research-harness-workflow

Generate a minimal remote artifact pull manifest (results.json + train/logs + eval logs/metrics/vis) and place files under research_workspace/experiments/<ExpID>/remote_artifacts.

Generate a copy-paste remote (AutoDL/tmux) run snippet (git fetch/checkout/pull + scripts/train_eval_shutdown_autodl.sh). Use when you need a remote train→eval command block and must manually choose train/eval scripts from scripts/.

You MUST use this before any creative work - creating features, building components, adding functionality, or modifying behavior. Explores user intent, requirements and design before implementation.

Analyze experiment results pulled by autodl-remote-pull-manifest (remote_artifacts) and generate a Hypothesis→Evidence→Next-Experiment (H→E→N) report. Use when you want evidence-first analysis from logs/metrics/vis and an actionable next-step plan to improve <主指标>.

Ingest pulled remote artifacts under research_workspace/experiments/<ExpID>/remote_artifacts, summarize latest eval (summary.json), and update research_workspace/00-实验记录.md (index row + detailed section). Use when you want to record <主指标>/<辅助指标> into 00-实验记录.md after pulling remote artifacts.

Use when the input is a user-approved design doc or implementation plan and the agent needs to turn it into an executable issues CSV before handing off to CSV execution.

Use when executing an existing task CSV and the agent needs to push all actionable rows to closed-loop completion without stopping between issues.

Use when the input is a complex task description that needs decomposition, persistence, and recovery before execution through a generated .mission CSV.

Use when the user wants to continue, resume, or recover interrupted mission work and the agent needs to locate unfinished CSV artifacts before execution resumes.

Use when a task may be a user-approved design/plan document, an existing task CSV, a markdown execution or investigation document, a long-running task description, or a recovery request and needs routing to the right mission sub-skill.

Review implemented research code after local lightweight validation and before the first training, evaluation, remote run, or experiment-result generation. Use for PRERUN-REVIEW rows, train/eval launch gates, AutoDL/tmux runs, or any task where wrong code would waste training time or contaminate conclusions.

Commit completed remote experiment results in the nested research_workspace repository. Use after pulling remote artifacts, ingesting metrics, updating research_workspace/00-实验记录.md or research_workspace/STATE.md, generating experiment analysis/HEN/next-step/search-evidence files, or whenever current ExpID research outputs should be consolidated into one linked research_workspace commit.

Use when encountering any bug, test failure, or unexpected behavior, before proposing fixes

Generate a minimal remote artifact pull manifest (results.json + train/logs + eval logs/metrics/vis) and place files under research_workspace/experiments/<ExpID>/remote_artifacts.

Use when a research CSV/Spec needs a remote AutoDL train→eval command from train/eval intent, constraints, and scripts available under scripts/.

Analyze experiment results pulled by autodl-remote-pull-manifest (remote_artifacts) and generate a Hypothesis→Evidence→Next-Experiment (H→E→N) report. Use when you want evidence-first analysis from logs/metrics/vis and an actionable next-step plan to improve <主指标>.

Ingest pulled remote artifacts under research_workspace/experiments/<ExpID>/remote_artifacts, summarize latest eval (summary.json), and update research_workspace/00-实验记录.md (index row + detailed section). Use when you want to record <主指标>/<辅助指标> into 00-实验记录.md after pulling remote artifacts.

Use when the input is a user-approved design doc or implementation plan and the agent needs to turn it into an executable issues CSV before handing off to CSV execution.

Use when executing an existing task CSV and the agent needs to push all actionable rows to closed-loop completion without stopping between issues.