Research Result Commit

Overview

Use this workflow to turn a completed experiment result ingest into one auditable research_workspace commit. Keep issues/, docs/, source code, configs, scripts, and tests in the main repository; this skill only handles the nested research_workspace repository.

Workflow

1. Identify the current ExpID.

   • Prefer the user-provided ExpID or the active mission/CSV context.
   • If absent, infer it from changed paths under research_workspace/experiments/<ExpID>/.
   • If multiple ExpIDs changed and the intended scope is unclear, stop and ask for the minimal clarification.

2. Inspect both repository boundaries.

   • In the main repository, note the associated branch and commit when the experiment depends on current main-repo code or workflow: git branch --show-current and git rev-parse --short HEAD.
   • Run git -C research_workspace rev-parse --show-toplevel. Its resolved path must be exactly <PROJECT_ROOT>/research_workspace, with its own Git metadata. A successful Git command alone does not establish independence: Git searches parent directories.
   • If it resolves to the code repository, do not stage or commit research files through this skill. Initialize a dedicated repository with git init -b main research_workspace, preserve the existing files, and confirm its top-level directory. If the code repository already tracks research data, preserve its history and complete that tracking migration before committing experiment results. A template checkout may distribute tracked examples; it is not an initialized project research repository.
   • In the verified research repository, run git status --short and git diff --stat.
   • Do not stage or commit issues/, docs/, source code, configs, scripts, or tests from this workflow.

3. Select only files belonging to the current experiment result.

   • Always consider experiments/<ExpID>/, EXPERIMENTS.csv, and STATE.md.
   • Include other research_workspace files only when they are clearly tied to the current ExpID, such as command records, search evidence, analysis drafts visualization notes, or next-step documents.
   • Exclude unrelated dirty files and unrelated experiments, even if they are already present in research_workspace.

4. Stage deliberately.

   • Use explicit git add paths, for example:

cd <PROJECT_ROOT>/research_workspace
git add experiments/<ExpID> EXPERIMENTS.csv STATE.md
git add <other-current-ExpID-research-files>
git diff --cached --stat

5. Commit once.
   • Use one commit for the pulled artifacts, records, analysis, search evidence, and state updates for this ExpID.
   • Use the project commit style. Template:

📃 docs(exp): 记录 <ExpID> 远程结果与分析

Why:
- 固化本次远程训练/评估的日志、指标、分析和可恢复状态

Why this works:
- 当前 ExpID 的 artifacts/analysis/search evidence 保存结论证据
- EXPERIMENTS.csv 与 STATE.md 记录实验状态
- 关联主仓库: <branch> @ <commit>

Remaining:
- <下一步、已知限制或 validation_gap>

6. Report the result.
   • Return the research_workspace commit SHA.
   • Mention any excluded dirty files or validation_gap.