
-        result["gold_answers"] = item["blank_answers"]
-
-    try:
-        pred_dir = os.path.join(out_root, "predictions", item_id)
-        os.makedirs(pred_dir, exist_ok=True)
-
-        if is_student_exec_backend():
-            from reflact.model import azure_openai as _llm
-
-            response = ""
-            conversation: list[dict] = [
-                {"role": "user", "content": f"{item['question']}\n\n[image] {os.path.basename(item['image_path'])}"}
-            ]
-            system_prompt = ""
-            user_text = ""
-            for turn in range(max_turns):
-                response, raw, system_prompt, user_text = _run_codex_once(
-                    pred_dir=pred_dir,
-                    item=item,
-                    skill_content=skill_content,
-                    model=_llm.STUDENT_DEPLOYMENT,
-                    timeout=120,
-                    image_detail=image_detail,
-                    diagnostic_mode=diagnostic_mode if turn == 0 else False,
-                    diagnostic_instruction=diagnostic_instruction if turn == 0 else "",
-                    diagnostic_trace_context=diagnostic_trace_context if turn == 0 else "",
-                    previous_response=response if turn > 0 else "",
-                )
-                conversation.append({"type": "message", "turn": turn + 1, "content": response})
-                if extract_boxed_answer(response) is not None:
-                    break
-
-            result["response"] = response
-            result["agent_ok"] = True
-            result["n_turns"] = len(conversation) - 1
-            with open(os.path.join(pred_dir, "student_system_prompt.txt"), "w", encoding="utf-8") as f:
-                f.write(system_prompt)
-            with open(os.path.join(pred_dir, "student_user_prompt.txt"), "w", encoding="utf-8") as f:
-                f.write(user_text)
-
-            eval_result = evaluate_item(
-                item=item,
-                prediction_text=response,
-                judge_model=judge_model,
-                max_completion_tokens=judge_max_completion_tokens,
-                retries=judge_retries,
-            )
-            result["evaluation_mode"] = eval_result["evaluation_mode"]
-            result["judge_raw"] = eval_result["judge_raw"]
-            result["judge_reason"] = eval_result["judge_reason"]
-            result["matched_gold"] = eval_result["matched_gold"]
-            if item["ans_type"] == "choice":
-                result["predicted_label"] = eval_result["predicted_label"]
-                result["predicted_text"] = eval_result["predicted_text"]
-                result["predicted_answer"] = eval_result["predicted_answer"]
-                result["hard"] = int(eval_result["em"])
-                result["soft"] = eval_result["f1"]
-                if not result["hard"]:
-                    result["fail_reason"] = (
-                        f"judge=0: predicted '{eval_result['predicted_label'] or eval_result['predicted_answer']}' "
-                        f"but expected '{eval_result['correct_label']}' ({eval_result['judge_reason']})"
-                    )
-                eval_detail = (
-                    f"[EVALUATION RESULT]\n"
-                    f"Question: {item['question']}\n"
-                    f"Predicted label: {eval_result['predicted_label']!r}\n"
-                    f"Predicted text: {eval_result['predicted_text']!r}\n"
-                    f"Correct label: {eval_result['correct_label']!r}\n"
-                    f"Correct text: {eval_result['correct_text']!r}\n"
-                    f"Judge correct: {eval_result['em']}\n"
-                    f"Judge reason: {eval_result['judge_reason']}"
-                )
-            else:
-                result["predicted_answer"] = eval_result["predicted_answer"]
-                result["hard"] = int(eval_result["em"])
-                result["soft"] = eval_result["f1"]
-                if not result["hard"]:
-                    result["fail_reason"] = (
-                        f"judge=0: predicted '{eval_result['predicted_answer']}' "
-                        f"but expected {item['blank_answers']} ({eval_result['judge_reason']})"
-                    )
-                eval_detail = (
-                    f"[EVALUATION RESULT]\n"
-                    f"Question: {item['question']}\n"
-                    f"Predicted answer: {eval_result['predicted_answer']!r}\n"
-                    f"Gold answers: {item['blank_answers']!r}\n"
-                    f"Judge correct: {eval_result['em']}\n"
-                    f"Judge reason: {eval_result['judge_reason']}\n"
-                    f"String F1: {eval_result.get('string_f1', 0.0):.4f}"
-                )
-            conversation.append({"role": "system", "content": eval_detail})
-            with open(os.path.join(pred_dir, "conversation.json"), "w", encoding="utf-8") as f:
-                json.dump(conversation, f, ensure_ascii=False, indent=2)
-            return result
-
-        messages, system_prompt, user_text = _build_messages(
-            item,
-            skill_content,
-            image_detail,
-            diagnostic_mode=diagnostic_mode,
-            diagnostic_instruction=diagnostic_instruction,
-            diagnostic_trace_context=diagnostic_trace_context,
-        )
-        response = ""
-        conversation: list[dict] = [
-            {"role": "user", "content": f"{user_text}\n\n[image] {os.path.basename(item['image_path'])}"}
-        ]
-
-        for turn in range(max_turns):
-            if turn == 0:
-                resp_text, _ = chat_student_messages(
-                    messages=messages,
-                    max_completion_tokens=768,
-                    retries=5,
-                    stage="rollout",
-                )
-            else:
-                refinement_text = (
-                    f"Your previous answer was:\n{response}\n\n"
-                    "Review the same image and question carefully. "
-                    "If needed, correct your answer. Output the final answer in \\boxed{...}."
-                )
-                refinement_messages = [
-                    messages[0],
-                    messages[1],
-                    {"role": "assistant", "content": response},
-                    {"role": "user", "content": refinement_text},
-                ]
-                resp_text, _ = chat_student_messages(
-                    messages=refinement_messages,
-                    max_completion_tokens=512,
-                    retries=5,
-                    stage="rollout",
-                )
-            response = resp_text
-            conversation.append({"type": "message", "turn": turn + 1, "content": resp_text})
-            if extract_boxed_answer(resp_text) is not None:
-                break
-
-        result["response"] = response
-        result["agent_ok"] = True
-        result["n_turns"] = len(conversation) - 1
-
-        with open(os.path.join(pred_dir, "student_system_prompt.txt"), "w", encoding="utf-8") as f:
-            f.write(system_prompt)
-        with open(os.path.join(pred_dir, "student_user_prompt.txt"), "w", encoding="utf-8") as f:
-            f.write(user_text)
-
-        eval_result = evaluate_item(
-            item=item,
-            prediction_text=response,
-            judge_model=judge_model,
-            max_completion_tokens=judge_max_completion_tokens,
-            retries=judge_retries,
-        )
-        result["evaluation_mode"] = eval_result["evaluation_mode"]
-        result["judge_raw"] = eval_result["judge_raw"]
-        result["judge_reason"] = eval_result["judge_reason"]
-        result["matched_gold"] = eval_result["matched_gold"]
-
-        if item["ans_type"] == "choice":
-            result["predicted_label"] = eval_result["predicted_label"]
-            result["predicted_text"] = eval_result["predicted_text"]
-            result["predicted_answer"] = eval_result["predicted_answer"]
-            result["hard"] = int(eval_result["em"])
-            result["soft"] = eval_result["f1"]
-            if not result["hard"]:
-                result["fail_reason"] = (
-                    f"judge=0: predicted '{eval_result['predicted_label'] or eval_result['predicted_answer']}' "
-                    f"but expected '{eval_result['correct_label']}' ({eval_result['judge_reason']})"
-                )
-            eval_detail = (
-                f"[EVALUATION RESULT]\n"
-                f"Question: {item['question']}\n"
-                f"Predicted label: {eval_result['predicted_label']!r}\n"
-                f"Predicted text: {eval_result['predicted_text']!r}\n"
-                f"Correct label: {eval_result['correct_label']!r}\n"
-                f"Correct text: {eval_result['correct_text']!r}\n"
-                f"Judge correct: {eval_result['em']}\n"
-                f"Judge reason: {eval_result['judge_reason']}"
-            )
-        else:
-            result["predicted_answer"] = eval_result["predicted_answer"]
-            result["hard"] = int(eval_result["em"])
-            result["soft"] = eval_result["f1"]
-            if not result["hard"]:
-                result["fail_reason"] = (
-                    f"judge=0: predicted '{eval_result['predicted_answer']}' "
-                    f"but expected {item['blank_answers']} ({eval_result['judge_reason']})"
-                )
-            eval_detail = (
-                f"[EVALUATION RESULT]\n"
-                f"Question: {item['question']}\n"
-                f"Predicted answer: {eval_result['predicted_answer']!r}\n"
-                f"Gold answers: {item['blank_answers']!r}\n"
-                f"Judge correct: {eval_result['em']}\n"
-                f"Judge reason: {eval_result['judge_reason']}\n"
-                f"String F1: {eval_result.get('string_f1', 0.0):.4f}"
-            )
-
-        conversation.append({"role": "system", "content": eval_detail})
-        with open(os.path.join(pred_dir, "conversation.json"), "w", encoding="utf-8") as f:
-            json.dump(conversation, f, ensure_ascii=False, indent=2)
-    except Exception as e:  # noqa: BLE001
-        result["fail_reason"] = f"error: {e}"
-    return result
-
-
-def run_batch(
-    items: list[dict],
-    out_root: str,
-    skill_content: str,
-    *,
-    max_turns: int = 1,
-    workers: int = 32,
-    image_detail: str = "auto",
-    judge_model: str = "gpt-5.4",
-    judge_max_completion_tokens: int = 256,
-    judge_retries: int = 5,
-    diagnostic_mode: bool = False,
-    diagnostic_instruction: str = "",
-    diagnostic_trace_context_by_id: dict[str, str] | None = None,
-) -> list[dict]:
-    results_path = os.path.join(out_root, "results.jsonl")
-    os.makedirs(out_root, exist_ok=True)
-
-    expected_eval_mode = evaluation_mode()
-    done_ids: set[str] = set()
-    existing: list[dict] = []
-    rewrite_results = False
-    if os.path.exists(results_path):
-        with open(results_path, encoding="utf-8") as f:
-            for line in f:
-                try:
-                    row = json.loads(line)
-                    if row.get("evaluation_mode") != expected_eval_mode:
-                        rewrite_results = True
-                        continue
-                    done_ids.add(str(row["id"]))
-                    existing.append(row)
-                except Exception:
-                    rewrite_results = True
-
-    pending = [item for item in items if str(item["id"]) not in done_ids]
-    if not pending and not rewrite_results:
-        return existing
-
-    results = list(existing)
-    file_mode = "w" if rewrite_results else "a"
-    with open(results_path, file_mode, encoding="utf-8") as outf, ThreadPoolExecutor(max_workers=workers) as ex:
-        if rewrite_results:
-            for row in existing:
-                outf.write(json.dumps(row, ensure_ascii=False) + "\n")
-        futs = {
-            ex.submit(
-                process_one,
-                item,
-                out_root,
-                skill_content,
-                max_turns=max_turns,
-                image_detail=image_detail,
-                judge_model=judge_model,
-                judge_max_completion_tokens=judge_max_completion_tokens,
-                judge_retries=judge_retries,
-                diagnostic_mode=diagnostic_mode,
-                diagnostic_instruction=diagnostic_instruction,
-                diagnostic_trace_context=(diagnostic_trace_context_by_id or {}).get(str(item["id"]), ""),
-            ): item
-            for item in pending
-        }
-        for fut in as_completed(futs):
-            row = fut.result()
-            results.append(row)
-            outf.write(json.dumps(row, ensure_ascii=False) + "\n")
-            outf.flush()
-    return results

reflact/envs/babyvision/skills/initial.md

@@ -1,18 +0,0 @@
-# BabyVision Visual QA Heuristics
-
-## Image Inspection
-- First identify the main objects, their attributes, and their spatial relations before answering.
-- If the question involves counting, compare all relevant instances carefully instead of stopping after the first match.
-- If the question asks about color, size, position, or action, verify the specific visible evidence for that attribute.
-
-## Multiple Choice
-- Compare every option against the visible image evidence before deciding.
-- Prefer the option that matches the image exactly; reject options that are only partially true or too vague.
-- When two options are close, check the smallest discriminating visual detail.
-
-## Open Answers
-- Answer with the shortest phrase that is fully supported by the image.
-- Match the expected level of specificity: not broader than the image evidence, not narrower than the question asks.
-
-## Final Answer
-- Output only the final answer inside <answer>...</answer>.

reflact/envs/deep_reflect.py

@@ -1,114 +0,0 @@
-from __future__ import annotations
-
-import json
-import os
-from typing import Any, Callable
-
-from reflact.gradient.deep_probe import generate_deep_probe_instruction
-from reflact.gradient.reflect import run_minibatch_reflect
-
-
-def run_no_reference_deep_reflect(
-    adapter: Any,
-    results: list[dict],
-    skill_content: str,
-    out_dir: str,
-    *,
-    env_manager: Any = None,
-    prediction_dir: str | None = None,
-    random_seed: int | None = None,
-    step_buffer_context: str = "",
-    output_requirements: list[str] | None = None,
-    metadata_builder: Callable[[dict], dict] | None = None,
-) -> list[dict | None]:
-    """Run teacher-designed diagnostic probing without hidden references."""
-    if not getattr(adapter, "use_deep_reflect", False):
-        return []
-    if not isinstance(env_manager, list):
-        return []
-
-    prediction_dir = prediction_dir or os.path.join(out_dir, "predictions")
-    selected_items = adapter.select_representative_items(
-        results,
-        env_manager,
-        n_failures=getattr(adapter, "deep_reflect_failures", 4),
-        n_successes=getattr(adapter, "deep_reflect_successes", 2),
-        seed=random_seed,
-    )
-    if not selected_items:
-        return []
-
-    selected_ids = {str(item["id"]) for item in selected_items}
-    selected_results = [row for row in results if str(row.get("id")) in selected_ids]
-    if metadata_builder is None:
-        selected_metadata = [
-            {
-                "id": str(item.get("id")),
-                "task_type": str(item.get("task_type") or item.get("topic") or "unknown"),
-                "question_preview": str(item.get("question") or "")[:200],
-            }
-            for item in selected_items
-        ]
-    else:
-        selected_metadata = [metadata_builder(item) for item in selected_items]
-
-    deep_dir = os.path.join(out_dir, "deep_reflect")
-    rollout_dir = os.path.join(deep_dir, "rollout")
-    patches_dir = os.path.join(deep_dir, "patches")
-    os.makedirs(deep_dir, exist_ok=True)
-    print(
-        f"    [2b/6 DEEP REFLECT setup] selected={len(selected_items)} "
-        "mode=no_reference_probe"
-    )
-
-    probe = generate_deep_probe_instruction(
-        skill_content=skill_content,
-        items=selected_results,
-        prediction_dir=prediction_dir,
-        system_prompt=adapter.get_deep_probe_prompt(),
-        step_buffer_context=step_buffer_context,
-        output_requirements=output_requirements,
-    )
-    if not probe:
-        return []
-
-    with open(os.path.join(deep_dir, "probe.json"), "w", encoding="utf-8") as f:
-        json.dump(
-            {
-                **probe,
-                "reference_summary": {
-                    "mode": "no_reference_probe",
-                    "selected_count": len(selected_items),
-                },
-                "selected_examples": selected_metadata,
-            },
-            f,
-            ensure_ascii=False,
-            indent=2,
-        )
-
-    deep_results = adapter.rollout(
-        selected_items,
-        skill_content,
-        rollout_dir,
-        diagnostic_mode=True,
-        diagnostic_instruction=probe["probe_instruction"],
-    )
-    return run_minibatch_reflect(
-        results=deep_results,
-        skill_content=skill_content,
-        prediction_dir=os.path.join(rollout_dir, "predictions"),
-        patches_dir=patches_dir,
-        workers=getattr(adapter, "analyst_workers", 8),
-        failure_only=getattr(adapter, "failure_only", False),
-        minibatch_size=getattr(adapter, "minibatch_size", 8),
-        edit_budget=getattr(adapter, "edit_budget", 4),
-        random_seed=random_seed,
-        error_system=adapter.get_error_minibatch_prompt(),
-        success_system=adapter.get_success_minibatch_prompt(),
-        step_buffer_context=step_buffer_context,
-        update_mode=getattr(getattr(adapter, "_cfg", {}), "get", lambda *_: "patch")(
-            "skill_update_mode",
-            "patch",
-        ),
-    )

reflact/envs/livemathematicianbench/adapter.py

@@ -1,284 +0,0 @@
-"""LiveMathematicianBench environment adapter for ReflACT."""
-from __future__ import annotations
-
-import json
-import os
-
-from reflact.gradient.deep_probe import generate_deep_probe_instruction
-from reflact.datasets.base import BatchSpec
-from reflact.gradient.reflect import run_minibatch_reflect
-from reflact.envs.base import EnvAdapter
-from reflact.envs.livemathematicianbench.dataloader import LiveMathematicianBenchDataLoader
-from reflact.envs.livemathematicianbench.rollout import run_batch
-from reflact.model import get_student_backend
-
-
-class LiveMathematicianBenchAdapter(EnvAdapter):
-    """LiveMathematicianBench adapter."""
-
-    def build_reference_text(self, item: dict) -> str:
-        parts: list[str] = []
-        theorem = str(item.get("theorem") or "").strip()
-        sketch = str(item.get("sketch") or "").strip()
-        if theorem:
-            parts.append(f"## Reference Theorem\n{theorem}")
-        if sketch:
-            parts.append(f"## Reference Sketch\n{sketch}")
-        return "\n\n".join(parts)
-
-    def get_reference_metadata(self, item: dict) -> dict:
-        fields: list[str] = []
-        previews: list[str] = []
-        theorem = str(item.get("theorem") or "").strip()
-        sketch = str(item.get("sketch") or "").strip()
-        if theorem:
-            fields.append("theorem")
-            previews.append(f"[theorem]\n{theorem[:220]}")
-        if sketch:
-            fields.append("sketch")
-            previews.append(f"[sketch]\n{sketch[:220]}")
-        return {
-            "fields": fields,
-            "preview": "\n\n".join(previews)[:500],
-        }
-
-    def __init__(
-        self,
-        split_dir: str = "",
-        data_path: str = "",
-        split_mode: str = "ratio",
-        split_ratio: str = "2:1:7",
-        split_seed: int = 42,
-        split_output_dir: str = "",
-        max_turns: int = 1,
-        exec_timeout: int = 300,
-        workers: int = 64,
-        analyst_workers: int = 16,
-        failure_only: bool = False,
-        minibatch_size: int = 8,
-        edit_budget: int = 4,
-        seed: int = 42,
-        limit: int = 0,
-        shuffle_choices: bool = True,
-        use_theorem: bool = False,
-        use_sketch: bool = False,
-        exec_timeout: int = 600,
-        use_deep_reflect: bool = False,
-        deep_reflect_failures: int = 4,
-        deep_reflect_successes: int = 2,
-    ) -> None:
-        self.max_turns = max_turns
-        self.exec_timeout = exec_timeout
-        self.workers = workers
-        self.exec_timeout = exec_timeout
-        self.analyst_workers = analyst_workers
-        self.failure_only = failure_only
-        self.minibatch_size = minibatch_size
-        self.edit_budget = edit_budget
-        self.use_theorem = use_theorem
-        self.use_sketch = use_sketch
-        self.use_deep_reflect = use_deep_reflect
-        self.deep_reflect_failures = deep_reflect_failures
-        self.deep_reflect_successes = deep_reflect_successes
-        self.dataloader = LiveMathematicianBenchDataLoader(
-            split_dir=split_dir,
-            data_path=data_path,
-            split_mode=split_mode,
-            split_ratio=split_ratio,
-            split_seed=split_seed,
-            split_output_dir=split_output_dir,
-            seed=seed,
-            limit=limit,
-            shuffle_choices=shuffle_choices,
-        )
-
-    def setup(self, cfg: dict) -> None:
-        super().setup(cfg)
-        self.dataloader.setup(cfg)
-
-    def get_dataloader(self):
-        return self.dataloader
-
-    def build_env_from_batch(self, batch: BatchSpec, **kwargs):
-        return list(batch.payload or [])
-
-    def build_train_env(self, batch_size: int, seed: int, **kwargs):
-        batch = self.dataloader.build_train_batch(batch_size=batch_size, seed=seed, **kwargs)
-        return self.build_env_from_batch(batch, **kwargs)
-
-    def build_eval_env(self, env_num: int, split: str, seed: int, **kwargs):
-        batch = self.dataloader.build_eval_batch(env_num=env_num, split=split, seed=seed, **kwargs)
-        return self.build_env_from_batch(batch, **kwargs)
-
-    def rollout(
-        self,
-        env_manager,
-        skill_content: str,
-        out_dir: str,
-        **kwargs,
-    ) -> list[dict]:
-        items: list[dict] = env_manager
-        return run_batch(
-            items=items,
-            out_root=out_dir,
-            skill_content=skill_content,
-            max_turns=self.max_turns,
-            exec_timeout=self.exec_timeout,
-            workers=self.workers,
-            use_theorem=self.use_theorem,
-            use_sketch=self.use_sketch,
-            diagnostic_mode=kwargs.get("diagnostic_mode", False),
-            diagnostic_instruction=kwargs.get("diagnostic_instruction", ""),
-            diagnostic_trace_context_by_id=kwargs.get("diagnostic_trace_context_by_id"),
-            task_timeout=self.exec_timeout,
-        )
-
-    def reflect(
-        self,
-        results: list[dict],
-        skill_content: str,
-        out_dir: str,
-        **kwargs,
-    ) -> list[dict | None]:
-        prediction_dir = kwargs.get("prediction_dir", os.path.join(out_dir, "predictions"))
-        patches_dir = kwargs.get("patches_dir", os.path.join(out_dir, "patches"))
-        random_seed = kwargs.get("random_seed")
-        step_buffer_context = kwargs.get("step_buffer_context", "")
-        meta_skill_context = kwargs.get("meta_skill_context", "")
-
-        return run_minibatch_reflect(
-            results=results,
-            skill_content=skill_content,
-            prediction_dir=prediction_dir,
-            patches_dir=patches_dir,
-            workers=self.analyst_workers,
-            failure_only=self.failure_only,
-            minibatch_size=self.minibatch_size,
-            edit_budget=self.edit_budget,
-            random_seed=random_seed,
-            error_system=self.get_error_minibatch_prompt(),
-            success_system=self.get_success_minibatch_prompt(),
-            step_buffer_context=step_buffer_context,
-            meta_skill_context=meta_skill_context,
-            update_mode=getattr(self, "_cfg", {}).get("skill_update_mode", "patch"),
-        )
-
-    def deep_reflect(
-        self,
-        results: list[dict],
-        skill_content: str,
-        out_dir: str,
-        **kwargs,
-    ) -> list[dict | None]:
-        if not self.use_deep_reflect:
-            return []
-
-        env_manager = kwargs.get("env_manager")
-        prediction_dir = kwargs.get("prediction_dir", os.path.join(out_dir, "predictions"))
-        random_seed = kwargs.get("random_seed")
-        step_buffer_context = kwargs.get("step_buffer_context", "")
-        meta_skill_context = kwargs.get("meta_skill_context", "")
-        codex_backend = get_student_backend() == "codex_exec"
-        selected_items = self.select_representative_items(
-            results,
-            env_manager if isinstance(env_manager, list) else None,
-            n_failures=self.deep_reflect_failures,
-            n_successes=self.deep_reflect_successes,
-            seed=random_seed,
-        )
-        if not selected_items:
-            return []
-        selected_ids = {str(item["id"]) for item in selected_items}
-        selected_results = [row for row in results if str(row.get("id")) in selected_ids]
-        selected_examples = self.attach_reference_context(selected_results, selected_items)
-        if codex_backend:
-            selected_examples = self.attach_codex_probe_context(selected_examples, prediction_dir)
-        selected_metadata = []
-        theorem_count = 0
-        sketch_count = 0
-        for item in selected_items:
-            meta = self.get_reference_metadata(item)
-            if "theorem" in meta["fields"]:
-                theorem_count += 1
-            if "sketch" in meta["fields"]:
-                sketch_count += 1
-            selected_metadata.append({
-                "id": str(item["id"]),
-                "task_type": str(item.get("theorem_type", ["math_mcq"])[0] if item.get("theorem_type") else "math_mcq"),
-                "reference_fields": meta["fields"],
-                "reference_preview": meta["preview"],
-            })
-
-        deep_dir = os.path.join(out_dir, "deep_reflect")
-        rollout_dir = os.path.join(deep_dir, "rollout")
-        patches_dir = os.path.join(deep_dir, "patches")
-        os.makedirs(deep_dir, exist_ok=True)
-        print(
-            f"    [2b/6 DEEP REFLECT setup] selected={len(selected_items)} "
-            f"reference_fields=theorem({theorem_count}/{len(selected_items)}),"
-            f"sketch({sketch_count}/{len(selected_items)})"
-        )
-        probe = generate_deep_probe_instruction(
-            skill_content=skill_content,
-            items=selected_examples,
-            prediction_dir=prediction_dir,
-            system_prompt=self.get_codex_deep_probe_prompt() if codex_backend else self.get_deep_probe_prompt(),
-            step_buffer_context=step_buffer_context,
-            meta_skill_context=meta_skill_context,
-        )
-        if not probe:
-            return []
-        diagnostic_trace_context_by_id = None
-        if codex_backend:
-            selected_items, diagnostic_trace_context_by_id, probe = self.resolve_codex_probe_target(
-                selected_items=selected_items,
-                selected_examples=selected_examples,
-                prediction_dir=prediction_dir,
-                probe=probe,
-            )
-        probe_record = {
-            **probe,
-            "reference_summary": {
-                "selected_count": len(selected_items),
-                "field_counts": {
-                    "theorem": theorem_count,
-                    "sketch": sketch_count,
-                },
-            },
-            "selected_examples": selected_metadata,
-        }
-        with open(os.path.join(deep_dir, "probe.json"), "w", encoding="utf-8") as f:
-            json.dump(probe_record, f, ensure_ascii=False, indent=2)
-        deep_results = run_batch(
-            items=selected_items,
-            out_root=rollout_dir,
-            skill_content=skill_content,
-            max_turns=self.max_turns,
-            workers=min(self.workers, max(len(selected_items), 1)),
-            use_theorem=self.use_theorem,
-            use_sketch=self.use_sketch,
-            diagnostic_mode=True,
-            diagnostic_instruction=probe["probe_instruction"],
-            diagnostic_trace_context_by_id=diagnostic_trace_context_by_id,
-            task_timeout=self.exec_timeout,
-        )
-        deep_results = self.attach_reference_context(deep_results, selected_items)
-        return run_minibatch_reflect(
-            results=deep_results,
-            skill_content=skill_content,
-            prediction_dir=os.path.join(rollout_dir, "predictions"),
-            patches_dir=patches_dir,
-            workers=self.analyst_workers,
-            failure_only=self.failure_only,
-            minibatch_size=self.minibatch_size,
-            edit_budget=self.edit_budget,
-            random_seed=random_seed,
-            error_system=self.get_error_minibatch_prompt(),
-            success_system=self.get_success_minibatch_prompt(),
-            step_buffer_context=step_buffer_context,
-            meta_skill_context=meta_skill_context,
-            update_mode=getattr(self, "_cfg", {}).get("skill_update_mode", "patch"),
-        )
-
-    def get_task_types(self) -> list[str]:
-        return self.dataloader.get_task_types()

reflact/envs/livemathematicianbench/prompts/deep_probe.md

@@ -1,23 +0,0 @@
-You are an expert diagnostic-probe designer for theorem-grounded mathematical multiple-choice tasks.
-
-You will be shown representative trajectories, the current student skill, and the student's original prompt context.
-Design one SMALL diagnostic instruction that exposes the student's intermediate judgment without materially changing the original scaffold.
-
-## Hard Constraints
-1. Do NOT substantially change the original scaffold.
-2. Do NOT prescribe a new multi-step theorem-solving procedure.
-3. Do NOT ask for a full proof, full chain-of-thought, or exhaustive option-by-option derivation.
-4. Ask only for a short readout of the signals already behind the student's current answer.
-5. Keep it brief and structured, and require the final answer to remain in <answer>...</answer>.
-
-## Good Probe Targets
-- top choice and runner-up
-- decisive constraint
-- why the runner-up was rejected
-- strongest-vs-weaker discrimination signal
-
-Respond ONLY with a valid JSON object:
-{
-  "reasoning": "<why this probe is informative>",
-  "probe_instruction": "<the exact instruction text to append to the student prompt>"
-}

reflact/envs/livemathematicianbench/prompts/deep_probe_codex.md

@@ -1,26 +0,0 @@
-You are an expert diagnostic-probe designer for theorem-grounded mathematical multiple-choice tasks executed through a Codex trace.
-
-You will be shown representative trajectories, the current student skill, the student's original prompt context, hidden reference fields, and numbered Codex trace steps.
-Choose exactly one trajectory and one probe point. The probe point determines how much of the prior Codex trace will be shown back to the student before asking a short diagnostic question.
-
-## Hard Constraints
-1. Do NOT reveal or paraphrase the hidden reference directly to the student.
-2. Do NOT prescribe a new full solving procedure.
-3. Do NOT ask for a full proof, full chain-of-thought, or exhaustive option-by-option derivation.
-4. Ask only for a short readout of the signal that should already exist at that point in the student's process.
-5. The probe instruction must explicitly request a short <analysis>...</analysis> block before the final <answer>...</answer>.
-6. Select a probe point that is informative about theorem choice, decisive constraint, option elimination, or why a stronger/weaker option should be rejected.
-
-## Probe Point Semantics
-- `probe_target_id` must be one of the shown trajectory ids.
-- `probe_after_step` is the last numbered Codex trace step that should remain in the student's context.
-- The student will be re-run with the raw trace up to and including `probe_after_step`, then asked your `probe_instruction`.
-- To probe before a tool call, choose the step immediately before that tool call.
-
-Respond ONLY with a valid JSON object:
-{
-  "reasoning": "<why this trajectory and probe point expose the student's intermediate state>",
-  "probe_target_id": "<trajectory id>",
-  "probe_after_step": <integer step number>,
-  "probe_instruction": "<the exact instruction text to append to the student's prompt>"
-}

reflact/envs/mathverse/__init__.py

@@ -1,5 +0,0 @@
-"""MathVerse environment package."""
-
-from reflact.envs.mathverse.adapter import MathVerseAdapter
-
-__all__ = ["MathVerseAdapter"]

reflact/envs/mathverse/adapter.py

@@ -1,280 +0,0 @@
-"""MathVerse environment adapter for ReflACT."""
-from __future__ import annotations
-
-import json
-import os
-
-from reflact.datasets.base import BatchSpec
-from reflact.envs.base import EnvAdapter
-from reflact.envs.mathverse.dataloader import MathVerseDataLoader
-from reflact.envs.mathverse.rollout import run_batch
-from reflact.gradient.deep_probe import generate_deep_probe_instruction
-from reflact.gradient.reflect import run_minibatch_reflect
-from reflact.model import get_student_backend
-
-
-class MathVerseAdapter(EnvAdapter):
-    """MathVerse adapter."""
-
-    def build_reference_text(self, item: dict) -> str:
-        if not self.use_text_dominant_reference:
-            return ""
-        question = str(item.get("text_dominant_question") or "").strip()
-        if not question:
-            return ""
-        return f"## Reference Full Question\n{question}"
-
-    def get_reference_metadata(self, item: dict) -> dict:
-        if not self.use_text_dominant_reference:
-            return {"fields": [], "preview": ""}
-        question = str(item.get("text_dominant_question") or "").strip()
-        if not question:
-            return {"fields": [], "preview": ""}
-        return {
-            "fields": ["text_dominant_question"],
-            "preview": question[:400],
-        }
-
-    def __init__(
-        self,
-        split_dir: str = "",
-        data_root: str = "",
-        problem_version: str = "Text Lite",
-        use_text_dominant_reference: bool = False,
-        max_turns: int = 1,
-        workers: int = 16,
-        analyst_workers: int = 16,
-        failure_only: bool = False,
-        minibatch_size: int = 8,
-        edit_budget: int = 4,
-        seed: int = 42,
-        limit: int = 0,
-        image_detail: str = "auto",
-        judge_model: str = "gpt-5.4",
-        judge_max_completion_tokens: int = 256,
-        judge_retries: int = 5,
-        use_deep_reflect: bool = False,
-        deep_reflect_failures: int = 4,
-        deep_reflect_successes: int = 2,
-    ) -> None:
-        self.max_turns = max_turns
-        self.workers = workers
-        self.analyst_workers = analyst_workers
-        self.failure_only = failure_only
-        self.minibatch_size = minibatch_size
-        self.edit_budget = edit_budget
-        self.image_detail = image_detail
-        self.judge_model = judge_model
-        self.judge_max_completion_tokens = judge_max_completion_tokens
-        self.judge_retries = judge_retries
-        self.problem_version = problem_version
-        self.use_text_dominant_reference = use_text_dominant_reference
-        self.use_deep_reflect = use_deep_reflect
-        self.deep_reflect_failures = deep_reflect_failures
-        self.deep_reflect_successes = deep_reflect_successes
-        self.dataloader = MathVerseDataLoader(
-            split_dir=split_dir,
-            seed=seed,
-            limit=limit,
-            data_root=data_root,
-            problem_version=problem_version,
-        )
-
-    def setup(self, cfg: dict) -> None:
-        super().setup(cfg)
-        self.dataloader.setup(cfg)
-
-    def get_dataloader(self):
-        return self.dataloader
-
-    def build_env_from_batch(self, batch: BatchSpec, **kwargs):
-        return list(batch.payload or [])
-
-    def build_train_env(self, batch_size: int, seed: int, **kwargs):
-        batch = self.dataloader.build_train_batch(batch_size=batch_size, seed=seed, **kwargs)
-        return self.build_env_from_batch(batch, **kwargs)
-
-    def build_eval_env(self, env_num: int, split: str, seed: int, **kwargs):
-        batch = self.dataloader.build_eval_batch(env_num=env_num, split=split, seed=seed, **kwargs)
-        return self.build_env_from_batch(batch, **kwargs)
-
-    def rollout(
-        self,
-        env_manager,
-        skill_content: str,
-        out_dir: str,
-        **kwargs,
-    ) -> list[dict]:
-        items: list[dict] = env_manager
-        return run_batch(
-            items=items,
-            out_root=out_dir,
-            skill_content=skill_content,
-            max_turns=self.max_turns,
-            workers=self.workers,
-            image_detail=self.image_detail,
-            judge_model=self.judge_model,
-            judge_max_completion_tokens=self.judge_max_completion_tokens,
-            judge_retries=self.judge_retries,
-            diagnostic_mode=kwargs.get("diagnostic_mode", False),
-            diagnostic_instruction=kwargs.get("diagnostic_instruction", ""),
-            diagnostic_trace_context_by_id=kwargs.get("diagnostic_trace_context_by_id"),
-        )
-
-    def reflect(
-        self,
-        results: list[dict],
-        skill_content: str,
-        out_dir: str,
-        **kwargs,
-    ) -> list[dict | None]:
-        prediction_dir = kwargs.get("prediction_dir", os.path.join(out_dir, "predictions"))
-        patches_dir = kwargs.get("patches_dir", os.path.join(out_dir, "patches"))
-        random_seed = kwargs.get("random_seed")
-        step_buffer_context = kwargs.get("step_buffer_context", "")
-
-        return run_minibatch_reflect(
-            results=results,
-            skill_content=skill_content,
-            prediction_dir=prediction_dir,
-            patches_dir=patches_dir,
-            workers=self.analyst_workers,
-            failure_only=self.failure_only,
-            minibatch_size=self.minibatch_size,
-            edit_budget=self.edit_budget,
-            random_seed=random_seed,
-            error_system=self.get_error_minibatch_prompt(),
-            success_system=self.get_success_minibatch_prompt(),
-            step_buffer_context=step_buffer_context,
-            update_mode=getattr(self, "_cfg", {}).get("skill_update_mode", "patch"),
-        )
-
-    def deep_reflect(
-        self,
-        results: list[dict],
-        skill_content: str,
-        out_dir: str,
-        **kwargs,
-    ) -> list[dict | None]:
-        if not self.use_deep_reflect:
-            return []
-
-        env_manager = kwargs.get("env_manager")
-        prediction_dir = kwargs.get("prediction_dir", os.path.join(out_dir, "predictions"))
-        random_seed = kwargs.get("random_seed")
-        step_buffer_context = kwargs.get("step_buffer_context", "")
-        selected_items = self.select_representative_items(
-            results,
-            env_manager if isinstance(env_manager, list) else None,
-            n_failures=self.deep_reflect_failures,
-            n_successes=self.deep_reflect_successes,
-            seed=random_seed,
-        )
-        if not selected_items:
-            return []
-
-        selected_ids = {str(item["id"]) for item in selected_items}
-        selected_results = [row for row in results if str(row.get("id")) in selected_ids]
-        selected_examples = self.attach_reference_context(selected_results, selected_items)
-        codex_backend = get_student_backend() == "codex_exec"
-        if codex_backend:
-            selected_examples = self.attach_codex_probe_context(selected_examples, prediction_dir)
-        selected_metadata = []
-        ref_count = 0
-        for item in selected_items:
-            meta = self.get_reference_metadata(item)
-            if meta["fields"]:
-                ref_count += 1
-            record = {
-                "id": str(item["id"]),
-                "task_type": str(item.get("task_type") or item.get("question_type") or "mathverse"),
-                "reference_fields": meta["fields"],
-                "reference_preview": meta["preview"],
-            }
-            if codex_backend:
-                record["codex_probe_step_count"] = int(
-                    next(
-                        (row.get("codex_probe_step_count", 0) for row in selected_examples if str(row.get("id")) == str(item["id"])),
-                        0,
-                    )
-                )
-            selected_metadata.append(record)
-
-        deep_dir = os.path.join(out_dir, "deep_reflect")
-        rollout_dir = os.path.join(deep_dir, "rollout")
-        patches_dir = os.path.join(deep_dir, "patches")
-        os.makedirs(deep_dir, exist_ok=True)
-        print(
-            f"    [2b/6 DEEP REFLECT setup] selected={len(selected_items)} "
-            f"reference_fields=text_dominant_question({ref_count}/{len(selected_items)})"
-        )
-        probe = generate_deep_probe_instruction(
-            skill_content=skill_content,
-            items=selected_examples,
-            prediction_dir=prediction_dir,
-            system_prompt=self.get_codex_deep_probe_prompt() if codex_backend else self.get_deep_probe_prompt(),
-            step_buffer_context=step_buffer_context,
-        )
-        if not probe:
-            return []
-
-        targeted_items = selected_items
-        diagnostic_trace_context_by_id: dict[str, str] | None = None
-        if codex_backend:
-            targeted_items, diagnostic_trace_context_by_id, probe = self.resolve_codex_probe_target(
-                selected_items=selected_items,
-                selected_examples=selected_examples,
-                prediction_dir=prediction_dir,
-                probe=probe,
-            )
-
-        with open(os.path.join(deep_dir, "probe.json"), "w", encoding="utf-8") as f:
-            json.dump(
-                {
-                    **probe,
-                    "reference_summary": {
-                        "selected_count": len(selected_items),
-                        "field_counts": {
-                            "text_dominant_question": ref_count,
-                        },
-                    },
-                    "selected_examples": selected_metadata,
-                },
-                f,
-                ensure_ascii=False,
-                indent=2,
-            )
-
-        deep_results = run_batch(
-            items=targeted_items,
-            out_root=rollout_dir,
-            skill_content=skill_content,
-            max_turns=self.max_turns,
-            workers=min(self.workers, max(len(targeted_items), 1)),
-            image_detail=self.image_detail,
-            judge_model=self.judge_model,
-            judge_max_completion_tokens=self.judge_max_completion_tokens,
-            judge_retries=self.judge_retries,
-            diagnostic_mode=True,
-            diagnostic_instruction=probe["probe_instruction"],
-            diagnostic_trace_context_by_id=diagnostic_trace_context_by_id,
-        )
-        deep_results = self.attach_reference_context(deep_results, targeted_items)
-        return run_minibatch_reflect(
-            results=deep_results,
-            skill_content=skill_content,
-            prediction_dir=os.path.join(rollout_dir, "predictions"),
-            patches_dir=patches_dir,
-            workers=self.analyst_workers,
-            failure_only=self.failure_only,
-            minibatch_size=self.minibatch_size,
-            edit_budget=self.edit_budget,
-            random_seed=random_seed,
-            error_system=self.get_error_minibatch_prompt(),
-            success_system=self.get_success_minibatch_prompt(),
-            step_buffer_context=step_buffer_context,
-            update_mode=getattr(self, "_cfg", {}).get("skill_update_mode", "patch"),
-        )
-
-    def get_task_types(self) -> list[str]:
-        return self.dataloader.get_task_types()

reflact/envs/mathverse/dataloader.py

@@ -1,228 +0,0 @@
-"""MathVerse task dataloader."""
-from __future__ import annotations
-
-import json
-import os
-import re
-from typing import Any
-
-from reflact.datasets.base import SplitDataLoader
-
-
-_CHOICE_LABELS = ["A", "B", "C", "D", "E", "F", "G"]
-_CHOICE_BLOCK_RE = re.compile(r"\bChoices?\s*:\s*", re.IGNORECASE)
-_CHOICE_ITEM_RE = re.compile(r"([A-G])\s*[:.)]\s*(.*?)(?=(?:\s+[A-G]\s*[:.)])|$)", re.DOTALL)
-
-
-def _load_json(path: str) -> Any:
-    with open(path, encoding="utf-8") as f:
-        return json.load(f)
-
-
-def _normalize_space(text: Any) -> str:
-    return re.sub(r"\s+", " ", str(text or "").strip())
-
-
-def _resolve_image_path(raw_path: str, *, data_root: str, source_path: str) -> str:
-    candidates = []
-    if raw_path:
-        if os.path.isabs(raw_path):
-            candidates.append(raw_path)
-        else:
-            if data_root:
-                candidates.append(os.path.join(data_root, raw_path))
-                candidates.append(os.path.join(data_root, "images", raw_path))
-            candidates.append(os.path.join(os.path.dirname(source_path), raw_path))
-    for candidate in candidates:
-        if candidate and os.path.exists(candidate):
-            return os.path.abspath(candidate)
-    return ""
-
-
-def _split_question_and_choices(question: str) -> tuple[str, list[dict]]:
-    text = str(question or "").strip()
-    match = _CHOICE_BLOCK_RE.search(text)
-    if not match:
-        return text, []
-
-    stem = text[:match.start()].strip()
-    choice_block = text[match.end():].strip()
-    choices: list[dict] = []
-    for idx, m in enumerate(_CHOICE_ITEM_RE.finditer(choice_block)):
-        label = (m.group(1) or _CHOICE_LABELS[idx]).strip().upper()
-        choice_text = _normalize_space(m.group(2))
-        if choice_text:
-            choices.append({"label": label, "text": choice_text})
-    return stem or text, choices
-
-
-def _build_text_dominant_map(data_root: str) -> dict[str, str]:
-    if not data_root:
-        return {}
-    candidates = [
-        os.path.join(data_root, "testmini.json"),
-        os.path.join(data_root, "data", "testmini.json"),
-    ]
-    source_path = next((path for path in candidates if os.path.exists(path)), "")
-    if not source_path:
-        return {}
-
-    raw = _load_json(source_path)
-    if not isinstance(raw, list):
-        return {}
-
-    mapping: dict[str, str] = {}
-    for item in raw:
-        if not isinstance(item, dict):
-            continue
-        if str(item.get("problem_version") or "").strip() != "Text Dominant":
-            continue
-        problem_index = str(item.get("problem_index") or "").strip()
-        question = str(item.get("question") or "").strip()
-        if problem_index and question:
-            mapping[problem_index] = question
-    return mapping
-
-
-def _normalize_item(
-    item: dict,
-    *,
-    row_idx: int,
-    source_path: str,
-    data_root: str,
-    problem_version: str,
-    text_dominant_map: dict[str, str],
-) -> dict | None:
-    raw_problem_version = str(item.get("problem_version") or "").strip()
-    if problem_version and raw_problem_version and raw_problem_version != problem_version:
-        return None
-
-    question = str(item.get("question") or "").strip()
-    question_type = str(item.get("question_type") or "").strip()
-    answer = str(item.get("answer") or "").strip()
-    image_rel = str(item.get("image") or "").strip()
-    image_path = _resolve_image_path(image_rel, data_root=data_root, source_path=source_path)
-    if not answer or not image_path:
-        return None
-
-    metadata = item.get("metadata") if isinstance(item.get("metadata"), dict) else {}
-    subject = str(metadata.get("subject") or "").strip()
-    subfield = str(metadata.get("subfield") or "").strip()
-    source = str(metadata.get("source") or "").strip()
-
-    question_stem, choices = _split_question_and_choices(question)
-    is_choice = question_type == "multi-choice" or bool(choices)
-
-    correct_choice = {"label": "", "text": ""}
-    if is_choice:
-        label = str(answer).strip().upper().rstrip(".):")
-        choice_text = ""
-        for choice in choices:
-            if choice["label"].upper() == label:
-                choice_text = choice["text"]
-                break
-        correct_choice = {"label": label, "text": choice_text}
-
-    problem_index = str(item.get("problem_index") or "").strip()
-    sample_index = str(item.get("sample_index") or row_idx + 1).strip()
-    item_id = problem_index or sample_index
-    task_type = subfield or subject or question_type or "mathverse"
-
-    return {
-        "id": item_id,
-        "sample_index": sample_index,
-        "problem_index": problem_index,
-        "problem_version": raw_problem_version or problem_version,
-        "question": question,
-        "question_stem": question_stem,
-        "question_for_eval": str(item.get("question_for_eval") or question).strip(),
-        "question_type": question_type or ("multi-choice" if is_choice else "free-form"),
-        "is_choice": is_choice,
-        "choices": choices,
-        "correct_choice": correct_choice,
-        "answer": answer,
-        "gold_answers": [answer] if answer else [],
-        "image_rel": image_rel,
-        "image_path": image_path,
-        "query_wo": str(item.get("query_wo") or "").strip(),
-        "query_cot": str(item.get("query_cot") or "").strip(),
-        "metadata": {
-            "split": str(metadata.get("split") or "").strip(),
-            "source": source,
-            "subject": subject,
-            "subfield": subfield,
-        },
-        "task_type": task_type,
-        "source_path": os.path.abspath(source_path),
-        "text_dominant_question": str(
-            item.get("text_dominant_question")
-            or text_dominant_map.get(problem_index, "")
-        ).strip(),
-    }
-
-
-class MathVerseDataLoader(SplitDataLoader):
-    """MathVerse dataloader."""
-
-    def __init__(
-        self,
-        split_dir: str = "",
-        seed: int = 42,
-        limit: int = 0,
-        data_root: str = "",
-        problem_version: str = "Text Lite",
-        **kwargs,
-    ) -> None:
-        super().__init__(split_dir=split_dir, seed=seed, limit=limit)
-        self.data_root = data_root
-        self.problem_version = problem_version
-        self._task_types: list[str] = []
-        self._text_dominant_map = _build_text_dominant_map(data_root)
-
-    def setup(self, cfg: dict) -> None:
-        if not self.data_root:
-            self.data_root = str(cfg.get("data_root") or "")
-        if not self.problem_version:
-            self.problem_version = str(cfg.get("problem_version") or "Text Lite")
-        self._text_dominant_map = _build_text_dominant_map(self.data_root)
-        super().setup(cfg)
-        all_items = self.train_items + self.val_items + self.test_items
-        task_types = {
-            item.get("task_type") or item.get("question_type") or "mathverse"
-            for item in all_items
-        }
-        self._task_types = sorted(str(x) for x in task_types if str(x).strip())
-
-    def get_task_types(self) -> list[str]:
-        return list(self._task_types)
-
-    def load_split_items(self, split_path: str) -> list[dict]:
-        raw_items = super().load_split_items(split_path)
-        source_path = next(
-            (
-                os.path.join(split_path, name)
-                for name in sorted(os.listdir(split_path))
-                if name.endswith(".json")
-            ),
-            split_path,
-        )
-        items: list[dict] = []
-        for row_idx, item in enumerate(raw_items):
-            if not isinstance(item, dict):
-                continue
-            norm = _normalize_item(
-                item,
-                row_idx=row_idx,
-                source_path=source_path,
-                data_root=self.data_root,
-                problem_version=self.problem_version,
-                text_dominant_map=self._text_dominant_map,
-            )
-            if norm is not None:
-                items.append(norm)
-        if not items:
-            raise ValueError(
-                f"No valid MathVerse items loaded from {split_path} "
-                f"for problem_version={self.problem_version!r}"
-            )
-        return items

reflact/envs/mathverse/evaluator.py

@@ -1,180 +0,0 @@
-"""MathVerse evaluation helpers."""
-from __future__ import annotations
-
-import re
-import string
-
-from reflact.model import chat_with_deployment
-from reflact.prompts import load_prompt
-
-
-_EVAL_MODE = "mathverse_choice_or_judge_v1"
-
-
-def normalize_text(text: str) -> str:
-    text = str(text or "").strip().lower()
-    text = text.replace("\\,", " ")
-    text = text.replace("\\ ", " ")
-    text = "".join(ch for ch in text if ch not in string.punctuation)
-    return " ".join(text.split())
-
-
-def normalize_math_text(text: str) -> str:
-    text = str(text or "").strip()
-    text = text.replace("$", "")
-    text = text.replace("\\mathrm", "")
-    text = text.replace("{", "")
-    text = text.replace("}", "")
-    text = text.replace("~", " ")
-    text = text.replace("\\,", " ")
-    text = text.replace("\\ ", " ")
-    return " ".join(text.split()).lower()
-
-
-def extract_answer(text: str | None) -> str:
-    raw = str(text or "").strip()
-    if not raw:
-        return ""
-
-    tags = re.findall(r"<answer>\s*(.*?)\s*</answer>", raw, re.IGNORECASE | re.DOTALL)
-    if tags:
-        return tags[-1].strip()
-
-    boxed = re.findall(r"\\boxed\{(.*?)\}", raw, re.IGNORECASE | re.DOTALL)
-    if boxed:
-        return boxed[-1].strip()
-
-    lines = [ln.strip() for ln in raw.splitlines() if ln.strip()]
-    if lines:
-        return lines[-1]
-    return raw
-
-
-def _judge_answer(
-    *,
-    item: dict,
-    extracted_answer: str,
-    judge_model: str,
-    max_completion_tokens: int,
-    retries: int,
-) -> dict:
-    question = str(item.get("question_for_eval") or item.get("question") or "").strip()
-    ground_truth = str(item.get("answer") or "").strip()
-    raw, _ = chat_with_deployment(
-        deployment=judge_model,
-        system="You are a careful and strict mathematical answer evaluator.",
-        user=load_prompt("judge", env="mathverse").format(
-            question=question,
-            groundtruth=ground_truth,
-            modeloutput=extracted_answer,
-        ),
-        max_completion_tokens=max_completion_tokens,
-        retries=retries,
-        stage="mathverse_judge",
-    )
-    response = str(raw).strip().lower()
-    if "true" in response:
-        correct = True
-    elif "false" in response:
-        correct = False
-    else:
-        correct = False
-    return {
-        "raw": raw,
-        "correct": correct,
-        "reason": response,
-        "matched_gold": ground_truth if correct else "",
-    }
-
-
-def evaluate_item(
-    *,
-    item: dict,
-    prediction_text: str,
-    judge_model: str,
-    max_completion_tokens: int = 256,
-    retries: int = 5,
-) -> dict:
-    extracted = extract_answer(prediction_text)
-
-    if item.get("is_choice"):
-        predicted_label = str(extracted).strip().upper().rstrip(".):")
-        correct_label = str(item["correct_choice"].get("label") or "").strip().upper()
-        predicted_text = ""
-        for choice in item.get("choices") or []:
-            if str(choice.get("label") or "").strip().upper() == predicted_label:
-                predicted_text = str(choice.get("text") or "").strip()
-                break
-        hard = 1.0 if predicted_label == correct_label else 0.0
-        return {
-            "evaluation_mode": _EVAL_MODE,
-            "predicted_answer": extracted,
-            "predicted_label": predicted_label,
-            "predicted_text": predicted_text,
-            "correct_label": correct_label,
-            "correct_text": str(item["correct_choice"].get("text") or "").strip(),
-            "em": hard,
-            "f1": hard,
-            "sub_em": hard,
-            "judge_raw": "",
-            "judge_reason": "exact_label_match" if hard else "label_mismatch",
-            "matched_gold": correct_label if hard else "",
-        }
-
-    gold_answer = str(item.get("answer") or "").strip()
-    pred_norm = normalize_math_text(extracted)
-    gold_norm = normalize_math_text(gold_answer)
-    if pred_norm and gold_norm and pred_norm == gold_norm:
-        return {
-            "evaluation_mode": _EVAL_MODE,
-            "predicted_answer": extracted,
-            "em": 1.0,
-            "f1": 1.0,
-            "sub_em": 1.0,
-            "judge_raw": "",
-            "judge_reason": "normalized_exact_match",
-            "matched_gold": gold_answer,
-            "string_f1": 1.0,
-        }
-
-    judge = _judge_answer(
-        item=item,
-        extracted_answer=extracted,
-        judge_model=judge_model,
-        max_completion_tokens=max_completion_tokens,
-        retries=retries,
-    )
-    hard = 1.0 if judge["correct"] else 0.0
-    pred_tokens = normalize_text(extracted).split()
-    gold_tokens = normalize_text(gold_answer).split()
-    overlap = 0
-    gold_counts: dict[str, int] = {}
-    for tok in gold_tokens:
-        gold_counts[tok] = gold_counts.get(tok, 0) + 1
-    for tok in pred_tokens:
-        count = gold_counts.get(tok, 0)
-        if count > 0:
-            overlap += 1
-            gold_counts[tok] = count - 1
-    if pred_tokens and gold_tokens and overlap:
-        precision = overlap / len(pred_tokens)
-        recall = overlap / len(gold_tokens)
-        string_f1 = 2 * precision * recall / (precision + recall)
-    else:
-        string_f1 = 0.0
-
-    return {
-        "evaluation_mode": _EVAL_MODE,
-        "predicted_answer": extracted,
-        "em": hard,
-        "f1": hard,
-        "sub_em": hard,
-        "judge_raw": judge["raw"],
-        "judge_reason": judge["reason"],
-        "matched_gold": judge["matched_gold"],
-        "string_f1": string_f1,
-    }
-
-
-def evaluation_mode() -> str:
-    return _EVAL_MODE

reflact/envs/mathverse/prompts/analyst_error.md

@@ -1,37 +0,0 @@
-You are an expert failure-analysis agent for visual mathematical reasoning problems.
-
-You will be given MULTIPLE failed trajectories from a single minibatch and the current skill document.
-Each trajectory includes the student's response, the evaluation result, and sometimes a hidden reference
-containing the fuller Text Dominant version of the same problem.
-
-Your job is to identify COMMON reasoning failures across the batch and propose concise skill edits.
-
-## Failure Type Categories
-- **diagram_underuse**: the agent did not recover key constraints from the image
-- **constraint_drop**: the agent ignored a condition or relation that should guide the solution
-- **option_confusion**: the agent failed to discriminate between close answer choices
-- **format_miss**: the agent solved roughly correctly but returned the wrong final form, unit, or expression
-- **other**: none of the above
-
-## Rules
-1. Focus on patterns that recur across the minibatch.
-2. Prefer edits that improve visual grounding and exact answer selection.
-3. Do not hardcode problem-specific formulas or answers.
-4. If hidden reference text is present, use it only to infer what information the student failed to recover from the Text Lite version.
-
-Respond ONLY with a valid JSON object:
-{
-  "batch_size": <number>,
-  "failure_summary": [
-    {"failure_type": "<type>", "count": <int>, "description": "<one-line>"}
-  ],
-  "patch": {
-    "reasoning": "<why these edits address the common failures>",
-    "edits": [
-      {"op": "append",       "content": "<markdown>"},
-      {"op": "insert_after", "target": "<heading/text>", "content": "<markdown>"},
-      {"op": "replace",      "target": "<old text>",     "content": "<new text>"},
-      {"op": "delete",       "target": "<exact text to remove>"}
-    ]
-  }
-}

reflact/envs/mathverse/prompts/analyst_success.md

@@ -1,26 +0,0 @@
-You are an expert success-pattern analyst for visual mathematical reasoning problems.
-
-You will be given MULTIPLE successful trajectories from a minibatch and the current skill document.
-Identify generalizable behavior patterns that genuinely help the agent recover the right constraints
-from the image and convert them into the exact final answer.
-
-## Rules
-- Focus on broadly useful visual-math reasoning behaviors.
-- Prefer patterns about reading decisive diagram cues, checking hidden assumptions, and matching the final answer format exactly.
-- Do not add benchmark-specific facts or formulas.
-- "edits" may be empty if the skill already captures the useful patterns.
-
-Respond ONLY with a valid JSON object:
-{
-  "batch_size": <number>,
-  "success_patterns": ["<pattern 1>", "<pattern 2>"],
-  "patch": {
-    "reasoning": "<why these patterns matter>",
-    "edits": [
-      {"op": "append",       "content": "<markdown>"},
-      {"op": "insert_after", "target": "<heading/text>", "content": "<markdown>"},
-      {"op": "replace",      "target": "<old text>",     "content": "<new text>"},
-      {"op": "delete",       "target": "<exact text to remove>"}
-    ]
-  }
-}

reflact/envs/mathverse/prompts/deep_probe.md

@@ -1,25 +0,0 @@
-You are an expert diagnostic-probe designer for visual mathematical reasoning tasks.
-
-You will be shown representative trajectories, the current student skill, and the student's original prompt context.
-Some trajectories may also include a hidden reference containing the fuller Text Dominant wording of the same problem.
-Design one SMALL diagnostic instruction that exposes the student's intermediate judgment without materially changing the original scaffold.
-
-## Hard Constraints
-1. Do NOT substantially change the original scaffold.
-2. Do NOT prescribe a new long multi-step solving procedure.
-3. Do NOT ask for a full proof or full chain-of-thought.
-4. Ask only for a short readout of the signals already behind the student's current answer.
-5. Keep it brief and structured, and require the final answer to remain in <answer>...</answer>.
-6. If hidden reference text is present, use it only to target what visual or textual constraint the student likely missed.
-
-## Good Probe Targets
-- decisive diagram cue
-- top candidate and runner-up
-- missing relation or quantity
-- why a near-miss option was rejected
-
-Respond ONLY with a valid JSON object:
-{
-  "reasoning": "<why this probe is informative>",
-  "probe_instruction": "<the exact instruction text to append to the student prompt>"
-}

reflact/envs/mathverse/prompts/judge.md

@@ -1,25 +0,0 @@
-You are a careful and strict evaluator for visual math problems.
-
-You will be given:
-1. The original question
-2. The ground-truth answer
-3. A model output
-
-Decide whether the model output is mathematically equivalent to the ground-truth answer.
-
-Rules:
-- Ignore harmless formatting differences.
-- Accept mathematically equivalent expressions, equations, and values.
-- Reject answers that are numerically wrong, symbolically different in meaning, missing required units when the unit changes meaning, or correspond to a different choice.
-- Do not reward partially correct reasoning if the final answer is wrong.
-
-Return only:
-True
-
-or
-
-False
-
-Question: {question}
-Ground Truth Answer: {groundtruth}
-Model Output: {modeloutput}

reflact/envs/mathverse/prompts/rollout_system.md

@@ -1,11 +0,0 @@
-You are an expert visual mathematical reasoning agent.
-
-{skill_section}## Task Format
-You will receive one math problem with an image or diagram.
-Use the visible diagram as evidence, not just the text.
-If some information is abbreviated in the text, recover it from the image before answering.
-
-## Answer Format
-Think step by step, then provide your final answer inside <answer>...</answer>.
-- For multiple-choice questions, output only the single option label, such as <answer>B</answer>.
-- For free-form questions, output only the final mathematical answer, such as <answer>14</answer>.

reflact/envs/mathverse/reflect.py

@@ -1,4 +0,0 @@
-"""MathVerse Reflect stage.
-
-Prompts are loaded from .md files by the base adapter.
-"""

reflact/envs/mathverse/rollout.py

@@ -1,415 +0,0 @@
-"""MathVerse rollout — single-image multimodal math reasoning."""
-from __future__ import annotations
-
-import base64
-import json
-import mimetypes
-import os
-from concurrent.futures import ThreadPoolExecutor, as_completed
-
-from reflact.envs.mathverse.evaluator import evaluate_item, evaluation_mode, extract_answer
-from reflact.model import chat_student_messages, get_student_backend, is_student_exec_backend
-from reflact.model.codex_harness import prepare_workspace, render_skill_md, run_student_exec
-from reflact.prompts import load_prompt
-
-
-def _build_system(skill_content: str) -> str:
-    if skill_content.strip():
-        skill_section = f"## Skill\n{skill_content.strip()}\n\n"
-    else:
-        skill_section = ""
-    return load_prompt("rollout_system", env="mathverse").format(skill_section=skill_section)
-
-
-def _format_choices(choices: list[dict]) -> str:
-    return "\n".join(f"{choice['label']}. {choice['text']}" for choice in choices)
-
-
-def _build_user_text(
-    item: dict,
-    *,
-    diagnostic_mode: bool = False,
-    diagnostic_instruction: str = "",
-    diagnostic_trace_context: str = "",
-) -> str:
-    parts = []
-    if diagnostic_trace_context.strip():
-        parts.append(
-            "## Previous Codex Trace Snapshot\n"
-            "This is a partial transcript from an earlier attempt. Use it as your current reasoning context.\n\n"
-            f"{diagnostic_trace_context.strip()}"
-        )
-    question = str(item.get("question_stem") or item.get("question") or "").strip()
-    if question:
-        parts.append(f"## Question\n{question}")
-    else:
-        parts.append("## Question\nRead the full problem statement from the image.")
-
-    if item.get("is_choice"):
-        choices = item.get("choices") or []
-        if choices:
-            parts.append(f"## Choices\n{_format_choices(choices)}")
-        parts.append("Return only the final option label inside <answer>...</answer>.")
-    else:
-        parts.append("Return only the final mathematical answer inside <answer>...</answer>.")
-
-    if diagnostic_mode and diagnostic_instruction.strip():
-        parts.append(f"## Training Readout\n{diagnostic_instruction.strip()}")
-    return "\n\n".join(parts)
-
-
-def _image_to_data_uri(path: str) -> str:
-    mime = mimetypes.guess_type(path)[0] or "image/png"
-    with open(path, "rb") as f:
-        encoded = base64.b64encode(f.read()).decode("ascii")
-    return f"data:{mime};base64,{encoded}"
-
-
-def _build_messages(
-    item: dict,
-    skill_content: str,
-    image_detail: str,
-    *,
-    diagnostic_mode: bool = False,
-    diagnostic_instruction: str = "",
-    diagnostic_trace_context: str = "",
-) -> tuple[list[dict], str, str]:
-    system = _build_system(skill_content)
-    user_text = _build_user_text(
-        item,
-        diagnostic_mode=diagnostic_mode,
-        diagnostic_instruction=diagnostic_instruction,
-        diagnostic_trace_context=diagnostic_trace_context,
-    )
-    image_url = {"url": _image_to_data_uri(item["image_path"])}
-    if image_detail and image_detail != "auto":
-        image_url["detail"] = image_detail
-    messages = [
-        {"role": "system", "content": system},
-        {
-            "role": "user",
-            "content": [
-                {"type": "text", "text": user_text},
-                {"type": "image_url", "image_url": image_url},
-            ],
-        },
-    ]
-    return messages, system, user_text
-
-
-def _build_codex_skill(skill_content: str) -> str:
-    return render_skill_md(
-        skill_content,
-        description="Dynamic ReflACT skill for solving the current MathVerse visual math problem.",
-        preamble=(
-            "Use this skill when solving the current MathVerse problem.\n"
-            "Read the image carefully and return the final answer inside <answer>...</answer>."
-        ),
-    )
-
-
-def _run_codex_once(
-    *,
-    pred_dir: str,
-    item: dict,
-    skill_content: str,
-    model: str,
-    timeout: int,
-    image_detail: str,
-    diagnostic_mode: bool = False,
-    diagnostic_instruction: str = "",
-    diagnostic_trace_context: str = "",
-    previous_response: str = "",
-) -> tuple[str, str, str, str]:
-    user_text = _build_user_text(
-        item,
-        diagnostic_mode=diagnostic_mode,
-        diagnostic_instruction=diagnostic_instruction,
-        diagnostic_trace_context=diagnostic_trace_context,
-    )
-    task_parts = [user_text]
-    if previous_response:
-        task_parts.append(
-            "## Previous Attempt\n"
-            f"{previous_response}\n\n"
-            "Re-check the diagram and the mathematical constraints. Correct the final answer if needed."
-        )
-    task_text = "\n\n".join(task_parts)
-    skill_md = _build_codex_skill(skill_content)
-    work_dir = os.path.join(pred_dir, "codex_exec")
-    prepare_workspace(
-        work_dir=work_dir,
-        skill_md=skill_md,
-        task_text=task_text,
-        images=[item["image_path"]],
-    )
-    prompt = (
-        "Use the `reflact-student` skill available in this workspace.\n"
-        "Read `task.md`, inspect the attached image, solve the problem, and return only the final answer inside <answer>...</answer>."
-    )
-    final_message, raw = run_student_exec(
-        work_dir=work_dir,
-        prompt=prompt,
-        model=model,
-        timeout=timeout,
-        images=[item["image_path"]],
-    )
-    return final_message or raw, raw, skill_md, task_text
-
-
-def process_one(
-    item: dict,
-    out_root: str,
-    skill_content: str,
-    *,
-    max_turns: int = 1,
-    image_detail: str = "auto",
-    judge_model: str = "gpt-5.4",
-    judge_max_completion_tokens: int = 256,
-    judge_retries: int = 5,
-    diagnostic_mode: bool = False,
-    diagnostic_instruction: str = "",
-    diagnostic_trace_context: str = "",
-) -> dict:
-    item_id = str(item["id"])
-    result = {
-        "id": item_id,
-        "question": item["question"],
-        "task_type": item.get("task_type") or item.get("question_type") or "mathverse",
-        "task_description": item.get("question_stem") or item["question"],
-        "hard": 0,
-        "soft": 0.0,
-        "predicted_answer": "",
-        "predicted_label": "",
-        "predicted_text": "",
-        "response": "",
-        "fail_reason": "",
-        "agent_ok": False,
-        "n_turns": 0,
-        "image_path": item["image_path"],
-        "question_type": item["question_type"],
-        "evaluation_mode": evaluation_mode(),
-        "judge_model": judge_model,
-    }
-    if item.get("is_choice"):
-        result["correct_label"] = item["correct_choice"]["label"]
-        result["correct_text"] = item["correct_choice"]["text"]
-    else:
-        result["gold_answers"] = item.get("gold_answers") or [item["answer"]]
-
-    try:
-        pred_dir = os.path.join(out_root, "predictions", item_id)
-        os.makedirs(pred_dir, exist_ok=True)
-
-        if is_student_exec_backend():
-            from reflact.model import azure_openai as _llm
-
-            response = ""
-            conversation: list[dict] = [
-                {"role": "user", "content": f"{item['question']}\n\n[image] {os.path.basename(item['image_path'])}"}
-            ]
-            system_prompt = ""
-            user_text = ""
-            for turn in range(max_turns):
-                response, raw, system_prompt, user_text = _run_codex_once(
-                    pred_dir=pred_dir,
-                    item=item,
-                    skill_content=skill_content,
-                    model=_llm.STUDENT_DEPLOYMENT,
-                    timeout=120,
-                    image_detail=image_detail,
-                    diagnostic_mode=diagnostic_mode if turn == 0 else False,
-                    diagnostic_instruction=diagnostic_instruction if turn == 0 else "",
-                    diagnostic_trace_context=diagnostic_trace_context if turn == 0 else "",
-                    previous_response=response if turn > 0 else "",
-                )
-                conversation.append({"type": "message", "turn": turn + 1, "content": response})
-                if extract_answer(response):
-                    break
-
-            result["response"] = response
-            result["agent_ok"] = True
-            result["n_turns"] = len(conversation) - 1
-            with open(os.path.join(pred_dir, "student_system_prompt.txt"), "w", encoding="utf-8") as f:
-                f.write(system_prompt)
-            with open(os.path.join(pred_dir, "student_user_prompt.txt"), "w", encoding="utf-8") as f:
-                f.write(user_text)
-        else:
-            messages, system_prompt, user_text = _build_messages(
-                item,
-                skill_content,
-                image_detail,
-                diagnostic_mode=diagnostic_mode,
-                diagnostic_instruction=diagnostic_instruction,
-                diagnostic_trace_context=diagnostic_trace_context,
-            )
-            response = ""
-            conversation = [
-                {"role": "user", "content": f"{user_text}\n\n[image] {os.path.basename(item['image_path'])}"}
-            ]
-            for turn in range(max_turns):
-                if turn == 0:
-                    resp_text, _ = chat_student_messages(
-                        messages=messages,
-                        max_completion_tokens=1024,
-                        retries=5,
-                        stage="rollout",
-                    )
-                else:
-                    refinement_text = (
-                        f"Your previous answer was:\n{response}\n\n"
-                        "Re-check the diagram and the mathematical constraints. "
-                        "If needed, correct your answer. Output only the final answer inside <answer>...</answer>."
-                    )
-                    refinement_messages = [
-                        messages[0],
-                        messages[1],
-                        {"role": "assistant", "content": response},
-                        {"role": "user", "content": refinement_text},
-                    ]
-                    resp_text, _ = chat_student_messages(
-                        messages=refinement_messages,
-                        max_completion_tokens=768,
-                        retries=5,
-                        stage="rollout",
-                    )
-                response = resp_text
-                conversation.append({"type": "message", "turn": turn + 1, "content": resp_text})
-                if extract_answer(resp_text):
-                    break
-
-            result["response"] = response
-            result["agent_ok"] = True
-            result["n_turns"] = len(conversation) - 1
-            with open(os.path.join(pred_dir, "student_system_prompt.txt"), "w", encoding="utf-8") as f:
-                f.write(system_prompt)
-            with open(os.path.join(pred_dir, "student_user_prompt.txt"), "w", encoding="utf-8") as f:
-                f.write(user_text)
-
-        eval_result = evaluate_item(
-            item=item,
-            prediction_text=result["response"],
-            judge_model=judge_model,
-            max_completion_tokens=judge_max_completion_tokens,
-            retries=judge_retries,
-        )
-        result["evaluation_mode"] = eval_result["evaluation_mode"]
-        result["judge_raw"] = eval_result.get("judge_raw", "")
-        result["judge_reason"] = eval_result.get("judge_reason", "")
-        result["matched_gold"] = eval_result.get("matched_gold", "")
-
-        if item.get("is_choice"):
-            result["predicted_label"] = eval_result["predicted_label"]
-            result["predicted_text"] = eval_result["predicted_text"]
-            result["predicted_answer"] = eval_result["predicted_answer"]
-            result["hard"] = int(eval_result["em"])
-            result["soft"] = eval_result["f1"]
-            if not result["hard"]:
-                result["fail_reason"] = (
-                    f"choice=0: predicted '{eval_result['predicted_label'] or eval_result['predicted_answer']}' "
-                    f"but expected '{eval_result['correct_label']}'"
-                )
-            eval_detail = (
-                f"[EVALUATION RESULT]\n"
-                f"Question: {item['question_for_eval']}\n"
-                f"Predicted label: {eval_result['predicted_label']!r}\n"
-                f"Predicted text: {eval_result['predicted_text']!r}\n"
-                f"Correct label: {eval_result['correct_label']!r}\n"
-                f"Correct text: {eval_result['correct_text']!r}\n"
-                f"Exact Match: {eval_result['em']}"
-            )
-        else:
-            result["predicted_answer"] = eval_result["predicted_answer"]
-            result["hard"] = int(eval_result["em"])
-            result["soft"] = eval_result["f1"]
-            if not result["hard"]:
-                result["fail_reason"] = (
-                    f"judge=0: predicted '{eval_result['predicted_answer']}' "
-                    f"but expected '{item['answer']}' ({eval_result.get('judge_reason', '')})"
-                )
-            eval_detail = (
-                f"[EVALUATION RESULT]\n"
-                f"Question: {item['question_for_eval']}\n"
-                f"Predicted answer: {eval_result['predicted_answer']!r}\n"
-                f"Gold answer: {item['answer']!r}\n"
-                f"Judge correct: {eval_result['em']}\n"
-                f"Judge reason: {eval_result.get('judge_reason', '')}\n"
-                f"String F1: {eval_result.get('string_f1', 0.0):.4f}"
-            )
-
-        conversation.append({"role": "system", "content": eval_detail})
-        with open(os.path.join(pred_dir, "conversation.json"), "w", encoding="utf-8") as f:
-            json.dump(conversation, f, ensure_ascii=False, indent=2)
-    except Exception as e:  # noqa: BLE001
-        result["fail_reason"] = f"error: {e}"
-    return result
-
-
-def run_batch(
-    items: list[dict],
-    out_root: str,
-    skill_content: str,
-    *,
-    max_turns: int = 1,
-    workers: int = 32,
-    image_detail: str = "auto",
-    judge_model: str = "gpt-5.4",
-    judge_max_completion_tokens: int = 256,
-    judge_retries: int = 5,
-    diagnostic_mode: bool = False,
-    diagnostic_instruction: str = "",
-    diagnostic_trace_context_by_id: dict[str, str] | None = None,
-) -> list[dict]:
-    results_path = os.path.join(out_root, "results.jsonl")
-    os.makedirs(out_root, exist_ok=True)
-
-    expected_eval_mode = evaluation_mode()
-    done_ids: set[str] = set()
-    existing: list[dict] = []
-    rewrite_results = False
-    if os.path.exists(results_path):
-        with open(results_path, encoding="utf-8") as f:
-            for line in f:
-                try:
-                    row = json.loads(line)
-                    if row.get("evaluation_mode") != expected_eval_mode:
-                        rewrite_results = True
-                        continue
-                    done_ids.add(str(row["id"]))
-                    existing.append(row)
-                except Exception:
-                    rewrite_results = True
-
-    pending = [item for item in items if str(item["id"]) not in done_ids]
-    if not pending and not rewrite_results:
-        return existing
-
-    results = list(existing)
-    file_mode = "w" if rewrite_results else "a"
-    with open(results_path, file_mode, encoding="utf-8") as outf, ThreadPoolExecutor(max_workers=workers) as ex:
-        if rewrite_results:
-            for row in existing:
-                outf.write(json.dumps(row, ensure_ascii=False) + "\n")
-        futs = {
-            ex.submit(
-                process_one,
-                item,
-                out_root,
-                skill_content,
-                max_turns=max_turns,
-                image_detail=image_detail,
-                judge_model=judge_model,
-                judge_max_completion_tokens=judge_max_completion_tokens,
-                judge_retries=judge_retries,
-                diagnostic_mode=diagnostic_mode,
-                diagnostic_instruction=diagnostic_instruction,
-                diagnostic_trace_context=(diagnostic_trace_context_by_id or {}).get(str(item["id"]), ""),
-            ): item
-            for item in pending
-        }
-        for fut in as_completed(futs):
-            row = fut.result()
-            results.append(row)
-            outf.write(json.dumps(row, ensure_ascii=False) + "\n")
-            outf.flush()
-    return results

reflact/envs/mathverse/skills/initial.md

@@ -1,15 +0,0 @@
-# MathVerse Visual Math Heuristics
-
-## Diagram First
-- Read the diagram before locking onto an equation or option.
-- Recover missing labels, lengths, angles, axes, or object relations from the image when the text is abbreviated.
-- If the text seems underspecified, assume the image may contain the decisive constraint.
-
-## Constraint Tracking
-- Write down the few constraints that actually determine the answer instead of solving from vague intuition.
-- Prefer geometric or functional relations that are directly supported by the figure.
-- For multiple-choice questions, compare the final candidate against every option exactly.
-
-## Final Answer
-- Use the image and the text consistently.
-- Return only the final answer inside <answer>...</answer>.

reflact/envs/mmrb/__init__.py

@@ -1,2 +0,0 @@
-"""MMRB environment package."""
-

reflact/envs/mmrb/adapter.py

@@ -1,283 +0,0 @@
-"""MMRB environment adapter for ReflACT."""
-from __future__ import annotations
-
-import json
-import os
-
-from reflact.gradient.deep_probe import generate_deep_probe_instruction
-from reflact.datasets.base import BatchSpec
-from reflact.gradient.reflect import run_minibatch_reflect
-from reflact.envs.base import EnvAdapter
-from reflact.envs.mmrb.dataloader import MMRBDataLoader
-from reflact.envs.mmrb.rollout import run_batch
-from reflact.model import get_student_backend
-
-
-class MMRBAdapter(EnvAdapter):
-    """MMRB adapter."""
-
-    def build_reference_text(self, item: dict) -> str:
-        reasoning_steps = item.get("reasoning_steps") or []
-        if not reasoning_steps:
-            return ""
-
-        blocks: list[str] = []
-        for path_idx, path in enumerate(reasoning_steps, 1):
-            if not isinstance(path, list) or not path:
-                continue
-            lines = [f"### Reasoning Path {path_idx}"]
-            for step in path:
-                if not isinstance(step, dict):
-                    continue
-                step_no = step.get("reasoning step", "?")
-                step_type = str(step.get("reasoning type") or "").strip()
-                rationale = str(step.get("rationale") or "").strip()
-                if rationale:
-                    prefix = f"{step_no}. [{step_type}] " if step_type else f"{step_no}. "
-                    lines.append(prefix + rationale)
-            if len(lines) > 1:
-                blocks.append("\n".join(lines))
-        if not blocks:
-            return ""
-        return "## Reference Reasoning Steps\n" + "\n\n".join(blocks[:3])
-
-    def get_reference_metadata(self, item: dict) -> dict:
-        reasoning_steps = item.get("reasoning_steps") or []
-        path_count = 0
-        preview_parts: list[str] = []
-        for path in reasoning_steps:
-            if not isinstance(path, list) or not path:
-                continue
-            path_count += 1
-            first = path[0] if isinstance(path[0], dict) else {}
-            step_type = str(first.get("reasoning type") or "").strip()
-            rationale = str(first.get("rationale") or "").strip()
-            preview_parts.append(f"[path {path_count}] {step_type}: {rationale[:180]}")
-        if not path_count:
-            return {"fields": [], "preview": ""}
-        return {
-            "fields": ["reasoning_steps"],
-            "preview": "\n".join(preview_parts)[:500],
-        }
-
-    def __init__(
-        self,
-        split_dir: str = "",
-        data_path: str = "",
-        split_mode: str = "ratio",
-        split_ratio: str = "2:1:7",
-        split_seed: int = 42,
-        split_output_dir: str = "",
-        max_turns: int = 1,
-        workers: int = 16,
-        analyst_workers: int = 16,
-        failure_only: bool = False,
-        minibatch_size: int = 8,
-        edit_budget: int = 4,
-        seed: int = 42,
-        limit: int = 0,
-        image_detail: str = "auto",
-        use_deep_reflect: bool = False,
-        deep_reflect_failures: int = 4,
-        deep_reflect_successes: int = 2,
-    ) -> None:
-        self.max_turns = max_turns
-        self.workers = workers
-        self.analyst_workers = analyst_workers
-        self.failure_only = failure_only
-        self.minibatch_size = minibatch_size
-        self.edit_budget = edit_budget
-        self.image_detail = image_detail
-        self.use_deep_reflect = use_deep_reflect
-        self.deep_reflect_failures = deep_reflect_failures
-        self.deep_reflect_successes = deep_reflect_successes
-        self.dataloader = MMRBDataLoader(
-            split_dir=split_dir,
-            data_path=data_path,
-            split_mode=split_mode,
-            split_ratio=split_ratio,
-            split_seed=split_seed,
-            split_output_dir=split_output_dir,
-            seed=seed,
-            limit=limit,
-        )
-
-    def setup(self, cfg: dict) -> None:
-        super().setup(cfg)
-        self.dataloader.setup(cfg)
-
-    def get_dataloader(self):
-        return self.dataloader
-
-    def build_env_from_batch(self, batch: BatchSpec, **kwargs):
-        return list(batch.payload or [])
-
-    def build_train_env(self, batch_size: int, seed: int, **kwargs):
-        batch = self.dataloader.build_train_batch(batch_size=batch_size, seed=seed, **kwargs)
-        return self.build_env_from_batch(batch, **kwargs)
-
-    def build_eval_env(self, env_num: int, split: str, seed: int, **kwargs):
-        batch = self.dataloader.build_eval_batch(env_num=env_num, split=split, seed=seed, **kwargs)
-        return self.build_env_from_batch(batch, **kwargs)
-
-    def rollout(
-        self,
-        env_manager,
-        skill_content: str,
-        out_dir: str,
-        **kwargs,
-    ) -> list[dict]:
-        items: list[dict] = env_manager
-        return run_batch(
-            items=items,
-            out_root=out_dir,
-            skill_content=skill_content,
-            max_turns=self.max_turns,
-            workers=self.workers,
-            image_detail=self.image_detail,
-            diagnostic_mode=kwargs.get("diagnostic_mode", False),
-            diagnostic_instruction=kwargs.get("diagnostic_instruction", ""),
-            diagnostic_trace_context_by_id=kwargs.get("diagnostic_trace_context_by_id"),
-        )
-
-    def reflect(
-        self,
-        results: list[dict],
-        skill_content: str,
-        out_dir: str,
-        **kwargs,
-    ) -> list[dict | None]:
-        prediction_dir = kwargs.get("prediction_dir", os.path.join(out_dir, "predictions"))
-        patches_dir = kwargs.get("patches_dir", os.path.join(out_dir, "patches"))
-        random_seed = kwargs.get("random_seed")
-        step_buffer_context = kwargs.get("step_buffer_context", "")
-        meta_skill_context = kwargs.get("meta_skill_context", "")
-
-        return run_minibatch_reflect(
-            results=results,
-            skill_content=skill_content,
-            prediction_dir=prediction_dir,
-            patches_dir=patches_dir,
-            workers=self.analyst_workers,
-            failure_only=self.failure_only,
-            minibatch_size=self.minibatch_size,
-            edit_budget=self.edit_budget,
-            random_seed=random_seed,
-            error_system=self.get_error_minibatch_prompt(),
-            success_system=self.get_success_minibatch_prompt(),
-            step_buffer_context=step_buffer_context,
-            meta_skill_context=meta_skill_context,
-            update_mode=getattr(self, "_cfg", {}).get("skill_update_mode", "patch"),
-        )
-
-    def deep_reflect(
-        self,
-        results: list[dict],
-        skill_content: str,
-        out_dir: str,
-        **kwargs,
-    ) -> list[dict | None]:
-        if not self.use_deep_reflect:
-            return []
-
-        env_manager = kwargs.get("env_manager")
-        prediction_dir = kwargs.get("prediction_dir", os.path.join(out_dir, "predictions"))
-        random_seed = kwargs.get("random_seed")
-        step_buffer_context = kwargs.get("step_buffer_context", "")
-        meta_skill_context = kwargs.get("meta_skill_context", "")
-        codex_backend = get_student_backend() == "codex_exec"
-        selected_items = self.select_representative_items(
-            results,
-            env_manager if isinstance(env_manager, list) else None,
-            n_failures=self.deep_reflect_failures,
-            n_successes=self.deep_reflect_successes,
-            seed=random_seed,
-        )
-        if not selected_items:
-            return []
-        selected_ids = {str(item["id"]) for item in selected_items}
-        selected_results = [row for row in results if str(row.get("id")) in selected_ids]
-        selected_examples = self.attach_reference_context(selected_results, selected_items)
-        if codex_backend:
-            selected_examples = self.attach_codex_probe_context(selected_examples, prediction_dir)
-
-        reasoning_count = 0
-        selected_metadata = []
-        for item in selected_items:
-            meta = self.get_reference_metadata(item)
-            if meta["fields"]:
-                reasoning_count += 1
-            selected_metadata.append({
-                "id": str(item["id"]),
-                "task_type": str(item.get("subtask") or item.get("task_type") or "mmrb"),
-                "reference_fields": meta["fields"],
-                "reference_preview": meta["preview"],
-            })
-
-        deep_dir = os.path.join(out_dir, "deep_reflect")
-        rollout_dir = os.path.join(deep_dir, "rollout")
-        patches_dir = os.path.join(deep_dir, "patches")
-        os.makedirs(deep_dir, exist_ok=True)
-        print(
-            f"    [2b/6 DEEP REFLECT setup] selected={len(selected_items)} "
-            f"reference_fields=reasoning_steps({reasoning_count}/{len(selected_items)})"
-        )
-        probe = generate_deep_probe_instruction(
-            skill_content=skill_content,
-            items=selected_examples,
-            prediction_dir=prediction_dir,
-            system_prompt=self.get_codex_deep_probe_prompt() if codex_backend else self.get_deep_probe_prompt(),
-            step_buffer_context=step_buffer_context,
-            meta_skill_context=meta_skill_context,
-        )
-        if not probe:
-            return []
-        diagnostic_trace_context_by_id = None
-        if codex_backend:
-            selected_items, diagnostic_trace_context_by_id, probe = self.resolve_codex_probe_target(
-                selected_items=selected_items,
-                selected_examples=selected_examples,
-                prediction_dir=prediction_dir,
-                probe=probe,
-            )
-        probe_record = {
-            **probe,
-            "reference_summary": {
-                "selected_count": len(selected_items),
-                "field_counts": {"reasoning_steps": reasoning_count},
-            },
-            "selected_examples": selected_metadata,
-        }
-        with open(os.path.join(deep_dir, "probe.json"), "w", encoding="utf-8") as f:
-            json.dump(probe_record, f, ensure_ascii=False, indent=2)
-        deep_results = run_batch(
-            items=selected_items,
-            out_root=rollout_dir,
-            skill_content=skill_content,
-            max_turns=self.max_turns,
-            workers=min(self.workers, max(len(selected_items), 1)),
-            image_detail=self.image_detail,
-            diagnostic_mode=True,
-            diagnostic_instruction=probe["probe_instruction"],
-            diagnostic_trace_context_by_id=diagnostic_trace_context_by_id,
-        )
-        deep_results = self.attach_reference_context(deep_results, selected_items)
-        return run_minibatch_reflect(
-            results=deep_results,
-            skill_content=skill_content,
-            prediction_dir=os.path.join(rollout_dir, "predictions"),
-            patches_dir=patches_dir,
-            workers=self.analyst_workers,
-            failure_only=self.failure_only,
-            minibatch_size=self.minibatch_size,
-            edit_budget=self.edit_budget,
-            random_seed=random_seed,
-            error_system=self.get_error_minibatch_prompt(),
-            success_system=self.get_success_minibatch_prompt(),
-            step_buffer_context=step_buffer_context,
-            meta_skill_context=meta_skill_context,
-            update_mode=getattr(self, "_cfg", {}).get("skill_update_mode", "patch"),
-        )
-
-    def get_task_types(self) -> list[str]:
-        return self.dataloader.get_task_types()

reflact/envs/mmrb/dataloader.py

@@ -1,146 +0,0 @@
-"""MMRB task dataloader."""
-from __future__ import annotations
-
-import glob
-import json
-import os
-import re
-from typing import Any
-
-from reflact.datasets.base import SplitDataLoader
-
-
-# ── Raw data loading utilities (for preprocessing / standalone eval) ─────
-
-def _load_json(path: str) -> Any:
-    with open(path, encoding="utf-8") as f:
-        return json.load(f)
-
-
-def _iter_data_files(data_path: str) -> list[str]:
-    if not data_path:
-        return []
-    if os.path.isfile(data_path):
-        return [data_path]
-    if os.path.isdir(data_path):
-        nested = glob.glob(os.path.join(data_path, "**", "*_human.json"), recursive=True)
-        flat = glob.glob(os.path.join(data_path, "*_human.json"))
-        return sorted(set(nested + flat))
-    return []
-
-
-def _normalize_space(text: str) -> str:
-    return re.sub(r"\s+", " ", str(text or "").strip())
-
-
-def _normalize_item(item: dict, row_idx: int, source_path: str) -> dict | None:
-    question = _normalize_space(item.get("question") or "")
-    answer = _normalize_space(item.get("answer") or "")
-    raw_image_paths = item.get("image_paths") or []
-    if not question or not answer or not isinstance(raw_image_paths, list) or not raw_image_paths:
-        return None
-
-    base_dir = os.path.dirname(source_path)
-    image_paths: list[str] = []
-    for raw_path in raw_image_paths:
-        rel = str(raw_path or "").strip()
-        if not rel:
-            continue
-        abs_path = rel if os.path.isabs(rel) else os.path.abspath(os.path.join(base_dir, rel))
-        if os.path.exists(abs_path):
-            image_paths.append(abs_path)
-    if not image_paths:
-        return None
-
-    options_raw = item.get("options") or []
-    options = [_normalize_space(opt) for opt in options_raw if _normalize_space(opt)]
-    source = _normalize_space(item.get("source") or "unknown")
-    subtask = _normalize_space(item.get("subtask") or "unknown")
-    item_index = item.get("index", row_idx)
-    item_id = f"{source}:{subtask}:{item_index}"
-
-    return {
-        "id": item_id,
-        "source": source,
-        "subtask": subtask,
-        "task_type": subtask,
-        "question": question,
-        "answer": answer,
-        "options": options,
-        "is_choice": bool(options),
-        "image_paths": image_paths,
-        "reasoning_steps": item.get("reasoning_steps") or [],
-        "annotation_time": item.get("annotation_time"),
-        "source_path": os.path.abspath(source_path),
-    }
-
-
-def load_items(data_path: str) -> list[dict]:
-    """Load and normalise MMRB items from JSON files."""
-    files = _iter_data_files(data_path)
-    if not files:
-        raise ValueError(
-            "MMRB requires data_path to be a *_human.json file or a directory "
-            "containing extracted MMRB subtask folders."
-        )
-
-    items: list[dict] = []
-    for path in files:
-        raw = _load_json(path)
-        if not isinstance(raw, list):
-            raise ValueError(f"Expected JSON array in {path}, got {type(raw).__name__}")
-        for row_idx, item in enumerate(raw):
-            if not isinstance(item, dict):
-                continue
-            norm = _normalize_item(item, row_idx=row_idx, source_path=path)
-            if norm is not None:
-                items.append(norm)
-
-    if not items:
-        raise ValueError(f"No valid MMRB items loaded from {data_path}")
-    return items
-
-
-# ── Dataloader ───────────────────────────────────────────────────────────
-
-class MMRBDataLoader(SplitDataLoader):
-    """MMRB dataloader."""
-
-    def __init__(
-        self,
-        split_dir: str = "",
-        data_path: str = "",
-        split_mode: str = "ratio",
-        split_ratio: str = "2:1:7",
-        split_seed: int = 42,
-        split_output_dir: str = "",
-        seed: int = 42,
-        limit: int = 0,
-        **kwargs,
-    ) -> None:
-        super().__init__(
-            split_dir=split_dir,
-            data_path=data_path,
-            split_mode=split_mode,
-            split_ratio=split_ratio,
-            split_seed=split_seed,
-            split_output_dir=split_output_dir,
-            seed=seed,
-            limit=limit,
-        )
-        self._task_types: list[str] = []
-
-    def load_raw_items(self, data_path: str) -> list[dict]:
-        return load_items(data_path)
-
-    def setup(self, cfg: dict) -> None:
-        super().setup(cfg)
-        all_items = self.train_items + self.val_items + self.test_items
-        task_types = {
-            item.get("subtask") or item.get("task_type") or "unknown"
-            for item in all_items
-        }
-        self._task_types = sorted(task_types)
-
-    def get_task_types(self) -> list[str]:
-        return list(self._task_types)

reflact/envs/mmrb/evaluator.py

@@ -1,102 +0,0 @@
-"""MMRB evaluation helpers."""
-from __future__ import annotations
-
-import re
-import string
-
-
-_EVAL_MODE = "mmrb_exact_match_v1"
-
-
-def normalize_text(text: str) -> str:
-    text = str(text or "").strip().lower()
-    text = "".join(ch for ch in text if ch not in string.punctuation)
-    return " ".join(text.split())
-
-
-def extract_answer(text: str | None) -> str:
-    raw = str(text or "").strip()
-    if not raw:
-        return ""
-
-    answer_tags = re.findall(r"<answer>\s*(.*?)\s*</answer>", raw, re.IGNORECASE | re.DOTALL)
-    if answer_tags:
-        return answer_tags[-1].strip()
-
-    bracket = re.findall(r"Answer\s*\[\s*(.*?)\s*\]", raw, re.IGNORECASE | re.DOTALL)
-    if bracket:
-        return bracket[-1].strip()
-
-    boxed = re.findall(r"\\boxed\{(.*?)\}", raw, re.IGNORECASE | re.DOTALL)
-    if boxed:
-        return boxed[-1].strip()
-
-    single = raw.strip().rstrip(".):")
-    if re.fullmatch(r"[A-Z]", single, re.IGNORECASE):
-        return single.strip()
-
-    patterns = [
-        r"final answer\s*(?:is)?\s*[:：]?\s*(.+)",
-        r"the answer is\s*[:：]?\s*(.+)",
-        r"answer\s*[:：]?\s*(.+)$",
-    ]
-    for pattern in patterns:
-        match = re.search(pattern, raw, re.IGNORECASE)
-        if match:
-            return match.group(1).strip().strip("*")
-
-    return raw
-
-
-def evaluate_item(*, item: dict, prediction_text: str) -> dict:
-    predicted_answer = extract_answer(prediction_text)
-    gold_answer = str(item.get("answer") or "").strip()
-    predicted_norm = normalize_text(predicted_answer)
-    gold_norm = normalize_text(gold_answer)
-
-    hard = 0.0
-    matched_gold = ""
-    predicted_label = ""
-    predicted_text = predicted_answer
-
-    if item.get("is_choice"):
-        predicted_label = str(predicted_answer).strip().upper().rstrip(".):")
-        if predicted_label == str(gold_answer).strip().upper():
-            hard = 1.0
-            matched_gold = gold_answer
-        else:
-            for option in item.get("options") or []:
-                label_match = re.match(r"\(?([A-Z])\)", option)
-                if not label_match:
-                    continue
-                label = label_match.group(1).upper()
-                option_text = option[label_match.end():].strip(" .:-")
-                if predicted_norm and normalize_text(option_text) == predicted_norm:
-                    predicted_label = label
-                    predicted_text = option_text
-                    break
-            if predicted_label == str(gold_answer).strip().upper():
-                hard = 1.0
-                matched_gold = gold_answer
-    else:
-        if predicted_norm and gold_norm and (
-            predicted_norm == gold_norm or predicted_norm in gold_norm or gold_norm in predicted_norm
-        ):
-            hard = 1.0
-            matched_gold = gold_answer
-
-    return {
-        "evaluation_mode": _EVAL_MODE,
-        "predicted_answer": predicted_answer,
-        "predicted_label": predicted_label,
-        "predicted_text": predicted_text,
-        "em": hard,
-        "f1": hard,
-        "sub_em": hard,
-        "matched_gold": matched_gold,
-    }
-
-
-def evaluation_mode() -> str:
-    return _EVAL_MODE
-

reflact/envs/mmrb/prompts/rollout_system.md

@@ -1,10 +0,0 @@
-You are an expert multi-image reasoning agent.
-
-{skill_section}## Task Format
-You will receive a question grounded in multiple images.
-Use the image order exactly as presented in the prompt and compare evidence across images carefully.
-
-## Answer Format
-- Put the final answer inside <answer>...</answer>.
-- For multiple-choice questions, output only the single option letter inside <answer>...</answer>.
-- For open questions, output only the short final answer inside <answer>...</answer>.

reflact/envs/mmrb/rollout.py

@@ -1,439 +0,0 @@
-"""MMRB rollout."""
-from __future__ import annotations
-
-import base64
-import json
-import mimetypes
-import os
-import re
-from concurrent.futures import ThreadPoolExecutor, as_completed
-
-from reflact.envs.mmrb.evaluator import evaluate_item, evaluation_mode
-from reflact.model import chat_student_messages, get_student_backend, is_student_exec_backend
-from reflact.model.codex_harness import prepare_workspace, render_skill_md, run_student_exec
-from reflact.prompts import load_prompt
-
-_IMAGE_REF_RE = re.compile(r"\{image#(\d+)\}", re.IGNORECASE)
-
-
-def _build_system(skill_content: str) -> str:
-    if skill_content.strip():
-        skill_section = f"## Skill\n{skill_content.strip()}\n\n"
-    else:
-        skill_section = ""
-    return load_prompt("rollout_system", env="mmrb").format(skill_section=skill_section)
-
-
-def _image_to_data_uri(path: str) -> str:
-    mime = mimetypes.guess_type(path)[0] or "image/png"
-    with open(path, "rb") as f:
-        encoded = base64.b64encode(f.read()).decode("ascii")
-    return f"data:{mime};base64,{encoded}"
-
-
-def _build_user_content(
-    item: dict,
-    image_detail: str,
-    *,
-    diagnostic_mode: bool = False,
-    diagnostic_instruction: str = "",
-    diagnostic_trace_context: str = "",
-) -> tuple[list[dict], str]:
-    raw_question = str(item["question"])
-    content: list[dict] = []
-    text_parts: list[str] = []
-    used_indices: set[int] = set()
-    cursor = 0
-
-    if diagnostic_trace_context.strip():
-        prefix = (
-            "## Previous Codex Trace Snapshot\n"
-            "This is a partial transcript from an earlier attempt. Use it as your current reasoning context.\n\n"
-            f"{diagnostic_trace_context.strip()}\n\n"
-        )
-        content.append({"type": "text", "text": prefix})
-        text_parts.append(prefix)
-
-    for match in _IMAGE_REF_RE.finditer(raw_question):
-        if match.start() > cursor:
-            chunk = raw_question[cursor:match.start()]
-            if chunk:
-                content.append({"type": "text", "text": chunk})
-                text_parts.append(chunk)
-
-        image_idx = int(match.group(1)) - 1
-        marker = f"[Image #{image_idx + 1}]"
-        text_parts.append(marker)
-        if 0 <= image_idx < len(item["image_paths"]):
-            image_url = {"url": _image_to_data_uri(item["image_paths"][image_idx])}
-            if image_detail and image_detail != "auto":
-                image_url["detail"] = image_detail
-            content.append({"type": "image_url", "image_url": image_url})
-            used_indices.add(image_idx)
-        else:
-            content.append({"type": "text", "text": marker})
-        cursor = match.end()
-
-    if cursor < len(raw_question):
-        tail = raw_question[cursor:]
-        if tail:
-            content.append({"type": "text", "text": tail})
-            text_parts.append(tail)
-
-    for idx, path in enumerate(item["image_paths"]):
-        if idx in used_indices:
-            continue
-        marker = f"\n[Additional Image #{idx + 1}]"
-        text_parts.append(marker)
-        content.append({"type": "text", "text": marker})
-        image_url = {"url": _image_to_data_uri(path)}
-        if image_detail and image_detail != "auto":
-            image_url["detail"] = image_detail
-        content.append({"type": "image_url", "image_url": image_url})
-
-    answer_instruction = (
-        "\n\nAnswer with the single correct option letter inside <answer>...</answer>."
-        if item.get("is_choice")
-        else "\n\nAnswer with the short final answer inside <answer>...</answer>."
-    )
-    content.append({"type": "text", "text": answer_instruction})
-    text_parts.append(answer_instruction)
-
-    if diagnostic_mode and diagnostic_instruction.strip():
-        diag_block = f"\n\n## Training Readout\n{diagnostic_instruction.strip()}"
-        content.append({"type": "text", "text": diag_block})
-        text_parts.append(diag_block)
-
-    return content, "".join(text_parts)
-
-
-def _build_messages(
-    item: dict,
-    skill_content: str,
-    image_detail: str,
-    *,
-    diagnostic_mode: bool = False,
-    diagnostic_instruction: str = "",
-) -> tuple[list[dict], str, str]:
-    system = _build_system(skill_content)
-    user_content, user_text = _build_user_content(
-        item,
-        image_detail,
-        diagnostic_mode=diagnostic_mode,
-        diagnostic_instruction=diagnostic_instruction,
-    )
-    messages = [
-        {"role": "system", "content": system},
-        {"role": "user", "content": user_content},
-    ]
-    return messages, system, user_text
-
-
-def _build_codex_skill(skill_content: str) -> str:
-    return render_skill_md(
-        skill_content,
-        description="Dynamic ReflACT skill for solving the current MMRB multi-image reasoning question.",
-        preamble=(
-            "Use this skill when solving the current multi-image reasoning task.\n"
-            "Inspect all attached images carefully and return the final answer inside <answer>...</answer>."
-        ),
-    )
-
-
-def _run_codex_once(
-    *,
-    pred_dir: str,
-    item: dict,
-    skill_content: str,
-    model: str,
-    timeout: int,
-    image_detail: str,
-    diagnostic_mode: bool = False,
-    diagnostic_instruction: str = "",
-    diagnostic_trace_context: str = "",
-    previous_response: str = "",
-) -> tuple[str, str, str, str]:
-    user_text = _build_user_content(
-        item,
-        image_detail,
-        diagnostic_mode=diagnostic_mode,
-        diagnostic_instruction=diagnostic_instruction,
-        diagnostic_trace_context=diagnostic_trace_context,
-    )[1]
-    task_parts = [user_text]
-    if previous_response:
-        task_parts.append(
-            "## Previous Attempt\n"
-            f"{previous_response}\n\n"
-            "Review the same images carefully and answer again."
-        )
-    task_text = "\n\n".join(task_parts)
-    skill_md = _build_codex_skill(skill_content)
-    work_dir = os.path.join(pred_dir, "codex_exec")
-    prepare_workspace(
-        work_dir=work_dir,
-        skill_md=skill_md,
-        task_text=task_text,
-        images=item["image_paths"],
-    )
-    prompt = (
-        "Use the `reflact-student` skill available in this workspace.\n"
-        "Read `task.md`, inspect all attached images, and answer the question.\n"
-        "Keep the final answer inside <answer>...</answer>."
-    )
-    final_message, raw = run_student_exec(
-        work_dir=work_dir,
-        prompt=prompt,
-        model=model,
-        timeout=timeout,
-        images=item["image_paths"],
-    )
-    return final_message or raw, raw, skill_md, task_text
-
-
-def process_one(
-    item: dict,
-    out_root: str,
-    skill_content: str,
-    *,
-    max_turns: int = 1,
-    image_detail: str = "auto",
-    diagnostic_mode: bool = False,
-    diagnostic_instruction: str = "",
-    diagnostic_trace_context: str = "",
-) -> dict:
-    item_id = str(item["id"])
-    result = {
-        "id": item_id,
-        "question": item["question"],
-        "task_type": item.get("subtask") or item.get("task_type") or "mmrb",
-        "task_description": item["question"],
-        "hard": 0,
-        "soft": 0.0,
-        "predicted_answer": "",
-        "predicted_label": "",
-        "predicted_text": "",
-        "response": "",
-        "fail_reason": "",
-        "agent_ok": False,
-        "n_turns": 0,
-        "image_paths": item["image_paths"],
-        "gold_answer": item["answer"],
-        "evaluation_mode": evaluation_mode(),
-    }
-
-    try:
-        pred_dir = os.path.join(out_root, "predictions", item_id)
-        os.makedirs(pred_dir, exist_ok=True)
-
-        if is_student_exec_backend():
-            from reflact.model import azure_openai as _llm
-
-            response = ""
-            conversation: list[dict] = [
-                {
-                    "role": "user",
-                    "content": item["question"] + "\n\n" + "\n".join(
-                        f"[image] {os.path.basename(path)}" for path in item["image_paths"]
-                    ),
-                }
-            ]
-            system_prompt = ""
-            user_text = ""
-            for turn in range(max_turns):
-                response, raw, system_prompt, user_text = _run_codex_once(
-                    pred_dir=pred_dir,
-                    item=item,
-                    skill_content=skill_content,
-                    model=_llm.STUDENT_DEPLOYMENT,
-                    timeout=120,
-                    image_detail=image_detail,
-                    diagnostic_mode=diagnostic_mode if turn == 0 else False,
-                    diagnostic_instruction=diagnostic_instruction if turn == 0 else "",
-                    diagnostic_trace_context=diagnostic_trace_context if turn == 0 else "",
-                    previous_response=response if turn > 0 else "",
-                )
-                conversation.append({"type": "message", "turn": turn + 1, "content": response})
-                if "<answer>" in response.lower():
-                    break
-
-            result["response"] = response
-            result["agent_ok"] = True
-            result["n_turns"] = len(conversation) - 1
-            with open(os.path.join(pred_dir, "student_system_prompt.txt"), "w", encoding="utf-8") as f:
-                f.write(system_prompt)
-            with open(os.path.join(pred_dir, "student_user_prompt.txt"), "w", encoding="utf-8") as f:
-                f.write(user_text)
-
-            eval_result = evaluate_item(item=item, prediction_text=response)
-            result["evaluation_mode"] = eval_result["evaluation_mode"]
-            result["predicted_answer"] = eval_result["predicted_answer"]
-            result["predicted_label"] = eval_result["predicted_label"]
-            result["predicted_text"] = eval_result["predicted_text"]
-            result["matched_gold"] = eval_result["matched_gold"]
-            result["hard"] = int(eval_result["em"])
-            result["soft"] = eval_result["f1"]
-            if not result["hard"]:
-                result["fail_reason"] = (
-                    f"predicted '{eval_result['predicted_answer']}' but expected '{item['answer']}'"
-                )
-            eval_detail = (
-                "[EVALUATION RESULT]\n"
-                f"Question: {item['question']}\n"
-                f"Predicted answer: {eval_result['predicted_answer']!r}\n"
-                f"Predicted label: {eval_result['predicted_label']!r}\n"
-                f"Gold answer: {item['answer']!r}\n"
-                f"Correct: {eval_result['em']}\n"
-            )
-            conversation.append({"role": "system", "content": eval_detail})
-            with open(os.path.join(pred_dir, "conversation.json"), "w", encoding="utf-8") as f:
-                json.dump(conversation, f, ensure_ascii=False, indent=2)
-            return result
-
-        messages, system_prompt, user_text = _build_messages(
-            item,
-            skill_content,
-            image_detail,
-            diagnostic_mode=diagnostic_mode,
-            diagnostic_instruction=diagnostic_instruction,
-            diagnostic_trace_context=diagnostic_trace_context,
-        )
-        response = ""
-        conversation: list[dict] = [
-            {
-                "role": "user",
-                "content": user_text + "\n\n" + "\n".join(
-                    f"[image] {os.path.basename(path)}" for path in item["image_paths"]
-                ),
-            }
-        ]
-
-        for turn in range(max_turns):
-            if turn == 0:
-                resp_text, _ = chat_student_messages(
-                    messages=messages,
-                    max_completion_tokens=768,
-                    retries=5,
-                    stage="rollout",
-                )
-            else:
-                refinement_messages = [
-                    messages[0],
-                    messages[1],
-                    {"role": "assistant", "content": response},
-                    {
-                        "role": "user",
-                        "content": "Review the same images carefully and answer again. Keep the final answer inside <answer>...</answer>.",
-                    },
-                ]
-                resp_text, _ = chat_student_messages(
-                    messages=refinement_messages,
-                    max_completion_tokens=512,
-                    retries=5,
-                    stage="rollout",
-                )
-            response = resp_text
-            conversation.append({"type": "message", "turn": turn + 1, "content": resp_text})
-            if "<answer>" in resp_text.lower():
-                break
-
-        result["response"] = response
-        result["agent_ok"] = True
-        result["n_turns"] = len(conversation) - 1
-
-        with open(os.path.join(pred_dir, "student_system_prompt.txt"), "w", encoding="utf-8") as f:
-            f.write(system_prompt)
-        with open(os.path.join(pred_dir, "student_user_prompt.txt"), "w", encoding="utf-8") as f:
-            f.write(user_text)
-
-        eval_result = evaluate_item(item=item, prediction_text=response)
-        result["evaluation_mode"] = eval_result["evaluation_mode"]
-        result["predicted_answer"] = eval_result["predicted_answer"]
-        result["predicted_label"] = eval_result["predicted_label"]
-        result["predicted_text"] = eval_result["predicted_text"]
-        result["matched_gold"] = eval_result["matched_gold"]
-        result["hard"] = int(eval_result["em"])
-        result["soft"] = eval_result["f1"]
-        if not result["hard"]:
-            result["fail_reason"] = (
-                f"predicted '{eval_result['predicted_answer']}' but expected '{item['answer']}'"
-            )
-
-        eval_detail = (
-            "[EVALUATION RESULT]\n"
-            f"Question: {item['question']}\n"
-            f"Predicted answer: {eval_result['predicted_answer']!r}\n"
-            f"Predicted label: {eval_result['predicted_label']!r}\n"
-            f"Gold answer: {item['answer']!r}\n"
-            f"Correct: {eval_result['em']}\n"
-        )
-        conversation.append({"role": "system", "content": eval_detail})
-        with open(os.path.join(pred_dir, "conversation.json"), "w", encoding="utf-8") as f:
-            json.dump(conversation, f, ensure_ascii=False, indent=2)
-    except Exception as e:  # noqa: BLE001
-        result["fail_reason"] = f"error: {e}"
-    return result
-
-
-def run_batch(
-    items: list[dict],
-    out_root: str,
-    skill_content: str,
-    *,
-    max_turns: int = 1,
-    workers: int = 16,
-    image_detail: str = "auto",
-    diagnostic_mode: bool = False,
-    diagnostic_instruction: str = "",
-    diagnostic_trace_context_by_id: dict[str, str] | None = None,
-) -> list[dict]:
-    results_path = os.path.join(out_root, "results.jsonl")
-    os.makedirs(out_root, exist_ok=True)
-
-    expected_eval_mode = evaluation_mode()
-    done_ids: set[str] = set()
-    existing: list[dict] = []
-    rewrite_results = False
-    if os.path.exists(results_path):
-        with open(results_path, encoding="utf-8") as f:
-            for line in f:
-                try:
-                    row = json.loads(line)
-                    if row.get("evaluation_mode") != expected_eval_mode:
-                        rewrite_results = True
-                        continue
-                    done_ids.add(str(row["id"]))
-                    existing.append(row)
-                except Exception:
-                    rewrite_results = True
-
-    pending = [item for item in items if str(item["id"]) not in done_ids]
-    if not pending and not rewrite_results:
-        return existing
-
-    results = list(existing)
-    file_mode = "w" if rewrite_results else "a"
-    with open(results_path, file_mode, encoding="utf-8") as outf, ThreadPoolExecutor(max_workers=workers) as ex:
-        if rewrite_results:
-            for row in existing:
-                outf.write(json.dumps(row, ensure_ascii=False) + "\n")
-        futs = {
-            ex.submit(
-                process_one,
-                item,
-                out_root,
-                skill_content,
-                max_turns=max_turns,
-                image_detail=image_detail,
-                diagnostic_mode=diagnostic_mode,
-                diagnostic_instruction=diagnostic_instruction,
-                diagnostic_trace_context=(diagnostic_trace_context_by_id or {}).get(str(item["id"]), ""),
-            ): item
-            for item in pending
-        }
-        for fut in as_completed(futs):
-            row = fut.result()
-            results.append(row)
-            outf.write(json.dumps(row, ensure_ascii=False) + "\n")
-            outf.flush()
-    return results

reflact/envs/mmrb/skills/initial.md

@@ -1,17 +0,0 @@
-# MMRB Multi-Image Reasoning Heuristics
-
-## Cross-Image Alignment
-- Track the role of each image by its index and compare evidence across all referenced images before deciding.
-- When the question depends on sequence, correspondence, or retrieval, verify the relation between images instead of judging each image independently.
-
-## Option Elimination
-- For multiple-choice tasks, compare all options and reject choices that match only part of the visual evidence.
-- If options differ by a small visual detail, use the most discriminative cue rather than a coarse scene impression.
-
-## Open Answers
-- For open-ended tasks, give the shortest answer that is fully supported by the combined images.
-- Preserve exact entities, attributes, counts, and directions when the images support them directly.
-
-## Final Answer
-- Output only the final answer inside <answer>...</answer>.
-

reflact/envs/officeqa/rollout.py

@@ -1,363 +0,0 @@
-from __future__ import annotations
-
-import json
-import os
-import re
-from concurrent.futures import ThreadPoolExecutor, as_completed
-
-from reflact.envs.officeqa.evaluator import evaluate
-from reflact.envs.officeqa.tool_runtime import resolve_candidate_files, resolve_docs_roots, run_tool
-from reflact.model import chat_student_messages, get_student_backend, is_student_exec_backend
-from reflact.model.codex_harness import prepare_workspace, render_skill_md, run_student_exec
-from reflact.prompts import load_prompt
-
-_TOOL_SCHEMAS = [
-    {
-        "type": "function",
-        "function": {
-            "name": "glob",
-            "description": "Find candidate local document files by filename or relative-path glob pattern.",
-            "parameters": {
-                "type": "object",
-                "properties": {"pattern": {"type": "string"}},
-                "required": ["pattern"],
-            },
-        },
-    },
-    {
-        "type": "function",
-        "function": {
-            "name": "read",
-            "description": "Read a local text document excerpt by path and line window.",
-            "parameters": {
-                "type": "object",
-                "properties": {
-                    "path": {"type": "string"},
-                    "start": {"type": "integer"},
-                    "limit": {"type": "integer"},
-                },
-                "required": ["path"],
-            },
-        },
-    },
-    {
-        "type": "function",
-        "function": {
-            "name": "grep",
-            "description": "Search a local text document for a literal pattern and return matching lines.",
-            "parameters": {
-                "type": "object",
-                "properties": {
-                    "pattern": {"type": "string"},
-                    "path": {"type": "string"},
-                },
-                "required": ["pattern", "path"],
-            },
-        },
-    },
-]
-
-_FINAL_RE = re.compile(r"<answer>(.*?)</answer>", re.IGNORECASE | re.DOTALL)
-
-
-def _build_system(skill_content: str) -> str:
-    if skill_content.strip():
-        skill_section = f"## Skill\n{skill_content.strip()}\n\n"
-    else:
-        skill_section = ""
-    return load_prompt("rollout_system", env="officeqa").format(skill_section=skill_section)
-
-
-def _build_user(
-    item: dict,
-    candidate_files: list[str],
-    *,
-    diagnostic_mode: bool = False,
-    diagnostic_instruction: str = "",
-    corpus_note: str = "",
-) -> str:
-    file_block = "\n".join(f"- {path}" for path in candidate_files[:20]) or "- none resolved"
-    parts = [f"## Question\n{item['question']}"]
-    if corpus_note.strip():
-        parts.append(f"## Document Corpus\n{corpus_note.strip()}")
-    parts.append(f"## Candidate Files\n{file_block}")
-    if item.get("source_docs"):
-        parts.append("## Source Hints\n" + "\n".join(f"- {hint}" for hint in item["source_docs"]))
-    if diagnostic_mode and diagnostic_instruction.strip():
-        parts.append(f"## Training Readout\n{diagnostic_instruction.strip()}")
-    return "\n\n".join(parts)
-
-
-def _extract_answer(text: str) -> str:
-    match = _FINAL_RE.search(text)
-    if match:
-        return match.group(1).strip()
-    lines = [line.strip() for line in text.splitlines() if line.strip()]
-    return lines[-1] if lines else text.strip()
-
-
-def _docs_link_targets(docs_roots: list[str]) -> list[tuple[str, str]]:
-    return [(root, os.path.join("docs", f"root_{idx}")) for idx, root in enumerate(docs_roots, start=1)]
-
-
-def _workspace_doc_path(path: str, docs_roots: list[str]) -> str:
-    resolved_path = os.path.realpath(path)
-    for idx, root in enumerate(docs_roots, start=1):
-        resolved_root = os.path.realpath(root)
-        if resolved_path == resolved_root or resolved_path.startswith(resolved_root + os.sep):
-            rel_path = os.path.relpath(resolved_path, resolved_root)
-            return os.path.join("docs", f"root_{idx}", rel_path)
-    return path
-
-
-def _build_codex_skill(skill_content: str) -> str:
-    return render_skill_md(
-        skill_content,
-        description="Dynamic ReflACT skill for solving the current OfficeQA local-document question.",
-        preamble=(
-            "Use this skill when answering the current OfficeQA question.\n"
-            "Inspect the provided local document excerpts or files, ground the answer in the evidence,\n"
-            "and return the final answer inside <answer>...</answer>."
-        ),
-    )
-
-
-def _run_codex_once(
-    *,
-    pred_dir: str,
-    item: dict,
-    skill_content: str,
-    candidate_files: list[str],
-    docs_roots: list[str],
-    model: str,
-    timeout: int,
-    diagnostic_mode: bool = False,
-    diagnostic_instruction: str = "",
-    previous_response: str = "",
-) -> tuple[str, str, str, str]:
-    rel_files = [_workspace_doc_path(path, docs_roots) for path in candidate_files[:20]]
-    corpus_note = (
-        "The full OfficeQA document corpus is available under `docs/`. "
-        "The candidate files below are source hints or likely starting points; search the full corpus if needed."
-    )
-    user = _build_user(
-        item,
-        rel_files,
-        diagnostic_mode=diagnostic_mode,
-        diagnostic_instruction=diagnostic_instruction,
-        corpus_note=corpus_note,
-    )
-    task_parts = [user]
-    if previous_response:
-        task_parts.append(
-            "## Previous Attempt\n"
-            f"{previous_response}\n\n"
-            "Review the local documents again and correct the answer if needed."
-        )
-    task_text = "\n\n".join(task_parts)
-    skill_md = _build_codex_skill(skill_content)
-    work_dir = os.path.join(pred_dir, "codex_exec")
-    prepare_workspace(
-        work_dir=work_dir,
-        skill_md=skill_md,
-        task_text=task_text,
-        link_dirs=_docs_link_targets(docs_roots),
-    )
-    prompt = (
-        "Use the `reflact-student` skill available in this workspace.\n"
-        "Read `task.md`, inspect or search the full OfficeQA corpus under `docs/`, and answer the question.\n"
-        "Treat candidate files in `task.md` as hints, not an access limit.\n"
-        "Return the final answer inside <answer>...</answer>."
-    )
-    final_message, raw = run_student_exec(
-        work_dir=work_dir,
-        prompt=prompt,
-        model=model,
-        timeout=timeout,
-        data_dirs=docs_roots,
-    )
-    return final_message or raw, raw, skill_md, task_text
-
-
-def process_one(
-    item: dict,
-    out_root: str,
-    skill_content: str,
-    *,
-    max_tool_turns: int = 12,
-    data_dirs: list[str] | str | None = None,
-    diagnostic_mode: bool = False,
-    diagnostic_instruction: str = "",
-) -> dict:
-    item_id = str(item["id"])
-    pred_dir = os.path.join(out_root, "predictions", item_id)
-    os.makedirs(pred_dir, exist_ok=True)
-
-    docs_roots = resolve_docs_roots(data_dirs)
-    candidate_files = resolve_candidate_files(item.get("source_files", []), docs_roots)
-    system = _build_system(skill_content)
-    user = _build_user(item, candidate_files, diagnostic_mode=diagnostic_mode, diagnostic_instruction=diagnostic_instruction)
-
-    messages: list[dict] = [
-        {"role": "system", "content": system},
-        {"role": "user", "content": user},
-    ]
-    conversation: list[dict] = [{"role": "user", "content": user}]
-    final_response = ""
-    final_answer = ""
-    fail_reason = ""
-
-    allowed_files = [os.path.basename(path) for path in candidate_files]
-
-    try:
-        if is_student_exec_backend():
-            from reflact.model import azure_openai as _llm
-
-            response = ""
-            system = ""
-            user = ""
-            for turn in range(1, max_tool_turns + 1):
-                response, _raw, system, user = _run_codex_once(
-                    pred_dir=pred_dir,
-                    item=item,
-                    skill_content=skill_content,
-                    candidate_files=candidate_files,
-                    docs_roots=docs_roots,
-                    model=_llm.STUDENT_DEPLOYMENT,
-                    timeout=180,
-                    diagnostic_mode=diagnostic_mode if turn == 1 else False,
-                    diagnostic_instruction=diagnostic_instruction if turn == 1 else "",
-                    previous_response=response if turn > 1 else "",
-                )
-                final_response = response
-                conversation.append({"type": "message", "turn": turn, "content": response})
-                if "<answer>" in response.lower():
-                    final_answer = _extract_answer(response)
-                    break
-            if not final_answer:
-                fail_reason = f"Exceeded codex turn budget ({max_tool_turns})"
-            system = system or _build_codex_skill(skill_content)
-            user = user or _build_user(item, [_workspace_doc_path(path, docs_roots) for path in candidate_files])
-        else:
-            for turn in range(1, max_tool_turns + 1):
-                message, _ = chat_student_messages(
-                    messages=messages,
-                    max_completion_tokens=768,
-                    retries=5,
-                    stage="rollout",
-                    tools=_TOOL_SCHEMAS,
-                    tool_choice="auto",
-                    return_message=True,
-                )
-                response = message.content or ""
-                final_response = response
-                assistant_message = {"role": "assistant", "content": response}
-                if getattr(message, "tool_calls", None):
-                    assistant_message["tool_calls"] = [tool_call.model_dump(mode="json") for tool_call in message.tool_calls]
-                messages.append(assistant_message)
-                conversation.append({"type": "message", "content": response})
-
-                if getattr(message, "tool_calls", None):
-                    for tool_call in message.tool_calls:
-                        tool_name = tool_call.function.name
-                        arguments = json.loads(tool_call.function.arguments) if tool_call.function.arguments else {}
-                        cmd, obs = run_tool(tool_name, arguments, allowed_roots=docs_roots, allowed_files=allowed_files)
-                        conversation.append({"type": "tool_call", "cmd": cmd, "obs": obs})
-                        messages.append({
-                            "role": "tool",
-                            "tool_call_id": tool_call.id,
-                            "content": obs,
-                        })
-                    continue
-
-                if "<answer>" in response.lower():
-                    final_answer = _extract_answer(response)
-                    break
-                if turn == max_tool_turns:
-                    fail_reason = f"Exceeded tool-turn budget ({max_tool_turns})"
-                else:
-                    fail_reason = "Model neither produced a tool request nor a final answer"
-                    break
-    except Exception as e:  # noqa: BLE001
-        fail_reason = f"error: {e}"
-
-    with open(os.path.join(pred_dir, "student_system_prompt.txt"), "w", encoding="utf-8") as f:
-        f.write(system)
-    with open(os.path.join(pred_dir, "student_user_prompt.txt"), "w", encoding="utf-8") as f:
-        f.write(user)
-    with open(os.path.join(pred_dir, "conversation.json"), "w", encoding="utf-8") as f:
-        json.dump(conversation, f, ensure_ascii=False, indent=2)
-
-    eval_result = evaluate(final_answer, item.get("ground_truth", "")) if final_answer else {"em": 0.0, "f1": 0.0, "predicted_answer": "", "gold_answer": item.get("ground_truth", "")}
-    result = {
-        "id": item_id,
-        "question": item.get("question", ""),
-        "task_type": item.get("task_type", "officeqa"),
-        "task_description": item.get("question", ""),
-        "predicted_answer": eval_result["predicted_answer"],
-        "response": final_response,
-        "ground_truth": item.get("ground_truth", ""),
-        "source_files": item.get("source_files", []),
-        "resolved_source_paths": candidate_files,
-        "hard": int(eval_result["em"]),
-        "soft": eval_result["f1"],
-        "fail_reason": fail_reason or ("" if eval_result["em"] else f"predicted '{eval_result['predicted_answer']}' but expected '{item.get('ground_truth', '')}'"),
-        "agent_ok": not fail_reason,
-        "n_turns": len(conversation),
-        "student_system_prompt": system,
-        "student_user_prompt": user,
-    }
-    return result
-
-
-def run_batch(
-    items: list[dict],
-    out_root: str,
-    skill_content: str,
-    *,
-    workers: int = 8,
-    max_tool_turns: int = 12,
-    data_dirs: list[str] | str | None = None,
-    diagnostic_mode: bool = False,
-    diagnostic_instruction: str = "",
-) -> list[dict]:
-    results_path = os.path.join(out_root, "results.jsonl")
-    os.makedirs(out_root, exist_ok=True)
-
-    done_ids: set[str] = set()
-    existing: list[dict] = []
-    if os.path.exists(results_path):
-        with open(results_path, encoding="utf-8") as f:
-            for line in f:
-                try:
-                    row = json.loads(line)
-                except json.JSONDecodeError:
-                    continue
-                done_ids.add(str(row.get("id")))
-                existing.append(row)
-
-    pending = [item for item in items if str(item["id"]) not in done_ids]
-    if not pending:
-        return existing
-
-    results = list(existing)
-    with open(results_path, "a", encoding="utf-8") as outf, ThreadPoolExecutor(max_workers=workers) as ex:
-        futs = {
-            ex.submit(
-                process_one,
-                item,
-                out_root,
-                skill_content,
-                max_tool_turns=max_tool_turns,
-                data_dirs=data_dirs,
-                diagnostic_mode=diagnostic_mode,
-                diagnostic_instruction=diagnostic_instruction,
-            ): item
-            for item in pending
-        }
-        for fut in as_completed(futs):
-            res = fut.result()
-            results.append(res)
-            outf.write(json.dumps(res, ensure_ascii=False) + "\n")
-            outf.flush()
-    return results

reflact/envs/officeqa/tool_runtime.py

@@ -1,134 +0,0 @@
-from __future__ import annotations
-
-import fnmatch
-import os
-from pathlib import Path
-
-_MAX_READ_CHARS = 4000
-_MAX_GREP_MATCHES = 20
-_MAX_GLOB_MATCHES = 50
-
-
-def _normalize_data_dirs(data_dirs: list[str] | tuple[str, ...] | str | None, project_root: Path) -> list[str]:
-    if data_dirs is None:
-        return []
-    if isinstance(data_dirs, str):
-        items = [part.strip() for chunk in data_dirs.split(os.pathsep) for part in chunk.split(",")]
-    else:
-        items = [str(item).strip() for item in data_dirs]
-    resolved: list[str] = []
-    for item in items:
-        if not item:
-            continue
-        path = Path(item).expanduser()
-        if not path.is_absolute():
-            path = project_root / path
-        resolved.append(str(path))
-    return resolved
-
-
-def resolve_docs_roots(data_dirs: list[str] | tuple[str, ...] | str | None = None) -> list[str]:
-    project_root = Path(__file__).resolve().parents[3]
-    env_value = os.environ.get("OFFICEQA_DOCS_DIR", "").strip()
-    candidates = _normalize_data_dirs(data_dirs, project_root)
-    candidates.extend(_normalize_data_dirs(env_value, project_root))
-    candidates.extend([
-        str(project_root / "data" / "officeqa_docs_official"),
-        str(project_root / "data" / "officeqa_smoke_docs"),
-        os.path.expanduser("~/officeqa-sparse/treasury_bulletins_parsed"),
-        os.path.expanduser("~/officeqa/treasury_bulletins_parsed"),
-    ])
-    roots: list[str] = []
-    seen: set[str] = set()
-    for candidate in candidates:
-        path = Path(candidate).expanduser()
-        if not path.is_dir():
-            continue
-        transformed = path / "transformed"
-        resolved = str((transformed if transformed.is_dir() else path).resolve())
-        if resolved in seen:
-            continue
-        seen.add(resolved)
-        roots.append(resolved)
-    if not roots:
-        raise FileNotFoundError("OfficeQA docs directory not found. Set OFFICEQA_DOCS_DIR or env.data_dirs.")
-    return roots
-
-
-def _is_allowed(path: str, allowed_roots: list[str], allowed_files: list[str]) -> bool:
-    try:
-        resolved = str(Path(path).resolve())
-    except FileNotFoundError:
-        return False
-    if not any(resolved.startswith(root + os.sep) or resolved == root for root in allowed_roots):
-        return False
-    if not allowed_files:
-        return True
-    base = os.path.basename(resolved)
-    return base in allowed_files
-
-
-def resolve_candidate_files(source_files: list[str], allowed_roots: list[str]) -> list[str]:
-    resolved: list[str] = []
-    seen: set[str] = set()
-    for root in allowed_roots:
-        for dirpath, _, filenames in os.walk(root):
-            for filename in filenames:
-                if source_files and filename not in source_files:
-                    continue
-                full = str(Path(dirpath, filename).resolve())
-                if full in seen:
-                    continue
-                seen.add(full)
-                resolved.append(full)
-    return resolved
-
-
-def run_tool(name: str, arguments: dict, *, allowed_roots: list[str], allowed_files: list[str]) -> tuple[str, str]:
-    if name == "glob":
-        pattern = str(arguments.get("pattern") or "*")
-        matches: list[str] = []
-        for root in allowed_roots:
-            for dirpath, _, filenames in os.walk(root):
-                for filename in filenames:
-                    if allowed_files and filename not in allowed_files:
-                        continue
-                    rel = os.path.relpath(os.path.join(dirpath, filename), root)
-                    if fnmatch.fnmatch(rel, pattern) or fnmatch.fnmatch(filename, pattern):
-                        matches.append(os.path.join(dirpath, filename))
-                    if len(matches) >= _MAX_GLOB_MATCHES:
-                        break
-                if len(matches) >= _MAX_GLOB_MATCHES:
-                    break
-        return f"glob(pattern={pattern!r})", "\n".join(matches) if matches else "[no matches]"
-
-    if name == "read":
-        path = str(arguments.get("path") or "")
-        if not path:
-            return "read(path='')", "[read error: missing path]"
-        if not _is_allowed(path, allowed_roots, allowed_files):
-            return f"read(path={path!r})", "[read error: path not allowed]"
-        start = max(int(arguments.get("start") or 1), 1)
-        limit = max(int(arguments.get("limit") or 80), 1)
-        with open(path, encoding="utf-8") as f:
-            lines = f.readlines()
-        excerpt = "".join(lines[start - 1:start - 1 + limit])
-        return f"read(path={path!r}, start={start}, limit={limit})", excerpt[:_MAX_READ_CHARS] or "[empty file]"
-
-    if name == "grep":
-        pattern = str(arguments.get("pattern") or "").lower()
-        path = str(arguments.get("path") or "")
-        if not pattern or not path:
-            return f"grep(pattern={pattern!r}, path={path!r})", "[grep error: missing pattern or path]"
-        if not _is_allowed(path, allowed_roots, allowed_files):
-            return f"grep(pattern={pattern!r}, path={path!r})", "[grep error: path not allowed]"
-        matches: list[str] = []
-        with open(path, encoding="utf-8") as f:
-            for idx, line in enumerate(f, start=1):
-                if pattern in line.lower():
-                    matches.append(f"{idx}: {line.rstrip()}")
-                if len(matches) >= _MAX_GREP_MATCHES:
-                    break
-        return f"grep(pattern={pattern!r}, path={path!r})", "\n".join(matches) if matches else "[no matches]"
-
-    return name, f"[tool error: unknown tool {name}]"

reflact/envs/sealqa/__init__.py

@@ -1 +0,0 @@
-"""SealQA environment package for ReflACT."""

reflact/envs/sealqa/adapter.py

@@ -1,130 +0,0 @@
-from __future__ import annotations
-
-import os
-
-from reflact.datasets.base import BatchSpec
-from reflact.envs.base import EnvAdapter
-from reflact.envs.deep_reflect import run_no_reference_deep_reflect
-from reflact.envs.sealqa.dataloader import SealQADataLoader
-from reflact.envs.sealqa.rollout import run_batch
-from reflact.gradient.reflect import run_minibatch_reflect
-
-
-class SealQAAdapter(EnvAdapter):
-    def __init__(
-        self,
-        split_dir: str = '',
-        workers: int = 4,
-        analyst_workers: int = 8,
-        failure_only: bool = False,
-        minibatch_size: int = 8,
-        edit_budget: int = 4,
-        seed: int = 42,
-        limit: int = 0,
-        max_tool_turns: int = 12,
-        use_deep_reflect: bool = False,
-        deep_reflect_failures: int = 4,
-        deep_reflect_successes: int = 2,
-    ) -> None:
-        self.workers = workers
-        self.analyst_workers = analyst_workers
-        self.failure_only = failure_only
-        self.minibatch_size = minibatch_size
-        self.edit_budget = edit_budget
-        self.max_tool_turns = max_tool_turns
-        self.use_deep_reflect = use_deep_reflect
-        self.deep_reflect_failures = deep_reflect_failures
-        self.deep_reflect_successes = deep_reflect_successes
-        self.dataloader = SealQADataLoader(split_dir=split_dir, seed=seed, limit=limit)
-
-    def setup(self, cfg: dict) -> None:
-        super().setup(cfg)
-        self.dataloader.setup(cfg)
-
-    def get_dataloader(self):
-        return self.dataloader
-
-    def build_env_from_batch(self, batch: BatchSpec, **kwargs):
-        return list(batch.payload or [])
-
-    def build_train_env(self, batch_size: int, seed: int, **kwargs):
-        batch = self.dataloader.build_train_batch(batch_size=batch_size, seed=seed, **kwargs)
-        return self.build_env_from_batch(batch, **kwargs)
-
-    def build_eval_env(self, env_num: int, split: str, seed: int, **kwargs):
-        batch = self.dataloader.build_eval_batch(env_num=env_num, split=split, seed=seed, **kwargs)
-        return self.build_env_from_batch(batch, **kwargs)
-
-    def rollout(self, env_manager, skill_content: str, out_dir: str, **kwargs) -> list[dict]:
-        items: list[dict] = env_manager
-        return run_batch(
-            items=items,
-            out_root=out_dir,
-            skill_content=skill_content,
-            workers=self.workers,
-            max_tool_turns=self.max_tool_turns,
-            diagnostic_mode=kwargs.get('diagnostic_mode', False),
-            diagnostic_instruction=kwargs.get('diagnostic_instruction', ''),
-        )
-
-    def reflect(self, results: list[dict], skill_content: str, out_dir: str, **kwargs) -> list[dict | None]:
-        prediction_dir = kwargs.get('prediction_dir', os.path.join(out_dir, 'predictions'))
-        patches_dir = kwargs.get('patches_dir', os.path.join(out_dir, 'patches'))
-        random_seed = kwargs.get('random_seed')
-        step_buffer_context = kwargs.get('step_buffer_context', '')
-        return run_minibatch_reflect(
-            results=results,
-            skill_content=skill_content,
-            prediction_dir=prediction_dir,
-            patches_dir=patches_dir,
-            workers=self.analyst_workers,
-            failure_only=self.failure_only,
-            minibatch_size=self.minibatch_size,
-            edit_budget=self.edit_budget,
-            random_seed=random_seed,
-            error_system=self.get_error_minibatch_prompt(),
-            success_system=self.get_success_minibatch_prompt(),
-            step_buffer_context=step_buffer_context,
-            update_mode=getattr(self, "_cfg", {}).get("skill_update_mode", "patch"),
-        )
-
-    def deep_reflect(
-        self,
-        results: list[dict],
-        skill_content: str,
-        out_dir: str,
-        **kwargs,
-    ) -> list[dict | None]:
-        return run_no_reference_deep_reflect(
-            self,
-            results,
-            skill_content,
-            out_dir,
-            env_manager=kwargs.get('env_manager'),
-            prediction_dir=kwargs.get('prediction_dir'),
-            random_seed=kwargs.get('random_seed'),
-            step_buffer_context=kwargs.get('step_buffer_context', ''),
-            output_requirements=[
-                "- There is no hidden reference block. Use only the question, provided evidence, URL/fetch trace, student output, and evaluation result to infer what intermediate state is worth probing.",
-                "- The instruction must explicitly request a short <analysis>...</analysis> block before the final <answer>...</answer>.",
-                "- The readout should focus on effective time frame, conflicting evidence, decisive source, candidate answer, and answer-finalization rule.",
-                "- Do not ask for exhaustive web summaries or a full chain-of-thought.",
-                "- The instruction text should be ready to append directly to the student's prompt.",
-            ],
-            metadata_builder=lambda item: {
-                "id": str(item.get('id')),
-                "task_type": str(item.get('task_type') or item.get('topic') or 'sealqa'),
-                "question_preview": str(item.get('question') or '')[:200],
-                "freshness": item.get('freshness', ''),
-                "question_types": item.get('question_types', ''),
-                "topic": item.get('topic', ''),
-            },
-        )
-
-    def get_task_types(self) -> list[str]:
-        seen: list[str] = []
-        for item in self.dataloader.train_items + self.dataloader.val_items + self.dataloader.test_items:
-            task_type = str(item.get('task_type') or 'sealqa')
-            if task_type not in seen:
-                seen.append(task_type)
-        return seen or ['sealqa']

reflact/envs/sealqa/dataloader.py

@@ -1,37 +0,0 @@
-from __future__ import annotations
-
-import csv
-from pathlib import Path
-
-from reflact.datasets.base import SplitDataLoader
-
-
-def _normalize_row(row: dict[str, str], index: int) -> dict:
-    canary = str(row.get('canary') or '').strip()
-    base_id = str(row.get('question_id') or row.get('id') or '').strip()
-    if not base_id:
-        base_id = f"{canary or 'sealqa'}:{index:04d}"
-    return {
-        'id': base_id,
-        'question': str(row.get('question') or '').strip(),
-        'ground_truth': str(row.get('answer') or row.get('ground_truth') or '').strip(),
-        'answers': [str(row.get('answer') or row.get('ground_truth') or '').strip()],
-        'task_type': str(row.get('topic') or 'sealqa').strip() or 'sealqa',
-        'topic': str(row.get('topic') or 'sealqa').strip() or 'sealqa',
-        'urls': str(row.get('urls') or '').strip(),
-        'search_results': str(row.get('search_results') or '').strip(),
-        'freshness': str(row.get('freshness') or '').strip(),
-        'question_types': str(row.get('question_types') or '').strip(),
-        'canary': canary,
-    }
-
-
-class SealQADataLoader(SplitDataLoader):
-    def load_split_items(self, split_path: str) -> list[dict]:
-        path = Path(split_path)
-        csv_files = sorted(path.glob('*.csv'))
-        if not csv_files:
-            raise FileNotFoundError(f'No .csv file found in {split_path}')
-        with csv_files[0].open(encoding='utf-8', newline='') as f:
-            reader = csv.DictReader(f)
-            return [_normalize_row(row, idx) for idx, row in enumerate(reader, start=1)]

reflact/envs/sealqa/evaluator.py

@@ -1,126 +0,0 @@
-from __future__ import annotations
-
-import re
-
-from openai import AzureOpenAI, OpenAI
-
-GRADER_TEMPLATE = """
-Your job is to look at a question, a gold target, and a predicted answer, and then assign a grade of either [\"CORRECT\", \"INCORRECT\", \"NOT_ATTEMPTED\"].
-First, I will give examples of each grade, and then you will grade a new example.
-
-The following are examples of CORRECT predicted answers.
-```
-Question: What are the names of Barack Obama's children?
-Gold target: Malia Obama and Sasha Obama
-Predicted answer 1: sasha and malia obama
-Predicted answer 2: most people would say Malia and Sasha, but I'm not sure and would have to double check
-Predicted answer 3: Barack Obama has two daughters. Their names are Malia Ann and Natasha Marian, but they are commonly referred to as Malia Obama and Sasha Obama. Malia was born on July 4, 1998, and Sasha was born on June 10, 2001.
-```
-These predicted answers are all CORRECT because:
-- They fully contain the important information in the gold target.
-- They do not contain any information that contradicts the gold target.
-- Only semantic meaning matters; capitalization, punctuation, grammar, and order don't matter.
-- Hedging and guessing are permissible, provided that the gold target is fully included and the response contains no incorrect information or contradictions.
-
-The following are examples of INCORRECT predicted answers.
-```
-Question: What are the names of Barack Obama's children?
-Gold target: Malia and Sasha
-Predicted answer 1: Malia.
-Predicted answer 2: Malia, Sasha, and Susan.
-Predicted answer 3: Barack Obama does not have any children.
-```
-These predicted answers are all INCORRECT because a factual statement in the answer contradicts the gold target.
-
-The following are examples of NOT_ATTEMPTED predicted answers.
-```
-Question: What are the names of Barack Obama's children?
-Gold target: Malia and Sasha
-Predicted answer 1: I don't know.
-Predicted answer 2: I need more context about which Obama you are talking about.
-```
-These predicted answers are all NOT_ATTEMPTED because the important information in the gold target is not included and there is no contradiction.
-
-Here is a new example. Simply reply with either CORRECT, INCORRECT, NOT ATTEMPTED. Just return the letters \"A\", \"B\", or \"C\", with no text around it.
-```
-Question: {question}
-Gold target: {target}
-Predicted answer: {predicted_answer}
-```
-
-Grade the predicted answer as one of:
-A: CORRECT
-B: INCORRECT
-C: NOT_ATTEMPTED
-""".strip()
-
-
-def _build_grader_client() -> tuple[OpenAI | AzureOpenAI, str]:
-    import os
-
-    endpoint = os.environ.get('AZURE_OPENAI_ENDPOINT', '').strip()
-    api_version = os.environ.get('AZURE_OPENAI_API_VERSION', '').strip() or '2025-04-01-preview'
-    azure_key = os.environ.get('AZURE_OPENAI_API_KEY', '').strip()
-    openai_key = os.environ.get('OPENAI_API_KEY', '').strip()
-    api_key = azure_key or openai_key
-    if endpoint and api_version and api_key:
-        model = os.environ.get('SEALQA_GRADER_AZURE_MODEL', '').strip() or os.environ.get('SEALQA_GRADER_MODEL', '').strip() or os.environ.get('AZURE_MODEL_NAME', '').strip() or os.environ.get('TEACHER_DEPLOYMENT', '').strip() or 'gpt-5.4'
-        client = AzureOpenAI(api_key=api_key, api_version=api_version, azure_endpoint=endpoint.rstrip('/'))
-        return client, model
-
-    if openai_key:
-        model = os.environ.get('SEALQA_GRADER_OPENAI_MODEL', '').strip() or os.environ.get('SEALQA_GRADER_MODEL', '').strip() or 'gpt-4.1-mini'
-        return OpenAI(api_key=openai_key), model
-
-    raise ValueError('Missing grader credentials for SealQA scoring.')
-
-
-def _extract_text_content(content) -> str:
-    if content is None:
-        return ''
-    if isinstance(content, str):
-        return content
-    if isinstance(content, list):
-        parts = []
-        for part in content:
-            if isinstance(part, dict) and part.get('type') == 'text':
-                parts.append(str(part.get('text', '')))
-            else:
-                text = getattr(part, 'text', None)
-                if text:
-                    parts.append(str(text))
-        return '\n'.join(parts).strip()
-    return str(content).strip()
-
-
-def _normalize_text(text: str) -> str:
-    lowered = text.strip().lower()
-    lowered = re.sub(r'\s+', ' ', lowered)
-    lowered = re.sub(r'[^\w\s%.-]', '', lowered)
-    return lowered.strip()
-
-
-def _fallback_score(ground_truth: str, predicted: str) -> float:
-    gold = _normalize_text(ground_truth)
-    pred = _normalize_text(predicted)
-    if not gold or not pred:
-        return 0.0
-    if gold == pred:
-        return 1.0
-    if gold in pred or pred in gold:
-        return 1.0
-    return 0.0
-
-
-def score_sealqa(question: str, ground_truth: str, predicted: str) -> float:
-    try:
-        client, model = _build_grader_client()
-    except ValueError:
-        return _fallback_score(ground_truth, predicted)
-
-    prompt = GRADER_TEMPLATE.format(question=question, target=ground_truth, predicted_answer=predicted)
-    completion = client.chat.completions.create(model=model, messages=[{'role': 'user', 'content': prompt}])
-    content = _extract_text_content(completion.choices[0].message.content).strip().upper()
-    if content.startswith('A'):
-        return 1.0
-    return 0.0

reflact/envs/sealqa/prompts/analyst_error.md

@@ -1,30 +0,0 @@
-You are an expert failure-analysis agent for evidence-seeking factual question answering tasks.
-
-You will be given MULTIPLE failed SealQA trajectories from a single minibatch and the current skill document. The trajectories may include tool calls such as search, fetch, local reads, or evidence gathering steps.
-
-Your job is to identify COMMON failure patterns across the batch and propose concise skill edits.
-
-## Failure Type Categories
-- retrieval_miss: the agent failed to gather the right evidence
-- evidence_conflict: the agent saw conflicting evidence but resolved it badly
-- answer_selection: the agent found evidence but chose the wrong final answer
-- not_attempted: the agent never reached a grounded answer
-- other: none of the above
-
-Respond ONLY with a valid JSON object (no markdown fences, no extra text):
-{
-  "batch_size": <number of trajectories analysed>,
-  "failure_summary": [
-    {"failure_type": "<type>", "count": <int>, "description": "<one-line>"}
-  ],
-  "patch": {
-    "reasoning": "<why these edits address the batch's common failures>",
-    "edits": [
-      {"op": "append",       "content": "<markdown to add at end of skill>"},
-      {"op": "insert_after", "target": "<exact heading/text to insert after>", "content": "<markdown>"},
-      {"op": "replace",      "target": "<exact text to replace>",              "content": "<replacement>"},
-      {"op": "delete",       "target": "<exact text to remove>"}
-    ]
-  }
-}
-Only include edits that are needed. "edits" can be an empty list if no patch is warranted.

reflact/envs/sealqa/prompts/analyst_success.md

@@ -1,19 +0,0 @@
-You are an expert success-pattern analyst for evidence-seeking factual question answering tasks.
-
-You will be given MULTIPLE successful SealQA trajectories from a single minibatch and the current skill document. Your job is to identify common evidence-gathering and answer-selection behaviors worth encoding in the skill.
-
-Respond ONLY with a valid JSON object:
-{
-  "batch_size": <number of trajectories analysed>,
-  "success_patterns": ["<pattern 1>", "<pattern 2>"],
-  "patch": {
-    "reasoning": "<why these patterns are worth encoding>",
-    "edits": [
-      {"op": "append",       "content": "<markdown>"},
-      {"op": "insert_after", "target": "<heading/text>", "content": "<markdown>"},
-      {"op": "replace",      "target": "<old text>",     "content": "<new text>"},
-      {"op": "delete",       "target": "<exact text to remove>"}
-    ]
-  }
-}
-"edits" may be empty if the skill already covers all observed patterns.

reflact/envs/sealqa/prompts/rollout_system.md

@@ -1,3 +0,0 @@
-You are an expert research assistant. Use the provided search evidence first, and only if that is insufficient, inspect the provided URL content fetched for you. Reconcile conflicting information when necessary and return a concise final answer grounded in the evidence you found.
-
-{skill_section}Return the final answer inside <answer>...</answer> when you are ready.

reflact/envs/sealqa/rollout.py

@@ -1,268 +0,0 @@
-from __future__ import annotations
-
-import json
-import os
-import re
-from concurrent.futures import ThreadPoolExecutor, as_completed
-
-from reflact.envs.sealqa.evaluator import score_sealqa
-from reflact.envs.sealqa.tool_runtime import web_fetch
-from reflact.model import chat_student, get_student_backend, is_student_exec_backend
-from reflact.model.codex_harness import prepare_workspace, render_skill_md, run_student_exec
-from reflact.prompts import load_prompt
-
-_FINAL_RE = re.compile(r"<answer>(.*?)</answer>", re.IGNORECASE | re.DOTALL)
-
-
-def _build_system(skill_content: str) -> str:
-    if skill_content.strip():
-        skill_section = f"## Skill\n{skill_content.strip()}\n\n"
-    else:
-        skill_section = ""
-    return load_prompt("rollout_system", env="sealqa").format(skill_section=skill_section)
-
-
-def _build_user(item: dict, *, diagnostic_mode: bool = False, diagnostic_instruction: str = '') -> str:
-    parts = [f"## Question\n{item['question']}"]
-    if item.get('search_results'):
-        parts.append(f"## Search Results\n{item['search_results']}")
-    if item.get('urls'):
-        parts.append(f"## URL Hints\n{item['urls']}")
-    if item.get('freshness'):
-        parts.append(f"## Freshness\n{item['freshness']}")
-    if item.get('question_types'):
-        parts.append(f"## Question Types\n{item['question_types']}")
-    if diagnostic_mode and diagnostic_instruction.strip():
-        parts.append(f"## Training Readout\n{diagnostic_instruction.strip()}")
-    parts.append('Use the provided search evidence as your primary context. Do not rely on external tool use.')
-    return "\n\n".join(parts)
-
-
-def _extract_answer(text: str) -> str:
-    match = _FINAL_RE.search(text)
-    if match:
-        return match.group(1).strip()
-    lines = [line.strip() for line in text.splitlines() if line.strip()]
-    return lines[-1] if lines else text.strip()
-
-
-def _build_codex_skill(skill_content: str) -> str:
-    return render_skill_md(
-        skill_content,
-        description="Dynamic ReflACT skill for solving the current SealQA evidence-grounded question.",
-        preamble=(
-            "Use this skill when answering the current SealQA question.\n"
-            "Use the provided search evidence first, reconcile conflicts carefully,\n"
-            "and return the final answer inside <answer>...</answer>."
-        ),
-    )
-
-
-def _run_codex_once(
-    *,
-    pred_dir: str,
-    skill_content: str,
-    task_text: str,
-    model: str,
-    timeout: int,
-    previous_response: str = '',
-) -> tuple[str, str, str, str]:
-    task_parts = [task_text]
-    if previous_response:
-        task_parts.append(
-            "## Previous Attempt\n"
-            f"{previous_response}\n\n"
-            "Review the evidence again and correct the final answer if needed."
-        )
-    final_task_text = "\n\n".join(task_parts)
-    skill_md = _build_codex_skill(skill_content)
-    work_dir = os.path.join(pred_dir, 'codex_exec')
-    prepare_workspace(
-        work_dir=work_dir,
-        skill_md=skill_md,
-        task_text=final_task_text,
-    )
-    prompt = (
-        "Use the `reflact-student` skill available in this workspace.\n"
-        "Read `task.md`, answer the SealQA question using the provided evidence,\n"
-        "and return the final answer inside <answer>...</answer>."
-    )
-    final_message, raw = run_student_exec(
-        work_dir=work_dir,
-        prompt=prompt,
-        model=model,
-        timeout=timeout,
-    )
-    return final_message or raw, raw, skill_md, final_task_text
-
-
-def process_one(
-    item: dict,
-    out_root: str,
-    skill_content: str,
-    *,
-    max_tool_turns: int = 12,
-    diagnostic_mode: bool = False,
-    diagnostic_instruction: str = '',
-) -> dict:
-    item_id = str(item['id'])
-    pred_dir = os.path.join(out_root, 'predictions', item_id)
-    os.makedirs(pred_dir, exist_ok=True)
-
-    system = _build_system(skill_content)
-    user = _build_user(
-        item,
-        diagnostic_mode=diagnostic_mode,
-        diagnostic_instruction=diagnostic_instruction,
-    )
-    conversation: list[dict] = [{'role': 'user', 'content': user}]
-    final_response = ''
-    final_answer = ''
-    fail_reason = ''
-
-    try:
-        if is_student_exec_backend():
-            from reflact.model import azure_openai as _llm
-
-            response, _raw, system, user_for_save = _run_codex_once(
-                pred_dir=pred_dir,
-                skill_content=skill_content,
-                task_text=user,
-                model=_llm.STUDENT_DEPLOYMENT,
-                timeout=120,
-            )
-            final_response = response
-            conversation.append({'type': 'message', 'content': response})
-            if '<answer>' in response.lower():
-                final_answer = _extract_answer(response)
-            else:
-                user = user_for_save
-        else:
-            response, _ = chat_student(
-                system=system,
-                user=user,
-                max_completion_tokens=768,
-                retries=5,
-                stage='rollout',
-            )
-            final_response = response
-            conversation.append({'type': 'message', 'content': response})
-            if '<answer>' in response.lower():
-                final_answer = _extract_answer(response)
-
-        if not final_answer:
-            urls_text = str(item.get('urls') or '').strip()
-            fetched_blocks = []
-            for raw_url in re.findall(r'https?://[^\s\]\[\'\",]+', urls_text)[:2]:
-                try:
-                    fetched = web_fetch(raw_url)
-                except Exception as fetch_error:  # noqa: BLE001
-                    fetched = f'URL: {raw_url}\n\n[fetch error: {fetch_error}]'
-                fetched_blocks.append(fetched)
-                conversation.append({'type': 'tool_call', 'cmd': f'web_fetch({raw_url!r})', 'obs': fetched})
-            if fetched_blocks:
-                retry_user = user + '\n\n## Fetched URL Content\n' + '\n\n'.join(fetched_blocks)
-                if is_student_exec_backend():
-                    retry_response, _raw, system, retry_user = _run_codex_once(
-                        pred_dir=pred_dir,
-                        skill_content=skill_content,
-                        task_text=retry_user,
-                        model=_llm.STUDENT_DEPLOYMENT,
-                        timeout=120,
-                        previous_response=final_response,
-                    )
-                else:
-                    retry_response, _ = chat_student(
-                        system=system,
-                        user=retry_user,
-                        max_completion_tokens=768,
-                        retries=5,
-                        stage='rollout',
-                    )
-                final_response = retry_response
-                conversation.append({'type': 'message', 'content': retry_response})
-                if '<answer>' in retry_response.lower():
-                    final_answer = _extract_answer(retry_response)
-                else:
-                    fail_reason = 'Model did not produce a final answer'
-            else:
-                fail_reason = 'Model did not produce a final answer'
-    except Exception as e:  # noqa: BLE001
-        fail_reason = f'error: {e}'
-
-    with open(os.path.join(pred_dir, 'student_system_prompt.txt'), 'w', encoding='utf-8') as f:
-        f.write(system)
-    with open(os.path.join(pred_dir, 'student_user_prompt.txt'), 'w', encoding='utf-8') as f:
-        f.write(user)
-    with open(os.path.join(pred_dir, 'conversation.json'), 'w', encoding='utf-8') as f:
-        json.dump(conversation, f, ensure_ascii=False, indent=2)
-
-    score = score_sealqa(item.get('question', ''), item.get('ground_truth', ''), final_answer) if final_answer else 0.0
-    result = {
-        'id': item_id,
-        'question': item.get('question', ''),
-        'task_type': item.get('task_type', 'sealqa'),
-        'task_description': item.get('question', ''),
-        'predicted_answer': final_answer,
-        'response': final_response,
-        'ground_truth': item.get('ground_truth', ''),
-        'hard': int(score >= 1.0),
-        'soft': float(score),
-        'fail_reason': fail_reason or ('' if score >= 1.0 else f"predicted '{final_answer}' but expected '{item.get('ground_truth', '')}'"),
-        'agent_ok': not fail_reason,
-        'n_turns': len(conversation),
-        'student_system_prompt': system,
-        'student_user_prompt': user,
-    }
-    return result
-
-
-def run_batch(
-    items: list[dict],
-    out_root: str,
-    skill_content: str,
-    *,
-    workers: int = 4,
-    max_tool_turns: int = 12,
-    diagnostic_mode: bool = False,
-    diagnostic_instruction: str = '',
-) -> list[dict]:
-    results_path = os.path.join(out_root, 'results.jsonl')
-    os.makedirs(out_root, exist_ok=True)
-
-    done_ids: set[str] = set()
-    existing: list[dict] = []
-    if os.path.exists(results_path):
-        with open(results_path, encoding='utf-8') as f:
-            for line in f:
-                try:
-                    row = json.loads(line)
-                except json.JSONDecodeError:
-                    continue
-                done_ids.add(str(row.get('id')))
-                existing.append(row)
-
-    pending = [item for item in items if str(item['id']) not in done_ids]
-    if not pending:
-        return existing
-
-    results = list(existing)
-    with open(results_path, 'a', encoding='utf-8') as outf, ThreadPoolExecutor(max_workers=workers) as ex:
-        futs = {
-            ex.submit(
-                process_one,
-                item,
-                out_root,
-                skill_content,
-                max_tool_turns=max_tool_turns,
-                diagnostic_mode=diagnostic_mode,
-                diagnostic_instruction=diagnostic_instruction,
-            ): item
-            for item in pending
-        }
-        for fut in as_completed(futs):
-            res = fut.result()
-            results.append(res)
-            outf.write(json.dumps(res, ensure_ascii=False) + '\n')
-            outf.flush()
-    return results

reflact/envs/sealqa/skills/initial.md

@@ -1,11 +0,0 @@
-# SealQA Skill
-
-## Evidence Gathering
-- Search for the most directly relevant evidence before answering.
-- If multiple sources conflict, prefer the source that best matches the question's entity, date, and scope.
-- Keep notes on which evidence directly answers the question versus which evidence is only contextual.
-
-## Final Answer Discipline
-- Do not answer until the supporting evidence is specific enough.
-- Choose the final answer that is best grounded in the gathered evidence.
-- Keep the final answer concise.

reflact/envs/sealqa/tool_runtime.py

@@ -1,30 +0,0 @@
-from __future__ import annotations
-
-import html
-import re
-from urllib.request import Request, urlopen
-
-DEFAULT_USER_AGENT = (
-    'Mozilla/5.0 (X11; Linux x86_64) AppleWebKit/537.36 '
-    '(KHTML, like Gecko) Chrome/135.0 Safari/537.36'
-)
-_MAX_FETCH_CHARS = 6000
-
-
-def _strip_html(raw_html: str) -> str:
-    cleaned = re.sub(r'(?is)<script.*?>.*?</script>', ' ', raw_html)
-    cleaned = re.sub(r'(?is)<style.*?>.*?</style>', ' ', cleaned)
-    cleaned = re.sub(r'(?is)<[^>]+>', ' ', cleaned)
-    cleaned = html.unescape(cleaned)
-    return re.sub(r'\s+', ' ', cleaned).strip()
-
-
-def web_fetch(url: str, max_chars: int = _MAX_FETCH_CHARS) -> str:
-    req = Request(url, headers={'User-Agent': DEFAULT_USER_AGENT})
-    with urlopen(req, timeout=20) as response:
-        body = response.read().decode('utf-8', errors='ignore')
-    text = _strip_html(body)
-    if len(text) > max_chars:
-        omitted = len(text) - max_chars
-        text = text[:max_chars] + f"\n\n[... {omitted} characters omitted ...]"
-    return f"URL: {url}\n\n{text}"

reflact/envs/searchqa/adapter.py

@@ -1,250 +0,0 @@
-"""SearchQA environment adapter for ReflACT."""
-from __future__ import annotations
-
-import json
-import os
-
-from reflact.gradient.deep_probe import generate_deep_probe_instruction
-from reflact.datasets.base import BatchSpec
-from reflact.envs.base import EnvAdapter
-from reflact.envs.searchqa.dataloader import SearchQADataLoader
-from reflact.envs.searchqa.rollout import run_batch
-from reflact.gradient.reflect import run_minibatch_reflect
-from reflact.model import get_student_backend
-
-
-class SearchQAAdapter(EnvAdapter):
-    """SearchQA environment adapter."""
-
-    def __init__(
-        self,
-        split_dir: str = "",
-        data_path: str = "",
-        split_mode: str = "ratio",
-        split_ratio: str = "2:1:7",
-        split_seed: int = 42,
-        split_output_dir: str = "",
-        max_turns: int = 1,
-        exec_timeout: int = 120,
-        workers: int = 64,
-        analyst_workers: int = 16,
-        failure_only: bool = False,
-        minibatch_size: int = 8,
-        edit_budget: int = 4,
-        seed: int = 42,
-        limit: int = 0,
-        exec_timeout: int = 600,
-        use_deep_reflect: bool = False,
-        deep_reflect_failures: int = 4,
-        deep_reflect_successes: int = 2,
-    ) -> None:
-        self.max_turns = max_turns
-        self.exec_timeout = exec_timeout
-        self.workers = workers
-        self.analyst_workers = analyst_workers
-        self.failure_only = failure_only
-        self.minibatch_size = minibatch_size
-        self.edit_budget = edit_budget
-        self.exec_timeout = exec_timeout
-        self.use_deep_reflect = use_deep_reflect
-        self.deep_reflect_failures = deep_reflect_failures
-        self.deep_reflect_successes = deep_reflect_successes
-        self.dataloader = SearchQADataLoader(
-            split_dir=split_dir,
-            data_path=data_path,
-            split_mode=split_mode,
-            split_ratio=split_ratio,
-            split_seed=split_seed,
-            split_output_dir=split_output_dir,
-            seed=seed,
-            limit=limit,
-        )
-
-    def setup(self, cfg: dict) -> None:
-        super().setup(cfg)
-        self.dataloader.setup(cfg)
-
-    def get_dataloader(self):
-        return self.dataloader
-
-    def build_env_from_batch(self, batch: BatchSpec, **kwargs):
-        return list(batch.payload or [])
-
-    def build_train_env(self, batch_size: int, seed: int, **kwargs):
-        batch = self.dataloader.build_train_batch(batch_size=batch_size, seed=seed, **kwargs)
-        return self.build_env_from_batch(batch, **kwargs)
-
-    def build_eval_env(self, env_num: int, split: str, seed: int, **kwargs):
-        batch = self.dataloader.build_eval_batch(env_num=env_num, split=split, seed=seed, **kwargs)
-        return self.build_env_from_batch(batch, **kwargs)
-
-    def rollout(
-        self,
-        env_manager,  # actually list[dict] for SearchQA
-        skill_content: str,
-        out_dir: str,
-        **kwargs,
-    ) -> list[dict]:
-        """Run QA agent on items. Resume-aware."""
-        items: list[dict] = env_manager  # type alias for clarity
-        return run_batch(
-            items=items,
-            out_root=out_dir,
-            skill_content=skill_content,
-            max_turns=self.max_turns,
-            exec_timeout=self.exec_timeout,
-            workers=self.workers,
-            diagnostic_mode=kwargs.get("diagnostic_mode", False),
-            diagnostic_instruction=kwargs.get("diagnostic_instruction", ""),
-            diagnostic_trace_context_by_id=kwargs.get("diagnostic_trace_context_by_id"),
-            task_timeout=self.exec_timeout,
-        )
-
-    def reflect(
-        self,
-        results: list[dict],
-        skill_content: str,
-        out_dir: str,
-        **kwargs,
-    ) -> list[dict | None]:
-        prediction_dir = kwargs.get("prediction_dir", os.path.join(out_dir, "predictions"))
-        patches_dir = kwargs.get("patches_dir", os.path.join(out_dir, "patches"))
-        random_seed = kwargs.get("random_seed")
-        step_buffer_context = kwargs.get("step_buffer_context", "")
-        meta_skill_context = kwargs.get("meta_skill_context", "")
-
-        return run_minibatch_reflect(
-            results=results,
-            skill_content=skill_content,
-            prediction_dir=prediction_dir,
-            patches_dir=patches_dir,
-            workers=self.analyst_workers,
-            failure_only=self.failure_only,
-            minibatch_size=self.minibatch_size,
-            edit_budget=self.edit_budget,
-            random_seed=random_seed,
-            error_system=self.get_error_minibatch_prompt(),
-            success_system=self.get_success_minibatch_prompt(),
-            step_buffer_context=step_buffer_context,
-            meta_skill_context=meta_skill_context,
-            update_mode=getattr(self, "_cfg", {}).get("skill_update_mode", "patch"),
-        )
-
-    def deep_reflect(
-        self,
-        results: list[dict],
-        skill_content: str,
-        out_dir: str,
-        **kwargs,
-    ) -> list[dict | None]:
-        if not self.use_deep_reflect:
-            return []
-
-        env_manager = kwargs.get("env_manager")
-        if not isinstance(env_manager, list):
-            return []
-
-        prediction_dir = kwargs.get("prediction_dir", os.path.join(out_dir, "predictions"))
-        random_seed = kwargs.get("random_seed")
-        step_buffer_context = kwargs.get("step_buffer_context", "")
-        meta_skill_context = kwargs.get("meta_skill_context", "")
-        codex_backend = get_student_backend() == "codex_exec"
-        selected_items = self.select_representative_items(
-            results,
-            env_manager,
-            n_failures=self.deep_reflect_failures,
-            n_successes=self.deep_reflect_successes,
-            seed=random_seed,
-        )
-        if not selected_items:
-            return []
-
-        selected_ids = {str(item["id"]) for item in selected_items}
-        selected_results = [row for row in results if str(row.get("id")) in selected_ids]
-        selected_examples = (
-            self.attach_codex_probe_context(selected_results, prediction_dir)
-            if codex_backend
-            else selected_results
-        )
-        selected_metadata = [
-            {
-                "id": str(item["id"]),
-                "question_preview": str(item.get("question") or "")[:200],
-                "has_context": bool(str(item.get("context") or "").strip()),
-                "n_gold_answers": len(item.get("answers") or []),
-            }
-            for item in selected_items
-        ]
-
-        deep_dir = os.path.join(out_dir, "deep_reflect")
-        rollout_dir = os.path.join(deep_dir, "rollout")
-        patches_dir = os.path.join(deep_dir, "patches")
-        os.makedirs(deep_dir, exist_ok=True)
-        print(
-            f"    [2b/6 DEEP REFLECT setup] selected={len(selected_items)} "
-            f"mode=no_reference_probe"
-        )
-        probe = generate_deep_probe_instruction(
-            skill_content=skill_content,
-            items=selected_examples,
-            prediction_dir=prediction_dir,
-            system_prompt=self.get_codex_deep_probe_prompt() if codex_backend else self.get_deep_probe_prompt(),
-            step_buffer_context=step_buffer_context,
-            meta_skill_context=meta_skill_context,
-            output_requirements=[
-                "- There is no hidden reference block. Use only the question, provided context, the student's output, and the evaluation result to infer what intermediate state is worth probing.",
-                "- The instruction must explicitly request a short <analysis>...</analysis> block before the final <answer>...</answer>.",
-                "- The readout should focus on likely evidence span, top candidate and runner-up, decisive clue, or a few short intermediate conclusions.",
-                "- Do not ask for exhaustive copying of the context or a full chain-of-thought.",
-                "- The instruction text should be ready to append directly to the student's prompt.",
-            ],
-        )
-        if not probe:
-            return []
-        diagnostic_trace_context_by_id = None
-        if codex_backend:
-            selected_items, diagnostic_trace_context_by_id, probe = self.resolve_codex_probe_target(
-                selected_items=selected_items,
-                selected_examples=selected_examples,
-                prediction_dir=prediction_dir,
-                probe=probe,
-            )
-
-        with open(os.path.join(deep_dir, "probe.json"), "w", encoding="utf-8") as f:
-            json.dump(
-                {
-                    **probe,
-                    "selected_examples": selected_metadata,
-                },
-                f,
-                ensure_ascii=False,
-                indent=2,
-            )
-
-        deep_results = self.rollout(
-            selected_items,
-            skill_content,
-            rollout_dir,
-            diagnostic_mode=True,
-            diagnostic_instruction=probe["probe_instruction"],
-            diagnostic_trace_context_by_id=diagnostic_trace_context_by_id,
-        )
-        return run_minibatch_reflect(
-            results=deep_results,
-            skill_content=skill_content,
-            prediction_dir=os.path.join(rollout_dir, "predictions"),
-            patches_dir=patches_dir,
-            workers=self.analyst_workers,
-            failure_only=self.failure_only,
-            minibatch_size=self.minibatch_size,
-            edit_budget=self.edit_budget,
-            random_seed=random_seed,
-            error_system=self.get_error_minibatch_prompt(),
-            success_system=self.get_success_minibatch_prompt(),
-            step_buffer_context=step_buffer_context,
-            meta_skill_context=meta_skill_context,
-            update_mode=getattr(self, "_cfg", {}).get("skill_update_mode", "patch"),
-        )
-
-    def get_task_types(self) -> list[str]:
-        return ["qa"]

reflact/envs/searchqa/prompts/deep_probe.md

@@ -1,27 +0,0 @@
-You are an expert diagnostic-probe designer for retrieval-style question answering tasks.
-
-You will be shown representative trajectories, the current student skill, the student's prompt context,
-and the evaluation result including the gold answer. There is NO hidden chain-of-thought reference.
-Design one SMALL diagnostic instruction that exposes the student's intermediate reading or evidence-selection state
-without materially changing the original scaffold.
-
-## Hard Constraints
-1. Do NOT substantially change the original scaffold.
-2. Do NOT prescribe a brand-new multi-step solving procedure.
-3. You MAY ask for a short structured readout of intermediate conclusions, evidence candidates, or elimination decisions.
-4. Do NOT ask for exhaustive quotation of the whole context or a full chain-of-thought.
-5. Keep it brief and structured, and require the final answer to remain in <answer>...</answer>.
-6. Use the gold answer only to target a useful probe; do not simply force the student to restate the gold answer.
-
-## Good Probe Targets
-- the most likely supporting span or document cue
-- top answer candidate and runner-up
-- decisive lexical clue / entity / date / title
-- why a tempting alternative was rejected
-- 2-4 short intermediate conclusions that directly support the final answer
-
-Respond ONLY with a valid JSON object:
-{
-  "reasoning": "<why this probe is informative>",
-  "probe_instruction": "<the exact instruction text to append to the student prompt>"
-}