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>
feat(slow-update): add config-controlled gated / force-injected modes
2026-05-31 09:01:25 +00:00 · 2026-05-31 02:02:23 +00:00 · 2026-05-30 15:28:17 +00:00 · 2026-05-30 15:15:15 +00:00 · 2026-05-30 15:01:34 +00:00 · 2026-05-30 08:09:03 +00:00
165 changed files with 17174 additions and 10600 deletions
@@ -1,24 +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) ──────────────────
-AZURE_OPENAI_ENDPOINT=https://your-resource.openai.azure.com/
-AZURE_OPENAI_API_VERSION=2024-12-01-preview
+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
-AZURE_OPENAI_API_KEY=
-#   Option 2: Azure CLI (set auth_mode=azure_cli in config)
-#   Option 3: Managed Identity (set auth_mode=managed_identity + client_id in config)
+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 (alternative to Azure) ────────────────────────────────────
-# OPENAI_API_KEY=sk-...
+# ── 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) ─────────────────────
-# ANTHROPIC_API_KEY=sk-ant-...
+# export ANTHROPIC_API_KEY=sk-ant-...

 # ── Qwen Local Model (for qwen_chat backend) ────────────────────────
-# QWEN_CHAT_BASE_URL=http://localhost:8000/v1
-# QWEN_CHAT_MODEL=Qwen/Qwen3.5-4B
-
-# ── Ray (optional, for distributed rollout) ──────────────────────────
-# RAY_ADDRESS=auto
+# export QWEN_CHAT_BASE_URL=http://localhost:8000/v1
+# export QWEN_CHAT_MODEL=Qwen/Qwen3.5-4B
@@ -39,3 +39,4 @@ docs/reflact_conda_env_export.yml
 docs/reflact_overview.html
 docs/render_ablation_paper_tables.py
 docs/让*
+.gradio/
@@ -1,112 +1,54 @@
 # SkillOpt: Executive Strategy for Self-Evolving Agent Skills

-> ⚠️ **This is a preliminary draft release. A formal open-source release will follow.**
+*Train agent skills like you train neural networks — with epochs, (mini-)batchsize, learning rates, and validation gates — but without touching model weights.*

-[![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)
+[![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)

-*Train agent skills like you train neural networks — with epochs, learning rates, and validation gates — but without touching model weights.*
+## 🎬 SkillOpt Demo Video
+
+https://github.com/user-attachments/assets/eb12d3bc-371c-467f-904d-91b61f339ed7
+
+<p align="center">
+  <a href="https://youtu.be/JUBMDTCiM0M"><b>▶ Watch the full demo on YouTube</b></a>
+</p>

 ---

-## What is SkillOpt?
+## Documentation

-SkillOpt is a framework for optimizing a natural-language **skill document** through iterative rollout, reflection, editing, and gated validation.
+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.

-It does **not** fine-tune model parameters. Instead, it treats the skill document as the optimization target:
+Because GitHub shows raw source for `.html` files instead of rendering them,
+open the guide one of these ways:

- 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
-
-| Deep Learning | SkillOpt |
-|---|---|
-| Model weights | Skill document (Markdown) |
-| Forward pass | Rollout (student executes tasks) |
-| Loss computation | Reflect (teacher analyzes trajectories) |
-| Gradient | Edit patches (proposed skill improvements) |
-| Gradient clipping | Edit ranking & selection (`learning_rate`) |
-| Weight update | Patch application to skill document |
-| Validation | Gated evaluation on held-out split |
-| Learning rate schedule | `lr_scheduler`: cosine, linear decay |
-| Epochs | Multi-epoch training with slow update & meta skill |
+- **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).

 ---

-## Method Overview
+## Install

-### Optimization Target
-
-Each run maintains a mutable markdown skill document. The framework repeatedly improves that document instead of changing model parameters.
-
-This gives a training-style loop for prompt / policy optimization:
-
-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.
-
-### Per-Step Pipeline
-
-Every training step executes the following pipeline in `skillopt/engine/trainer.py`:
-
-1. **Rollout**
-   The student model runs a batch of tasks using the current skill.
-
-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. 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
-
-#### 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.
-
-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.
-
---
-
-## Quick Start
-
-### Install
+**Requirements:** Python 3.10+

 ```bash
-git clone https://github.com/AgenticOpt/SkillOpt.git
+git clone https://github.com/microsoft/SkillOpt.git
 cd SkillOpt
 pip install -e .
+
+# For ALFWorld benchmark (optional):
+pip install -e ".[alfworld]"
+alfworld-download
 ```

 ### Configure API Credentials
@@ -117,18 +59,31 @@ cp .env.example .env
 source .env
 ```

-**Azure OpenAI** (API key or managed identity):
+**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"
-# Or use managed identity: set azure_openai_auth_mode=managed_identity in config
+# Option 2: Azure CLI auth (no API key needed)
+export AZURE_OPENAI_AUTH_MODE="azure_cli"
 ```

-**OpenAI** directly:
+> **Note:** `AZURE_OPENAI_ENDPOINT` is required for all three modes (`api_key`, `azure_cli`,
+> `openai_compatible`). Without it, all LLM calls will fail.
+
+**OpenAI-compatible endpoints**:
 ```bash
-export OPENAI_API_KEY="sk-..."
+export AZURE_OPENAI_ENDPOINT="https://api.openai.com/v1"
+export AZURE_OPENAI_API_KEY="sk-..."
+export AZURE_OPENAI_AUTH_MODE="openai_compatible"
 ```

+This routes all calls through the plain OpenAI Python client (no Azure auth, no `api-version`
+header).
+
+> **Note:** SkillOpt reuses the `AZURE_OPENAI_*` env var names even in this mode — there is no
+> separate `OPENAI_API_KEY` knob.
+
 **Anthropic Claude**:
 ```bash
 export ANTHROPIC_API_KEY="sk-ant-..."
@@ -140,237 +95,159 @@ export QWEN_CHAT_BASE_URL="http://localhost:8000/v1"
 export QWEN_CHAT_MODEL="Qwen/Qwen3.5-4B"
 ```

-### Run Training
-
-```bash
-python scripts/train.py --config configs/searchqa/default.yaml
-```
-
 ---

-## Configuration
+## Data Preparation

-SkillOpt uses a hierarchical YAML configuration system. Each benchmark config inherits from `configs/_base_/default.yaml`.
+SkillOpt expects data in a **split directory** with `train/`, `val/`, `test/` subdirectories, each containing a JSON file (e.g., `items.json`).

-### Configuration Structure
-
-```yaml
-model:
-  teacher_backend: openai_chat     # openai_chat | claude_chat | qwen_chat
-  student_backend: openai_chat     # openai_chat | claude_chat | codex_exec | qwen_chat
-  teacher: gpt-5.5                 # teacher model deployment name
-  student: gpt-5.5                 # student model deployment name
-  reasoning_effort: medium         # low | medium | high
-
-train:
-  num_epochs: 4
-  batch_size: 40
-  seed: 42
-
-gradient:
-  minibatch_size: 8                # trajectories per reflection call
-  analyst_workers: 16              # parallel reflection workers
-  use_deep_reflect: false          # deep multi-turn probing
-  deep_reflect_failures: 4
-  deep_reflect_successes: 2
-
-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
-  skill_update_mode: patch         # patch | rewrite_from_suggestions | full_rewrite_minibatch
-  use_slow_update: true
-  use_meta_skill: true
-  use_meta_reflect: false
-
-evaluation:
-  use_gate: true                   # gated validation (always recommended)
-
-env:
-  name: ""                         # benchmark name
-  skill_init: ""                   # path to initial skill document
-  split_mode: ratio                # ratio | split_dir
-  split_ratio: "2:1:7"            # train:val:test
+```
+data/my_split/
+├── train/items.json
+├── val/items.json
+└── test/items.json
 ```

-### CLI Overrides
+Each JSON file is an array of task items. The required fields depend on the benchmark. For example, SearchQA items look like:

-Override any config key from the command line:
-
-```bash
-python scripts/train.py \
-  --config configs/searchqa/default.yaml \
-  --cfg-options model.teacher_backend=openai_chat \
-                model.student_backend=codex_exec \
-                train.batch_size=40 \
-                optimizer.learning_rate=4
-
-# Legacy flat overrides also work for common keys:
-python scripts/train.py \
-  --config configs/searchqa/default.yaml \
-  --backend azure_openai \
-  --teacher_model gpt-5.5 \
-  --student_model gpt-5.5 \
-  --reasoning_effort medium
+```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.

-## Model Backends
+> **Note:** Benchmark datasets are not included in this repository. Prepare your own data following the format above.

-All model access goes through the unified backend router in `skillopt/model/`.
-
-| Backend | Use case | Config key |
-|---|---|---|
-| `openai_chat` | Azure OpenAI / OpenAI API | teacher / student |
-| `claude_chat` | Anthropic Claude | teacher / student |
-| `codex_exec` | Codex execution harness | student only |
-| `qwen_chat` | Local Qwen via vLLM | teacher / student |
-
-Separate teacher/student endpoints are supported:
-
-```yaml
-model:
-  teacher_backend: openai_chat
-  student_backend: codex_exec
-  teacher: gpt-5.5
-  student: gpt-5.5-codex
-```
-
---
-
-## Data Splits
-
-SkillOpt supports two split modes:
-
-**Ratio split** — auto-generate from raw data:
-```bash
-python scripts/train.py \
-  --config configs/searchqa/default.yaml \
-  --split_mode ratio \
-  --data_path /path/to/searchqa_data.json
-```
-
-**Pre-split directory** — consume prepared splits:
-```bash
-python scripts/train.py \
-  --config configs/searchqa/default.yaml \
-  --split_mode split_dir \
-  --split_dir /path/to/searchqa_split
-```
-
---
-
-## Supported Benchmarks
+### Supported Benchmarks

 | Benchmark | Type | Config |
 |---|---|---|
 | SearchQA | QA | `configs/searchqa/default.yaml` |
-| SpreadsheetBench | Code generation | `configs/spreadsheetbench/default.yaml` |
 | ALFWorld | Embodied agent | `configs/alfworld/default.yaml` |
 | DocVQA | Document QA | `configs/docvqa/default.yaml` |
-| OfficeQA | Tool-augmented QA | `configs/officeqa/default.yaml` |
-| SealQA | Tool-augmented QA | `configs/sealqa/default.yaml` |
-| BabyVision | Vision QA | `configs/babyvision/default.yaml` |
 | LiveMathematicianBench | Math | `configs/livemathematicianbench/default.yaml` |
-| MathVerse | Multimodal math | `configs/mathverse/default.yaml` |
-| MMRB | Multimodal reasoning | `configs/mmrb/default.yaml` |
-| SWEBench | Software engineering | `configs/swebench/default.yaml` |
+| SpreadsheetBench | Code generation | `configs/spreadsheetbench/default.yaml` |
+| OfficeQA | Tool-augmented QA | `configs/officeqa/default.yaml` |

 ---

-## Running Training
+## Quick Start

-Basic training:
-
-```bash
-python scripts/train.py --config configs/searchqa/default.yaml
-```
-
-Exec harness (Codex student):
+### Training

 ```bash
+# Minimal example — train on SearchQA:
 python scripts/train.py \
-  --config configs/searchqa/default.yaml \
-  --teacher_backend openai_chat \
-  --student_backend codex_exec \
-  --teacher_model gpt-5.5 \
-  --student_model gpt-5.5-codex \
-  --use_deep_reflect true \
-  --skill_update_mode rewrite_from_suggestions
-```
+    --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

-SWEBench:
-
-```bash
+# Train on LiveMathematicianBench:
 python scripts/train.py \
-  --config configs/swebench/default.yaml \
-  --cfg-options env.dataset_name=lite env.split_ratio=2:1:7
+    --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
+
+# Train on ALFWorld:
+python scripts/train.py \
+    --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
 ```

+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 specific skill without training:
+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 skillopt/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/
 ```

---
+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`.

-## Output Structure
+| Split | Description |
+|---|---|
+| `valid_unseen` | Test set |
+| `valid_seen` | Validation set |
+| `train` | Training set |
+| `all` | All splits combined (default) |

-Each run writes a structured output directory:
+### Output Structure
+
+Each run writes to a structured output directory:

 ```
 outputs/<run_name>/
 ├── config.json              # Flattened runtime config
-├── history.json             # Per-step history records
-├── runtime_state.json       # Resume state (for auto-resume)
-├── best_skill.md            # Current best validated skill
+├── 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
-│   ├── merged_patch.json
-│   ├── ranked_edits.json
-│   ├── candidate_skill.md
-│   ├── edit_apply_report.json
-│   ├── rewrite_result.json  # when rewrite mode is enabled
-│   └── selection_eval/
-├── slow_update/epoch_XX/
-├── meta_skill/epoch_XX/
-└── meta_reflect/epoch_XX/
+├── steps/step_XXXX/        # Per-step artifacts (patches, evals)
+├── slow_update/epoch_XX/   # Slow update logs
+└── meta_skill/epoch_XX/    # Meta skill logs
 ```

-### Resume Behavior
-
-The trainer resumes from `runtime_state.json` when present. That state tracks:
-
- Last completed step
- Current skill path and score
- Best skill path and score
- Origin tags for current and best skill
+Re-running the same command auto-resumes from the last completed step.

 ---

-## Extending SkillOpt
+## Community-contributed configs

-### Add a New Benchmark
+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.

-1. Create `skillopt/envs/<your_env>/` with:
-   - `adapter.py` — implements `EnvAdapter`
-   - `dataloader.py` — data loading logic
-   - `rollout.py` — student execution logic
-   - `skills/initial.md` — initial skill document
-2. Add a config at `configs/<your_env>/default.yaml`
-3. Register in `skillopt/envs/__init__.py`
-
-See `skillopt/envs/_template/` for a scaffold.
-
-### Add a New Model Backend
-
-Implement a backend in `skillopt/model/` following the interface in `skillopt/model/common.py`, then register it in `skillopt/model/router.py`.
+- **`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.

 ---

@@ -383,36 +260,29 @@ pip install -e ".[webui]"
 python -m skillopt_webui.app
 ```

-Provides browser-based config selection, training launch, and real-time log monitoring.
-
---
-
-## Minimal Setup
+| Flag | Default | Description |
+|---|---|---|
+| `--port` | 7860 | Server port |
+| `--host` | `0.0.0.0` | Bind address |
+| `--share` | off | Create a public Gradio share link |

 ```bash
-conda create -n skillopt python=3.11
-conda activate skillopt
-pip install -e .
+# With public share link (useful for remote servers)
+python -m skillopt_webui.app --share
 ```

-Depending on the benchmark, you may also need:
-
-```bash
-pip install datasets gymnasium numpy
-```
-
-For SWEBench, you also need a working Docker environment plus the SWE-bench harness dependencies.
-
 ---

 ## Citation

 ```bibtex
-@article{skillopt2026,
-  title={SkillOpt: Executive Strategy for Self-Evolving Agent Skills},
-  author={SkillOpt Team},
-  year={2026}
+@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}
 }
 ```
-
-
@@ -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 -->
@@ -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 -->
@@ -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 -->
@@ -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 -->
@@ -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 -->
@@ -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 -->
@@ -3,10 +3,10 @@

 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
+  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: ""   # e.g. "https://your-resource.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: ""   # e.g. "https://your-resource.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: ""
@@ -10,7 +10,6 @@ gradient:

 optimizer:
  learning_rate: 4
-  use_meta_reflect: false

 evaluation:
  sel_env_num: 0
@@ -20,11 +19,11 @@ env:
  name: alfworld
  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
@@ -1,21 +0,0 @@
-_base_: ../_base_/default.yaml
-
-train:
-  batch_size: 64
-  accumulation: 1
-
-env:
-  name: babyvision
-  skill_init: skillopt/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
@@ -18,11 +18,11 @@ env:
  name: docvqa
  skill_init: skillopt/envs/docvqa/skills/initial.md
  split_mode: split_dir
-  split_ratio: "2:1:7"
  split_dir: data/docvqa/splits
  data_path: ""
  split_output_dir: ""
  max_turns: 1
+  max_completion_tokens: 16384
  workers: 16
  image_detail: auto
  limit: 0
@@ -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
@@ -9,11 +9,11 @@ env:
  name: livemathematicianbench
  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
@@ -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: skillopt/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
@@ -1,18 +0,0 @@
-_base_: ../_base_/default.yaml
-
-train:
-  batch_size: 128
-  accumulation: 1
-
-env:
-  name: mmrb
-  skill_init: skillopt/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
@@ -23,7 +23,7 @@ env:
    - data/officeqa_docs_official
  workers: 4
  max_tool_turns: 24
-  max_completion_tokens: 10000
+  max_completion_tokens: 16384
  search_mode: offline
  max_queries_per_turn: 4
  search_api_url: http://apisix.westus2.cloudapp.azure.com/search_tool/search
@@ -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: skillopt/envs/sealqa/skills/initial.md
-  split_dir: data/sealqa_split
-  workers: 4
-  max_tool_turns: 12
-  limit: 0
@@ -23,10 +23,10 @@ env:
  name: searchqa
  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
@@ -23,12 +23,12 @@ env:
  name: spreadsheetbench
  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
@@ -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: skillopt/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
@@ -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"
+  }
+]
@@ -25,8 +25,8 @@ Benchmark configs inherit from `_base_/default.yaml` and override specific value
 ```yaml
 model:
  backend: azure_openai          # azure_openai | openai_chat | claude_code_exec | qwen
-  teacher: gpt-5.5               # Teacher model (for reflection)
-  student: gpt-5.5               # Student model (for rollout)
+  optimizer: gpt-5.5               # Optimizer model (for reflection)
+  target: gpt-5.5               # Target model (for rollout)
 ```

 ### Training
@@ -7,9 +7,9 @@ SkillOpt is designed around a core insight: **optimizing natural-language prompt
 | Deep Learning | SkillOpt | Description |
 |---|---|---|
 | **Model weights** | Skill document (Markdown) | The thing being optimized |
-| **Forward pass** | Rollout | Student executes tasks using current skill |
+| **Forward pass** | Rollout | Target executes tasks using current skill |
 | **Loss function** | Task evaluator | Scores task execution quality |
-| **Backpropagation** | Reflect | Teacher analyzes failures → edit patches |
+| **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 |
@@ -21,7 +21,7 @@ SkillOpt is designed around a core insight: **optimizing natural-language prompt
 | **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 teacher strategy memory |
+| **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 |
@@ -44,7 +44,7 @@ From our experiments, these DL intuitions transfer well:
    - **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** — teacher benefits from cross-epoch strategy notes
+    - **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
@@ -33,7 +33,7 @@ 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 teacher memory)
+  use_meta_skill: true    # (cross-epoch optimizer memory)

 gradient:
  analyst_workers: 16     # (parallel reflection workers)
@@ -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.
@@ -76,7 +76,7 @@ class MyBenchmarkEnv(EnvAdapter):
        Args:
            item: The data item to process
            skill: Current skill document content
-            model: The student model instance
+            model: The target model instance
            
        Returns:
            TaskResult with prediction, score, and trajectory
@@ -70,7 +70,7 @@ Track your skill's evolution through:
    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 teacher accumulates strategy memory
+    4. **Enable meta skill** (`use_meta_skill: true`) so the optimizer accumulates strategy memory

 ## Next Steps

@@ -10,8 +10,8 @@ SkillOpt's core insight: **optimizing natural-language skill documents follows t
 │                                                         │
 │  for epoch in epochs:                                   │
 │    for step in steps:                                   │
-│      1. Rollout   — Student executes tasks              │
-│      2. Reflect   — Teacher analyzes trajectories       │
+│      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          │
@@ -27,7 +27,7 @@ SkillOpt's core insight: **optimizing natural-language skill documents follows t

 ### 1. Rollout (Forward Pass)

-The **student** model executes tasks using the current skill document as its prompt. Each task produces a trajectory and a score.
+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
@@ -37,7 +37,7 @@ scores = evaluate(predictions, ground_truth)

 ### 2. Reflect (Backward Pass)

-The **teacher** model analyzes failed trajectories and produces **edit patches** — structured suggestions for improving the skill document.
+The **optimizer** model analyzes failed trajectories and produces **edit patches** — structured suggestions for improving the skill document.

 Two modes:

@@ -84,7 +84,7 @@ At the end of each epoch (starting from epoch 2), the system performs a **longit

 ### Meta Skill

-A **meta-skill memory** accumulates high-level strategy notes across the entire training run. At the end of each epoch, the teacher reflects on what changed between epochs and produces a compact memory that is provided as additional context during future reflection steps.
+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

@@ -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>
@@ -26,7 +26,7 @@ hide:
 <div class="pipeline-stage" id="stage-rollout">
 <div class="stage-icon">🎯</div>
 <div class="stage-label">Rollout</div>
-<div class="stage-desc">Student executes tasks</div>
+<div class="stage-desc">Target executes tasks</div>
 </div>

 <div class="pipeline-arrow"><div class="flow-line"></div></div>
@@ -34,7 +34,7 @@ hide:
 <div class="pipeline-stage" id="stage-reflect">
 <div class="stage-icon">🔍</div>
 <div class="stage-label">Reflect</div>
-<div class="stage-desc">Teacher analyzes trajectories</div>
+<div class="stage-desc">Optimizer analyzes trajectories</div>
 </div>

 <div class="pipeline-arrow"><div class="flow-line"></div></div>
@@ -88,8 +88,8 @@ SkillOpt brings the familiar deep-learning training paradigm to agentic prompt o
 | Deep Learning | SkillOpt | 
 |---|---|
 | Model weights | Skill document (Markdown) |
-| Forward pass | Rollout (student executes tasks) |
-| Loss / gradient | Reflect (teacher produces edit patches) |
+| 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 |
@@ -7,8 +7,8 @@ Complete reference for all SkillOpt configuration parameters.
 | Parameter | Type | Default | Description |
 |---|---|---|---|
 | `model.backend` | str | `azure_openai` | Backend: `azure_openai` / `openai_chat` / `claude_code_exec` / `qwen` |
-| `model.teacher` | str | `gpt-5.5` | Teacher model (for reflection & slow update) |
-| `model.student` | str | `gpt-5.5` | Student model (for rollout execution) |
+| `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`)
@@ -40,7 +40,7 @@ Complete reference for all SkillOpt configuration parameters.
 | `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 teacher-side strategy memory |
+| `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`)
@@ -49,6 +49,7 @@ nav:
    - 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
@@ -1,53 +0,0 @@
-#!/usr/bin/env python3
-"""Download BabyVision from Hugging Face and convert it to local meta_data.jsonl + images/ format."""
-from __future__ import annotations
-
-import argparse
-import json
-import os
-from pathlib import Path
-
-
-def parse_args() -> argparse.Namespace:
-    p = argparse.ArgumentParser(description=__doc__)
-    p.add_argument("--out_dir", type=str, required=True)
-    p.add_argument("--dataset", type=str, default="UnipatAI/BabyVision")
-    p.add_argument("--split", type=str, default="train")
-    return p.parse_args()
-
-
-def main() -> None:
-    args = parse_args()
-
-    try:
-        from datasets import load_dataset
-    except ImportError as exc:  # pragma: no cover
-        raise SystemExit("Please install `datasets` first: pip install datasets pillow") from exc
-
-    out_dir = Path(args.out_dir).resolve()
-    images_dir = out_dir / "images"
-    meta_path = out_dir / "meta_data.jsonl"
-    images_dir.mkdir(parents=True, exist_ok=True)
-
-    dataset = load_dataset(args.dataset, split=args.split)
-    with open(meta_path, "w", encoding="utf-8") as outf:
-        for idx, row in enumerate(dataset):
-            image = row.get("image")
-            if image is None:
-                continue
-            task_id = str(row.get("taskId") or row.get("id") or idx + 1)
-            image_name = f"{task_id}.png"
-            image_path = images_dir / image_name
-            image.save(image_path)
-
-            record = dict(row)
-            record["image"] = image_name
-            outf.write(json.dumps(record, ensure_ascii=False) + "\n")
-
-    print(f"Saved BabyVision to {out_dir}")
-    print(f"Metadata: {meta_path}")
-    print(f"Images:   {images_dir}")
-
-
-if __name__ == "__main__":
-    main()
@@ -1,5 +1,5 @@
 #!/usr/bin/env python3
-"""ReflACT eval-only: run a single skill on a dataset without training.
+"""SkillOpt eval-only: run a single skill on a dataset without training.

 Usage
 -----
@@ -29,10 +29,10 @@ from skillopt.model import (
    configure_claude_code_exec,
    configure_codex_exec,
    set_reasoning_effort,
-    set_student_backend,
-    set_student_deployment,
-    set_teacher_backend,
-    set_teacher_deployment,
+    set_target_backend,
+    set_target_deployment,
+    set_optimizer_backend,
+    set_optimizer_deployment,
 )
 from skillopt.model.common import default_model_for_backend, normalize_backend_name

@@ -126,7 +126,7 @@ _BOOL = lambda x: str(x).lower() in ("true", "1", "yes")  # noqa: E731


 def parse_args() -> argparse.Namespace:
-    p = argparse.ArgumentParser(description="ReflACT eval-only")
+    p = argparse.ArgumentParser(description="SkillOpt eval-only")
    p.add_argument("--config", type=str, required=True)
    p.add_argument("--skill", type=str, required=True,
                   help="Path to skill .md file to evaluate")
@@ -138,10 +138,10 @@ def parse_args() -> argparse.Namespace:
    p.add_argument("--env", type=str)
    p.add_argument("--backend", type=str,
                   choices=["azure_openai", "codex", "codex_exec", "claude", "claude_chat", "claude_code_exec"])
-    p.add_argument("--teacher_model", type=str)
-    p.add_argument("--student_model", type=str)
-    p.add_argument("--teacher_backend", type=str)
-    p.add_argument("--student_backend", type=str)
+    p.add_argument("--optimizer_model", type=str)
+    p.add_argument("--target_model", type=str)
+    p.add_argument("--optimizer_backend", type=str)
+    p.add_argument("--target_backend", type=str)
    p.add_argument("--reasoning_effort", type=str,
                   choices=["", "low", "medium", "high", "xhigh", "max"])
    p.add_argument("--azure_endpoint", type=str)
@@ -153,18 +153,18 @@ def parse_args() -> argparse.Namespace:
    p.add_argument("--azure_openai_auth_mode", type=str)
    p.add_argument("--azure_openai_ad_scope", type=str)
    p.add_argument("--azure_openai_managed_identity_client_id", type=str)
-    p.add_argument("--teacher_azure_openai_endpoint", type=str)
-    p.add_argument("--teacher_azure_openai_api_version", type=str)
-    p.add_argument("--teacher_azure_openai_api_key", type=str)
-    p.add_argument("--teacher_azure_openai_auth_mode", type=str)
-    p.add_argument("--teacher_azure_openai_ad_scope", type=str)
-    p.add_argument("--teacher_azure_openai_managed_identity_client_id", type=str)
-    p.add_argument("--student_azure_openai_endpoint", type=str)
-    p.add_argument("--student_azure_openai_api_version", type=str)
-    p.add_argument("--student_azure_openai_api_key", type=str)
-    p.add_argument("--student_azure_openai_auth_mode", type=str)
-    p.add_argument("--student_azure_openai_ad_scope", type=str)
-    p.add_argument("--student_azure_openai_managed_identity_client_id", type=str)
+    p.add_argument("--optimizer_azure_openai_endpoint", type=str)
+    p.add_argument("--optimizer_azure_openai_api_version", type=str)
+    p.add_argument("--optimizer_azure_openai_api_key", type=str)
+    p.add_argument("--optimizer_azure_openai_auth_mode", type=str)
+    p.add_argument("--optimizer_azure_openai_ad_scope", type=str)
+    p.add_argument("--optimizer_azure_openai_managed_identity_client_id", type=str)
+    p.add_argument("--target_azure_openai_endpoint", type=str)
+    p.add_argument("--target_azure_openai_api_version", type=str)
+    p.add_argument("--target_azure_openai_api_key", type=str)
+    p.add_argument("--target_azure_openai_auth_mode", type=str)
+    p.add_argument("--target_azure_openai_ad_scope", type=str)
+    p.add_argument("--target_azure_openai_managed_identity_client_id", type=str)
    p.add_argument("--codex_exec_path", type=str)
    p.add_argument("--codex_exec_sandbox", type=str)
    p.add_argument("--codex_exec_profile", type=str)
@@ -214,10 +214,10 @@ def main() -> None:
            from skillopt.config import apply_overrides
            _MAP = {
                "backend": "model.backend",
-                "teacher_model": "model.teacher",
-                "student_model": "model.student",
-                "teacher_backend": "model.teacher_backend",
-                "student_backend": "model.student_backend",
+                "optimizer_model": "model.optimizer",
+                "target_model": "model.target",
+                "optimizer_backend": "model.optimizer_backend",
+                "target_backend": "model.target_backend",
                "reasoning_effort": "model.reasoning_effort",
                "azure_endpoint": "model.azure_endpoint",
                "azure_api_version": "model.azure_api_version",
@@ -228,18 +228,18 @@ def main() -> None:
                "azure_openai_auth_mode": "model.azure_openai_auth_mode",
                "azure_openai_ad_scope": "model.azure_openai_ad_scope",
                "azure_openai_managed_identity_client_id": "model.azure_openai_managed_identity_client_id",
-                "teacher_azure_openai_endpoint": "model.teacher_azure_openai_endpoint",
-                "teacher_azure_openai_api_version": "model.teacher_azure_openai_api_version",
-                "teacher_azure_openai_api_key": "model.teacher_azure_openai_api_key",
-                "teacher_azure_openai_auth_mode": "model.teacher_azure_openai_auth_mode",
-                "teacher_azure_openai_ad_scope": "model.teacher_azure_openai_ad_scope",
-                "teacher_azure_openai_managed_identity_client_id": "model.teacher_azure_openai_managed_identity_client_id",
-                "student_azure_openai_endpoint": "model.student_azure_openai_endpoint",
-                "student_azure_openai_api_version": "model.student_azure_openai_api_version",
-                "student_azure_openai_api_key": "model.student_azure_openai_api_key",
-                "student_azure_openai_auth_mode": "model.student_azure_openai_auth_mode",
-                "student_azure_openai_ad_scope": "model.student_azure_openai_ad_scope",
-                "student_azure_openai_managed_identity_client_id": "model.student_azure_openai_managed_identity_client_id",
+                "optimizer_azure_openai_endpoint": "model.optimizer_azure_openai_endpoint",
+                "optimizer_azure_openai_api_version": "model.optimizer_azure_openai_api_version",
+                "optimizer_azure_openai_api_key": "model.optimizer_azure_openai_api_key",
+                "optimizer_azure_openai_auth_mode": "model.optimizer_azure_openai_auth_mode",
+                "optimizer_azure_openai_ad_scope": "model.optimizer_azure_openai_ad_scope",
+                "optimizer_azure_openai_managed_identity_client_id": "model.optimizer_azure_openai_managed_identity_client_id",
+                "target_azure_openai_endpoint": "model.target_azure_openai_endpoint",
+                "target_azure_openai_api_version": "model.target_azure_openai_api_version",
+                "target_azure_openai_api_key": "model.target_azure_openai_api_key",
+                "target_azure_openai_auth_mode": "model.target_azure_openai_auth_mode",
+                "target_azure_openai_ad_scope": "model.target_azure_openai_ad_scope",
+                "target_azure_openai_managed_identity_client_id": "model.target_azure_openai_managed_identity_client_id",
                "codex_exec_path": "model.codex_exec_path",
                "codex_exec_sandbox": "model.codex_exec_sandbox",
                "codex_exec_profile": "model.codex_exec_profile",
@@ -288,7 +288,7 @@ def main() -> None:
                explicit_backend = str(option).split("=", 1)[1].strip()
                break

-    backend = normalize_backend_name(cfg.get("model_backend") or cfg.get("student_backend") or "azure_openai")
+    backend = normalize_backend_name(cfg.get("model_backend") or cfg.get("target_backend") or "azure_openai")

    def _has_model_override(dotted_key: str, legacy_key: str) -> bool:
        if getattr(args, legacy_key, None) is not None:
@@ -303,43 +303,43 @@ def main() -> None:
        backend = normalize_backend_name(explicit_backend)
        cfg["model_backend"] = backend
        if backend in {"claude", "claude_chat"}:
-            cfg.setdefault("teacher_backend", "claude_chat")
-            cfg.setdefault("student_backend", "claude_chat")
+            cfg.setdefault("optimizer_backend", "claude_chat")
+            cfg.setdefault("target_backend", "claude_chat")
        elif backend in {"codex", "codex_exec"}:
-            cfg.setdefault("teacher_backend", "openai_chat")
-            cfg.setdefault("student_backend", "codex_exec")
+            cfg.setdefault("optimizer_backend", "openai_chat")
+            cfg.setdefault("target_backend", "codex_exec")
        elif backend == "claude_code_exec":
-            cfg.setdefault("teacher_backend", "openai_chat")
-            cfg.setdefault("student_backend", "claude_code_exec")
+            cfg.setdefault("optimizer_backend", "openai_chat")
+            cfg.setdefault("target_backend", "claude_code_exec")
        else:
-            cfg.setdefault("teacher_backend", "openai_chat")
-            cfg.setdefault("student_backend", "openai_chat")
+            cfg.setdefault("optimizer_backend", "openai_chat")
+            cfg.setdefault("target_backend", "openai_chat")
    else:
-        cfg.setdefault("teacher_backend", "openai_chat")
-        cfg.setdefault("student_backend", "openai_chat")
+        cfg.setdefault("optimizer_backend", "openai_chat")
+        cfg.setdefault("target_backend", "openai_chat")

-    if cfg.get("teacher_backend") == "claude_chat":
+    if cfg.get("optimizer_backend") == "claude_chat":
        if (
-            str(cfg.get("teacher_model", "") or "").strip() in _OPENAI_DEFAULT_MODEL_SENTINELS
-            and not _has_model_override("model.teacher", "teacher_model")
+            str(cfg.get("optimizer_model", "") or "").strip() in _OPENAI_DEFAULT_MODEL_SENTINELS
+            and not _has_model_override("model.optimizer", "optimizer_model")
        ):
-            cfg["teacher_model"] = default_model_for_backend("claude_chat")
-    if cfg.get("student_backend") == "claude_chat":
+            cfg["optimizer_model"] = default_model_for_backend("claude_chat")
+    if cfg.get("target_backend") == "claude_chat":
        if (
-            str(cfg.get("student_model", "") or "").strip() in _OPENAI_DEFAULT_MODEL_SENTINELS
-            and not _has_model_override("model.student", "student_model")
+            str(cfg.get("target_model", "") or "").strip() in _OPENAI_DEFAULT_MODEL_SENTINELS
+            and not _has_model_override("model.target", "target_model")
        ):
-            cfg["student_model"] = default_model_for_backend("claude_chat")
-    if cfg.get("student_backend") == "claude_code_exec":
+            cfg["target_model"] = default_model_for_backend("claude_chat")
+    if cfg.get("target_backend") == "claude_code_exec":
        if (
-            str(cfg.get("student_model", "") or "").strip() in _OPENAI_DEFAULT_MODEL_SENTINELS
-            and not _has_model_override("model.student", "student_model")
+            str(cfg.get("target_model", "") or "").strip() in _OPENAI_DEFAULT_MODEL_SENTINELS
+            and not _has_model_override("model.target", "target_model")
        ):
-            cfg["student_model"] = default_model_for_backend("claude_chat")
+            cfg["target_model"] = default_model_for_backend("claude_chat")

    if not cfg.get("out_root"):
        env = cfg.get("env", "unknown")
-        model = cfg.get("student_model", "unknown").replace("/", "-")
+        model = cfg.get("target_model", "unknown").replace("/", "-")
        ts = datetime.datetime.now().strftime("%Y%m%d_%H%M%S")
        cfg["out_root"] = os.path.join("outputs", f"eval_{env}_{model}_{ts}")

@@ -362,27 +362,27 @@ def main() -> None:
        auth_mode=cfg.get("azure_openai_auth_mode") or None,
        ad_scope=cfg.get("azure_openai_ad_scope") or None,
        managed_identity_client_id=cfg.get("azure_openai_managed_identity_client_id") or None,
-        teacher_endpoint=cfg.get("teacher_azure_openai_endpoint") or None,
-        teacher_api_version=cfg.get("teacher_azure_openai_api_version") or None,
-        teacher_api_key=cfg.get("teacher_azure_openai_api_key") or None,
-        teacher_auth_mode=cfg.get("teacher_azure_openai_auth_mode") or None,
-        teacher_ad_scope=cfg.get("teacher_azure_openai_ad_scope") or None,
-        teacher_managed_identity_client_id=(
-            cfg.get("teacher_azure_openai_managed_identity_client_id") or None
+        optimizer_endpoint=cfg.get("optimizer_azure_openai_endpoint") or None,
+        optimizer_api_version=cfg.get("optimizer_azure_openai_api_version") or None,
+        optimizer_api_key=cfg.get("optimizer_azure_openai_api_key") or None,
+        optimizer_auth_mode=cfg.get("optimizer_azure_openai_auth_mode") or None,
+        optimizer_ad_scope=cfg.get("optimizer_azure_openai_ad_scope") or None,
+        optimizer_managed_identity_client_id=(
+            cfg.get("optimizer_azure_openai_managed_identity_client_id") or None
        ),
-        student_endpoint=cfg.get("student_azure_openai_endpoint") or None,
-        student_api_version=cfg.get("student_azure_openai_api_version") or None,
-        student_api_key=cfg.get("student_azure_openai_api_key") or None,
-        student_auth_mode=cfg.get("student_azure_openai_auth_mode") or None,
-        student_ad_scope=cfg.get("student_azure_openai_ad_scope") or None,
-        student_managed_identity_client_id=(
-            cfg.get("student_azure_openai_managed_identity_client_id") or None
+        target_endpoint=cfg.get("target_azure_openai_endpoint") or None,
+        target_api_version=cfg.get("target_azure_openai_api_version") or None,
+        target_api_key=cfg.get("target_azure_openai_api_key") or None,
+        target_auth_mode=cfg.get("target_azure_openai_auth_mode") or None,
+        target_ad_scope=cfg.get("target_azure_openai_ad_scope") or None,
+        target_managed_identity_client_id=(
+            cfg.get("target_azure_openai_managed_identity_client_id") or None
        ),
    )
-    set_teacher_backend(cfg.get("teacher_backend", "openai_chat"))
-    set_student_backend(cfg.get("student_backend", "openai_chat"))
-    set_teacher_deployment(cfg.get("teacher_model", default_model_for_backend(backend)))
-    set_student_deployment(cfg.get("student_model", default_model_for_backend(backend)))
+    set_optimizer_backend(cfg.get("optimizer_backend", "openai_chat"))
+    set_target_backend(cfg.get("target_backend", "openai_chat"))
+    set_optimizer_deployment(cfg.get("optimizer_model", default_model_for_backend(backend)))
+    set_target_deployment(cfg.get("target_model", default_model_for_backend(backend)))
    configure_codex_exec(
        path=cfg.get("codex_exec_path", "codex"),
        sandbox=cfg.get("codex_exec_sandbox", "workspace-write"),
@@ -1,28 +1,26 @@
 #!/usr/bin/env bash
 # ──────────────────────────────────────────────────────────────────────────────
-# ReflACT — ALFWorld training launch script
+# SkillOpt — ALFWorld training launch script
+#
+# Prerequisites:
+#   pip install -e ".[alfworld]"
+#   pip install alfworld[full] && alfworld-download
 #
 # Usage:
 #   bash scripts/run_alfworld.sh
 #   bash scripts/run_alfworld.sh --num_epochs 2 --edit_budget 6
+#   bash scripts/run_alfworld.sh --split_dir /path/to/alfworld_split
 # ──────────────────────────────────────────────────────────────────────────────
 set -euo pipefail

-# ── Paths ────────────────────────────────────────────────────────────────────
-WORKSPACE="${WORKSPACE:-$(cd "$(dirname "$0")/../.." && pwd)}"
 SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
 PROJECT_ROOT="$(dirname "${SCRIPT_DIR}")"

-# Activate conda environment
-export PATH="${WORKSPACE}/miniconda3/envs/skillopt/bin:${WORKSPACE}/miniconda3/bin:${PATH}"
-
-# ALFWorld data — uses ~/.cache/alfworld by default (standard alfworld location)
-export ALFWORLD_DATA="${ALFWORLD_DATA:-${HOME}/.cache/alfworld}"
-
-# Ensure ReflACT is importable
 export PYTHONPATH="${PROJECT_ROOT}:${PYTHONPATH:-}"

-# ── Verify ALFWorld data exists ──────────────────────────────────────────────
+# ALFWorld data — uses ~/.cache/alfworld by default
+export ALFWORLD_DATA="${ALFWORLD_DATA:-${HOME}/.cache/alfworld}"
+
 if [ ! -d "${ALFWORLD_DATA}/json_2.1.1" ]; then
    echo "ERROR: ALFWorld data not found at ${ALFWORLD_DATA}/json_2.1.1"
    echo ""
@@ -34,25 +32,17 @@ if [ ! -d "${ALFWORLD_DATA}/json_2.1.1" ]; then
    exit 1
 fi

-# ── Azure OpenAI credentials ────────────────────────────────────────────────
-export AZURE_OPENAI_ENDPOINT="${AZURE_OPENAI_ENDPOINT:?Set AZURE_OPENAI_ENDPOINT}"
-export AZURE_OPENAI_API_KEY="${AZURE_OPENAI_API_KEY:?Set AZURE_OPENAI_API_KEY}"
-export AZURE_OPENAI_API_VERSION="${AZURE_OPENAI_API_VERSION:-2025-04-01-preview}"
+OPTIMIZER_MODEL="${OPTIMIZER_MODEL:-gpt-5.5}"
+TARGET_MODEL="${TARGET_MODEL:-gpt-5.5}"

-# ── Model configuration ─────────────────────────────────────────────────────
-export TEACHER_DEPLOYMENT="${TEACHER_DEPLOYMENT:-gpt-5.5}"
-export STUDENT_DEPLOYMENT="${STUDENT_DEPLOYMENT:-gpt-5.5}"
-
-# ── Output directory ─────────────────────────────────────────────────────────
 TIMESTAMP=$(date +%Y%m%d_%H%M%S)
-DEFAULT_OUT_ROOT="${PROJECT_ROOT}/outputs/skillopt_alfworld_${STUDENT_DEPLOYMENT}_${TIMESTAMP}"
+DEFAULT_OUT_ROOT="${PROJECT_ROOT}/outputs/skillopt_alfworld_${TARGET_MODEL}_${TIMESTAMP}"

-# ── Run ──────────────────────────────────────────────────────────────────────
 echo "============================================================"
-echo "  ReflACT — Reflective Agent Tuning (ALFWorld)"
+echo "  SkillOpt — ALFWorld Training"
 echo "============================================================"
-echo "  Teacher:       ${TEACHER_DEPLOYMENT}"
-echo "  Student:       ${STUDENT_DEPLOYMENT}"
+echo "  Optimizer:       ${OPTIMIZER_MODEL}"
+echo "  Target:       ${TARGET_MODEL}"
 echo "  ALFWORLD_DATA: ${ALFWORLD_DATA}"
 echo "  Output:        ${DEFAULT_OUT_ROOT}"
 echo "============================================================"
@@ -60,7 +50,9 @@ echo "============================================================"
 cd "${PROJECT_ROOT}"

 python scripts/train.py \
-    --config configs/alfworld_default.yaml \
+    --config configs/alfworld/default.yaml \
+    --optimizer_model "${OPTIMIZER_MODEL}" \
+    --target_model "${TARGET_MODEL}" \
    --out_root "${DEFAULT_OUT_ROOT}" \
    "$@"

@@ -1,41 +1,38 @@
 #!/usr/bin/env bash
 # ──────────────────────────────────────────────────────────────────────────────
-# ReflACT — SearchQA training launch script
+# SkillOpt — SearchQA training launch script
 #
 # Usage:
 #   bash scripts/run_searchqa.sh
-#   bash scripts/run_searchqa.sh --data_path data/searchqa_train_2000.json
 #   bash scripts/run_searchqa.sh --num_epochs 2 --edit_budget 6
+#   bash scripts/run_searchqa.sh --split_dir /path/to/searchqa_split
 # ──────────────────────────────────────────────────────────────────────────────
 set -euo pipefail

-# ── Paths ────────────────────────────────────────────────────────────────────
 SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
 PROJECT_ROOT="$(dirname "${SCRIPT_DIR}")"

-# Ensure ReflACT is importable
 export PYTHONPATH="${PROJECT_ROOT}:${PYTHONPATH:-}"

-# ── Model configuration ─────────────────────────────────────────────────────
-export TEACHER_DEPLOYMENT="${TEACHER_DEPLOYMENT:-gpt-5.5}"
-export STUDENT_DEPLOYMENT="${STUDENT_DEPLOYMENT:-gpt-5.5}"
+OPTIMIZER_MODEL="${OPTIMIZER_MODEL:-gpt-5.5}"
+TARGET_MODEL="${TARGET_MODEL:-gpt-5.5}"

-# ── Output directory ─────────────────────────────────────────────────────────
 TIMESTAMP=$(date +%Y%m%d_%H%M%S)
-DEFAULT_OUT_ROOT="${PROJECT_ROOT}/outputs/skillopt_searchqa_${STUDENT_DEPLOYMENT}_${TIMESTAMP}"
+DEFAULT_OUT_ROOT="${PROJECT_ROOT}/outputs/skillopt_searchqa_${TARGET_MODEL}_${TIMESTAMP}"

-# ── Run ──────────────────────────────────────────────────────────────────────
 echo "============================================================"
-echo "  ReflACT — Reflective Agent Tuning (SearchQA)"
+echo "  SkillOpt — SearchQA Training"
 echo "============================================================"
-echo "  Teacher:  ${TEACHER_DEPLOYMENT}"
-echo "  Student:  ${STUDENT_DEPLOYMENT}"
+echo "  Optimizer:  ${OPTIMIZER_MODEL}"
+echo "  Target:  ${TARGET_MODEL}"
 echo "============================================================"

 cd "${PROJECT_ROOT}"

 python scripts/train.py \
-    --config configs/searchqa_default.yaml \
+    --config configs/searchqa/default.yaml \
+    --optimizer_model "${OPTIMIZER_MODEL}" \
+    --target_model "${TARGET_MODEL}" \
    --out_root "${DEFAULT_OUT_ROOT}" \
    "$@"

@@ -1,46 +1,37 @@
 #!/usr/bin/env bash
 # ──────────────────────────────────────────────────────────────────────────────
-# ReflACT — SpreadsheetBench training launch script
+# SkillOpt — SpreadsheetBench training launch script
 #
 # Usage:
-#   bash scripts/run_spreadsheetbench.sh \
-#       --data_root /path/to/data \
-#       --jsonl_path /path/to/benchmark.jsonl
-#
-#   bash scripts/run_spreadsheetbench.sh \
-#       --data_root /path/to/data \
-#       --jsonl_path /path/to/benchmark.jsonl \
-#       --num_epochs 2 --edit_budget 6
+#   bash scripts/run_spreadsheetbench.sh --split_dir /path/to/split --data_root /path/to/data
+#   bash scripts/run_spreadsheetbench.sh --num_epochs 2 --edit_budget 6
 # ──────────────────────────────────────────────────────────────────────────────
 set -euo pipefail

-# ── Paths ────────────────────────────────────────────────────────────────────
 SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
 PROJECT_ROOT="$(dirname "${SCRIPT_DIR}")"

-# Ensure ReflACT is importable
 export PYTHONPATH="${PROJECT_ROOT}:${PYTHONPATH:-}"

-# ── Model configuration ─────────────────────────────────────────────────────
-export TEACHER_DEPLOYMENT="${TEACHER_DEPLOYMENT:-gpt-5.5}"
-export STUDENT_DEPLOYMENT="${STUDENT_DEPLOYMENT:-gpt-5.5}"
+OPTIMIZER_MODEL="${OPTIMIZER_MODEL:-gpt-5.5}"
+TARGET_MODEL="${TARGET_MODEL:-gpt-5.5}"

-# ── Output directory ─────────────────────────────────────────────────────────
 TIMESTAMP=$(date +%Y%m%d_%H%M%S)
-DEFAULT_OUT_ROOT="${PROJECT_ROOT}/outputs/skillopt_spreadsheetbench_${STUDENT_DEPLOYMENT}_${TIMESTAMP}"
+DEFAULT_OUT_ROOT="${PROJECT_ROOT}/outputs/skillopt_spreadsheetbench_${TARGET_MODEL}_${TIMESTAMP}"

-# ── Run ──────────────────────────────────────────────────────────────────────
 echo "============================================================"
-echo "  ReflACT — Reflective Agent Tuning (SpreadsheetBench)"
+echo "  SkillOpt — SpreadsheetBench Training"
 echo "============================================================"
-echo "  Teacher:  ${TEACHER_DEPLOYMENT}"
-echo "  Student:  ${STUDENT_DEPLOYMENT}"
+echo "  Optimizer:  ${OPTIMIZER_MODEL}"
+echo "  Target:  ${TARGET_MODEL}"
 echo "============================================================"

 cd "${PROJECT_ROOT}"

 python scripts/train.py \
-    --config configs/spreadsheetbench_default.yaml \
+    --config configs/spreadsheetbench/default.yaml \
+    --optimizer_model "${OPTIMIZER_MODEL}" \
+    --target_model "${TARGET_MODEL}" \
    --out_root "${DEFAULT_OUT_ROOT}" \
    "$@"

@@ -1,5 +1,5 @@
 #!/usr/bin/env python3
-"""ReflACT unified training entry point.
+"""SkillOpt unified training entry point.

 Usage
 -----
@@ -125,7 +125,7 @@ _BOOL = lambda x: x.lower() in ("true", "1", "yes")  # noqa: E731

 def parse_args() -> argparse.Namespace:
    p = argparse.ArgumentParser(
-        description="ReflACT: Reflective Agent Tuning",
+        description="SkillOpt: Executive Strategy for Self-Evolving Agent Skills",
        formatter_class=argparse.RawDescriptionHelpFormatter,
        epilog=__doc__,
    )
@@ -138,10 +138,10 @@ def parse_args() -> argparse.Namespace:
    p.add_argument("--env", type=str)
    p.add_argument("--backend", type=str,
                   choices=["azure_openai", "codex", "codex_exec", "claude", "claude_chat", "claude_code_exec", "qwen", "qwen_chat"])
-    p.add_argument("--teacher_model", type=str)
-    p.add_argument("--student_model", type=str)
-    p.add_argument("--teacher_backend", type=str)
-    p.add_argument("--student_backend", type=str)
+    p.add_argument("--optimizer_model", type=str)
+    p.add_argument("--target_model", type=str)
+    p.add_argument("--optimizer_backend", type=str)
+    p.add_argument("--target_backend", type=str)
    p.add_argument("--reasoning_effort", type=str,
                   choices=["", "low", "medium", "high", "xhigh", "max"])
    p.add_argument("--rewrite_reasoning_effort", type=str)
@@ -155,18 +155,18 @@ def parse_args() -> argparse.Namespace:
    p.add_argument("--azure_openai_auth_mode", type=str)
    p.add_argument("--azure_openai_ad_scope", type=str)
    p.add_argument("--azure_openai_managed_identity_client_id", type=str)
-    p.add_argument("--teacher_azure_openai_endpoint", type=str)
-    p.add_argument("--teacher_azure_openai_api_version", type=str)
-    p.add_argument("--teacher_azure_openai_api_key", type=str)
-    p.add_argument("--teacher_azure_openai_auth_mode", type=str)
-    p.add_argument("--teacher_azure_openai_ad_scope", type=str)
-    p.add_argument("--teacher_azure_openai_managed_identity_client_id", type=str)
-    p.add_argument("--student_azure_openai_endpoint", type=str)
-    p.add_argument("--student_azure_openai_api_version", type=str)
-    p.add_argument("--student_azure_openai_api_key", type=str)
-    p.add_argument("--student_azure_openai_auth_mode", type=str)
-    p.add_argument("--student_azure_openai_ad_scope", type=str)
-    p.add_argument("--student_azure_openai_managed_identity_client_id", type=str)
+    p.add_argument("--optimizer_azure_openai_endpoint", type=str)
+    p.add_argument("--optimizer_azure_openai_api_version", type=str)
+    p.add_argument("--optimizer_azure_openai_api_key", type=str)
+    p.add_argument("--optimizer_azure_openai_auth_mode", type=str)
+    p.add_argument("--optimizer_azure_openai_ad_scope", type=str)
+    p.add_argument("--optimizer_azure_openai_managed_identity_client_id", type=str)
+    p.add_argument("--target_azure_openai_endpoint", type=str)
+    p.add_argument("--target_azure_openai_api_version", type=str)
+    p.add_argument("--target_azure_openai_api_key", type=str)
+    p.add_argument("--target_azure_openai_auth_mode", type=str)
+    p.add_argument("--target_azure_openai_ad_scope", type=str)
+    p.add_argument("--target_azure_openai_managed_identity_client_id", type=str)
    p.add_argument("--qwen_chat_base_url", type=str)
    p.add_argument("--qwen_chat_api_key", type=str)
    p.add_argument("--qwen_chat_temperature", type=float)
@@ -187,7 +187,7 @@ def parse_args() -> argparse.Namespace:
    p.add_argument("--claude_code_exec_use_sdk", type=str)
    p.add_argument("--claude_code_exec_effort", type=str)
    p.add_argument("--claude_code_exec_max_thinking_tokens", type=int)
-    p.add_argument("--codex_trace_to_teacher", type=_BOOL)
+    p.add_argument("--codex_trace_to_optimizer", type=_BOOL)
    p.add_argument("--skill_init", type=str)
    p.add_argument("--num_epochs", type=int)
    p.add_argument("--train_size", type=int)
@@ -212,8 +212,6 @@ def parse_args() -> argparse.Namespace:
    p.add_argument("--analyst_workers", type=int)
    p.add_argument("--failure_only", type=_BOOL)
    p.add_argument("--minibatch_size", type=int)
-    p.add_argument("--use_meta_reflect", type=_BOOL)
-    p.add_argument("--meta_edit_budget", type=int)
    p.add_argument("--skill_update_mode", type=str,
                   choices=[
                       "patch",
@@ -224,9 +222,6 @@ def parse_args() -> argparse.Namespace:
                       "full_rewrite_minibatch",
                       "minibatch_full_rewrite",
                   ])
-    p.add_argument("--use_deep_reflect", type=_BOOL)
-    p.add_argument("--deep_reflect_failures", type=int)
-    p.add_argument("--deep_reflect_successes", type=int)
    p.add_argument("--use_slow_update", type=_BOOL)
    p.add_argument("--slow_update_samples", type=int)
    p.add_argument("--longitudinal_pair_policy", type=str,
@@ -260,10 +255,10 @@ def parse_args() -> argparse.Namespace:

 _LEGACY_TO_STRUCTURED: dict[str, str] = {
    "backend": "model.backend",
-    "teacher_model": "model.teacher",
-    "student_model": "model.student",
-    "teacher_backend": "model.teacher_backend",
-    "student_backend": "model.student_backend",
+    "optimizer_model": "model.optimizer",
+    "target_model": "model.target",
+    "optimizer_backend": "model.optimizer_backend",
+    "target_backend": "model.target_backend",
    "reasoning_effort": "model.reasoning_effort",
    "rewrite_reasoning_effort": "model.rewrite_reasoning_effort",
    "rewrite_max_completion_tokens": "model.rewrite_max_completion_tokens",
@@ -276,18 +271,18 @@ _LEGACY_TO_STRUCTURED: dict[str, str] = {
    "azure_openai_auth_mode": "model.azure_openai_auth_mode",
    "azure_openai_ad_scope": "model.azure_openai_ad_scope",
    "azure_openai_managed_identity_client_id": "model.azure_openai_managed_identity_client_id",
-    "teacher_azure_openai_endpoint": "model.teacher_azure_openai_endpoint",
-    "teacher_azure_openai_api_version": "model.teacher_azure_openai_api_version",
-    "teacher_azure_openai_api_key": "model.teacher_azure_openai_api_key",
-    "teacher_azure_openai_auth_mode": "model.teacher_azure_openai_auth_mode",
-    "teacher_azure_openai_ad_scope": "model.teacher_azure_openai_ad_scope",
-    "teacher_azure_openai_managed_identity_client_id": "model.teacher_azure_openai_managed_identity_client_id",
-    "student_azure_openai_endpoint": "model.student_azure_openai_endpoint",
-    "student_azure_openai_api_version": "model.student_azure_openai_api_version",
-    "student_azure_openai_api_key": "model.student_azure_openai_api_key",
-    "student_azure_openai_auth_mode": "model.student_azure_openai_auth_mode",
-    "student_azure_openai_ad_scope": "model.student_azure_openai_ad_scope",
-    "student_azure_openai_managed_identity_client_id": "model.student_azure_openai_managed_identity_client_id",
+    "optimizer_azure_openai_endpoint": "model.optimizer_azure_openai_endpoint",
+    "optimizer_azure_openai_api_version": "model.optimizer_azure_openai_api_version",
+    "optimizer_azure_openai_api_key": "model.optimizer_azure_openai_api_key",
+    "optimizer_azure_openai_auth_mode": "model.optimizer_azure_openai_auth_mode",
+    "optimizer_azure_openai_ad_scope": "model.optimizer_azure_openai_ad_scope",
+    "optimizer_azure_openai_managed_identity_client_id": "model.optimizer_azure_openai_managed_identity_client_id",
+    "target_azure_openai_endpoint": "model.target_azure_openai_endpoint",
+    "target_azure_openai_api_version": "model.target_azure_openai_api_version",
+    "target_azure_openai_api_key": "model.target_azure_openai_api_key",
+    "target_azure_openai_auth_mode": "model.target_azure_openai_auth_mode",
+    "target_azure_openai_ad_scope": "model.target_azure_openai_ad_scope",
+    "target_azure_openai_managed_identity_client_id": "model.target_azure_openai_managed_identity_client_id",
    "qwen_chat_base_url": "model.qwen_chat_base_url",
    "qwen_chat_api_key": "model.qwen_chat_api_key",
    "qwen_chat_temperature": "model.qwen_chat_temperature",
@@ -308,7 +303,7 @@ _LEGACY_TO_STRUCTURED: dict[str, str] = {
    "claude_code_exec_use_sdk": "model.claude_code_exec_use_sdk",
    "claude_code_exec_effort": "model.claude_code_exec_effort",
    "claude_code_exec_max_thinking_tokens": "model.claude_code_exec_max_thinking_tokens",
-    "codex_trace_to_teacher": "model.codex_trace_to_teacher",
+    "codex_trace_to_optimizer": "model.codex_trace_to_optimizer",
    "num_epochs": "train.num_epochs",
    "train_size": "train.train_size",
    "steps_per_epoch": "train.steps_per_epoch",
@@ -320,16 +315,11 @@ _LEGACY_TO_STRUCTURED: dict[str, str] = {
    "analyst_workers": "gradient.analyst_workers",
    "max_analyst_rounds": "gradient.max_analyst_rounds",
    "failure_only": "gradient.failure_only",
-    "use_deep_reflect": "gradient.use_deep_reflect",
-    "deep_reflect_failures": "gradient.deep_reflect_failures",
-    "deep_reflect_successes": "gradient.deep_reflect_successes",
    "edit_budget": "optimizer.learning_rate",
    "min_edit_budget": "optimizer.min_learning_rate",
    "lr_scheduler": "optimizer.lr_scheduler",
    "lr_control_mode": "optimizer.lr_control_mode",
    "skill_update_mode": "optimizer.skill_update_mode",
-    "use_meta_reflect": "optimizer.use_meta_reflect",
-    "meta_edit_budget": "optimizer.meta_learning_rate",
    "use_slow_update": "optimizer.use_slow_update",
    "slow_update_samples": "optimizer.slow_update_samples",
    "longitudinal_pair_policy": "optimizer.longitudinal_pair_policy",
@@ -387,7 +377,7 @@ def load_config(args: argparse.Namespace) -> dict:
                explicit_backend = str(option).split("=", 1)[1].strip()
                break

-    backend = normalize_backend_name(flat.get("model_backend") or flat.get("student_backend") or "azure_openai")
+    backend = normalize_backend_name(flat.get("model_backend") or flat.get("target_backend") or "azure_openai")

    def _has_model_override(dotted_key: str, legacy_key: str) -> bool:
        if getattr(args, legacy_key, None) is not None:
@@ -402,53 +392,53 @@ def load_config(args: argparse.Namespace) -> dict:
        backend = normalize_backend_name(explicit_backend)
        flat["model_backend"] = backend
        if backend in {"claude", "claude_chat"}:
-            flat.setdefault("teacher_backend", "claude_chat")
-            flat.setdefault("student_backend", "claude_chat")
+            flat.setdefault("optimizer_backend", "claude_chat")
+            flat.setdefault("target_backend", "claude_chat")
        elif backend in {"codex", "codex_exec"}:
-            flat.setdefault("teacher_backend", "openai_chat")
-            flat.setdefault("student_backend", "codex_exec")
+            flat.setdefault("optimizer_backend", "openai_chat")
+            flat.setdefault("target_backend", "codex_exec")
        elif backend == "claude_code_exec":
-            flat.setdefault("teacher_backend", "openai_chat")
-            flat.setdefault("student_backend", "claude_code_exec")
+            flat.setdefault("optimizer_backend", "openai_chat")
+            flat.setdefault("target_backend", "claude_code_exec")
        elif backend in {"qwen", "qwen_chat"}:
-            flat.setdefault("teacher_backend", "openai_chat")
-            flat.setdefault("student_backend", "qwen_chat")
+            flat.setdefault("optimizer_backend", "openai_chat")
+            flat.setdefault("target_backend", "qwen_chat")
        else:
-            flat.setdefault("teacher_backend", "openai_chat")
-            flat.setdefault("student_backend", "openai_chat")
+            flat.setdefault("optimizer_backend", "openai_chat")
+            flat.setdefault("target_backend", "openai_chat")
    else:
-        flat.setdefault("teacher_backend", "openai_chat")
-        flat.setdefault("student_backend", "openai_chat")
+        flat.setdefault("optimizer_backend", "openai_chat")
+        flat.setdefault("target_backend", "openai_chat")

-    if flat.get("teacher_backend") == "claude_chat":
+    if flat.get("optimizer_backend") == "claude_chat":
        if (
-            str(flat.get("teacher_model", "") or "").strip() in _OPENAI_DEFAULT_MODEL_SENTINELS
-            and not _has_model_override("model.teacher", "teacher_model")
+            str(flat.get("optimizer_model", "") or "").strip() in _OPENAI_DEFAULT_MODEL_SENTINELS
+            and not _has_model_override("model.optimizer", "optimizer_model")
        ):
-            flat["teacher_model"] = default_model_for_backend("claude_chat")
-    if flat.get("student_backend") == "claude_chat":
+            flat["optimizer_model"] = default_model_for_backend("claude_chat")
+    if flat.get("target_backend") == "claude_chat":
        if (
-            str(flat.get("student_model", "") or "").strip() in _OPENAI_DEFAULT_MODEL_SENTINELS
-            and not _has_model_override("model.student", "student_model")
+            str(flat.get("target_model", "") or "").strip() in _OPENAI_DEFAULT_MODEL_SENTINELS
+            and not _has_model_override("model.target", "target_model")
        ):
-            flat["student_model"] = default_model_for_backend("claude_chat")
-    if flat.get("student_backend") == "claude_code_exec":
+            flat["target_model"] = default_model_for_backend("claude_chat")
+    if flat.get("target_backend") == "claude_code_exec":
        if (
-            str(flat.get("student_model", "") or "").strip() in _OPENAI_DEFAULT_MODEL_SENTINELS
-            and not _has_model_override("model.student", "student_model")
+            str(flat.get("target_model", "") or "").strip() in _OPENAI_DEFAULT_MODEL_SENTINELS
+            and not _has_model_override("model.target", "target_model")
        ):
-            flat["student_model"] = default_model_for_backend("claude_chat")
-    if flat.get("student_backend") == "qwen_chat":
+            flat["target_model"] = default_model_for_backend("claude_chat")
+    if flat.get("target_backend") == "qwen_chat":
        if (
-            str(flat.get("student_model", "") or "").strip() in _OPENAI_DEFAULT_MODEL_SENTINELS
-            and not _has_model_override("model.student", "student_model")
+            str(flat.get("target_model", "") or "").strip() in _OPENAI_DEFAULT_MODEL_SENTINELS
+            and not _has_model_override("model.target", "target_model")
        ):
-            flat["student_model"] = default_model_for_backend("qwen_chat")
+            flat["target_model"] = default_model_for_backend("qwen_chat")

    # Auto-generate output root
    if not flat.get("out_root"):
        env = flat.get("env", "unknown")
-        model = flat.get("teacher_model", "unknown").replace("/", "-")
+        model = flat.get("optimizer_model", "unknown").replace("/", "-")
        ts = datetime.datetime.now().strftime("%Y%m%d_%H%M%S")
        flat["out_root"] = os.path.join("outputs", f"skillopt_{env}_{model}_{ts}")

@@ -463,13 +453,13 @@ def main() -> None:
    cfg = load_config(args)

    print(f"\n{'='*60}")
-    print(f"  ReflACT — Reflective Agent Tuning")
+    print(f"  SkillOpt — Executive Strategy for Self-Evolving Agent Skills")
    print(f"{'='*60}")
    print(f"  env:            {cfg.get('env')}")
-    print(f"  teacher_model:  {cfg.get('teacher_model')}")
-    print(f"  student_model:  {cfg.get('student_model')}")
-    print(f"  teacher_backend:{cfg.get('teacher_backend', 'openai_chat')}")
-    print(f"  student_backend:{cfg.get('student_backend', 'openai_chat')}")
+    print(f"  optimizer_model:  {cfg.get('optimizer_model')}")
+    print(f"  target_model:  {cfg.get('target_model')}")
+    print(f"  optimizer_backend:{cfg.get('optimizer_backend', 'openai_chat')}")
+    print(f"  target_backend:{cfg.get('target_backend', 'openai_chat')}")
    print(f"  reasoning:      {cfg.get('reasoning_effort') or 'off'}")
    print(f"  rewrite_effort: {cfg.get('rewrite_reasoning_effort') or 'off'}")
    print(f"  epochs:         {cfg.get('num_epochs')}")
@@ -482,8 +472,8 @@ def main() -> None:
    print(f"  min_edit_budget:{cfg.get('min_edit_budget', 2)}")
    print(f"  minibatch_size: {cfg.get('minibatch_size')}")
    print(f"  seed:           {cfg.get('seed')}")
-    print(f"  meta_reflect:   {cfg.get('use_meta_reflect', False)}")
    print(f"  meta_skill:     {cfg.get('use_meta_skill', False)}")
+    print(f"  slow_update:    {cfg.get('use_slow_update', False)}")
    print(f"  out_root:       {cfg.get('out_root')}")
    print(f"{'='*60}\n")

@@ -0,0 +1 @@
+<svg id="logomark" xmlns="http://www.w3.org/2000/svg" viewBox="0 0 74.492 100.25"><g id="tiny_-_black" data-name="tiny - black"><path d="M586.72,255.616a3.377,3.377,0,0,1,.448.031,5.917,5.917,0,0,1,3.581,2.79c.454,1.116.314,2.023-1.315,4.141L563.168,293.6l-8.558-10.047,29.348-26.616a4.406,4.406,0,0,1,2.762-1.321m0-1.5a5.766,5.766,0,0,0-3.69,1.643l-.041.032-.038.035L553.6,282.442l-1.077.977.943,1.107,8.558,10.047,1.145,1.344,1.141-1.348,26.267-31.022.022-.027.022-.028c1.574-2.046,2.327-3.622,1.516-5.619a7.309,7.309,0,0,0-4.779-3.714,5.083,5.083,0,0,0-.64-.043Z" transform="translate(-526.086 -245.559)"/><path d="M553.423,284.593l8.977,10.558L597.911,337.9c.873,1.093,1.419,2.186,1.047,3.418a4.092,4.092,0,0,1-2.721,2.837,3.557,3.557,0,0,1-1.045.159,4,4,0,0,1-2.687-1.124L548.01,300.808c-3.5-3.5-2.971-8.151.436-11.558l4.977-4.657m.124-2.17L552.4,283.5l-4.976,4.656c-4.192,4.191-4.372,9.816-.473,13.714l44.521,42.4a5.485,5.485,0,0,0,3.722,1.538,5.1,5.1,0,0,0,1.483-.224,5.59,5.59,0,0,0,3.719-3.838,5.176,5.176,0,0,0-1.31-4.788l-35.53-42.767-8.988-10.571-1.019-1.2Z" transform="translate(-526.086 -245.559)"/><path d="M562.4,295.151l9.556,11.5,5.761-5.356a7.926,7.926,0,0,0,.041-11.743l-43.7-41.923s-1.671-2.029-3.437-2.071a4.49,4.49,0,0,0-4.23,2.718c-.688,1.651-.194,2.809,1.315,4.97l29.306,35.565Z" transform="translate(-526.086 -245.559)"/><path d="M553.7,306.223l-17.116,21.024c-1.255,1.337-2.032,3.683-1.331,5.367a4.587,4.587,0,0,0,4.287,2.841,4.087,4.087,0,0,0,3.082-1.523l20.328-18.9Z" transform="translate(-526.086 -245.559)"/><path d="M592.074,250.547" transform="translate(-526.086 -245.559)" fill="#fff" stroke="#000" stroke-miterlimit="10" stroke-width="0.25"/></g></svg>
@@ -21,7 +21,6 @@ from skillopt.types import (  # noqa: F401
    FailureSummaryEntry,
    GateAction,
    GateResult,
-    MetaReflectResult,
    Patch,
    RawPatch,
    RolloutResult,
@@ -30,10 +30,10 @@ _STRUCTURED_SECTIONS = frozenset({

 _FLATTEN_MAP: dict[str, str] = {
    "model.backend": "model_backend",
-    "model.teacher": "teacher_model",
-    "model.student": "student_model",
-    "model.teacher_backend": "teacher_backend",
-    "model.student_backend": "student_backend",
+    "model.optimizer": "optimizer_model",
+    "model.target": "target_model",
+    "model.optimizer_backend": "optimizer_backend",
+    "model.target_backend": "target_backend",
    "model.reasoning_effort": "reasoning_effort",
    "model.rewrite_reasoning_effort": "rewrite_reasoning_effort",
    "model.rewrite_max_completion_tokens": "rewrite_max_completion_tokens",
@@ -51,7 +51,7 @@ _FLATTEN_MAP: dict[str, str] = {
    "model.claude_code_exec_use_sdk": "claude_code_exec_use_sdk",
    "model.claude_code_exec_effort": "claude_code_exec_effort",
    "model.claude_code_exec_max_thinking_tokens": "claude_code_exec_max_thinking_tokens",
-    "model.codex_trace_to_teacher": "codex_trace_to_teacher",
+    "model.codex_trace_to_optimizer": "codex_trace_to_optimizer",
    "model.azure_endpoint": "azure_endpoint",
    "model.azure_api_version": "azure_api_version",
    "model.azure_api_key": "azure_api_key",
@@ -61,18 +61,18 @@ _FLATTEN_MAP: dict[str, str] = {
    "model.azure_openai_auth_mode": "azure_openai_auth_mode",
    "model.azure_openai_ad_scope": "azure_openai_ad_scope",
    "model.azure_openai_managed_identity_client_id": "azure_openai_managed_identity_client_id",
-    "model.teacher_azure_openai_endpoint": "teacher_azure_openai_endpoint",
-    "model.teacher_azure_openai_api_version": "teacher_azure_openai_api_version",
-    "model.teacher_azure_openai_api_key": "teacher_azure_openai_api_key",
-    "model.teacher_azure_openai_auth_mode": "teacher_azure_openai_auth_mode",
-    "model.teacher_azure_openai_ad_scope": "teacher_azure_openai_ad_scope",
-    "model.teacher_azure_openai_managed_identity_client_id": "teacher_azure_openai_managed_identity_client_id",
-    "model.student_azure_openai_endpoint": "student_azure_openai_endpoint",
-    "model.student_azure_openai_api_version": "student_azure_openai_api_version",
-    "model.student_azure_openai_api_key": "student_azure_openai_api_key",
-    "model.student_azure_openai_auth_mode": "student_azure_openai_auth_mode",
-    "model.student_azure_openai_ad_scope": "student_azure_openai_ad_scope",
-    "model.student_azure_openai_managed_identity_client_id": "student_azure_openai_managed_identity_client_id",
+    "model.optimizer_azure_openai_endpoint": "optimizer_azure_openai_endpoint",
+    "model.optimizer_azure_openai_api_version": "optimizer_azure_openai_api_version",
+    "model.optimizer_azure_openai_api_key": "optimizer_azure_openai_api_key",
+    "model.optimizer_azure_openai_auth_mode": "optimizer_azure_openai_auth_mode",
+    "model.optimizer_azure_openai_ad_scope": "optimizer_azure_openai_ad_scope",
+    "model.optimizer_azure_openai_managed_identity_client_id": "optimizer_azure_openai_managed_identity_client_id",
+    "model.target_azure_openai_endpoint": "target_azure_openai_endpoint",
+    "model.target_azure_openai_api_version": "target_azure_openai_api_version",
+    "model.target_azure_openai_api_key": "target_azure_openai_api_key",
+    "model.target_azure_openai_auth_mode": "target_azure_openai_auth_mode",
+    "model.target_azure_openai_ad_scope": "target_azure_openai_ad_scope",
+    "model.target_azure_openai_managed_identity_client_id": "target_azure_openai_managed_identity_client_id",
    "model.qwen_chat_base_url": "qwen_chat_base_url",
    "model.qwen_chat_api_key": "qwen_chat_api_key",
    "model.qwen_chat_temperature": "qwen_chat_temperature",
@@ -89,22 +89,21 @@ _FLATTEN_MAP: dict[str, str] = {
    "gradient.merge_batch_size": "merge_batch_size",
    "gradient.analyst_workers": "analyst_workers",
    "gradient.failure_only": "failure_only",
-    "gradient.use_deep_reflect": "use_deep_reflect",
-    "gradient.deep_reflect_failures": "deep_reflect_failures",
-    "gradient.deep_reflect_successes": "deep_reflect_successes",
    "gradient.max_analyst_rounds": "max_analyst_rounds",
    "optimizer.learning_rate": "edit_budget",
    "optimizer.min_learning_rate": "min_edit_budget",
    "optimizer.lr_scheduler": "lr_scheduler",
    "optimizer.lr_control_mode": "lr_control_mode",
    "optimizer.skill_update_mode": "skill_update_mode",
-    "optimizer.use_meta_reflect": "use_meta_reflect",
    "optimizer.meta_learning_rate": "meta_edit_budget",
    "optimizer.use_slow_update": "use_slow_update",
    "optimizer.slow_update_samples": "slow_update_samples",
+    "optimizer.slow_update_gate_with_selection": "slow_update_gate_with_selection",
    "optimizer.longitudinal_pair_policy": "longitudinal_pair_policy",
    "optimizer.use_meta_skill": "use_meta_skill",
    "evaluation.use_gate": "use_gate",
+    "evaluation.gate_metric": "gate_metric",
+    "evaluation.gate_mixed_weight": "gate_mixed_weight",
    "evaluation.sel_env_num": "sel_env_num",
    "evaluation.test_env_num": "test_env_num",
    "evaluation.eval_test": "eval_test",
@@ -31,7 +31,7 @@ optimizer:
  learning_rate: 4                # Max edits per step (edit budget)
  lr_scheduler: cosine            # cosine | linear | constant | autonomous
  use_slow_update: true           # Epoch-boundary momentum
-  use_meta_skill: true            # Cross-epoch teacher memory
+  use_meta_skill: true            # Cross-epoch optimizer memory

 # ── Evaluation ───────────────────────────────────
 evaluation:
@@ -41,5 +41,5 @@ evaluation:
 # ── Model ────────────────────────────────────────
 model:
  backend: azure_openai           # azure_openai | openai_chat | claude_code_exec | qwen
-  teacher: gpt-5.5
-  student: gpt-5.5
+  optimizer: gpt-4o
+  target: gpt-4o
@@ -4,7 +4,7 @@ Benchmark Environment Template
 Copy this file and implement the TODO sections to add a new benchmark.

 The EnvAdapter is responsible for:
-1. Executing tasks using the student model + current skill document
+1. Executing tasks using the target model + current skill document
 2. Evaluating predictions against ground truth
 3. Returning structured results for the training loop
 """
@@ -25,12 +25,12 @@ class TemplateBenchmarkEnv(EnvAdapter):

    async def execute(self, item, skill: str, model):
        """
-        Execute a single task with the student model.
+        Execute a single task with the target model.

        Args:
            item: DataItem with .id, .input, .ground_truth, .metadata
            skill: Current skill document content (Markdown string)
-            model: Student model backend instance
+            model: Target model backend instance

        Returns:
            TaskResult with prediction, score, and trajectory
@@ -38,7 +38,7 @@ class TemplateBenchmarkEnv(EnvAdapter):
        # Step 1: Build the prompt combining skill + task input
        prompt = self.build_prompt(item, skill)

-        # Step 2: Call the student model
+        # Step 2: Call the target model
        # TODO: Customize the message format for your benchmark
        messages = [
            {"role": "system", "content": skill},
@@ -9,7 +9,6 @@ from dataclasses import dataclass
 import json
 import os

-from skillopt.gradient.deep_probe import generate_deep_probe_instruction
 from skillopt.datasets.base import BatchSpec
 from skillopt.envs.base import EnvAdapter
 from skillopt.envs.alfworld.dataloader import ALFWorldDataLoader
@@ -83,20 +82,16 @@ class ALFWorldAdapter(EnvAdapter):
        failure_only: bool = False,
        minibatch_size: int = 8,
        edit_budget: int = 4,
-        use_deep_reflect: bool = False,
-        deep_reflect_failures: int = 4,
-        deep_reflect_successes: int = 2,
+        max_completion_tokens: int = 16384,
    ) -> None:
        self.max_steps = max_steps
        self.workers = max(int(workers or 1), 1)
        self.max_api_workers = max_api_workers
+        self.max_completion_tokens = int(max_completion_tokens)
        self.analyst_workers = analyst_workers
        self.failure_only = failure_only
        self.minibatch_size = minibatch_size
        self.edit_budget = edit_budget
-        self.use_deep_reflect = use_deep_reflect
-        self.deep_reflect_failures = deep_reflect_failures
-        self.deep_reflect_successes = deep_reflect_successes
        self.dataloader = ALFWorldDataLoader(
            split_dir=split_dir,
            data_path=data_path,
@@ -357,6 +352,7 @@ class ALFWorldAdapter(EnvAdapter):
                max_steps=self.max_steps,
                out_root=out_dir,
                max_api_workers=self.max_api_workers,
+                max_completion_tokens=self.max_completion_tokens,
                result_ids=getattr(env_manager, "_skillopt_result_ids", None),
            )

@@ -419,6 +415,7 @@ class ALFWorldAdapter(EnvAdapter):
                    max_steps=self.max_steps,
                    out_root=out_dir,
                    max_api_workers=min(self.max_api_workers, chunk_size),
+                    max_completion_tokens=self.max_completion_tokens,
                    diagnostic_mode=diagnostic_mode,
                    diagnostic_instruction=diagnostic_instruction,
                    result_ids=chunk_ids,
@@ -457,129 +454,6 @@ class ALFWorldAdapter(EnvAdapter):
            meta_skill_context=meta_skill_context,
        )

-    def deep_reflect(
-        self,
-        results: list[dict],
-        skill_content: str,
-        out_dir: str,
-        **kwargs,
-    ) -> list[dict | None]:
-        if not self.use_deep_reflect:
-            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", "")
-        selected_items = self.select_representative_items(
-            results,
-            results,
-            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)
-
-        field_counts: dict[str, int] = {}
-        selected_metadata: list[dict] = []
-        for item in selected_items:
-            meta = self.get_reference_metadata(item)
-            for field in meta["fields"]:
-                field_counts[field] = field_counts.get(field, 0) + 1
-            selected_metadata.append({
-                "id": str(item["id"]),
-                "task_type": str(item.get("task_type") or "alfworld"),
-                "gamefile": str(item.get("gamefile") or ""),
-                "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)
-        field_summary = ", ".join(
-            f"{field}({count}/{len(selected_items)})"
-            for field, count in sorted(field_counts.items())
-        ) or "none"
-        print(
-            f"    [2b/6 DEEP REFLECT setup] selected={len(selected_items)} "
-            f"reference_fields={field_summary}"
-        )
-        probe = generate_deep_probe_instruction(
-            skill_content=skill_content,
-            items=selected_examples,
-            prediction_dir=prediction_dir,
-            system_prompt=self.get_deep_probe_prompt(),
-            step_buffer_context=step_buffer_context,
-            meta_skill_context=meta_skill_context,
-            output_requirements=[
-                "- Some trajectories may include a hidden Reference block. Use it to target the student's latent subgoal, missing precondition, or next-step intent, but do not reveal or paraphrase that reference to the student.",
-                "- The instruction must request a brief diagnostic readout inside the existing <think>...</think> block.",
-                "- The student must still output exactly one admissible action inside <action>...</action>.",
-                "- Do not ask for exhaustive inventories, full plans, or long chain-of-thought.",
-                "- The instruction text should be ready to append directly to the student's prompt.",
-            ],
-        )
-        if not probe:
-            return []
-
-        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": field_counts,
-                    },
-                    "selected_examples": selected_metadata,
-                },
-                f,
-                ensure_ascii=False,
-                indent=2,
-            )
-
-        gamefiles = [str(item.get("gamefile") or "") for item in selected_items]
-        if any(not gamefile for gamefile in gamefiles):
-            return []
-        eval_dataset, is_train = self._infer_dataset_from_gamefile(gamefiles[0])
-        deep_env = ALFWorldBatchRun(
-            env_num=len(selected_items),
-            eval_dataset=eval_dataset,
-            seed=random_seed or 42,
-            is_train=is_train,
-            specific_gamefiles=gamefiles,
-            workers=min(self.workers, max(len(selected_items), 1)),
-            result_ids=[str(item["id"]) for item in selected_items],
-        )
-        deep_results = self._run_batch(
-            deep_env,
-            skill_content=skill_content,
-            out_dir=rollout_dir,
-            diagnostic_mode=True,
-            diagnostic_instruction=probe["probe_instruction"],
-        )
-        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,
-        )

    def get_task_types(self) -> list[str]:
        return list(TASKS)
@@ -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>"
-}
@@ -11,11 +11,10 @@ import json
 import os
 import re
 import sys
-import time
 import concurrent.futures
 import numpy as np

-from skillopt.model import chat_student
+from skillopt.model import chat_target

 # ── Constants ─────────────────────────────────────────────────────────────────

@@ -134,7 +133,7 @@ def run_alfworld_batch(
    out_root: str = "",
    max_api_workers: int = 8,
    temperature: float = 0.4,
-    max_completion_tokens: int = 2048,
+    max_completion_tokens: int = 16384,
    diagnostic_mode: bool = False,
    diagnostic_instruction: str = "",
    result_ids: list[str] | None = None,
@@ -206,17 +205,16 @@ def run_alfworld_batch(

        # Call API in parallel
        actions = ["None"] * env_num
-        action_timeout = 180

        def call_api(idx):
            try:
-                response, _ = chat_student(
+                response, _ = chat_target(
                    system="You are an expert agent operating in the ALFRED Embodied Environment.",
                    user=prompts[idx],
                    max_completion_tokens=max_completion_tokens,
                    retries=5,
                    stage="rollout",
-                    timeout=120,
+                    timeout=None,
                )
                response = (response or "").strip()
                if not response:
@@ -230,7 +228,6 @@ def run_alfworld_batch(
        executor = concurrent.futures.ThreadPoolExecutor(max_workers=max_api_workers)
        try:
            futures = {executor.submit(call_api, i): i for i in active_indices}
-            started_at = {future: time.time() for future in futures}
            pending_futs = set(futures)
            while pending_futs:
                done, _ = concurrent.futures.wait(
@@ -238,11 +235,6 @@ def run_alfworld_batch(
                    timeout=5,
                    return_when=concurrent.futures.FIRST_COMPLETED,
                )
-                now = time.time()
-                timed_out = [
-                    future for future in pending_futs - done
-                    if now - started_at[future] >= action_timeout
-                ]
                for future in done:
                    pending_futs.remove(future)
                    try:
@@ -251,10 +243,6 @@ def run_alfworld_batch(
                        idx = futures[future]
                        response = "<think>error</think><action>look</action>"
                    actions[idx] = response
-                for future in timed_out:
-                    pending_futs.remove(future)
-                    idx = futures[future]
-                    actions[idx] = "<think>api timeout</think><action>look</action>"
        finally:
            executor.shutdown(wait=False, cancel_futures=True)

@@ -1 +0,0 @@
-"""BabyVision environment package for ReflACT."""
@@ -1,267 +0,0 @@
-"""BabyVision environment adapter for ReflACT."""
-from __future__ import annotations
-
-import json
-import os
-
-from skillopt.gradient.deep_probe import generate_deep_probe_instruction
-from skillopt.datasets.base import BatchSpec
-from skillopt.gradient.reflect import run_minibatch_reflect
-from skillopt.envs.base import EnvAdapter
-from skillopt.envs.babyvision.dataloader import BabyVisionDataLoader
-from skillopt.envs.babyvision.rollout import run_batch
-from skillopt.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()
@@ -1,214 +0,0 @@
-"""BabyVision task dataloader."""
-from __future__ import annotations
-
-import json
-import os
-from typing import Any
-
-from skillopt.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)
@@ -1,160 +0,0 @@
-"""BabyVision evaluation helpers using the official-style LLM judge."""
-from __future__ import annotations
-
-import re
-import string
-
-import regex
-
-from skillopt.model import chat_with_deployment
-from skillopt.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
@@ -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>"}
-    ]
-  }
-}
@@ -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>"}
-    ]
-  }
-}
@@ -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>"
-}
@@ -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}
@@ -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}}
@@ -1,4 +0,0 @@
-"""BabyVision Reflect stage.
-
-Prompts are now loaded from .md files by the base adapter.
-"""
@@ -1,483 +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 skillopt.envs.babyvision.evaluator import evaluate_item, evaluation_mode, extract_boxed_answer
-from skillopt.model import chat_student_messages, get_student_backend, is_student_exec_backend
-from skillopt.model.codex_harness import prepare_workspace, render_skill_md, run_student_exec
-from skillopt.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 `skillopt-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 skillopt.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
-
-    total = len(existing) + len(pending)
-    completed = len(existing)
-    correct_count = sum(1 for r in existing if r.get("hard", 0))
-    if existing:
-        print(f"    [rollout] resuming: {completed}/{total} already done", flush=True)
-
-    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)
-            completed += 1
-            if row.get("hard", 0):
-                correct_count += 1
-            acc = correct_count / completed if completed else 0
-            print(
-                f"    [rollout] {completed}/{total} "
-                f"(acc={acc:.3f}) id={row.get('id', '?')} "
-                f"hard={row.get('hard', '?')}",
-                flush=True,
-            )
-            outf.write(json.dumps(row, ensure_ascii=False) + "\n")
-            outf.flush()
-    return results
@@ -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>.
@@ -31,7 +31,6 @@ import os
 import random

 from skillopt.datasets.base import BaseDataLoader, BatchSpec
-from skillopt.model.codex_harness import extract_codex_trace_prefix, format_codex_trace_steps, parse_codex_raw
 from skillopt.prompts import load_prompt


@@ -60,24 +59,8 @@ class EnvAdapter(ABC):
        """Return whether this adapter requires Ray runtime initialization."""
        return False

-    def deep_reflect(
-        self,
-        results: list[dict],
-        skill_content: str,
-        out_dir: str,
-        **kwargs,
-    ) -> list[dict | None]:
-        """Optional deeper diagnostic reflection pass.
-
-        Default behavior is a no-op. Dataset-backed adapters may override this
-        to re-query the student on a small representative subset of the current
-        batch using minimally-perturbed diagnostic prompts that expose
-        intermediate reasoning state.
-        """
-        return []
-
    def build_reference_text(self, item: dict) -> str:
-        """Return hidden reference material for deep reflection, if any."""
+        """Return hidden reference material for reflection, if any."""
        return str(item.get("reference_text") or "").strip()

    def get_reference_metadata(self, item: dict) -> dict:
@@ -90,65 +73,6 @@ class EnvAdapter(ABC):
            "preview": reference_text[:400],
        }

-    def get_codex_deep_probe_prompt(self) -> str | None:
-        env_name = getattr(self, "_cfg", {}).get("env_name")
-        return load_prompt("deep_probe_codex", env=env_name)
-
-    def attach_codex_probe_context(
-        self,
-        results: list[dict],
-        prediction_dir: str,
-    ) -> list[dict]:
-        """Attach compact Codex step metadata for codex-aware deep reflection."""
-        enriched: list[dict] = []
-        for row in results:
-            merged = dict(row)
-            tid = str(row.get("id"))
-            raw_path = os.path.join(prediction_dir, tid, "codex_raw.txt")
-            if os.path.exists(raw_path):
-                with open(raw_path, encoding="utf-8") as f:
-                    raw = f.read()
-                parsed = parse_codex_raw(raw)
-                merged["codex_probe_trace_steps"] = format_codex_trace_steps(raw)
-                merged["codex_probe_step_count"] = len(parsed["steps"])
-            enriched.append(merged)
-        return enriched
-
-    def resolve_codex_probe_target(
-        self,
-        *,
-        selected_items: list[dict],
-        selected_examples: list[dict],
-        prediction_dir: str,
-        probe: dict,
-    ) -> tuple[list[dict], dict[str, str] | None, dict]:
-        """Resolve the teacher-selected codex probe target and raw trace prefix."""
-        target_id = str(probe.get("probe_target_id", "")).strip()
-        selected_id_set = {str(item["id"]) for item in selected_items}
-        if target_id not in selected_id_set:
-            target_id = str(selected_items[0]["id"])
-        target_item = next(item for item in selected_items if str(item["id"]) == target_id)
-        target_result = next(
-            (row for row in selected_examples if str(row.get("id")) == target_id),
-            None,
-        )
-        max_probe_step = int((target_result or {}).get("codex_probe_step_count", 0))
-        default_probe_step = max_probe_step - 1 if max_probe_step > 1 else max_probe_step
-        probe_after_step = int(probe.get("probe_after_step", default_probe_step))
-        if max_probe_step > 0:
-            probe_after_step = max(0, min(probe_after_step, max_probe_step))
-        else:
-            probe_after_step = 0
-        raw_path = os.path.join(prediction_dir, target_id, "codex_raw.txt")
-        trace_prefix = ""
-        if os.path.exists(raw_path):
-            with open(raw_path, encoding="utf-8") as f:
-                trace_prefix = extract_codex_trace_prefix(f.read(), after_step=probe_after_step)
-        updated_probe = dict(probe)
-        updated_probe["probe_target_id"] = target_id
-        updated_probe["probe_after_step"] = probe_after_step
-        return [target_item], {target_id: trace_prefix}, updated_probe
-
    def attach_reference_context(
        self,
        results: list[dict],
@@ -383,14 +307,3 @@ class EnvAdapter(ABC):
            if prompt is not None:
                return prompt
        return self._load_env_prompt("analyst_success")
-
-    def get_deep_probe_prompt(self) -> str | None:
-        return self._load_env_prompt("deep_probe")
-
-    def get_meta_reflect_prompt(self) -> str | None:
-        update_mode = getattr(self, "_cfg", {}).get("skill_update_mode", "patch")
-        if str(update_mode).strip().lower() == "rewrite_from_suggestions":
-            prompt = self._load_env_prompt("meta_reflect_rewrite")
-            if prompt is not None:
-                return prompt
-        return self._load_env_prompt("meta_reflect")
@@ -1,114 +0,0 @@
-from __future__ import annotations
-
-import json
-import os
-from typing import Any, Callable
-
-from skillopt.gradient.deep_probe import generate_deep_probe_instruction
-from skillopt.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",
-        ),
-    )
@@ -4,7 +4,6 @@ import os

 from skillopt.datasets.base import BatchSpec
 from skillopt.envs.base import EnvAdapter
-from skillopt.envs.deep_reflect import run_no_reference_deep_reflect
 from skillopt.envs.docvqa.dataloader import DocVQADataLoader
 from skillopt.envs.docvqa.rollout import run_batch
 from skillopt.gradient.reflect import run_minibatch_reflect
@@ -29,21 +28,17 @@ class DocVQAAdapter(EnvAdapter):
        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,
+        max_completion_tokens: int = 16384,
    ) -> None:
        self.max_turns = max_turns
        self.exec_timeout = exec_timeout
        self.workers = workers
+        self.max_completion_tokens = int(max_completion_tokens)
        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 = DocVQADataLoader(
            split_dir=split_dir,
            data_path=data_path,
@@ -83,6 +78,7 @@ class DocVQAAdapter(EnvAdapter):
            exec_timeout=self.exec_timeout,
            workers=self.workers,
            image_detail=self.image_detail,
+            max_completion_tokens=self.max_completion_tokens,
            diagnostic_mode=kwargs.get("diagnostic_mode", False),
            diagnostic_instruction=kwargs.get("diagnostic_instruction", ""),
            task_timeout=self.exec_timeout,
@@ -109,38 +105,6 @@ class DocVQAAdapter(EnvAdapter):
            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 document image prompt, 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 visual region, field/table/figure label, OCR text read, candidate answer, and answer-format normalization.",
-                "- Do not ask for exhaustive transcription 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 "docvqa"),
-                "question_preview": str(item.get("question") or "")[:200],
-                "image_path": item.get("image_path", ""),
-                "docId": item.get("docId", ""),
-                "page": item.get("ucsf_document_page_no", ""),
-            },
-        )

    def get_task_types(self) -> list[str]:
        seen: list[str] = []
@@ -6,8 +6,8 @@ import time
 from concurrent.futures import FIRST_COMPLETED, ThreadPoolExecutor, wait

 from skillopt.envs.docvqa.evaluator import evaluate
-from skillopt.model import chat_student_messages, get_student_backend, is_student_exec_backend
-from skillopt.model.codex_harness import prepare_workspace, render_skill_md, run_student_exec
+from skillopt.model import chat_target_messages, get_target_backend, is_target_exec_backend
+from skillopt.model.codex_harness import prepare_workspace, render_skill_md, run_target_exec
 from skillopt.prompts import load_prompt


@@ -112,11 +112,11 @@ def _run_codex_once(
        images=[item["image_path"]],
    )
    prompt = (
-        "Use the `skillopt-student` skill available in this workspace.\n"
+        "Use the `skillopt-target` skill available in this workspace.\n"
        "Read `task.md`, inspect the attached document image, and answer the DocVQA question.\n"
        "Return the final answer inside <answer>...</answer>."
    )
-    final_message, raw = run_student_exec(
+    final_message, raw = run_target_exec(
        work_dir=work_dir,
        prompt=prompt,
        model=model,
@@ -134,6 +134,7 @@ def process_one(
    max_turns: int = 1,
    exec_timeout: int = 120,
    image_detail: str = "auto",
+    max_completion_tokens: int = 16384,
    diagnostic_mode: bool = False,
    diagnostic_instruction: str = "",
 ) -> dict:
@@ -158,7 +159,7 @@ def process_one(
        system_prompt = ""
        user_text = ""
        conversation: list[dict] = []
-        if is_student_exec_backend():
+        if is_target_exec_backend():
            from skillopt.model import azure_openai as _llm

            conversation = [
@@ -172,7 +173,7 @@ def process_one(
                    pred_dir=os.path.join(out_root, "predictions", item_id),
                    item=item,
                    skill_content=skill_content,
-                    model=_llm.STUDENT_DEPLOYMENT,
+                    model=_llm.TARGET_DEPLOYMENT,
                    timeout=exec_timeout,
                    image_detail=image_detail,
                    diagnostic_mode=diagnostic_mode if turn == 0 else False,
@@ -198,9 +199,9 @@ def process_one(
            ]
            for turn in range(max_turns):
                if turn == 0:
-                    resp_text, _ = chat_student_messages(
+                    resp_text, _ = chat_target_messages(
                        messages=messages,
-                        max_completion_tokens=768,
+                        max_completion_tokens=max_completion_tokens,
                        retries=5,
                        stage="rollout",
                        timeout=exec_timeout,
@@ -212,9 +213,9 @@ def process_one(
                        {"role": "assistant", "content": response},
                        {"role": "user", "content": "Review the same image carefully and answer again. Keep the final answer inside <answer>...</answer>."},
                    ]
-                    resp_text, _ = chat_student_messages(
+                    resp_text, _ = chat_target_messages(
                        messages=refinement_messages,
-                        max_completion_tokens=512,
+                        max_completion_tokens=max_completion_tokens,
                        retries=5,
                        stage="rollout",
                        timeout=exec_timeout,
@@ -230,9 +231,9 @@ def process_one(

        pred_dir = os.path.join(out_root, "predictions", item_id)
        os.makedirs(pred_dir, exist_ok=True)
-        with open(os.path.join(pred_dir, "student_system_prompt.txt"), "w", encoding="utf-8") as f:
+        with open(os.path.join(pred_dir, "target_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:
+        with open(os.path.join(pred_dir, "target_user_prompt.txt"), "w", encoding="utf-8") as f:
            f.write(user_text)

        eval_result = evaluate(response, item.get("answers", []))
@@ -266,6 +267,7 @@ def run_batch(
    exec_timeout: int = 120,
    workers: int = 16,
    image_detail: str = "auto",
+    max_completion_tokens: int = 16384,
    diagnostic_mode: bool = False,
    diagnostic_instruction: str = "",
    task_timeout: int = 600,
@@ -325,6 +327,7 @@ def run_batch(
            max_turns=max_turns,
            exec_timeout=exec_timeout,
            image_detail=image_detail,
+            max_completion_tokens=max_completion_tokens,
            diagnostic_mode=diagnostic_mode,
            diagnostic_instruction=diagnostic_instruction,
        )
@@ -4,13 +4,12 @@ from __future__ import annotations
 import json
 import os

-from skillopt.gradient.deep_probe import generate_deep_probe_instruction
 from skillopt.datasets.base import BatchSpec
 from skillopt.gradient.reflect import run_minibatch_reflect
 from skillopt.envs.base import EnvAdapter
 from skillopt.envs.livemathematicianbench.dataloader import LiveMathematicianBenchDataLoader
 from skillopt.envs.livemathematicianbench.rollout import run_batch
-from skillopt.model import get_student_backend
+from skillopt.model import get_target_backend


 class LiveMathematicianBenchAdapter(EnvAdapter):
@@ -62,22 +61,18 @@ class LiveMathematicianBenchAdapter(EnvAdapter):
        shuffle_choices: bool = True,
        use_theorem: bool = False,
        use_sketch: bool = False,
-        use_deep_reflect: bool = False,
-        deep_reflect_failures: int = 4,
-        deep_reflect_successes: int = 2,
+        max_completion_tokens: int = 16384,
    ) -> None:
        self.max_turns = max_turns
        self.exec_timeout = exec_timeout
        self.workers = workers
+        self.max_completion_tokens = int(max_completion_tokens)
        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,
@@ -123,6 +118,7 @@ class LiveMathematicianBenchAdapter(EnvAdapter):
            max_turns=self.max_turns,
            exec_timeout=self.exec_timeout,
            workers=self.workers,
+            max_completion_tokens=self.max_completion_tokens,
            use_theorem=self.use_theorem,
            use_sketch=self.use_sketch,
            diagnostic_mode=kwargs.get("diagnostic_mode", False),
@@ -161,122 +157,6 @@ class LiveMathematicianBenchAdapter(EnvAdapter):
            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()
@@ -1,7 +1,7 @@
 You are an expert failure-analysis agent for theorem-grounded mathematical multiple-choice questions.

 You will be given MULTIPLE failed trajectories from a single minibatch and the current skill document.
-Each trajectory includes the student's response and an evaluation result showing the predicted option
+Each trajectory includes the target's response and an evaluation result showing the predicted option
 versus the correct option.

 Your job is to identify COMMON reasoning failures across the batch and propose concise skill edits.
@@ -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>"
-}
@@ -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>"
-}
@@ -7,8 +7,8 @@ import time
 from concurrent.futures import FIRST_COMPLETED, ThreadPoolExecutor, wait

 from skillopt.envs.livemathematicianbench.evaluator import evaluate
-from skillopt.model import chat_student, get_student_backend, is_student_exec_backend
-from skillopt.model.codex_harness import prepare_workspace, render_skill_md, run_student_exec
+from skillopt.model import chat_target, get_target_backend, is_target_exec_backend
+from skillopt.model.codex_harness import prepare_workspace, render_skill_md, run_target_exec
 from skillopt.prompts import load_prompt

 def _build_system(skill_content: str) -> str:
@@ -95,11 +95,11 @@ def _run_codex_once(
    work_dir = os.path.join(pred_dir, "codex_exec")
    prepare_workspace(work_dir=work_dir, skill_md=skill_md, task_text=task_text)
    prompt = (
-        "Use the `skillopt-student` skill available in this workspace.\n"
+        "Use the `skillopt-target` skill available in this workspace.\n"
        "Read `task.md` and solve the multiple-choice problem.\n"
        "Output only the final choice label inside <answer>...</answer>."
    )
-    final_message, raw = run_student_exec(
+    final_message, raw = run_target_exec(
        work_dir=work_dir,
        prompt=prompt,
        model=model,
@@ -119,7 +119,8 @@ def process_one(
    diagnostic_mode: bool = False,
    diagnostic_instruction: str = "",
    diagnostic_trace_context: str = "",
-    exec_timeout: int = 300,
+    exec_timeout: int | None = 300,
+    max_completion_tokens: int = 16384,
 ) -> dict:
    item_id = str(item["id"])
    result = {
@@ -142,8 +143,9 @@ def process_one(
    try:
        pred_dir = os.path.join(out_root, "predictions", item_id)
        os.makedirs(pred_dir, exist_ok=True)
+        llm_timeout = int(exec_timeout) if exec_timeout and int(exec_timeout) > 0 else None

-        if is_student_exec_backend():
+        if is_target_exec_backend():
            from skillopt.model import azure_openai as _llm

            conversation: list[dict] = []
@@ -155,8 +157,8 @@ def process_one(
                    pred_dir=pred_dir,
                    skill_content=skill_content,
                    item=item,
-                    model=_llm.STUDENT_DEPLOYMENT,
-                    timeout=exec_timeout,
+                    model=_llm.TARGET_DEPLOYMENT,
+                    timeout=llm_timeout,
                    use_theorem=use_theorem,
                    use_sketch=use_sketch,
                    diagnostic_mode=diagnostic_mode if turn == 0 else False,
@@ -172,9 +174,9 @@ def process_one(
            result["agent_ok"] = True
            result["n_turns"] = len(conversation)

-            with open(os.path.join(pred_dir, "student_system_prompt.txt"), "w", encoding="utf-8") as f:
+            with open(os.path.join(pred_dir, "target_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:
+            with open(os.path.join(pred_dir, "target_user_prompt.txt"), "w", encoding="utf-8") as f:
                f.write(user)

            eval_result = evaluate(response, item["correct_choice"], item["choices"])
@@ -216,13 +218,13 @@ def process_one(

        for turn in range(max_turns):
            if turn == 0:
-                resp_text, _ = chat_student(
+                resp_text, _ = chat_target(
                    system=system,
                    user=user,
-                    max_completion_tokens=16384,
+                    max_completion_tokens=max_completion_tokens,
                    retries=5,
                    stage="rollout",
-                    timeout=exec_timeout,
+                    timeout=llm_timeout,
                )
            else:
                refinement = (
@@ -230,13 +232,13 @@ def process_one(
                    "Re-evaluate the exact option wording. If needed, correct it. "
                    "Output only the final choice label inside <answer>...</answer>."
                )
-                resp_text, _ = chat_student(
+                resp_text, _ = chat_target(
                    system=system,
                    user=refinement,
-                    max_completion_tokens=16384,
+                    max_completion_tokens=max_completion_tokens,
                    retries=5,
                    stage="rollout",
-                    timeout=exec_timeout,
+                    timeout=llm_timeout,
                )
            response = resp_text
            conversation.append({"type": "message", "turn": turn + 1, "content": resp_text})
@@ -247,9 +249,9 @@ def process_one(
        result["agent_ok"] = True
        result["n_turns"] = len(conversation)

-        with open(os.path.join(pred_dir, "student_system_prompt.txt"), "w", encoding="utf-8") as f:
+        with open(os.path.join(pred_dir, "target_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:
+        with open(os.path.join(pred_dir, "target_user_prompt.txt"), "w", encoding="utf-8") as f:
            f.write(user)

        eval_result = evaluate(response, item["correct_choice"], item["choices"])
@@ -291,16 +293,22 @@ def run_batch(
    skill_content: str,
    *,
    max_turns: int = 1,
-    exec_timeout: int = 300,
+    exec_timeout: int | None = 300,
    workers: int = 64,
+    max_completion_tokens: int = 16384,
    use_theorem: bool = False,
    use_sketch: bool = False,
    diagnostic_mode: bool = False,
    diagnostic_instruction: str = "",
    diagnostic_trace_context_by_id: dict[str, str] | None = None,
-    task_timeout: int = 600,
+    task_timeout: int | None = 600,
 ) -> list[dict]:
-    task_timeout = max(int(task_timeout), int(exec_timeout) + 60)
+    exec_timeout_value = int(exec_timeout) if exec_timeout and int(exec_timeout) > 0 else 0
+    task_timeout_value = int(task_timeout) if task_timeout and int(task_timeout) > 0 else 0
+    if exec_timeout_value <= 0 or task_timeout_value <= 0:
+        task_timeout = None
+    else:
+        task_timeout = max(task_timeout_value, exec_timeout_value + 60)
    results_path = os.path.join(out_root, "results.jsonl")
    os.makedirs(out_root, exist_ok=True)

@@ -338,6 +346,7 @@ def run_batch(
            skill_content,
            max_turns=max_turns,
            exec_timeout=exec_timeout,
+            max_completion_tokens=max_completion_tokens,
            use_theorem=use_theorem,
            use_sketch=use_sketch,
            diagnostic_mode=diagnostic_mode,
@@ -382,6 +391,7 @@ def run_batch(
                now = time.time()
                timed_out = [
                    fut for fut in pending_futs - done
+                    if task_timeout is not None
                    if str(futs[fut]["id"]) in started_at
                    and now - started_at[str(futs[fut]["id"])] >= task_timeout
                ]
@@ -1,5 +0,0 @@
-"""MathVerse environment package."""
-
-from skillopt.envs.mathverse.adapter import MathVerseAdapter
-
-__all__ = ["MathVerseAdapter"]
@@ -1,280 +0,0 @@
-"""MathVerse environment adapter for ReflACT."""
-from __future__ import annotations
-
-import json
-import os
-
-from skillopt.datasets.base import BatchSpec
-from skillopt.envs.base import EnvAdapter
-from skillopt.envs.mathverse.dataloader import MathVerseDataLoader
-from skillopt.envs.mathverse.rollout import run_batch
-from skillopt.gradient.deep_probe import generate_deep_probe_instruction
-from skillopt.gradient.reflect import run_minibatch_reflect
-from skillopt.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()
@@ -1,228 +0,0 @@
-"""MathVerse task dataloader."""
-from __future__ import annotations
-
-import json
-import os
-import re
-from typing import Any
-
-from skillopt.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
@@ -1,180 +0,0 @@
-"""MathVerse evaluation helpers."""
-from __future__ import annotations
-
-import re
-import string
-
-from skillopt.model import chat_with_deployment
-from skillopt.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
@@ -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>"}
-    ]
-  }
-}
@@ -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>"}
-    ]
-  }
-}
@@ -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>"
-}
@@ -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}
@@ -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>.
@@ -1,4 +0,0 @@
-"""MathVerse Reflect stage.
-
-Prompts are loaded from .md files by the base adapter.
-"""
@@ -1,431 +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 skillopt.envs.mathverse.evaluator import evaluate_item, evaluation_mode, extract_answer
-from skillopt.model import chat_student_messages, get_student_backend, is_student_exec_backend
-from skillopt.model.codex_harness import prepare_workspace, render_skill_md, run_student_exec
-from skillopt.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 `skillopt-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 skillopt.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
-
-    total = len(existing) + len(pending)
-    completed = len(existing)
-    correct_count = sum(1 for r in existing if r.get("hard", 0))
-    if existing:
-        print(f"    [rollout] resuming: {completed}/{total} already done", flush=True)
-
-    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)
-            completed += 1
-            if row.get("hard", 0):
-                correct_count += 1
-            acc = correct_count / completed if completed else 0
-            print(
-                f"    [rollout] {completed}/{total} "
-                f"(acc={acc:.3f}) id={row.get('id', '?')} "
-                f"hard={row.get('hard', '?')}",
-                flush=True,
-            )
-            outf.write(json.dumps(row, ensure_ascii=False) + "\n")
-            outf.flush()
-    return results
@@ -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>.
@@ -1,2 +0,0 @@
-"""MMRB environment package."""
-
@@ -1,283 +0,0 @@
-"""MMRB environment adapter for ReflACT."""
-from __future__ import annotations
-
-import json
-import os
-
-from skillopt.gradient.deep_probe import generate_deep_probe_instruction
-from skillopt.datasets.base import BatchSpec
-from skillopt.gradient.reflect import run_minibatch_reflect
-from skillopt.envs.base import EnvAdapter
-from skillopt.envs.mmrb.dataloader import MMRBDataLoader
-from skillopt.envs.mmrb.rollout import run_batch
-from skillopt.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()
@@ -1,146 +0,0 @@
-"""MMRB task dataloader."""
-from __future__ import annotations
-
-import glob
-import json
-import os
-import re
-from typing import Any
-
-from skillopt.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)
@@ -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
-
@@ -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>.
@@ -1,455 +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 skillopt.envs.mmrb.evaluator import evaluate_item, evaluation_mode
-from skillopt.model import chat_student_messages, get_student_backend, is_student_exec_backend
-from skillopt.model.codex_harness import prepare_workspace, render_skill_md, run_student_exec
-from skillopt.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 `skillopt-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 skillopt.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
-
-    total = len(existing) + len(pending)
-    completed = len(existing)
-    correct_count = sum(1 for r in existing if r.get("hard", 0))
-    if existing:
-        print(f"    [rollout] resuming: {completed}/{total} already done", flush=True)
-
-    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)
-            completed += 1
-            if row.get("hard", 0):
-                correct_count += 1
-            acc = correct_count / completed if completed else 0
-            print(
-                f"    [rollout] {completed}/{total} "
-                f"(acc={acc:.3f}) id={row.get('id', '?')} "
-                f"hard={row.get('hard', '?')}",
-                flush=True,
-            )
-            outf.write(json.dumps(row, ensure_ascii=False) + "\n")
-            outf.flush()
-    return results
@@ -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>.
-
@@ -4,7 +4,6 @@ import os

 from skillopt.datasets.base import BatchSpec
 from skillopt.envs.base import EnvAdapter
-from skillopt.envs.deep_reflect import run_no_reference_deep_reflect
 from skillopt.envs.officeqa.dataloader import OfficeQADataLoader
 from skillopt.envs.officeqa.rollout import run_batch
 from skillopt.gradient.reflect import run_minibatch_reflect
@@ -27,7 +26,7 @@ class OfficeQAAdapter(EnvAdapter):
        seed: int = 42,
        limit: int = 0,
        max_tool_turns: int = 12,
-        max_completion_tokens: int = 64000,
+        max_completion_tokens: int = 16384,
        search_mode: str = "offline",
        max_queries_per_turn: int = 4,
        search_api_url: str = os.environ.get("OFFICEQA_SEARCH_API_URL", "http://localhost:8080/search_tool/search"),
@@ -37,11 +36,7 @@ class OfficeQAAdapter(EnvAdapter):
        search_timeout_seconds: int = 20,
        use_local_tools: bool = True,
        data_dirs: list[str] | str | None = None,
-        docs_dirs: list[str] | str | None = None,
-        use_deep_reflect: bool = False,
-        deep_reflect_failures: int = 4,
-        deep_reflect_successes: int = 2,
-    ) -> None:
+        docs_dirs: list[str] | str | None = None,    ) -> None:
        self.workers = workers
        self.analyst_workers = analyst_workers
        self.failure_only = failure_only
@@ -58,9 +53,6 @@ class OfficeQAAdapter(EnvAdapter):
        self.search_timeout_seconds = int(search_timeout_seconds)
        self.use_local_tools = bool(use_local_tools)
        self.data_dirs = data_dirs if data_dirs is not None else docs_dirs
-        self.use_deep_reflect = use_deep_reflect
-        self.deep_reflect_failures = deep_reflect_failures
-        self.deep_reflect_successes = deep_reflect_successes
        self.dataloader = OfficeQADataLoader(
            split_dir=split_dir,
            data_path=data_path,
@@ -133,37 +125,6 @@ class OfficeQAAdapter(EnvAdapter):
            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, candidate files, tool 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 selected document/file, evidence span or table, extracted value, units, and any date or fiscal-period normalization.",
-                "- Do not ask for exhaustive copying of source text 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 "officeqa"),
-                "question_preview": str(item.get("question") or "")[:200],
-                "source_files": item.get("source_files", []),
-                "source_docs": item.get("source_docs", []),
-            },
-        )

    def get_task_types(self) -> list[str]:
        seen: list[str] = []
@@ -6,16 +6,13 @@ from concurrent.futures import ThreadPoolExecutor, as_completed
 from skillopt.envs.officeqa.evaluator import evaluate
 from skillopt.envs.officeqa.tool_runtime import (
    build_oracle_parsed_pages_context,
+    custom_search,
    resolve_candidate_files,
    resolve_docs_roots,
    run_tool,
 )
-try:
-    from skillopt.envs.sealqa.tool_runtime import custom_search
-except ImportError:
-    custom_search = None  # type: ignore[assignment]
-from skillopt.model import chat_student_messages, get_student_backend, is_student_exec_backend
-from skillopt.model.codex_harness import prepare_workspace, render_skill_md, run_student_exec
+from skillopt.model import chat_target_messages, get_target_backend, is_target_exec_backend
+from skillopt.model.codex_harness import prepare_workspace, render_skill_md, run_target_exec
 from skillopt.prompts import load_prompt
 _TOOL_SCHEMAS = [
    {
@@ -299,12 +296,12 @@ def _run_codex_once(
        link_dirs=_docs_link_targets(docs_roots),
    )
    prompt = (
-        "Use the `skillopt-student` skill available in this workspace.\n"
+        "Use the `skillopt-target` 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(
+    final_message, raw = run_target_exec(
        work_dir=work_dir,
        prompt=prompt,
        model=model,
@@ -356,8 +353,8 @@ def _run_custom_search_process(
        raise ValueError("custom_search mode requires a non-empty search_api_url")
    if not os.environ.get(search_auth_env, "").strip():
        raise ValueError(f"custom_search mode requires auth token env var {search_auth_env}")
-    if get_student_backend() not in {"openai_chat", "qwen_chat"}:
-        raise ValueError("custom_search mode is only supported with student_backend='openai_chat' or 'qwen_chat'")
+    if get_target_backend() not in {"openai_chat", "qwen_chat"}:
+        raise ValueError("custom_search mode is only supported with target_backend='openai_chat' or 'qwen_chat'")
    system = _build_system(
        skill_content,
        search_mode=_CUSTOM_SEARCH_MODE,
@@ -385,7 +382,7 @@ def _run_custom_search_process(
    fail_reason = ""
    last_response_metadata: dict = {}
    for turn in range(1, max_tool_turns + 1):
-        message, _ = chat_student_messages(
+        message, _ = chat_target_messages(
            messages=messages,
            max_completion_tokens=max_completion_tokens,
            retries=5,
@@ -439,8 +436,8 @@ def _run_azure_search_process(
    diagnostic_mode: bool,
    diagnostic_instruction: str,
 ) -> tuple[str, str, str, str, list[dict], str, dict]:
-    if get_student_backend() != "openai_chat":
-        raise ValueError("azure_search mode is only supported with student_backend='openai_chat'")
+    if get_target_backend() != "openai_chat":
+        raise ValueError("azure_search mode is only supported with target_backend='openai_chat'")
    system = _build_system(skill_content, search_mode=_AZURE_SEARCH_MODE)
    user = _build_user(
        item,
@@ -453,7 +450,7 @@ def _run_azure_search_process(
        {"role": "user", "content": user},
    ]
    conversation: list[dict] = [{"role": "user", "content": user}]
-    message, _ = chat_student_messages(
+    message, _ = chat_target_messages(
        messages=messages,
        max_completion_tokens=max_completion_tokens,
        retries=5,
@@ -494,7 +491,7 @@ def _run_offline_no_tools_process(
        {"role": "user", "content": user},
    ]
    conversation: list[dict] = [{"role": "user", "content": user}]
-    message, _ = chat_student_messages(
+    message, _ = chat_target_messages(
        messages=messages,
        max_completion_tokens=max_completion_tokens,
        retries=5,
@@ -516,7 +513,7 @@ def process_one(
    skill_content: str,
    *,
    max_tool_turns: int = 12,
-    max_completion_tokens: int = 64000,
+    max_completion_tokens: int = 16384,
    search_mode: str = _DEFAULT_SEARCH_MODE,
    max_queries_per_turn: int = 4,
    search_api_url: str = "",
@@ -616,7 +613,7 @@ def process_one(
                candidate_files=candidate_files,
                oracle_context=oracle_context,
            )
-        elif is_student_exec_backend():
+        elif is_target_exec_backend():
            from skillopt.model import azure_openai as _llm
            response = ""
            system = ""
@@ -628,7 +625,7 @@ def process_one(
                    skill_content=skill_content,
                    candidate_files=candidate_files,
                    docs_roots=docs_roots,
-                    model=_llm.STUDENT_DEPLOYMENT,
+                    model=_llm.TARGET_DEPLOYMENT,
                    timeout=180,
                    diagnostic_mode=diagnostic_mode if turn == 1 else False,
                    diagnostic_instruction=diagnostic_instruction if turn == 1 else "",
@@ -650,9 +647,9 @@ def process_one(
                {"role": "user", "content": user},
            ]
            for turn in range(1, max_tool_turns + 1):
-                message, _ = chat_student_messages(
+                message, _ = chat_target_messages(
                    messages=messages,
-                    max_completion_tokens=768,
+                    max_completion_tokens=max_completion_tokens,
                    retries=5,
                    stage="rollout",
                    tools=_TOOL_SCHEMAS,
@@ -688,9 +685,9 @@ def process_one(
                    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:
+    with open(os.path.join(pred_dir, "target_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:
+    with open(os.path.join(pred_dir, "target_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)
@@ -714,8 +711,8 @@ def process_one(
        "agent_ok": not fail_reason,
        "n_turns": len(conversation),
        "last_finish_reason": last_response_metadata.get("finish_reason", ""),
-        "student_system_prompt": system,
-        "student_user_prompt": user,
+        "target_system_prompt": system,
+        "target_user_prompt": user,
    }
    return result
 def run_batch(
@@ -725,7 +722,7 @@ def run_batch(
    *,
    workers: int = 8,
    max_tool_turns: int = 12,
-    max_completion_tokens: int = 64000,
+    max_completion_tokens: int = 16384,
    search_mode: str = _DEFAULT_SEARCH_MODE,
    max_queries_per_turn: int = 4,
    search_api_url: str = "",
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
Cuzyoung	8acc2dd03e	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>	2026-05-31 09:01:25 +00:00
Cuzyoung	00602df9e9	feat(slow-update): add config-controlled gated / force-injected modes Add optimizer.slow_update_gate_with_selection to control how epoch-boundary slow-update guidance is applied: - false (default): force-injected - inject guidance into current & best unconditionally (unchanged behavior). - true: gated - evaluate the slow-update candidate on the selection set and accept/reject via the same validation gate as step-level updates (logic follows the SkillReflection ablation). Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>	2026-05-31 02:02:23 +00:00
hwq	42e555d28e	Update eval-only README example	2026-05-30 15:28:17 +00:00
hwq	933c0a4ab5	Add GPT-5.5 benchmark skills	2026-05-30 15:15:15 +00:00
hwq	1f75d022a5	y	2026-05-30 15:01:34 +00:00
Yif Yang	4f3a9bc055	docs: scope PR #25 gate_metric as opt-in example, not default Move the soft/mixed gate-metric configuration introduced in PR #25 out of the base default config and into a standalone example config so that default SkillOpt runs (and paper reproduction) remain bit-for-bit on the original hard gate. - configs/_base_/default.yaml: drop gate_metric / gate_mixed_weight keys. The trainer's cfg.get("gate_metric", "hard") fallback preserves the original behavior unchanged. - configs/examples/soft_gate.yaml: new standalone reference config with a header explaining when to consider it (small selection split with continuous rewards) and when not to (paper reproduction, large or binary-reward settings). - README.md: add a short "Community-contributed configs" section that clearly flags this as user-contributed and non-default.	2026-05-30 08:09:03 +00:00
Yif Yang	d190bf37c1	Merge pull request #25 from lvbaocheng/feature/gate-soft-metric Add configurable gate metric (hard / soft / mixed) for skill validation Default is `hard`, preserving exact pre-PR behavior — verified by 22 unit assertions on the gate module plus an end-to-end 8-step trainer-trajectory test that produces a bit-for-bit identical accept/reject sequence between the pre-PR and post-PR code paths under `gate_metric: hard`. Paper- reproduction results are unaffected. `soft` and `mixed` are opt-in via `evaluation.gate_metric` in the config and address small-selection-set runs where discrete hard accuracy is too coarse to distinguish candidate skills.	2026-05-30 08:01:39 +00:00
Yif Yang	02695bd813	Merge pull request #24 from lvbaocheng/fix/claude-cli-effort-flag fix(claude): use --effort instead of deprecated --thinking flag	2026-05-30 15:31:00 +08:00
Yif Yang	cf287cb608	Merge pull request #20 from 1s1x/fix-continuous-reward-scores fix: support continuous reward scores (int truncation + falsy float)	2026-05-30 15:30:15 +08:00
Huangzisu	dbc90bd755	fix(auth): let env vars override yaml for openai_compatible mode The yaml default `azure_openai_auth_mode: azure_cli` was silently overwriting `AZURE_OPENAI_AUTH_MODE` exported by the user, because `configure_clients()` treats any non-empty config value as an explicit override. Switching the three auth_mode defaults (shared / optimizer / target) to "" lets `_clean()` drop them and restores the intended fallback chain: yaml → env var → module default ("azure_cli"). Also update README and .env.example to document the openai_compatible mode introduced in `d5c5b61`, and remove the misleading `OPENAI_API_KEY` snippet — SkillOpt reuses the `AZURE_OPENAI_*` env vars in this mode.	2026-05-30 06:58:05 +00:00
lvbaocheng	5d7875cb2e	Add configurable gate metric (hard / soft / mixed) for skill validation The training gate currently always compares candidate vs. current/best using hard exact-match accuracy. On environments with a small held-out selection set (e.g. 3-6 items) or partial-credit scoring, hard accuracy is too coarse: candidate skills that meaningfully improve per-item soft scores get rejected because the discrete hard count does not move. Add three opt-in metrics so users can pick the one that matches their scoring function: - `gate_metric: hard` — original behavior (default, fully backward compatible). - `gate_metric: soft` — gate on the soft / F1 / partial-credit score. - `gate_metric: mixed` — `(1 - w) * hard + w * soft`, where `w` is set by `gate_mixed_weight` (default 0.5). Changes ------- - `skillopt/evaluation/gate.py`: extend `evaluate_gate` with `cand_soft`, `metric`, and `mixed_weight` keyword arguments; add a pure helper `select_gate_score(hard, soft, metric, mixed_weight)`. Defaults preserve the original `metric="hard"` behavior — existing callers that only pass `cand_hard` keep working unchanged. - `skillopt/evaluation/__init__.py`: export the new helper / type. - `skillopt/engine/trainer.py`: read `evaluation.gate_metric` and `evaluation.gate_mixed_weight` from the config (with safe defaults), pass both metrics into `evaluate_gate`, and project the baseline `current_score` / `best_score` into metric space so subsequent comparisons are consistent. Print the gate metric on the `[6/6 EVALUATE]` line so logs make the decision basis explicit. The selection cache still records both `(hard, soft)` so a metric change on resume is non-destructive. - `configs/_base_/default.yaml`: document and ship the new keys with backward-compatible defaults (`hard`, `0.5`). Backward compatibility ---------------------- - Default config does not change behavior: `gate_metric` defaults to `hard`, exactly matching the previous gate. - `evaluate_gate(...)` keeps its existing positional signature; the new parameters are keyword-only with safe defaults. - `step_record.json` gains optional `gate_metric` and `candidate_gate_score` fields; old records still load. Tested ------ - Unit-tested all three metrics + boundary `mixed_weight` values (0.0 / 1.0) and rejection of unknown metric strings. All six cases pass. - Verified `skillopt.engine.trainer` imports cleanly after the refactor.	2026-05-30 14:45:27 +08:00
lvbaocheng	2532043d25	fix(claude): use --effort instead of deprecated --thinking flag Claude Code CLI v2.x renamed the flag; passing --thinking low causes all rollout calls to fail on CLI 2.1.87+. Co-authored-by: Cursor <cursoragent@cursor.com>	2026-05-30 11:24:13 +08:00
zq	41be2f1803	fix(scoring): use float() instead of int() for continuous reward scores int() truncates smoothed composite scores (0.0-1.0) to 0, making all continuous reward values appear as failures. This broke SkillOpt training pipelines using SmoothedCompositeReward.	2026-05-30 07:47:41 +08:00
zq	a62ec857f1	fix(reflect): support continuous reward scores in failure filtering not r.get("hard") treats non-zero floats as success. Add explicit float threshold check (< 1e-9). Backward compatible with binary hard=0/1.	2026-05-29 19:04:42 +08:00
zq	afb552008b	fix(trainer): support continuous reward scores in bucket aggregation int() truncates any float in [0,1) to 0. Replace with float(). Also fix falsy float check in failure detection. Backward compatible with binary hard=0/1.	2026-05-29 19:03:52 +08:00
Yif Yang	75b5c7f31c	Merge pull request #16 from guilhermeleste/feat/pioneer-ai-provider-integration Add OpenAI-compatible backend support for Pioneer.ai and other providers	2026-05-29 10:14:32 +08:00
Yif Yang	74ea3a1a8f	Merge pull request #18 from yong2bba/docs/custom-env-smoke docs: add local environment smoke test guide	2026-05-29 10:12:55 +08:00
yongjin	657b987de6	docs: add local environment smoke test guide	2026-05-29 09:26:38 +09:00
hwq	2a40aa3c98	Add SearchQA id split	2026-05-28 11:29:59 +00:00
hwq	786d57b5cf	Make rollout completion tokens configurable	2026-05-28 09:45:47 +00:00
guilhermeleste	d5c5b61830	Add OpenAI-compatible backend support for Pioneer.ai and other providers - Add 'openai_compatible', 'compat', and 'openai' auth modes to azure_openai.py - Modify _make_client() to use OpenAI client (not AzureOpenAI) for compatible endpoints - Update type hints to support both AzureOpenAI and OpenAI clients - Auto-configure API version sentinel when using compatible modes - Add .env template for Pioneer.ai configuration This allows users to use Pioneer.ai or any OpenAI-compatible API endpoint as both optimizer and target backend without requiring Azure OpenAI. Resolves: Support for non-Azure OpenAI-compatible providers	2026-05-28 05:54:43 -03:00
Cuzyoung	99212e3956	docs: remove Star History section for now Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>	2026-05-26 08:12:51 +00:00
Cuzyoung	fc54c44e93	docs: add Star History chart to README Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>	2026-05-26 08:10:16 +00:00
Yif Yang	48adf5a69f	Update citation format in README.md	2026-05-26 02:56:58 +08:00
Yif Yang	b11e6dcfb9	Enhance training description in README Updated README to include '(mini-)batchsize' in the training description.	2026-05-26 02:35:10 +08:00
Yif Yang	4c1b74fce2	Update BibTeX entry in index.html	2026-05-25 14:30:01 +08:00
Yif Yang	db6443384a	Update BibTeX entry for SkillOpt publication	2026-05-25 14:28:13 +08:00
Huangzisu	2c7d9074fb	update webpage for arxiv link	2026-05-25 05:32:04 +00:00
Yif Yang	c98bcdd5b3	Update README.md	2026-05-25 13:27:40 +08:00
Yif Yang	0f6db9afc4	Update README.md	2026-05-25 13:26:55 +08:00
Yif Yang	5a36ac35ae	Merge pull request #7 from microsoft/users/GitHubPolicyService/a41a3ce1-e5a1-4e18-810b-cfb8d2d21c29 Adding Microsoft SECURITY.MD	2026-05-25 13:09:26 +08:00
Lliar-liar	5f4b228543	Soften average gain column styling	2026-05-24 19:45:10 +00:00
Lliar-liar	a9cad7a125	Use official arXiv logomark	2026-05-24 19:43:19 +00:00
Lliar-liar	5e968115f5	Align citation section with SkillLens	2026-05-24 19:39:16 +00:00
Cuzyoung	ded8c27c90	restore: bring back project page HTML and assets These were accidentally deleted in the cleanup commit. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>	2026-05-24 19:38:34 +00:00
Cuzyoung	f55a26414e	cleanup: remove unused benchmarks, deep_probe, meta_reflect Remove sealqa, babyvision, mathverse, mmrb, swebench envs and configs. Remove deep_probe, deep_reflect, meta_reflect modules and prompts. Remove download_babyvision script. These are not part of the core released benchmarks. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>	2026-05-24 19:36:48 +00:00
Lliar-liar	2df2542aec	Stabilize skill evolution layout	2026-05-24 19:36:08 +00:00
Lliar-liar	faa4ec6199	Align header and scroll effects with SkillLens	2026-05-24 19:31:24 +00:00
Cuzyoung	cff7ff6846	fix: rename remaining teacher/student refs, remove .gradio from repo - Fix teacher/student in deep_reflect, meta_reflect, sealqa, babyvision, mathverse, mmrb, swebench envs and prompt templates - Remove .gradio/certificate.pem from tracked files - Add .gradio/ to .gitignore Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>	2026-05-24 19:22:20 +00:00
Cuzyoung	7ae2d8766e	docs: restore clean README with Install/Data/QuickStart/WebUI/Citation only Keep remote project page header (badges, video), replace body with our streamlined 5-section README focused on reproducibility. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>	2026-05-24 19:19:19 +00:00
Lliar-liar	338a88d31c	Add model logos to results table	2026-05-24 19:18:57 +00:00
Cuzyoung	4a1b984d87	refactor: rename teacher/student to optimizer/target, remove best skills, fix slow update - Rename teacher -> optimizer, student -> target across all code, configs, docs, prompts - CLI: --teacher_model -> --optimizer_model, --student_model -> --target_model - Remove best_skill files, keep only initial skills - Fix slow update gate (force write into skill) - Fix SLOW_UPDATE marker stripping - Remove deep_reflect and meta_reflect mechanisms - Update .env.example with export prefix and azure_cli docs - Add endpoint empty validation in azure_openai.py Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>	2026-05-24 19:15:10 +00:00
Lliar-liar	6e165d5347	Add Microsoft favicon	2026-05-24 19:14:33 +00:00
Lliar-liar	dde7dc9dd8	Add SkillLens related project link	2026-05-24 19:12:27 +00:00
Lliar-liar	cd9a0a02b9	Restyle project page after SkillLens	2026-05-24 19:08:05 +00:00
Lliar-liar	607bf74a1b	Reorder hero evaluation stats	2026-05-24 18:52:05 +00:00
Lliar-liar	9605217e75	Use Microsoft logo in page header	2026-05-24 18:27:25 +00:00
Lliar-liar	c42d541828	Refine project links and citation section	2026-05-24 18:24:48 +00:00
Lliar-liar	2e05edc399	Add project links and citation section	2026-05-24 18:18:36 +00:00
Lliar-liar	6e7d5d0117	Clarify hero harness names	2026-05-24 18:15:35 +00:00
Yif Yang	441ccb9bda	Update README.md	2026-05-25 02:15:02 +08:00
Lliar-liar	88a99048a4	Align method comparison chart with page theme	2026-05-24 18:05:23 +00:00
Lliar-liar	bf2106808e	Remove method comparison implementation caption	2026-05-24 18:03:21 +00:00
Lliar-liar	ba0fa8c14b	Render method comparison from raw data	2026-05-24 18:00:08 +00:00
Lliar-liar	9012a79827	Add main results method comparison chart	2026-05-24 17:55:22 +00:00
Lliar-liar	c64fbcd4f8	Shorten hero target model label	2026-05-24 17:51:11 +00:00
Lliar-liar	6e1027f01a	Add harness count to hero badge	2026-05-24 17:48:32 +00:00
Lliar-liar	cd56a5fe7d	Make hero results badge more prominent	2026-05-24 17:43:29 +00:00
Lliar-liar	bbb250cc63	Clarify hero setting wins	2026-05-24 17:36:45 +00:00
Lliar-liar	5c45add28b	Update hero metrics to video results framing	2026-05-24 17:30:13 +00:00
Lliar-liar	e1896c691c	Improve ablation table layout	2026-05-24 17:23:28 +00:00
Lliar-liar	ec0841cccf	Remove duplicate GPT-5.5 results table	2026-05-24 17:15:24 +00:00
Lliar-liar	4019f1cbe7	Align webpage model terminology	2026-05-24 17:12:43 +00:00
Lliar-liar	cad3ab2d19	Simplify main results webpage table	2026-05-24 17:10:35 +00:00
Lliar-liar	9a064f7c97	Use YouTube teaser video	2026-05-24 14:59:26 +00:00
Lliar-liar	74cbe704fc	Polish project webpage copy	2026-05-24 14:55:44 +00:00
Lliar-liar	5862bbdc97	Add SkillOpt project webpage	2026-05-24 14:16:34 +00:00
				`@@ -0,0 +1 @@`
				<svg id="logomark" xmlns="http://www.w3.org/2000/svg" viewBox="0 0 74.492 100.25"><g id="tiny_-_black" data-name="tiny - black"><path d="M586.72,255.616a3.377,3.377,0,0,1,.448.031,5.917,5.917,0,0,1,3.581,2.79c.454,1.116.314,2.023-1.315,4.141L563.168,293.6l-8.558-10.047,29.348-26.616a4.406,4.406,0,0,1,2.762-1.321m0-1.5a5.766,5.766,0,0,0-3.69,1.643l-.041.032-.038.035L553.6,282.442l-1.077.977.943,1.107,8.558,10.047,1.145,1.344,1.141-1.348,26.267-31.022.022-.027.022-.028c1.574-2.046,2.327-3.622,1.516-5.619a7.309,7.309,0,0,0-4.779-3.714,5.083,5.083,0,0,0-.64-.043Z" transform="translate(-526.086 -245.559)"/><path d="M553.423,284.593l8.977,10.558L597.911,337.9c.873,1.093,1.419,2.186,1.047,3.418a4.092,4.092,0,0,1-2.721,2.837,3.557,3.557,0,0,1-1.045.159,4,4,0,0,1-2.687-1.124L548.01,300.808c-3.5-3.5-2.971-8.151.436-11.558l4.977-4.657m.124-2.17L552.4,283.5l-4.976,4.656c-4.192,4.191-4.372,9.816-.473,13.714l44.521,42.4a5.485,5.485,0,0,0,3.722,1.538,5.1,5.1,0,0,0,1.483-.224,5.59,5.59,0,0,0,3.719-3.838,5.176,5.176,0,0,0-1.31-4.788l-35.53-42.767-8.988-10.571-1.019-1.2Z" transform="translate(-526.086 -245.559)"/><path d="M562.4,295.151l9.556,11.5,5.761-5.356a7.926,7.926,0,0,0,.041-11.743l-43.7-41.923s-1.671-2.029-3.437-2.071a4.49,4.49,0,0,0-4.23,2.718c-.688,1.651-.194,2.809,1.315,4.97l29.306,35.565Z" transform="translate(-526.086 -245.559)"/><path d="M553.7,306.223l-17.116,21.024c-1.255,1.337-2.032,3.683-1.331,5.367a4.587,4.587,0,0,0,4.287,2.841,4.087,4.087,0,0,0,3.082-1.523l20.328-18.9Z" transform="translate(-526.086 -245.559)"/><path d="M592.074,250.547" transform="translate(-526.086 -245.559)" fill="#fff" stroke="#000" stroke-miterlimit="10" stroke-width="0.25"/></g></svg>
				`@@ -1 +0,0 @@`
				`"""BabyVision environment package for ReflACT."""`