Mode d’emploi
- Copiez la commande et examinez le paquet avant de l’installer dans votre environnement agent.
- Consultez le document original pour les outils, identifiants et environnements requis.
- Décrivez les entrées et le livrable souhaité ; comparez la sortie aux sources.
npx skills add https://github.com/doeixd/opencode-ralph-rlm --skill setup-opencode-ralph-rlmAvant de commencer
- L’installation n’inclut ni abonnement au modèle, ni quota API, ni accès aux fichiers.
- Étoiles et installations reflètent l’activité, pas un test indépendant de qualité ou de sécurité.
Rendu depuis le dépôt source en conservant titres, exemples, code, tableaux, liens et images.
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
- Inspect the target repository.
- Confirm the working directory is the project root.
- Identify the package manager and test command from lockfiles and
package.jsonscripts. - 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.
- 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 setupinstead. - Do not require Bun for end users; the published CLI is a Node binary.
- 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.
- Review generated files.
.opencode/plugins/ralph-worker.tsshould re-export@doeixd/opencode-ralph-rlm/worker-plugin..opencode/plugins/ralph-session-bridge.tsshould injectx-opencode-session-idanddirectory..opencode/plugins/ralph-autostart.tsstarts the provider automatically when OpenCode loads (so the user need not runserveby hand). Skipped ifsetup --no-autostartwas used; disable at runtime withRALPH_AUTOSTART=0..opencode/ralph.jsonshould have a realisticverify.command..opencode/ralph.jsonincludesplans: { 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-pathprints the active plan'sPLAN.mdlocation..opencode/ralph.jsonmay includefff.enabledandfff.scanTimeoutMs; leave the defaults unless the project cannot load optional native dependencies.opencode.jsonshould preserve existing providers and includeprovider["ralph-rlm"].
- Adjust
verify.commandif 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"].
- Ask whether to add Ralph guidance to the project agent instructions.
- Look for
.opencode/ralph.jsonagentMdPath, thenAGENT.md, thenAGENTS.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.
- 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.
- 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 loginor 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.
- 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/supervisormodel + plugins to take effect. - Check readiness:
curl http://127.0.0.1:8787/api/health—supervisor.readyshould betrue(it shows the resolvedmodel,sourcee.g.opencode-auth:google, and the providerversion). 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.
- 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 beforestart_loop) or via theinterview-and-create-planskill, rather than delegating a one-line goal cold.
Notes
- The setup command is intentionally conservative: it skips existing managed files unless
--forceis passed. npx @doeixd/opencode-ralph-rlm ...is the portable command. The binary installed by the package is namedopencode-ralph-rlm.- Worker search uses optional
@ff-labs/fff-nodeacceleration when available. If it is unavailable,rlm_grepfalls back automatically; setRALPH_FFF_DISABLED=1only when native search causes local issues. - Node can run some
.tsfiles directly in recent versions, but Node refuses TypeScript files undernode_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_KEYRALPH_SUPERVISOR_MODEL- Use
.opencode/ralph-provider.jsononly when repo-local supervisor/worker defaults are needed.