Homerail Shared

HomeRail Shared

Core Rules

Use these rules with all HomeRail skills.

• Treat this release as a local-source install. Do not assume npm packages or public GitHub URLs exist yet.

• Prefer placeholders such as <local-homerail-repo> and /path/to/HomeRail in user-facing instructions.

• Keep provider and integration secrets in the Manager encrypted secret store through CLI/UI configuration; never write API keys, tokens, or passwords into repo files, skill files, DAG templates, commits, or issue text.

• Use one shared HOMERAIL_HOME for Manager and Node. Default is $HOME/.homerail.

• Prefer hr config for local runtime settings and hr start to start Manager and Node together.

• Manager and Node run as local services. Manager is not expected to run inside the Worker Docker image.

• Node provisions Worker containers. Workers for one run should share ${HOMERAIL_HOME}/workspace/<run_id>.

• Local service ports (use this table and the CLI before falling back to low-level process probes):

  Port	Role	Default URL	Notes
  19191	Manager (HTTP + WS, what DAGs and the CLI talk to)	http://localhost:19191	This is the only port a DAG run / CLI hits. Bind via HOMERAIL_MANAGER_HOST / manager.host / hr start --host; public access via HOMERAIL_MANAGER_PUBLIC_URL / hr start --public-url.
  19192	Agent UI (HTTPS, browser)	https://localhost:19192	Browser-only. Do NOT send DAG/CLI traffic here. Config via HOMERAIL_UI_HOST / HOMERAIL_UI_PORT / ui.host / ui.port.
  19193	Agent UI HTTP fallback	http://localhost:19193	Fallback when HTTPS port is taken. Config via HOMERAIL_UI_HTTP_PORT / ui.httpPort.

  The Manager port is 19191, not 19192. 19192 is the Agent UI. Mixing these up is the most common startup mistake.

• For hr start --ui --public without explicit public URLs, Agent UI should bind and advertise the detected machine IP with HTTPS, and the browser should use same-origin Manager API/WS proxying for the HTTPS UI. Use HOMERAIL_PUBLIC_HOST only when automatic machine IP detection picks the wrong interface.

• For reverse-proxied public access, prefer explicit HTTPS/WSS public URLs through HOMERAIL_MANAGER_PUBLIC_URL / HOMERAIL_UI_PUBLIC_URL or --public-url / --ui-public-url.

• Do not hardcode Docker bridge addresses. Use Docker host.docker.internal / host-gateway support or explicit operator-provided callback settings.

Provider Boundaries

For Coding Plan / Agent Plan accounts:

• Never use direct-llm or Chat Completions as the worker runtime path.
• Kimi plans should use the Kimi Code CLI harness (kimi_code / kimi-code).
• Other provider plans should use the Claude Code compatible harness (claude-sdk) with an Anthropic-compatible endpoint.
• If a plan does not expose an Anthropic-compatible endpoint, fail explicitly instead of falling back to the Chat Completions URL.

For MiMo:

• DAG worker LLM access uses the MiMo token-plan Anthropic-compatible endpoint.
• MiMo ASR/API-billing endpoints are a separate voice path and must not be mixed into DAG worker model configuration.
• If a user provides a provider key, configure it with hr model configure --api-key-stdin or the interactive CLI/UI flow so Manager stores it encrypted. Plaintext env files are legacy import fallbacks only.

Local Source Layout

Expected source checkout:

textCopy code
<local-homerail-repo>/
  homerail_manager/
  homerail_node/
  homerail_worker/
  homerail_protocol/
  homerail_cli/
  agent-ui/
  assets/
  skills/

Use homerail_cli for the TypeScript CLI. Do not refer to the old homerail_cli_ts name.

Anti-Patterns (Don'ts)

These cause wasted investigation rounds during normal operation. The CLI already exposes everything below.

• Do not start service diagnosis with raw process probes or source spelunking. Run hr doctor (readiness) and hr runtime status (live state) first; they report manager/node/worker/model state. Use ps, lsof, direct curl, logs, or source reading only after the CLI output is insufficient or the user explicitly asks for low-level debugging.
• Do not start a fixed host Worker process. DAG Workers are provisioned on demand by a Node into Docker containers; there is no standing worker to launch. Starting a bare homerail_worker node process manually will not register correctly and bypasses the isolated-workspace contract every DAG depends on.
• Do not hand-craft WebSocket URLs or reverse-engineer the worker connection protocol. hr start handles Manager↔Node↔Worker wiring. If a Worker cannot reach the Manager, set the Manager worker-WS callback host (see homerail-install-ops troubleshooting), do not hardcode Docker bridge IPs.
• Do not rebuild packages before every DAG run. install:all / build / typecheck / test are only needed after changing source or on first install. A ready runtime runs DAGs directly via hr run.
• Do not read dag_runs/dag_handoffs tables via sqlite3 to get run output. Use hr dag handoffs <run_id> (with --content-limit 0 for untruncated content). The CLI is the supported read path.
• Do not treat an idle or running snapshot as success. Terminal status (completed/failed/cancelled) plus non-empty handoffs is the only success signal.

Validation Ladder

When validating a local install or change, prefer this order:

bashCopy code
npm run install:all
npm run build
npm run typecheck
npm test
npm run ci

For end-to-end readiness, also run hr config, hr start, hr doctor, and one public smoke DAG through homerail-install-ops.