Setup Opencode Ralph Rlm

Setup OpenCode Ralph RLM

Use this skill to install Ralph RLM into a target repository and make it usable from OpenCode as the ralph-rlm/supervisor model.

Workflow

1. Inspect the target repository.

   • Confirm the working directory is the project root.
   • Identify the package manager and test command from lockfiles and package.json scripts.
   • Check for existing opencode.json, .opencode/ralph.json, and .opencode/plugins/.
   • If the repo already has OpenCode or Ralph config, read references/config-files.md before editing.

2. Install the Ralph package if the project can accept a dev dependency.

   • Prefer the package manager already used by the repo.
   • Package name: @doeixd/opencode-ralph-rlm.
   • If the user only wants a one-time setup, use npx @doeixd/opencode-ralph-rlm setup instead.
   • Do not require Bun for end users; the published CLI is a Node binary.

3. Run setup from the project root.

   npx @doeixd/opencode-ralph-rlm setup --worktree .
   

   Use --dry-run first when the repo already has OpenCode config. Use --force only when the user explicitly wants generated Ralph files overwritten. See references/cli.md for command flags and expected output.

4. Review generated files.

   • .opencode/plugins/ralph-worker.ts should re-export @doeixd/opencode-ralph-rlm/worker-plugin.
   • .opencode/plugins/ralph-session-bridge.ts should inject x-opencode-session-id and directory.
   • .opencode/plugins/ralph-autostart.ts starts the provider automatically when OpenCode loads (so the user need not run serve by hand). Skipped if setup --no-autostart was used; disable at runtime with RALPH_AUTOSTART=0.
   • .opencode/ralph.json should have a realistic verify.command.
   • .opencode/ralph.json includes plans: { dir: ".ralph-rlm/plans", active: "default" } — the named-plan layout. Protocol files (PLAN.md, RLM_INSTRUCTIONS.md, …) will live under .ralph-rlm/plans/<name>/, not the repo root. opencode-ralph-rlm plan-path prints the active plan's PLAN.md location.
   • .opencode/ralph.json may include fff.enabled and fff.scanTimeoutMs; leave the defaults unless the project cannot load optional native dependencies.
   • opencode.json should preserve existing providers and include provider["ralph-rlm"].

5. Adjust verify.command if auto-detection guessed poorly.

   • Bun projects usually use ["bun", "run", "test"] or ["bun", "run", "verify"].
   • npm projects usually use ["npm", "test"].
   • Rust projects usually use ["cargo", "test"].
   • Python projects usually use ["python", "-m", "pytest"].

6. Ask whether to add Ralph guidance to the project agent instructions.

   • Look for .opencode/ralph.json agentMdPath, then AGENT.md, then AGENTS.md.
   • Ask the user before creating or editing the guidance file.
   • If they agree, append a short Ralph section; use references/agent-guidance.md for wording.

7. Run diagnostics.

   npx @doeixd/opencode-ralph-rlm doctor --worktree .
   

   A provider warning is expected until the provider is running. For failures, read references/troubleshooting.md.

8. Set up models — discover what the user already has first; only ask if nothing usable is found.

   There are two model decisions:

   • Supervisor — the LLM the provider calls for orchestration (planning interview, tool routing, status). Needs reliable function/tool-calling.
   • Worker — the model spawned OpenCode sessions use to write code. Workers run through OpenCode, so they already use OpenCode's providers/auth.

   Discover existing credentials/models before configuring anything:

   • opencode auth list — which providers are authenticated.
   • opencode models — available model ids (e.g. opencode/deepseek-v4-flash-free, anthropic/claude-*).

   Supervisor: if the user has a keyed provider authenticated in OpenCode (e.g. Google, OpenCode Zen), the provider auto-detects it — no separate key needed. So usually you do nothing here. Only set RALPH_SUPERVISOR_API_KEY (+ RALPH_SUPERVISOR_MODEL / RALPH_SUPERVISOR_BASE_URL) or .opencode/ralph-provider.json supervisor if the user wants a specific provider/model, or if no keyed provider is authenticated.

   Worker: leaving it unset is fine — workers use OpenCode's configured default model. To run coding on a free model, recommend (don't force) setting .opencode/ralph-provider.json worker.providerID + worker.modelID to a free OpenCode model such as opencode / deepseek-v4-flash-free. Confirm with the user before writing it.

   If nothing usable is found (no authenticated provider, no key): direct the user to authenticate before launching, e.g.:

   opencode auth login
   

   or set RALPH_SUPERVISOR_API_KEY. Do not start the provider expecting the supervisor to work until at least one credential path exists. See references/config-files.md.

9. Re-open OpenCode — the provider auto-starts; then verify the supervisor is ready.

   • With the auto-start plugin installed (default), the provider launches automatically when OpenCode loads — no manual serve needed. OpenCode loads providers and plugins at startup, so quit and re-open the OpenCode TUI for the ralph-rlm/supervisor model + plugins to take effect.
   • Check readiness: curl http://127.0.0.1:8787/api/health — supervisor.ready should be true (it shows the resolved model, source e.g. opencode-auth:google, and the provider version). If unreachable, the provider may still be starting (give it a moment) or auto-start was disabled — then start it manually:

     npx @doeixd/opencode-ralph-rlm serve --worktree .
     

   • If you just updated Ralph, restart so the new version loads. The provider is idempotent (one instance per port, reused), and serve's pre-flight refuses to start a duplicate — if it reports an existing provider of a different version, stop the old one first.
   • Then select the ralph-rlm/supervisor model.

10. Suggest planning before the first loop.

• A loop is only as good as its PLAN.md. Recommend planning the goal with the supervisor (it runs an interview before start_loop) or via the interview-and-create-plan skill, rather than delegating a one-line goal cold.

Notes

• The setup command is intentionally conservative: it skips existing managed files unless --force is passed.
• npx @doeixd/opencode-ralph-rlm ... is the portable command. The binary installed by the package is named opencode-ralph-rlm.
• Worker search uses optional @ff-labs/fff-node acceleration when available. If it is unavailable, rlm_grep falls back automatically; set RALPH_FFF_DISABLED=1 only when native search causes local issues.
• Node can run some .ts files directly in recent versions, but Node refuses TypeScript files under node_modules; the npm package should ship compiled JavaScript for the CLI.
• Do not hand-copy the legacy v0.1 plugin or tell users to call ralph_spawn_worker.
• For supervisor LLM credentials, prefer environment variables first:
  • RALPH_SUPERVISOR_API_KEY
  • RALPH_SUPERVISOR_MODEL
• Use .opencode/ralph-provider.json only when repo-local supervisor/worker defaults are needed.