> **⚠️ OUTPUT CONTRACT (READ FIRST):** `guides/output-contract.md` > Checklist-only output. No tool-call narration. No raw MCP/JSON/log dumps. > One checklist upfront, updated in place, shown again at end with a 1-line verdict. > **Sandbox contract:** This guide honors `RDC_TEST=1` per `guides/agent-bootstrap.md` § RDC_TEST Sandbox Contract. Destructive external calls short-circuit under the flag. # /rdc:setup — Intelligent Project Setup Scan the current project, detect all existing configuration, and generate (or update) `.rdc/config.json` and overlay guides. Works on both new and existing projects. --- ## Step 1 — Scan the project Read these files (all optional — use what exists): | File | What to extract | |------|----------------| | `package.json` | `name`, `description`, stack (next/vite/etc), monorepo? | | `pnpm-workspace.yaml` or `lerna.json` | monorepo structure | | `git remote get-url origin` | GitHub org + repo name | | `git branch -a` | main/develop/master detection | | `.env.local`, `apps/**/.env.local` | `NEXT_PUBLIC_SUPABASE_URL` → extract project ref | | `CLAUDE.md` | PROJECT_SCOPE, forbidden commands, branch rules, supabase ref | | `.claude/rules/*.md` | existing rule files, what systems are covered | | `docs/guides/agent-bootstrap.md` or `.rdc/guides/agent-bootstrap.md` | existing credentials pattern | | `.rdc/config.json` | already configured? just show current state + offer update | | `docs/guides/`, `docs/plans/`, `.rdc/guides/`, `.rdc/plans/` | existing directory structure (either layout) | | `apps/` or `sites/` | app names, ports from package.json `dev` scripts | | `supabase/migrations/` | schema exists? | | `tailwind.config.*` | Tailwind confirmed | | `vitest.config.*` or `jest.config.*` | test framework | | `.coolify/` or coolify deployment refs | deploy platform | Emit a structured summary of what you found: ``` Detected: Project name : GitHub : / Supabase ref : Branches : main=, dev= Stack : Monorepo : yes/no () Apps found : Guides dir : <.rdc/guides or docs/guides or "none"> Existing cfg : yes/no Hook scope : ``` --- ## Step 2 — Check for gaps List anything that was NOT found and ask the user to fill in: ``` Missing (please provide): [ ] Project description [ ] Deploy platform / dashboard URL [ ] Forbidden commands beyond defaults? [ ] Any satellite repos (other GitHub repos in this project)? ``` Keep questions to a minimum — only ask what can't be auto-detected. --- ## Step 3 — Generate .rdc/config.json Using all detected + user-supplied values, write `.rdc/config.json`: ```json { "name": "", "version": "1.0.0", "description": "", "rdc_skills_version": ">=0.2.0", "hook_scope": "", "git": { "org": "", "repo": "", "main_branch": "
", "dev_branch": "", "auto_commit_branch": "" }, "supabase": { "ref": "", "url": "https://.supabase.co", "mcp_server": "mcp__claude_ai_Supabase__execute_sql" }, "credentials": { "provider": "clauth", "daemon_url": "http://127.0.0.1:52437", "env_paths": [".env.local"] }, "repos": [ { "path": ".", "role": "primary", "description": "" } ], "paths": { "guides": ".rdc/guides", "plans": ".rdc/plans", "reports": ".rdc/reports", "research": ".rdc/research", "state": ".rdc/state", "systems": "docs/systems" }, "work_items": { "enabled": true }, "constraints": { "forbidden_commands": ["pnpm build"], "typecheck_command": "npx tsc --noEmit", "test_command": "npx vitest run", "never_push_to": ["main"] } } ``` If `.rdc/config.json` already exists, show a diff of proposed changes and ask before overwriting. --- ## Step 4 — Generate overlay guides Create `.rdc/guides/agent-bootstrap.md` if it doesn't exist: ```markdown # Agent Bootstrap — > Project overlay — extends rdc-skills base guide. ## Credentials curl -s http://127.0.0.1:52437/get/ ## Git Rules - Branch: `` for all work. `` = production. - GitHub org: - Auto-commit to `` after every logical block. ## Supabase Project ref: `` Use `mcp__claude_ai_Supabase__execute_sql` — no `project_id` needed. ## Work Items SELECT get_open_epics(); SELECT insert_work_item(p_title := '...', p_priority := 'high'); SELECT update_work_item_status(''::uuid, 'review', '["Implementation complete; ready for validator"]'::jsonb, '', 'agent'); Create work items BEFORE starting work. ## Completion Report After finishing work, output: ## Work Summary - Files changed: - Tests: pass/fail - Work items updated: - Committed: ``` If `docs/guides/` or `.rdc/guides/` already has an `agent-bootstrap.md`, read it and preserve any custom sections — only add missing sections, don't overwrite existing ones. --- ## Step 5 — Offer to move docs/ dirs If `docs/guides/`, `docs/plans/`, `docs/reports/`, or `docs/research/` exist, ask: ``` Found existing docs/ directories. Move to .rdc/ layout? docs/guides → .rdc/guides [Y/n] docs/plans → .rdc/plans [Y/n] docs/reports → .rdc/reports [Y/n] docs/research → .rdc/research [Y/n] ``` Move only the directories the user confirms. --- ## Step 6 — Scaffold CLAUDE.ai.md and relay dirs Create `.rdc/relay/from-claude-ai/`, `.rdc/relay/from-claude-code/`, and `.rdc/relay/RELAY.md`. Create `CLAUDE.ai.md` in the project root if it doesn't exist: ```markdown # CLAUDE.ai — Startup Guide for claude.ai Sessions > Read this at the start of every claude.ai session on this repo. ## What You Are You are a claude.ai session — the **creative studio** in this stack. You produce artifacts, React prototypes, plans, copy, and designs. Claude Code CLI is the **build engine** — it runs agents, manages git, enforces hooks, and ships to production. **This is co-working, not handoff.** You produce real artifacts; Claude Code adapts and ships them. ## Startup Sequence — every session ``` 1. fs_read CLAUDE.ai.md ← you are here 2. fs_read CLAUDE.md ← CLI routing, hard rules 3. SELECT get_open_epics() ← what's in flight 4. fs_list .rdc/relay/from-claude-code/ ← messages from Claude Code 5. fs_read apps//CLAUDE.md ← only if working in a specific app ``` ## Capabilities | Capability | Tool | |-----------|------| | Read/write/import repo files | `fs_read`, `fs_write`, `fs_write_chunk`, `fs_ingest_url`, `fs_import_git_files`, `fs_glob`, `fs_grep` | | Query Supabase | Supabase MCP `execute_sql` | | GitHub operations | Github Proxy MCP | | Credentials | clauth MCP `clauth_get` | | Trigger Claude Code | `monkey_dispatch` MCP | For filesystem decisions, load `rdc:fs-mcp`. It defines when to use direct FS writes, chunked writes, URL ingest, or GitHub-branch import into the dirty local monorepo. ## Two-Claude Relay Two channels to communicate with Claude Code: **File relay (async):** Write to `.rdc/relay/from-claude-ai/-.md`: ``` --- from: claude-ai to: claude-code type: task | artifact topic: epic_id: status: pending --- ``` **Live dispatch (immediate):** Fire Claude Code directly: ``` mcp__claude_ai_clauth__monkey_dispatch({ prompt: "Read .rdc/guides/agent-bootstrap.md. Your task: ...", cwd: "" }) ``` Check `.rdc/relay/from-claude-code/` at startup for results and questions from Claude Code. Full protocol: `.rdc/relay/RELAY.md` ## Hard Rules - **NEVER run `pnpm build`** — crashes machine - **NEVER commit to `main`** — always `` - **NEVER start work without a work item** - **NEVER ask for credentials** — use clauth MCP directly ``` If `CLAUDE.ai.md` already exists, read it first and add only the relay section if missing — do not overwrite existing content. --- ## Step 7 — Commit ```bash git add .rdc/ CLAUDE.ai.md git commit -m "feat(rdc): add .rdc/ config, guides, relay scaffold and CLAUDE.ai.md Generated by /rdc:setup scan. Co-Authored-By: Claude Sonnet 4.6 " if [ "$RDC_TEST" != "1" ]; then git push origin else echo "[RDC_TEST] skipping git push origin " fi ``` --- ## Output Report: - What was detected automatically - What was generated - What still needs manual attention (e.g. `deploy_dashboard_url`, `satellite_repos`) - Next step: `/rdc:status` to verify