{"owner":"microsoft","repo":"RD-Agent","hasSkills":true,"totalSkillsCount":11,"totalTokensCount":2873,"categories":["anthropic-skill","subagent-persona"],"hasMcp":false,"mcpConfig":null,"found":["rdagent/scenarios/rl/autorl_bench/agents/example_agent/config.yaml","rdagent/scenarios/rl/autorl_bench/agents/codex/AGENTS.md","rdagent/scenarios/rl/autorl_bench/agents/claude/AGENTS.md","rdagent/scenarios/rl/autorl_bench/core/skill.md","rdagent/scenarios/rl/autorl_bench/agents/codex/config.yaml","rdagent/scenarios/rl/autorl_bench/agents/claude/config.yaml","rdagent/scenarios/rl/autorl_bench/agents/gemini/AGENTS.md","rdagent/scenarios/rl/autorl_bench/agents/openhands/config.yaml","rdagent/scenarios/rl/autorl_bench/agents/gemini/config.yaml","rdagent/scenarios/rl/autorl_bench/agents/opencode/config.yaml","rdagent/scenarios/rl/autorl_bench/agents/rdagent/config.yaml"],"skills":{"rdagent/scenarios/rl/autorl_bench/agents/example_agent/config.yaml":"name: \"Example Agent\"\ndescription: \"GRPO 训练 + 评测\"\nstart: \"start.sh\"\nenv_vars:\n  TRAIN_RATIO: \"0.1\"\n  NUM_EPOCHS: \"1\"\n","rdagent/scenarios/rl/autorl_bench/agents/codex/AGENTS.md":"# AutoRL-Bench Agent Guidelines\n\n## Summary Maintenance (MANDATORY)\n\nYou MUST maintain a file called `summary.md` in the workspace root. Update it **after every training attempt and every submission**, not just at the end.\n\n### Format\n\n```markdown\n# 运行总结\n\n## Attempt N (YYYY-MM-DD HH:MM)\n- **状态**: ✅ 成功 / ❌ 失败\n- **Score**: X.XX | Improvement: +Y.YY | Best: Z.ZZ\n- **训练类型**: SFT / GRPO / PPO / DPO / ...\n- **超参数**: lr=X, epochs=Y, batch_size=Z, ...\n- **做了什么**: 简述本次尝试的策略和具体操作\n- **为什么**: 为什么选择这个方法/这些超参数\n- **问题/进步**: 遇到了什么问题，相比上次有什么改进\n- **关键代码**: 关键改动的代码片段（如有）\n- **下一步建议**: 基于本次结果，下一步打算怎么做\n```\n\n### Rules\n1. **Append only** — never overwrite previous attempts\n2. Analyze `code/train.py` source to extract training type and hyperparameters\n3. If training fails, extract root cause from error output\n4. \"做了什么\" and \"为什么\" are the most important fields — be thorough\n5. Update summary.md IMMEDIATELY after each submission result comes back\n6. Include the grading server response (score, improvement, best) verbatim\n","rdagent/scenarios/rl/autorl_bench/agents/claude/AGENTS.md":"# AutoRL-Bench Agent Guidelines\n\n## Summary Maintenance (MANDATORY)\n\nYou MUST maintain a file called `summary.md` in the workspace root. Update it **after every training attempt and every submission**, not just at the end.\n\n### Format\n\n```markdown\n# 运行总结\n\n## Attempt N (YYYY-MM-DD HH:MM)\n- **状态**: ✅ 成功 / ❌ 失败\n- **Score**: X.XX | Improvement: +Y.YY | Best: Z.ZZ\n- **训练类型**: SFT / GRPO / PPO / DPO / ...\n- **超参数**: lr=X, epochs=Y, batch_size=Z, ...\n- **做了什么**: 简述本次尝试的策略和具体操作\n- **为什么**: 为什么选择这个方法/这些超参数\n- **问题/进步**: 遇到了什么问题，相比上次有什么改进\n- **关键代码**: 关键改动的代码片段（如有）\n- **下一步建议**: 基于本次结果，下一步打算怎么做\n```\n\n### Rules\n1. **Append only** — never overwrite previous attempts\n2. Analyze `code/train.py` source to extract training type and hyperparameters\n3. If training fails, extract root cause from error output\n4. \"做了什么\" and \"为什么\" are the most important fields — be thorough\n5. Update summary.md IMMEDIATELY after each submission result comes back\n6. Include the grading server response (score, improvement, best) verbatim\n","rdagent/scenarios/rl/autorl_bench/core/skill.md":"你负责维护实验的累积运行总结文件 `reports/summary.md`。\n\n每轮实验结束后，你需要用 file_editor 在 `reports/summary.md` **末尾追加**一个新 section。\n\n如果 `reports/summary.md` 不存在，先创建文件，首行写 `# 运行总结`。\n\n每轮追加的 section 格式（严格遵守，不增删字段）：\n\n```\n## Iteration N (YYYY-MM-DD HH:MM, 耗时 Xs)\n- **状态**: ✅ 成功 / ❌ 失败 (exit_code=X)\n- **Score**: X | Improvement: X | Best: X (iter N)\n- **训练类型**: GRPO / SFT / PPO / copy_model / placeholder / unknown\n- **关键配置**: lr=X, epochs=X, batch=X, ...（从代码中提取）\n- **做了什么**: 具体策略（训练方法、reward 函数设计、数据处理等）\n- **为什么**: 为什么选择这个策略（基于上轮结果的推理）\n- **问题/进步**: 发现了什么问题，或相比上轮取得了什么进步\n- **关键代码**: 最能体现本轮策略的 3-5 行代码\n- **代码片段（上下文）**: 从 train.py 复制 15-40 行可运行上下文，标注函数名/行号范围；若与上轮相同写“与 Iteration N 相同，无变更”\n```\n\n数据来源建议（按优先级）：\n- Score/Improvement/Best: scores.json 或服务器返回\n- 状态/耗时/exit_code: run.log\n- 训练类型/关键配置/关键代码: code/train.py\n- 失败根因: agent.log + run.log\n\n规则：\n1. **追加**，不要覆盖已有内容\n2. 必须按 Iteration 递增写入：读取 `reports/summary.md` 中最后一个迭代号，当前必须是 N+1；若不满足先修正再写\n3. 分析 code下面的 源码，提取训练类型和超参数；若无法确定写 unknown，不留空\n4. 如果训练失败，必须给出可定位的根因（日志片段或错误类型），并标注 failure_type（如：code_error_runtime / rollout_logic_wrong / timeout_no_submission / copy_model_fallback / training_diverged / unknown）\n5. “做了什么”“为什么”是最重要字段，必须可复现、可检验，且“为什么”必须引用上轮证据（scores.json/run.log/agent.log）\n6. **问题/进步** 必须包含过程指标或失败类型（如：valid_submission_rate / first_valid_idx / time_to_first_improvement / time_used_ratio / failure_type）\n7. 长代码片段每轮最多 1 段，最多 40 行；优先贴本轮新增或改动处；若与上轮相同写“与 Iteration N 相同，无变更”\n8. 若与上一轮根因相同，避免整段重复，明确写出“新增证据/新增尝试/无新增”\n9. 同步追加 `reports/summary.jsonl`（同目录），每轮一行 JSON，字段建议：\n   - iteration, timestamp, duration_s, status, exit_code\n   - score, improvement\n   - train_type, failure_type\n   - 仅保留可量化字段（图表使用），不写 why/what/next_step 等长文本\n   - metrics 由 benchmark 后处理统一计算（不在 jsonl 中填写）\n   - 示例：\n     {\"iteration\": 3, \"timestamp\": \"2026-03-10 12:34\", \"duration_s\": 812, \"status\": \"success\", \"exit_code\": 0, \"score\": 65.0, \"improvement\": 20.0, \"train_type\": \"GRPO\", \"failure_type\": \"unknown\"}\n10. 写入后自检 section 完整性：上述字段必须齐全，缺失则补全后再结束\n","rdagent/scenarios/rl/autorl_bench/agents/codex/config.yaml":"name: \"Codex CLI Agent\"\ndescription: \"One-shot agent：给完整 prompt，Codex CLI 自主完成代码编写、训练、评测提交\"\nstart: \"start.sh\"\nenv_vars:\n  CODEX_MODEL: \"gpt-5.2\"            # 默认 LLM（可被环境变量覆盖）\n  CODEX_PROVIDER: \"litellm\"         # 默认走 litellm proxy（可被环境变量覆盖）\n  CODEX_TIMEOUT: \"36000\"            # Codex CLI 总超时（秒）= 10小时\n","rdagent/scenarios/rl/autorl_bench/agents/claude/config.yaml":"name: \"Claude CLI Agent\"\ndescription: \"One-shot agent：给完整 prompt，Claude Code CLI 自主完成代码编写、训练、评测提交\"\nstart: \"start.sh\"\nenv_vars:\n  CLAUDE_MODEL: \"claude-sonnet-4-6\"\n  CLAUDE_TIMEOUT: \"36000\"\n","rdagent/scenarios/rl/autorl_bench/agents/gemini/AGENTS.md":"# AutoRL-Bench Agent Guidelines\n\n## Summary Maintenance (MANDATORY)\n\nYou MUST maintain a file called `summary.md` in the workspace root. Update it **after every training attempt and every submission**, not just at the end.\n\n### Format\n\n```markdown\n# 运行总结\n\n## Attempt N (YYYY-MM-DD HH:MM)\n- **状态**: ✅ 成功 / ❌ 失败\n- **Score**: X.XX | Improvement: +Y.YY | Best: Z.ZZ\n- **训练类型**: SFT / GRPO / PPO / DPO / ...\n- **超参数**: lr=X, epochs=Y, batch_size=Z, ...\n- **做了什么**: 简述本次尝试的策略和具体操作\n- **为什么**: 为什么选择这个方法/这些超参数\n- **问题/进步**: 遇到了什么问题，相比上次有什么改进\n- **关键代码**: 关键改动的代码片段（如有）\n- **下一步建议**: 基于本次结果，下一步打算怎么做\n```\n\n### Rules\n1. **Append only** — never overwrite previous attempts\n2. Analyze `code/train.py` source to extract training type and hyperparameters\n3. If training fails, extract root cause from error output\n4. \"做了什么\" and \"为什么\" are the most important fields — be thorough\n5. Update summary.md IMMEDIATELY after each submission result comes back\n6. Include the grading server response (score, improvement, best) verbatim\n","rdagent/scenarios/rl/autorl_bench/agents/openhands/config.yaml":"name: \"OpenHands Agent\"\ndescription: \"固定阶段 pipeline：每轮 代码生成→训练→评测→反馈（参考 openhands-magic）\"\nstart: \"start.sh\"\nenv_vars:\n  MAX_ITERATIONS: \"30\"           # Pipeline 迭代次数（每轮=写代码+训练+评测）\n  TRAINING_TIMEOUT: \"36000\"      # 每轮训练超时（秒）= 10小时\n  MAX_AGENT_STEPS: \"20\"         # 每轮代码生成 agent 最大步数\n  LLM_MODEL: \"gpt-5.2\"\n","rdagent/scenarios/rl/autorl_bench/agents/gemini/config.yaml":"name: \"Gemini CLI Agent\"\ndescription: \"One-shot agent：给完整 prompt，Gemini CLI 自主完成代码编写、训练、评测提交\"\nstart: \"start.sh\"\nenv_vars:\n  GEMINI_MODEL: \"gemini-2.5-pro\"\n  GEMINI_TIMEOUT: \"36000\"\n","rdagent/scenarios/rl/autorl_bench/agents/opencode/config.yaml":"name: \"OpenCode Agent\"\ndescription: \"固定阶段 pipeline：代码生成→训练→评测→反馈（基于 opencode-rl）\"\nstart: \"start.sh\"\nenv_vars:\n  MAX_ITERATIONS: \"5\"\n  TRAINING_TIMEOUT: \"7200\"\n  MAX_AGENT_STEPS: \"25\"\n  MAX_RETRIES: \"20\"\n  STALE_TIMEOUT: \"1800\"\n  HTTP_TIMEOUT: \"600\"\n  EVAL_TIMEOUT: \"7200\"\n  MAX_STEPS: \"20\"\n","rdagent/scenarios/rl/autorl_bench/agents/rdagent/config.yaml":"name: \"RD-Agent\"\ndescription: \"RD-Agent RL Post-training Loop (自动假设生成 + 代码生成 + 验证迭代)\"\nstart: \"start.sh\"\nenv_vars:\n  STEP_N: \"200\"\n  LOOP_N: \"40\"\n"},"files":{"rdagent/scenarios/rl/autorl_bench/agents/example_agent/config.yaml":"name: \"Example Agent\"\ndescription: \"GRPO 训练 + 评测\"\nstart: \"start.sh\"\nenv_vars:\n  TRAIN_RATIO: \"0.1\"\n  NUM_EPOCHS: \"1\"\n","rdagent/scenarios/rl/autorl_bench/agents/codex/AGENTS.md":"# AutoRL-Bench Agent Guidelines\n\n## Summary Maintenance (MANDATORY)\n\nYou MUST maintain a file called `summary.md` in the workspace root. Update it **after every training attempt and every submission**, not just at the end.\n\n### Format\n\n```markdown\n# 运行总结\n\n## Attempt N (YYYY-MM-DD HH:MM)\n- **状态**: ✅ 成功 / ❌ 失败\n- **Score**: X.XX | Improvement: +Y.YY | Best: Z.ZZ\n- **训练类型**: SFT / GRPO / PPO / DPO / ...\n- **超参数**: lr=X, epochs=Y, batch_size=Z, ...\n- **做了什么**: 简述本次尝试的策略和具体操作\n- **为什么**: 为什么选择这个方法/这些超参数\n- **问题/进步**: 遇到了什么问题，相比上次有什么改进\n- **关键代码**: 关键改动的代码片段（如有）\n- **下一步建议**: 基于本次结果，下一步打算怎么做\n```\n\n### Rules\n1. **Append only** — never overwrite previous attempts\n2. Analyze `code/train.py` source to extract training type and hyperparameters\n3. If training fails, extract root cause from error output\n4. \"做了什么\" and \"为什么\" are the most important fields — be thorough\n5. Update summary.md IMMEDIATELY after each submission result comes back\n6. Include the grading server response (score, improvement, best) verbatim\n","rdagent/scenarios/rl/autorl_bench/agents/claude/AGENTS.md":"# AutoRL-Bench Agent Guidelines\n\n## Summary Maintenance (MANDATORY)\n\nYou MUST maintain a file called `summary.md` in the workspace root. Update it **after every training attempt and every submission**, not just at the end.\n\n### Format\n\n```markdown\n# 运行总结\n\n## Attempt N (YYYY-MM-DD HH:MM)\n- **状态**: ✅ 成功 / ❌ 失败\n- **Score**: X.XX | Improvement: +Y.YY | Best: Z.ZZ\n- **训练类型**: SFT / GRPO / PPO / DPO / ...\n- **超参数**: lr=X, epochs=Y, batch_size=Z, ...\n- **做了什么**: 简述本次尝试的策略和具体操作\n- **为什么**: 为什么选择这个方法/这些超参数\n- **问题/进步**: 遇到了什么问题，相比上次有什么改进\n- **关键代码**: 关键改动的代码片段（如有）\n- **下一步建议**: 基于本次结果，下一步打算怎么做\n```\n\n### Rules\n1. **Append only** — never overwrite previous attempts\n2. Analyze `code/train.py` source to extract training type and hyperparameters\n3. If training fails, extract root cause from error output\n4. \"做了什么\" and \"为什么\" are the most important fields — be thorough\n5. Update summary.md IMMEDIATELY after each submission result comes back\n6. Include the grading server response (score, improvement, best) verbatim\n","rdagent/scenarios/rl/autorl_bench/core/skill.md":"你负责维护实验的累积运行总结文件 `reports/summary.md`。\n\n每轮实验结束后，你需要用 file_editor 在 `reports/summary.md` **末尾追加**一个新 section。\n\n如果 `reports/summary.md` 不存在，先创建文件，首行写 `# 运行总结`。\n\n每轮追加的 section 格式（严格遵守，不增删字段）：\n\n```\n## Iteration N (YYYY-MM-DD HH:MM, 耗时 Xs)\n- **状态**: ✅ 成功 / ❌ 失败 (exit_code=X)\n- **Score**: X | Improvement: X | Best: X (iter N)\n- **训练类型**: GRPO / SFT / PPO / copy_model / placeholder / unknown\n- **关键配置**: lr=X, epochs=X, batch=X, ...（从代码中提取）\n- **做了什么**: 具体策略（训练方法、reward 函数设计、数据处理等）\n- **为什么**: 为什么选择这个策略（基于上轮结果的推理）\n- **问题/进步**: 发现了什么问题，或相比上轮取得了什么进步\n- **关键代码**: 最能体现本轮策略的 3-5 行代码\n- **代码片段（上下文）**: 从 train.py 复制 15-40 行可运行上下文，标注函数名/行号范围；若与上轮相同写“与 Iteration N 相同，无变更”\n```\n\n数据来源建议（按优先级）：\n- Score/Improvement/Best: scores.json 或服务器返回\n- 状态/耗时/exit_code: run.log\n- 训练类型/关键配置/关键代码: code/train.py\n- 失败根因: agent.log + run.log\n\n规则：\n1. **追加**，不要覆盖已有内容\n2. 必须按 Iteration 递增写入：读取 `reports/summary.md` 中最后一个迭代号，当前必须是 N+1；若不满足先修正再写\n3. 分析 code下面的 源码，提取训练类型和超参数；若无法确定写 unknown，不留空\n4. 如果训练失败，必须给出可定位的根因（日志片段或错误类型），并标注 failure_type（如：code_error_runtime / rollout_logic_wrong / timeout_no_submission / copy_model_fallback / training_diverged / unknown）\n5. “做了什么”“为什么”是最重要字段，必须可复现、可检验，且“为什么”必须引用上轮证据（scores.json/run.log/agent.log）\n6. **问题/进步** 必须包含过程指标或失败类型（如：valid_submission_rate / first_valid_idx / time_to_first_improvement / time_used_ratio / failure_type）\n7. 长代码片段每轮最多 1 段，最多 40 行；优先贴本轮新增或改动处；若与上轮相同写“与 Iteration N 相同，无变更”\n8. 若与上一轮根因相同，避免整段重复，明确写出“新增证据/新增尝试/无新增”\n9. 同步追加 `reports/summary.jsonl`（同目录），每轮一行 JSON，字段建议：\n   - iteration, timestamp, duration_s, status, exit_code\n   - score, improvement\n   - train_type, failure_type\n   - 仅保留可量化字段（图表使用），不写 why/what/next_step 等长文本\n   - metrics 由 benchmark 后处理统一计算（不在 jsonl 中填写）\n   - 示例：\n     {\"iteration\": 3, \"timestamp\": \"2026-03-10 12:34\", \"duration_s\": 812, \"status\": \"success\", \"exit_code\": 0, \"score\": 65.0, \"improvement\": 20.0, \"train_type\": \"GRPO\", \"failure_type\": \"unknown\"}\n10. 写入后自检 section 完整性：上述字段必须齐全，缺失则补全后再结束\n","rdagent/scenarios/rl/autorl_bench/agents/codex/config.yaml":"name: \"Codex CLI Agent\"\ndescription: \"One-shot agent：给完整 prompt，Codex CLI 自主完成代码编写、训练、评测提交\"\nstart: \"start.sh\"\nenv_vars:\n  CODEX_MODEL: \"gpt-5.2\"            # 默认 LLM（可被环境变量覆盖）\n  CODEX_PROVIDER: \"litellm\"         # 默认走 litellm proxy（可被环境变量覆盖）\n  CODEX_TIMEOUT: \"36000\"            # Codex CLI 总超时（秒）= 10小时\n","rdagent/scenarios/rl/autorl_bench/agents/claude/config.yaml":"name: \"Claude CLI Agent\"\ndescription: \"One-shot agent：给完整 prompt，Claude Code CLI 自主完成代码编写、训练、评测提交\"\nstart: \"start.sh\"\nenv_vars:\n  CLAUDE_MODEL: \"claude-sonnet-4-6\"\n  CLAUDE_TIMEOUT: \"36000\"\n","rdagent/scenarios/rl/autorl_bench/agents/gemini/AGENTS.md":"# AutoRL-Bench Agent Guidelines\n\n## Summary Maintenance (MANDATORY)\n\nYou MUST maintain a file called `summary.md` in the workspace root. Update it **after every training attempt and every submission**, not just at the end.\n\n### Format\n\n```markdown\n# 运行总结\n\n## Attempt N (YYYY-MM-DD HH:MM)\n- **状态**: ✅ 成功 / ❌ 失败\n- **Score**: X.XX | Improvement: +Y.YY | Best: Z.ZZ\n- **训练类型**: SFT / GRPO / PPO / DPO / ...\n- **超参数**: lr=X, epochs=Y, batch_size=Z, ...\n- **做了什么**: 简述本次尝试的策略和具体操作\n- **为什么**: 为什么选择这个方法/这些超参数\n- **问题/进步**: 遇到了什么问题，相比上次有什么改进\n- **关键代码**: 关键改动的代码片段（如有）\n- **下一步建议**: 基于本次结果，下一步打算怎么做\n```\n\n### Rules\n1. **Append only** — never overwrite previous attempts\n2. Analyze `code/train.py` source to extract training type and hyperparameters\n3. If training fails, extract root cause from error output\n4. \"做了什么\" and \"为什么\" are the most important fields — be thorough\n5. Update summary.md IMMEDIATELY after each submission result comes back\n6. Include the grading server response (score, improvement, best) verbatim\n","rdagent/scenarios/rl/autorl_bench/agents/openhands/config.yaml":"name: \"OpenHands Agent\"\ndescription: \"固定阶段 pipeline：每轮 代码生成→训练→评测→反馈（参考 openhands-magic）\"\nstart: \"start.sh\"\nenv_vars:\n  MAX_ITERATIONS: \"30\"           # Pipeline 迭代次数（每轮=写代码+训练+评测）\n  TRAINING_TIMEOUT: \"36000\"      # 每轮训练超时（秒）= 10小时\n  MAX_AGENT_STEPS: \"20\"         # 每轮代码生成 agent 最大步数\n  LLM_MODEL: \"gpt-5.2\"\n","rdagent/scenarios/rl/autorl_bench/agents/gemini/config.yaml":"name: \"Gemini CLI Agent\"\ndescription: \"One-shot agent：给完整 prompt，Gemini CLI 自主完成代码编写、训练、评测提交\"\nstart: \"start.sh\"\nenv_vars:\n  GEMINI_MODEL: \"gemini-2.5-pro\"\n  GEMINI_TIMEOUT: \"36000\"\n","rdagent/scenarios/rl/autorl_bench/agents/opencode/config.yaml":"name: \"OpenCode Agent\"\ndescription: \"固定阶段 pipeline：代码生成→训练→评测→反馈（基于 opencode-rl）\"\nstart: \"start.sh\"\nenv_vars:\n  MAX_ITERATIONS: \"5\"\n  TRAINING_TIMEOUT: \"7200\"\n  MAX_AGENT_STEPS: \"25\"\n  MAX_RETRIES: \"20\"\n  STALE_TIMEOUT: \"1800\"\n  HTTP_TIMEOUT: \"600\"\n  EVAL_TIMEOUT: \"7200\"\n  MAX_STEPS: \"20\"\n","rdagent/scenarios/rl/autorl_bench/agents/rdagent/config.yaml":"name: \"RD-Agent\"\ndescription: \"RD-Agent RL Post-training Loop (自动假设生成 + 代码生成 + 验证迭代)\"\nstart: \"start.sh\"\nenv_vars:\n  STEP_N: \"200\"\n  LOOP_N: \"40\"\n"},"items":[{"name":"config.yaml","path":"rdagent/scenarios/rl/autorl_bench/agents/claude/config.yaml","rawUrl":"https://raw.githubusercontent.com/microsoft/RD-Agent/HEAD/rdagent/scenarios/rl/autorl_bench/agents/claude/config.yaml","title":"Claude Skill","category":"anthropic-skill","format":"yaml","content":"name: \"Claude CLI Agent\"\ndescription: \"One-shot agent：给完整 prompt，Claude Code CLI 自主完成代码编写、训练、评测提交\"\nstart: \"start.sh\"\nenv_vars:\n  CLAUDE_MODEL: \"claude-sonnet-4-6\"\n  CLAUDE_TIMEOUT: \"36000\"\n","isInternal":false,"tokens":77,"sizeBytes":231},{"name":"config.yaml","path":"rdagent/scenarios/rl/autorl_bench/agents/codex/config.yaml","rawUrl":"https://raw.githubusercontent.com/microsoft/RD-Agent/HEAD/rdagent/scenarios/rl/autorl_bench/agents/codex/config.yaml","title":"Codex Skill","category":"anthropic-skill","format":"yaml","content":"name: \"Codex CLI Agent\"\ndescription: \"One-shot agent：给完整 prompt，Codex CLI 自主完成代码编写、训练、评测提交\"\nstart: \"start.sh\"\nenv_vars:\n  CODEX_MODEL: \"gpt-5.2\"            # 默认 LLM（可被环境变量覆盖）\n  CODEX_PROVIDER: \"litellm\"         # 默认走 litellm proxy（可被环境变量覆盖）\n  CODEX_TIMEOUT: \"36000\"            # Codex CLI 总超时（秒）= 10小时\n","isInternal":false,"tokens":140,"sizeBytes":410},{"name":"config.yaml","path":"rdagent/scenarios/rl/autorl_bench/agents/example_agent/config.yaml","rawUrl":"https://raw.githubusercontent.com/microsoft/RD-Agent/HEAD/rdagent/scenarios/rl/autorl_bench/agents/example_agent/config.yaml","title":"Example_agent Skill","category":"anthropic-skill","format":"yaml","content":"name: \"Example Agent\"\ndescription: \"GRPO 训练 + 评测\"\nstart: \"start.sh\"\nenv_vars:\n  TRAIN_RATIO: \"0.1\"\n  NUM_EPOCHS: \"1\"\n","isInternal":false,"tokens":48,"sizeBytes":125},{"name":"config.yaml","path":"rdagent/scenarios/rl/autorl_bench/agents/gemini/config.yaml","rawUrl":"https://raw.githubusercontent.com/microsoft/RD-Agent/HEAD/rdagent/scenarios/rl/autorl_bench/agents/gemini/config.yaml","title":"Gemini Skill","category":"anthropic-skill","format":"yaml","content":"name: \"Gemini CLI Agent\"\ndescription: \"One-shot agent：给完整 prompt，Gemini CLI 自主完成代码编写、训练、评测提交\"\nstart: \"start.sh\"\nenv_vars:\n  GEMINI_MODEL: \"gemini-2.5-pro\"\n  GEMINI_TIMEOUT: \"36000\"\n","isInternal":false,"tokens":74,"sizeBytes":223},{"name":"config.yaml","path":"rdagent/scenarios/rl/autorl_bench/agents/opencode/config.yaml","rawUrl":"https://raw.githubusercontent.com/microsoft/RD-Agent/HEAD/rdagent/scenarios/rl/autorl_bench/agents/opencode/config.yaml","title":"Opencode Skill","category":"anthropic-skill","format":"yaml","content":"name: \"OpenCode Agent\"\ndescription: \"固定阶段 pipeline：代码生成→训练→评测→反馈（基于 opencode-rl）\"\nstart: \"start.sh\"\nenv_vars:\n  MAX_ITERATIONS: \"5\"\n  TRAINING_TIMEOUT: \"7200\"\n  MAX_AGENT_STEPS: \"25\"\n  MAX_RETRIES: \"20\"\n  STALE_TIMEOUT: \"1800\"\n  HTTP_TIMEOUT: \"600\"\n  EVAL_TIMEOUT: \"7200\"\n  MAX_STEPS: \"20\"\n","isInternal":false,"tokens":117,"sizeBytes":334},{"name":"config.yaml","path":"rdagent/scenarios/rl/autorl_bench/agents/openhands/config.yaml","rawUrl":"https://raw.githubusercontent.com/microsoft/RD-Agent/HEAD/rdagent/scenarios/rl/autorl_bench/agents/openhands/config.yaml","title":"Openhands Skill","category":"anthropic-skill","format":"yaml","content":"name: \"OpenHands Agent\"\ndescription: \"固定阶段 pipeline：每轮 代码生成→训练→评测→反馈（参考 openhands-magic）\"\nstart: \"start.sh\"\nenv_vars:\n  MAX_ITERATIONS: \"30\"           # Pipeline 迭代次数（每轮=写代码+训练+评测）\n  TRAINING_TIMEOUT: \"36000\"      # 每轮训练超时（秒）= 10小时\n  MAX_AGENT_STEPS: \"20\"         # 每轮代码生成 agent 最大步数\n  LLM_MODEL: \"gpt-5.2\"\n","isInternal":false,"tokens":155,"sizeBytes":427},{"name":"config.yaml","path":"rdagent/scenarios/rl/autorl_bench/agents/rdagent/config.yaml","rawUrl":"https://raw.githubusercontent.com/microsoft/RD-Agent/HEAD/rdagent/scenarios/rl/autorl_bench/agents/rdagent/config.yaml","title":"Rdagent Skill","category":"anthropic-skill","format":"yaml","content":"name: \"RD-Agent\"\ndescription: \"RD-Agent RL Post-training Loop (自动假设生成 + 代码生成 + 验证迭代)\"\nstart: \"start.sh\"\nenv_vars:\n  STEP_N: \"200\"\n  LOOP_N: \"40\"\n","isInternal":false,"tokens":57,"sizeBytes":173},{"name":"skill.md","path":"rdagent/scenarios/rl/autorl_bench/core/skill.md","rawUrl":"https://raw.githubusercontent.com/microsoft/RD-Agent/HEAD/rdagent/scenarios/rl/autorl_bench/core/skill.md","title":"core","category":"anthropic-skill","format":"markdown","content":"你负责维护实验的累积运行总结文件 `reports/summary.md`。\n\n每轮实验结束后，你需要用 file_editor 在 `reports/summary.md` **末尾追加**一个新 section。\n\n如果 `reports/summary.md` 不存在，先创建文件，首行写 `# 运行总结`。\n\n每轮追加的 section 格式（严格遵守，不增删字段）：\n\n```\n## Iteration N (YYYY-MM-DD HH:MM, 耗时 Xs)\n- **状态**: ✅ 成功 / ❌ 失败 (exit_code=X)\n- **Score**: X | Improvement: X | Best: X (iter N)\n- **训练类型**: GRPO / SFT / PPO / copy_model / placeholder / unknown\n- **关键配置**: lr=X, epochs=X, batch=X, ...（从代码中提取）\n- **做了什么**: 具体策略（训练方法、reward 函数设计、数据处理等）\n- **为什么**: 为什么选择这个策略（基于上轮结果的推理）\n- **问题/进步**: 发现了什么问题，或相比上轮取得了什么进步\n- **关键代码**: 最能体现本轮策略的 3-5 行代码\n- **代码片段（上下文）**: 从 train.py 复制 15-40 行可运行上下文，标注函数名/行号范围；若与上轮相同写“与 Iteration N 相同，无变更”\n```\n\n数据来源建议（按优先级）：\n- Score/Improvement/Best: scores.json 或服务器返回\n- 状态/耗时/exit_code: run.log\n- 训练类型/关键配置/关键代码: code/train.py\n- 失败根因: agent.log + run.log\n\n规则：\n1. **追加**，不要覆盖已有内容\n2. 必须按 Iteration 递增写入：读取 `reports/summary.md` 中最后一个迭代号，当前必须是 N+1；若不满足先修正再写\n3. 分析 code下面的 源码，提取训练类型和超参数；若无法确定写 unknown，不留空\n4. 如果训练失败，必须给出可定位的根因（日志片段或错误类型），并标注 failure_type（如：code_error_runtime / rollout_logic_wrong / timeout_no_submission / copy_model_fallback / training_diverged / unknown）\n5. “做了什么”“为什么”是最重要字段，必须可复现、可检验，且“为什么”必须引用上轮证据（scores.json/run.log/agent.log）\n6. **问题/进步** 必须包含过程指标或失败类型（如：valid_submission_rate / first_valid_idx / time_to_first_improvement / time_used_ratio / failure_type）\n7. 长代码片段每轮最多 1 段，最多 40 行；优先贴本轮新增或改动处；若与上轮相同写“与 Iteration N 相同，无变更”\n8. 若与上一轮根因相同，避免整段重复，明确写出“新增证据/新增尝试/无新增”\n9. 同步追加 `reports/summary.jsonl`（同目录），每轮一行 JSON，字段建议：\n   - iteration, timestamp, duration_s, status, exit_code\n   - score, improvement\n   - train_type, failure_type\n   - 仅保留可量化字段（图表使用），不写 why/what/next_step 等长文本\n   - metrics 由 benchmark 后处理统一计算（不在 jsonl 中填写）\n   - 示例：\n     {\"iteration\": 3, \"timestamp\": \"2026-03-10 12:34\", \"duration_s\": 812, \"status\": \"success\", \"exit_code\": 0, \"score\": 65.0, \"improvement\": 20.0, \"train_type\": \"GRPO\", \"failure_type\": \"unknown\"}\n10. 写入后自检 section 完整性：上述字段必须齐全，缺失则补全后再结束\n","isInternal":false,"tokens":1083,"sizeBytes":3181},{"name":"AGENTS.md","path":"rdagent/scenarios/rl/autorl_bench/agents/claude/AGENTS.md","rawUrl":"https://raw.githubusercontent.com/microsoft/RD-Agent/HEAD/rdagent/scenarios/rl/autorl_bench/agents/claude/AGENTS.md","title":"Subagent: agents","category":"subagent-persona","format":"markdown","content":"# AutoRL-Bench Agent Guidelines\n\n## Summary Maintenance (MANDATORY)\n\nYou MUST maintain a file called `summary.md` in the workspace root. Update it **after every training attempt and every submission**, not just at the end.\n\n### Format\n\n```markdown\n# 运行总结\n\n## Attempt N (YYYY-MM-DD HH:MM)\n- **状态**: ✅ 成功 / ❌ 失败\n- **Score**: X.XX | Improvement: +Y.YY | Best: Z.ZZ\n- **训练类型**: SFT / GRPO / PPO / DPO / ...\n- **超参数**: lr=X, epochs=Y, batch_size=Z, ...\n- **做了什么**: 简述本次尝试的策略和具体操作\n- **为什么**: 为什么选择这个方法/这些超参数\n- **问题/进步**: 遇到了什么问题，相比上次有什么改进\n- **关键代码**: 关键改动的代码片段（如有）\n- **下一步建议**: 基于本次结果，下一步打算怎么做\n```\n\n### Rules\n1. **Append only** — never overwrite previous attempts\n2. Analyze `code/train.py` source to extract training type and hyperparameters\n3. If training fails, extract root cause from error output\n4. \"做了什么\" and \"为什么\" are the most important fields — be thorough\n5. Update summary.md IMMEDIATELY after each submission result comes back\n6. Include the grading server response (score, improvement, best) verbatim\n","isInternal":false,"tokens":374,"sizeBytes":1250},{"name":"AGENTS.md","path":"rdagent/scenarios/rl/autorl_bench/agents/codex/AGENTS.md","rawUrl":"https://raw.githubusercontent.com/microsoft/RD-Agent/HEAD/rdagent/scenarios/rl/autorl_bench/agents/codex/AGENTS.md","title":"Subagent: agents","category":"subagent-persona","format":"markdown","content":"# AutoRL-Bench Agent Guidelines\n\n## Summary Maintenance (MANDATORY)\n\nYou MUST maintain a file called `summary.md` in the workspace root. Update it **after every training attempt and every submission**, not just at the end.\n\n### Format\n\n```markdown\n# 运行总结\n\n## Attempt N (YYYY-MM-DD HH:MM)\n- **状态**: ✅ 成功 / ❌ 失败\n- **Score**: X.XX | Improvement: +Y.YY | Best: Z.ZZ\n- **训练类型**: SFT / GRPO / PPO / DPO / ...\n- **超参数**: lr=X, epochs=Y, batch_size=Z, ...\n- **做了什么**: 简述本次尝试的策略和具体操作\n- **为什么**: 为什么选择这个方法/这些超参数\n- **问题/进步**: 遇到了什么问题，相比上次有什么改进\n- **关键代码**: 关键改动的代码片段（如有）\n- **下一步建议**: 基于本次结果，下一步打算怎么做\n```\n\n### Rules\n1. **Append only** — never overwrite previous attempts\n2. Analyze `code/train.py` source to extract training type and hyperparameters\n3. If training fails, extract root cause from error output\n4. \"做了什么\" and \"为什么\" are the most important fields — be thorough\n5. Update summary.md IMMEDIATELY after each submission result comes back\n6. Include the grading server response (score, improvement, best) verbatim\n","isInternal":false,"tokens":374,"sizeBytes":1250},{"name":"AGENTS.md","path":"rdagent/scenarios/rl/autorl_bench/agents/gemini/AGENTS.md","rawUrl":"https://raw.githubusercontent.com/microsoft/RD-Agent/HEAD/rdagent/scenarios/rl/autorl_bench/agents/gemini/AGENTS.md","title":"Subagent: agents","category":"subagent-persona","format":"markdown","content":"# AutoRL-Bench Agent Guidelines\n\n## Summary Maintenance (MANDATORY)\n\nYou MUST maintain a file called `summary.md` in the workspace root. Update it **after every training attempt and every submission**, not just at the end.\n\n### Format\n\n```markdown\n# 运行总结\n\n## Attempt N (YYYY-MM-DD HH:MM)\n- **状态**: ✅ 成功 / ❌ 失败\n- **Score**: X.XX | Improvement: +Y.YY | Best: Z.ZZ\n- **训练类型**: SFT / GRPO / PPO / DPO / ...\n- **超参数**: lr=X, epochs=Y, batch_size=Z, ...\n- **做了什么**: 简述本次尝试的策略和具体操作\n- **为什么**: 为什么选择这个方法/这些超参数\n- **问题/进步**: 遇到了什么问题，相比上次有什么改进\n- **关键代码**: 关键改动的代码片段（如有）\n- **下一步建议**: 基于本次结果，下一步打算怎么做\n```\n\n### Rules\n1. **Append only** — never overwrite previous attempts\n2. Analyze `code/train.py` source to extract training type and hyperparameters\n3. If training fails, extract root cause from error output\n4. \"做了什么\" and \"为什么\" are the most important fields — be thorough\n5. Update summary.md IMMEDIATELY after each submission result comes back\n6. Include the grading server response (score, improvement, best) verbatim\n","isInternal":false,"tokens":374,"sizeBytes":1250}],"systemPromptSnippet":"<yakaai_skills repo=\"microsoft/RD-Agent\">\n<!-- File: rdagent/scenarios/rl/autorl_bench/agents/claude/config.yaml (Tokens: ~77 | Category: anthropic-skill) -->\nname: \"Claude CLI Agent\"\ndescription: \"One-shot agent：给完整 prompt，Claude Code CLI 自主完成代码编写、训练、评测提交\"\nstart: \"start.sh\"\nenv_vars:\n  CLAUDE_MODEL: \"claude-sonnet-4-6\"\n  CLAUDE_TIMEOUT: \"36000\"\n\n<!-- File: rdagent/scenarios/rl/autorl_bench/agents/codex/config.yaml (Tokens: ~140 | Category: anthropic-skill) -->\nname: \"Codex CLI Agent\"\ndescription: \"One-shot agent：给完整 prompt，Codex CLI 自主完成代码编写、训练、评测提交\"\nstart: \"start.sh\"\nenv_vars:\n  CODEX_MODEL: \"gpt-5.2\"            # 默认 LLM（可被环境变量覆盖）\n  CODEX_PROVIDER: \"litellm\"         # 默认走 litellm proxy（可被环境变量覆盖）\n  CODEX_TIMEOUT: \"36000\"            # Codex CLI 总超时（秒）= 10小时\n\n<!-- File: rdagent/scenarios/rl/autorl_bench/agents/example_agent/config.yaml (Tokens: ~48 | Category: anthropic-skill) -->\nname: \"Example Agent\"\ndescription: \"GRPO 训练 + 评测\"\nstart: \"start.sh\"\nenv_vars:\n  TRAIN_RATIO: \"0.1\"\n  NUM_EPOCHS: \"1\"\n\n<!-- File: rdagent/scenarios/rl/autorl_bench/agents/gemini/config.yaml (Tokens: ~74 | Category: anthropic-skill) -->\nname: \"Gemini CLI Agent\"\ndescription: \"One-shot agent：给完整 prompt，Gemini CLI 自主完成代码编写、训练、评测提交\"\nstart: \"start.sh\"\nenv_vars:\n  GEMINI_MODEL: \"gemini-2.5-pro\"\n  GEMINI_TIMEOUT: \"36000\"\n\n<!-- File: rdagent/scenarios/rl/autorl_bench/agents/opencode/config.yaml (Tokens: ~117 | Category: anthropic-skill) -->\nname: \"OpenCode Agent\"\ndescription: \"固定阶段 pipeline：代码生成→训练→评测→反馈（基于 opencode-rl）\"\nstart: \"start.sh\"\nenv_vars:\n  MAX_ITERATIONS: \"5\"\n  TRAINING_TIMEOUT: \"7200\"\n  MAX_AGENT_STEPS: \"25\"\n  MAX_RETRIES: \"20\"\n  STALE_TIMEOUT: \"1800\"\n  HTTP_TIMEOUT: \"600\"\n  EVAL_TIMEOUT: \"7200\"\n  MAX_STEPS: \"20\"\n\n<!-- File: rdagent/scenarios/rl/autorl_bench/agents/openhands/config.yaml (Tokens: ~155 | Category: anthropic-skill) -->\nname: \"OpenHands Agent\"\ndescription: \"固定阶段 pipeline：每轮 代码生成→训练→评测→反馈（参考 openhands-magic）\"\nstart: \"start.sh\"\nenv_vars:\n  MAX_ITERATIONS: \"30\"           # Pipeline 迭代次数（每轮=写代码+训练+评测）\n  TRAINING_TIMEOUT: \"36000\"      # 每轮训练超时（秒）= 10小时\n  MAX_AGENT_STEPS: \"20\"         # 每轮代码生成 agent 最大步数\n  LLM_MODEL: \"gpt-5.2\"\n\n<!-- File: rdagent/scenarios/rl/autorl_bench/agents/rdagent/config.yaml (Tokens: ~57 | Category: anthropic-skill) -->\nname: \"RD-Agent\"\ndescription: \"RD-Agent RL Post-training Loop (自动假设生成 + 代码生成 + 验证迭代)\"\nstart: \"start.sh\"\nenv_vars:\n  STEP_N: \"200\"\n  LOOP_N: \"40\"\n\n<!-- File: rdagent/scenarios/rl/autorl_bench/core/skill.md (Tokens: ~1083 | Category: anthropic-skill) -->\n你负责维护实验的累积运行总结文件 `reports/summary.md`。\n\n每轮实验结束后，你需要用 file_editor 在 `reports/summary.md` **末尾追加**一个新 section。\n\n如果 `reports/summary.md` 不存在，先创建文件，首行写 `# 运行总结`。\n\n每轮追加的 section 格式（严格遵守，不增删字段）：\n\n```\n## Iteration N (YYYY-MM-DD HH:MM, 耗时 Xs)\n- **状态**: ✅ 成功 / ❌ 失败 (exit_code=X)\n- **Score**: X | Improvement: X | Best: X (iter N)\n- **训练类型**: GRPO / SFT / PPO / copy_model / placeholder / unknown\n- **关键配置**: lr=X, epochs=X, batch=X, ...（从代码中提取）\n- **做了什么**: 具体策略（训练方法、reward 函数设计、数据处理等）\n- **为什么**: 为什么选择这个策略（基于上轮结果的推理）\n- **问题/进步**: 发现了什么问题，或相比上轮取得了什么进步\n- **关键代码**: 最能体现本轮策略的 3-5 行代码\n- **代码片段（上下文）**: 从 train.py 复制 15-40 行可运行上下文，标注函数名/行号范围；若与上轮相同写“与 Iteration N 相同，无变更”\n```\n\n数据来源建议（按优先级）：\n- Score/Improvement/Best: scores.json 或服务器返回\n- 状态/耗时/exit_code: run.log\n- 训练类型/关键配置/关键代码: code/train.py\n- 失败根因: agent.log + run.log\n\n规则：\n1. **追加**，不要覆盖已有内容\n2. 必须按 Iteration 递增写入：读取 `reports/summary.md` 中最后一个迭代号，当前必须是 N+1；若不满足先修正再写\n3. 分析 code下面的 源码，提取训练类型和超参数；若无法确定写 unknown，不留空\n4. 如果训练失败，必须给出可定位的根因（日志片段或错误类型），并标注 failure_type（如：code_error_runtime / rollout_logic_wrong / timeout_no_submission / copy_model_fallback / training_diverged / unknown）\n5. “做了什么”“为什么”是最重要字段，必须可复现、可检验，且“为什么”必须引用上轮证据（scores.json/run.log/agent.log）\n6. **问题/进步** 必须包含过程指标或失败类型（如：valid_submission_rate / first_valid_idx / time_to_first_improvement / time_used_ratio / failure_type）\n7. 长代码片段每轮最多 1 段，最多 40 行；优先贴本轮新增或改动处；若与上轮相同写“与 Iteration N 相同，无变更”\n8. 若与上一轮根因相同，避免整段重复，明确写出“新增证据/新增尝试/无新增”\n9. 同步追加 `reports/summary.jsonl`（同目录），每轮一行 JSON，字段建议：\n   - iteration, timestamp, duration_s, status, exit_code\n   - score, improvement\n   - train_type, failure_type\n   - 仅保留可量化字段（图表使用），不写 why/what/next_step 等长文本\n   - metrics 由 benchmark 后处理统一计算（不在 jsonl 中填写）\n   - 示例：\n     {\"iteration\": 3, \"timestamp\": \"2026-03-10 12:34\", \"duration_s\": 812, \"status\": \"success\", \"exit_code\": 0, \"score\": 65.0, \"improvement\": 20.0, \"train_type\": \"GRPO\", \"failure_type\": \"unknown\"}\n10. 写入后自检 section 完整性：上述字段必须齐全，缺失则补全后再结束\n\n<!-- File: rdagent/scenarios/rl/autorl_bench/agents/claude/AGENTS.md (Tokens: ~374 | Category: subagent-persona) -->\n# AutoRL-Bench Agent Guidelines\n\n## Summary Maintenance (MANDATORY)\n\nYou MUST maintain a file called `summary.md` in the workspace root. Update it **after every training attempt and every submission**, not just at the end.\n\n### Format\n\n```markdown\n# 运行总结\n\n## Attempt N (YYYY-MM-DD HH:MM)\n- **状态**: ✅ 成功 / ❌ 失败\n- **Score**: X.XX | Improvement: +Y.YY | Best: Z.ZZ\n- **训练类型**: SFT / GRPO / PPO / DPO / ...\n- **超参数**: lr=X, epochs=Y, batch_size=Z, ...\n- **做了什么**: 简述本次尝试的策略和具体操作\n- **为什么**: 为什么选择这个方法/这些超参数\n- **问题/进步**: 遇到了什么问题，相比上次有什么改进\n- **关键代码**: 关键改动的代码片段（如有）\n- **下一步建议**: 基于本次结果，下一步打算怎么做\n```\n\n### Rules\n1. **Append only** — never overwrite previous attempts\n2. Analyze `code/train.py` source to extract training type and hyperparameters\n3. If training fails, extract root cause from error output\n4. \"做了什么\" and \"为什么\" are the most important fields — be thorough\n5. Update summary.md IMMEDIATELY after each submission result comes back\n6. Include the grading server response (score, improvement, best) verbatim\n\n<!-- File: rdagent/scenarios/rl/autorl_bench/agents/codex/AGENTS.md (Tokens: ~374 | Category: subagent-persona) -->\n# AutoRL-Bench Agent Guidelines\n\n## Summary Maintenance (MANDATORY)\n\nYou MUST maintain a file called `summary.md` in the workspace root. Update it **after every training attempt and every submission**, not just at the end.\n\n### Format\n\n```markdown\n# 运行总结\n\n## Attempt N (YYYY-MM-DD HH:MM)\n- **状态**: ✅ 成功 / ❌ 失败\n- **Score**: X.XX | Improvement: +Y.YY | Best: Z.ZZ\n- **训练类型**: SFT / GRPO / PPO / DPO / ...\n- **超参数**: lr=X, epochs=Y, batch_size=Z, ...\n- **做了什么**: 简述本次尝试的策略和具体操作\n- **为什么**: 为什么选择这个方法/这些超参数\n- **问题/进步**: 遇到了什么问题，相比上次有什么改进\n- **关键代码**: 关键改动的代码片段（如有）\n- **下一步建议**: 基于本次结果，下一步打算怎么做\n```\n\n### Rules\n1. **Append only** — never overwrite previous attempts\n2. Analyze `code/train.py` source to extract training type and hyperparameters\n3. If training fails, extract root cause from error output\n4. \"做了什么\" and \"为什么\" are the most important fields — be thorough\n5. Update summary.md IMMEDIATELY after each submission result comes back\n6. Include the grading server response (score, improvement, best) verbatim\n\n<!-- File: rdagent/scenarios/rl/autorl_bench/agents/gemini/AGENTS.md (Tokens: ~374 | Category: subagent-persona) -->\n# AutoRL-Bench Agent Guidelines\n\n## Summary Maintenance (MANDATORY)\n\nYou MUST maintain a file called `summary.md` in the workspace root. Update it **after every training attempt and every submission**, not just at the end.\n\n### Format\n\n```markdown\n# 运行总结\n\n## Attempt N (YYYY-MM-DD HH:MM)\n- **状态**: ✅ 成功 / ❌ 失败\n- **Score**: X.XX | Improvement: +Y.YY | Best: Z.ZZ\n- **训练类型**: SFT / GRPO / PPO / DPO / ...\n- **超参数**: lr=X, epochs=Y, batch_size=Z, ...\n- **做了什么**: 简述本次尝试的策略和具体操作\n- **为什么**: 为什么选择这个方法/这些超参数\n- **问题/进步**: 遇到了什么问题，相比上次有什么改进\n- **关键代码**: 关键改动的代码片段（如有）\n- **下一步建议**: 基于本次结果，下一步打算怎么做\n```\n\n### Rules\n1. **Append only** — never overwrite previous attempts\n2. Analyze `code/train.py` source to extract training type and hyperparameters\n3. If training fails, extract root cause from error output\n4. \"做了什么\" and \"为什么\" are the most important fields — be thorough\n5. Update summary.md IMMEDIATELY after each submission result comes back\n6. Include the grading server response (score, improvement, best) verbatim\n</yakaai_skills>"}