omp-router

OhMyPi extension connecting to any OpenAI-compatible AI router (9router, OmniRoute, yardmaster, …) via its /v1 API: you add the router URL, its models appear in omp, and omp's native usage surfaces report the upstream providers' quotas as seen through the router.

Ported from pi-router + the pi-sub router adapter; the omp-native provider, discovery, login, and usage seams replace pi's hand-rolled machinery.

Install

From npm (published automatically on v* tags):

npm install -g @bacnh85/omp-router

Then load the installed package path in ~/.omp/agent/config.yml — run npm root -g to locate your global node_modules and point at the package:

# ~/.omp/agent/config.yml
extensions:
  - /usr/local/lib/node_modules/@bacnh85/omp-router   # $(npm root -g)/@bacnh85/omp-router

Or point the config at a checkout for development:

# ~/.omp/agent/config.yml
extensions:
  - ~/code/omp-extensions/omp-router

Quick start

# ~/.omp/agent/config.yml
extensions:
  - ~/code/omp-extensions/omp-router

Then in omp:

/router-url http://localhost:20128/v1   # store the endpoint
/login router                           # store the API key (masked prompt)
/router-model                           # browse/select a router model

A URL without the /v1 suffix is fine (e.g. https://yardmaster.bacnh.com) — it is normalized to <url>/v1 when the provider registers, so chat, model discovery, and usage all target the same OpenAI-compatible base.

The provider id is router; models show up as router/<router-model-id>. Models are discovered from GET <baseUrl>/models with a 24 h host-side cache (omp models refresh forces a pull).

Configuration

Precedence: env > project (trusted repos only) > global.

The API key is not stored in these files: use /login router (or omp login router), which stores it in omp's own credential store.

ROUTER_API_KEY / OMP_ROUTER_API_KEY env vars work as a fallback: discovery sends them directly, and at session start the extension registers the env value as an in-session auth override only when no credential resolves at all — so a later /login router always wins from the next session on. Do NOT pin the key under providers.router.apiKey in models.yml: that override shadows the stored /login credential (omp's config layer beats stored credentials), which manifests as requests signed with the literal config string and, ultimately, no loadable models.

Commands

Usage reporting (upstream providers via the router)

The extension registers an omp-native usage provider, so per-credential caching (5 min TTL), last-good retention, and history are handled by the host. Two router flavors are supported, tried in order:

1. Generic JSON — GET <baseUrl>/usage[?provider=<upstream>] (windows.session/weekly/monthly.remaining_pct + reset_at, credits.currency/balance). Unknown upstream slugs 404 and fall back to the aggregate report.
2. OmniRoute text — GET <origin>/api/usage/om-usage[?provider=] ("Provider quota" Session/Weekly + "Personal quota" Daily/Weekly budgets; countdowns are converted to absolute reset timestamps). Credit-based upstreams (e.g. deepseek) report "Unavailable" windows; the raw USD balance is then fetched from the management API (/api/v1/me/status → /api/usage/<connectionId>, needs a manage-scope key or ROUTER_MGMT_TOKEN/OMNIROUTE_MGMT_TOKEN).

The ?provider= slug is derived from the active model id's first segment, so selecting cmd/deepseek/deepseek-v4 reports the command-code upstream's quota. Aliases normalize (cmd→command-code, oc→opencode-go, ds→deepseek, glmcn→glm-cn); generic aliases (auto, combo, …) fall back to the aggregate snapshot.

Model mapping notes

• Curated context-window corrections (models.dev-verified) with 9router's default-floor de-poisoning: GLM-5.2/5.3 → 1M ctx, DeepSeek V4 → 1M, Kimi K3 → 1M, GLM-5.1/5/4.6/4.7 → 200k. Truthful above-floor router values are never overridden.
• Thinking efforts per detected family as omp thinking capability: deepseek/zai (GLM) → high+max, gemini-3 → requiresEffort, OpenAI → minimal…xhigh, claude-adaptive → low…max; command-code upstreams reject off/minimal (requiresEffort, no minimal).
• Vision: surgical transport-verified overrides/downgrades — the router's capabilities.vision flag is trusted only where proven (re-probe with upstream pi-router's probe-vision.mjs when your router image changes).
• /router-reasoning off drops thinking capability and compat.supportsReasoningEffort entirely.

Development

npm install
npm run typecheck
npm test

extensions/test/unit.test.ts covers config precedence/trust gating, model mapping, both usage flavors (mocked fetch), and the provider registration contract.