projtool-bootstrap

projtool-bootstrap

This skill activates the first time a user wants to bring a project under projtool. It is intentionally narrow: the user must mention projtool by name or use a clearly-projtool verb such as retrofit, install, set up, or onboard.

When You Should Not Act

• The project already has .claude/skills/ml-research-workflow/; that project-level skill owns day-to-day projtool work.
• The user is asking about agent/client skills in general, not projtool.
• The user is discussing ML workflow but has not asked to use projtool.

Main Flow

The full flow is in references/retrofit-flow.md. Follow that file, and trust the live MCP tool schema over this document if they ever disagree.

Briefly:

1. Check MCP availability. mcp__projtool__detect_project_structure should be callable. If not, tell the user to run projtool setup once from a shell and restart the active MCP client in this directory.
2. Detect. Call mcp__projtool__detect_project_structure and summarize the detected train script, config directory, invocation style, package manager, data path hint, git state, and docs repo status.
3. Negotiate v4 config. Build a ProposedConfig using the v4 schema: paths.code_dir, main_worktree_name, data_dir, docs_dir, output_dir, remote_workspace, outputs_remote, data_remote, plus paths.docs_repo_url. Do not use removed v3 fields such as docs_repo_path or outputs_local.
4. Write templates. Call mcp__projtool__retrofit_write_template with the agreed config. Surface conflicts and warnings exactly; do not claim success if conflicts remain.
5. Tell the user what changed. The raw git repo is converted into a v4 container layout: code/main/ holds the git worktree, while .proj-tool/, .claude/, .codex/, docs/, data/, and output/ live at container root.
6. Hand off — executably, not rhetorically. The project-level ml-research-workflow skill (just written to .claude/skills/) owns experiments, image builds, training, reports, and sync work from here. retrofit_write_template returns next_action + readiness — follow them:
   • If the user's ORIGINAL request continues into training / reproduction ("接入 projtool 并复现...", "set up projtool and run..."), attempt to load the ml-research-workflow skill NOW and continue the work there in the same session.
   • If this session cannot load a freshly-written project skill, say so explicitly and tell the user to restart the session before any remote work. NEVER continue with raw remote_exec / SSH clone / tmux — that is how projects end up with zero tracked runs (the weavetime failure mode).
   • First reproduction belongs on upstream main: instance_up("main") → sync_now("main") → run_script(experiment_id="main", ...) (or run_script(experiment_id="exp-<topic>", worktree="main") for a semantic run identity). new_experiment is for branching real research work, not for getting a sync session.