project-init

name	project-init
description	Use this skill to create a new MLClaw project. Triggers when the user wants to start tracking an ML model, set up a project workspace, or initialize project structure. Use for: '新建项目', 'create project', 'initialize', 'set up a new model'. Not for running inference/evaluation (use the stage-specific skills).

/project-init — Initialize MLClaw Project

One-time project setup. Creates directory structure, project.json, and stage config templates.

Ask one question at a time — multiple questions at once overwhelms users.

Follow the Workflow State Protocol from CLAUDE.md: push on entry, update step as you progress, pop on completion.

Conversation flow

1. Ask: What is the project name? (lowercase, underscores, e.g., rtdetr_detection)

2. Resolve workspace. Default = workspace_root: line at the top of CLAUDE.md (read it once, expand ~/). Confirm with user; if they pick a different path, replace the workspace_root: line in CLAUDE.md so subsequent invocations get the new default. If --workspace <path> was passed in the user's prompt, skip the confirm and use it directly. mkdir -p whichever path won — workspaces are cheap, accidentally creating one is harmless.

3. Ask: Which stages to enable? Options: data, exploration, training, evaluation, inference, deployment

4. For each enabled stage, ask one at a time: Code source for {stage}? (git URL / local path / skip for now)

5. Show a full summary with all fields and values (including defaults):

   Project: detection
   Workspace: ~/agent_space/mlclaw/projects
   Root: ~/agent_space/mlclaw/projects/detection
   Created: 2026-03-16
   
   Stages:
     data:        disabled
     training:    enabled  repo: https://github.com/...
     inference:   enabled  repo: skipped
     ...
   
   Create? Anything to change?
   

6. User confirms -> create everything. User wants changes -> update and show summary again.

Write project.json

Fixed keys, agent only modifies values. Template: lifecycle/project.json.

Each stage has: enabled, code_path (stages/{stage}/code), code_source (source, path, branch, commit, credentials). Paths section: stages, runs_pattern, artifacts_pattern, data_pattern.

Create project

MLCLAW_ROOT=$(python <repo>/lifecycle/scripts/shared/workspaces.py tool)
python "$MLCLAW_ROOT/lifecycle/scripts/project-init/init_project.py" '<project_json_string>' "$MLCLAW_ROOT"

Creates directories, copies templates, writes .gitignore, runs git init + initial commit. $HOME-relative paths (project root, workspace, each stage's code_source.path) are rewritten to ~/-prefixed form in project.json so the file survives rsync across machines.

Fallback: if script fails, manually create directories (stages/{stage}/code, runs, artifacts, data for each enabled stage), copy JSON templates from lifecycle/, write .gitignore, git init + commit.

Clone / Link code

The unified contract (see CLAUDE.md "Code Source Resolution") is code_dir = stages/{stage}/code/_source if exists else stages/{stage}/code. Per source mode:

• Git URL (code_source.source == "github"): git clone <code_source.path> into stages/{stage}/code/, record branch + commit (HEAD SHA) in project.json, remove .git so the code becomes plain files tracked by project git. No _source symlink for this mode.
• Local path (code_source.source == "local"): init_project.py already creates the symlink stages/{stage}/code/_source -> expanduser(code_source.path) during creation. Do not copy — the user iterates in their own repo, copy creates bidirectional sync friction. Reproducibility comes from code_snapshot.py at run-time, not from a project-wide copy. After rsync to a new machine the symlink will dangle (it stores an expanded absolute path); recreate with ln -sfn $(python -c "import json,os;print(os.path.expanduser(json.load(open('project.json'))['stages']['<stage>']['code_source']['path']))") stages/<stage>/code/_source.
• Server (code_source.source == "server"): scp into stages/{stage}/code/, no _source symlink.

Code modifications during runs stay in project git (for github/server) or are captured per-run via code_snapshot.py SHA + dirty patch (for local) — never pushed back to the original repo.

After creation

1. The chosen workspace was already upserted in step 2d — last_used is fresh and n_projects is now incremented to include this new project. If it wasn't yet upserted (unusual — only when the script was unavailable and you used the bootstrap fallback path), run python lifecycle/scripts/shared/workspaces.py upsert <workspace> now.
2. Tell user: Project created at {root}.
3. Downstream: offer /resources and available /{stage}-init skills for enabled stages. For stages without an init skill yet, note it's not available.