multirepo-space（多仓库空间）

multirepo-space

Tools

This SKILL orchestrates two tools:

• workspace-setup.sh — bundled script for workspace infrastructure (dirs, symlinks, settings.json, AGENTS.md/CLAUDE.md)
• agent-factory — npm CLI for creating AI agents (npx @tacuchi/agent-factory)

Quick routing

User intent → Action:

• "Create/scaffold workspace" → Full setup flow
• "Add a repo" → Add flow
• "Remove/detach repo" → Remove flow
• "Check/verify/health" → Status flow
• "Create an agent" → agent-factory create directly
• Single repo project → DO NOT use this Skill
• Monorepo (Nx/Turborepo) → DO NOT use this Skill

Pre-flight checklist

• Workspace dir exists? If not, mkdir first.
• New workspace or existing? → setup vs add.
• All repo paths absolute? Required — symlinks break with relative paths.
• Always run with --dry-run first and show the user the output before executing.

Security

Write commands (setup, add, remove) perform irreversible filesystem changes: symlinks, file overwrites (AGENTS.md, CLAUDE.md, settings.json), and rm -rf on skill dirs.

Before any write command:

1. Run the command with --dry-run and show the output to the user.
2. Ask for explicit approval: "Proceed with the above changes?"
3. Only after approval, re-run without --dry-run.

Never pass -y to workspace-setup.sh without prior user approval in the current session. agent-factory create writes files; confirm with the user before invoking.

Full setup flow

Step 1: Detect stacks

npx @tacuchi/agent-factory detect <repo_path> --json -q

Capture JSON. Extract alias, primaryTech, stackCsv, verifyCommands. Use stackCsv exactly as returned — pass full value to --stacks.

Step 2: Preview workspace infrastructure (dry-run)

bash "$SKILL_DIR/scripts/workspace-setup.sh" setup <workspace_path> <repo1> <repo2> ... \
  --stacks "stackCsv1|stackCsv2|..." --dry-run

Show output to user. Ask: "Proceed with these changes?" On approval:

bash "$SKILL_DIR/scripts/workspace-setup.sh" setup <workspace_path> <repo1> <repo2> ... \
  --stacks "stackCsv1|stackCsv2|..." -y

Creates: dirs, symlinks, settings.json, AGENTS.md, CLAUDE.md, config.

Step 3: Create specialist agents

Confirm with the user before running. Per repo:

npx @tacuchi/agent-factory create \
  --name repo-<alias> --role specialist --scope <repo_path> \
  --output <workspace_path> --target all -q

Auto-appends -agent suffix: repo-<alias> → repo-<alias>-agent.md.

Step 4: Create coordinator agent

Confirm with the user before running.

npx @tacuchi/agent-factory create \
  --name coordinator --role coordinator --model opus \
  --specialists "repo-<alias1>-agent,repo-<alias2>-agent,..." \
  --repo-count <N> --output <workspace_path> --target all -q

--specialists list must use full names with -agent suffix.

Add repo flow

Step 1: Detect stack

npx @tacuchi/agent-factory detect <repo_path> --json -q

Step 2: Preview add (dry-run)

bash "$SKILL_DIR/scripts/workspace-setup.sh" add <workspace_path> <repo_path> \
  --alias <alias> --stack-csv "<stackCsv>" --dry-run

Show output to user. Ask: "Proceed with these changes?" On approval:

bash "$SKILL_DIR/scripts/workspace-setup.sh" add <workspace_path> <repo_path> \
  --alias <alias> --stack-csv "<stackCsv>" -y

Step 3: Create specialist

Confirm with the user before running.

npx @tacuchi/agent-factory create \
  --name repo-<alias> --role specialist --scope <repo_path> \
  --output <workspace_path> --target claude -q

Step 4: Regenerate coordinator

Get updated specialist list (with -agent suffix) from existing symlinks. Confirm with the user before running.

npx @tacuchi/agent-factory create \
  --name coordinator --role coordinator --model opus \
  --specialists "<updated_csv_list_with_agent_suffix>" \
  --repo-count <updated_N> --output <workspace_path> --target all -q

Remove repo flow

Step 1: Preview removal (dry-run)

bash "$SKILL_DIR/scripts/workspace-setup.sh" remove <workspace_path> <alias> --dry-run

Show output to user. Ask: "Proceed with these changes?" On approval:

bash "$SKILL_DIR/scripts/workspace-setup.sh" remove <workspace_path> <alias> -y

Removes: symlink, agent files (.agents/,.claude/agents/,.agents/skills/), regenerates docs.

Step 2: Regenerate coordinator

Get updated specialist list from remaining symlinks. Confirm with the user before running.

npx @tacuchi/agent-factory create \
  --name coordinator --role coordinator --model opus \
  --specialists "<updated_csv_list>" --repo-count <updated_N> \
  --output <workspace_path> --target claude -q

Status flow

bash "$SKILL_DIR/scripts/workspace-setup.sh" status <workspace_path>
npx @tacuchi/agent-factory list <workspace_path>

Custom agent flow

To create additional agents (architecture, style, code-review, etc.). Confirm with the user before running.

npx @tacuchi/agent-factory create \
  --name <agent-name> \
  --role custom \
  --description "<short description>" \
  --instructions "<body text or path to .md file>" \
  --model sonnet \
  --output <workspace_path> \
  --target all -q

If detect returns "Generic"

Ask the user: "Stack was not auto-detected for [repo]. What is the main language/framework?" Then pass --stack-csv to the setup/add command so it persists in config.

Constraints

• Always run --dry-run and show output to user before any write command.
• Always confirm with the user before invoking agent-factory create.
• Never pass -y without explicit user approval in the current session.
• Always use absolute paths — symlinks break with relative.
• Do not modify settings.json manually.
• Single repo or monorepo (Nx/Turborepo) → do NOT use this skill.

Replace $SKILL_DIR with the absolute path to this skill's directory.

Agent hierarchy

coordinator-agent (opus) → repo-*-agent (sonnet) per repo. Coordinator delegates to specialists via Task. Specialists execute autonomously.

Compatibility

Supported targets (one at a time, never combined):

• claude → .claude/agents/*.md + .agents/skills/*/SKILL.md
• codex → .agents/*.md + .agents/skills/*/SKILL.md
• gemini → .gemini/agents/*.md + .agents/skills/*/SKILL.md
• crush → .crush.json + .agents/skills/*/SKILL.md
• warp → .agents/skills/*/SKILL.md
• AGENTS.md → Universal context file (20+ tools)