Feat/skill evolution project#919
Conversation
There was a problem hiding this comment.
Code Review
This pull request introduces a skill evolution workflow framework, including configuration files, a SearchQA task implementation, and utility functions. The review identifies critical correctness issues regarding index alignment when rollouts fail, which can be addressed by updating gather_with_semaphore to optionally retain None values. Additionally, high-severity issues were raised concerning shared state mutations from shallow-copying configurations, a persistent trajectories buffer that causes redundant reflections and potential overfitting, and incomplete TODO blocks for rejected updates and macro skill management. Other recommendations include fixing a configuration key mismatch, correcting a signature mismatch in BaseDataset.load_data, and reshuffling training data upon epoch reset.
Important
The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.
There was a problem hiding this comment.
Code Review
This pull request introduces a skill evolution workflow framework, including agent configurations, task definitions for SearchQA, and workflow orchestration utilities. The review feedback highlights several critical issues: a bug in gather_with_semaphore that filters out exceptions and causes silent data misalignment during evaluation; a potential crash in _extract_answer when handling non-string or null responses; infinite trajectory accumulation in trajectories_buffer leading to redundant reflections; an incomplete feedback loop due to an unimplemented TODO for rejected updates; a configuration key mismatch for the rejected update buffer size; and an abstract method signature mismatch for load_data in the base dataset class.
Important
The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.
Change Summary
The experimental results on a small-scale SearchQA dataset are as follows (20 for training, 10 for validation, and 70 for testing):
[INFO:ms_agent] Test Results: Initial Skills Score: 0.7286 (from ../../../results/msagent_searchqa_qwen36flash/workdir/test/init/skills) Last Skills Score: 0.7571 (from ../../../results/msagent_searchqa_qwen36flash/workdir/test/last/skills= ../../../results/msagent_searchqa_qwen36flash/workdir/epoch_01/step_final/skills) Best Skills Score: 0.8000 (from ../../../results/msagent_searchqa_qwen36flash/workdir/test/best/skills= ../../../results/msagent_searchqa_qwen36flash/workdir/epoch_01/step_0001/skills)Related issue number
Checklist
pre-commit installandpre-commit run --all-filesbefore git commit, and passed lint check.