CharlesYang030
244e346b83
- Skill optimization framework with training loop analogy - 11 benchmarks, 4 model backends (Azure OpenAI, Claude, Codex, Qwen) - WebUI for browser-based training control - Pluggable architecture for extending benchmarks and backends
1.2 KiB
1.2 KiB
SWE-bench Bug Fixing Skill
Overview
This skill guides agents in resolving real-world GitHub issues by producing correct patches.
Goal: Given a repository and an issue description, produce a minimal, correct git diff patch that resolves the issue without modifying test files.
Workflow
- Understand the issue. Read the problem statement carefully and restate the expected behavior before editing code.
- Locate relevant code. Use targeted search to identify the files, functions, and tests that encode the buggy behavior.
- Reproduce the issue. Build a small, local reproduction before changing source files when feasible.
- Implement the fix. Make the smallest source change that addresses the root cause.
- Verify the fix. Re-run the reproduction and any focused checks needed to confirm the change.
- Submit the patch. Generate a clean unified diff of only the source files you modified.
Key Rules
- Keep changes minimal and directly tied to the bug.
- Do not modify tests, fixtures, or unrelated configuration unless the issue explicitly requires it.
- Prefer understanding the code path before patching.
- Verify behavior after editing instead of relying on intuition.
- The final submission must be a valid unified diff.