omp-cliproxyapi-provider

Oh My Pi provider plugin that discovers models from CLIProxyAPI and registers them for use in omp. Ported from pi-cliproxyapi-provider. It supports catalog-driven OpenAI Fast mode and a status helper that shows elapsed runtime plus a TPS / token-usage toast after each agent turn.

What it does

1. Registers the cliproxyapi provider (OAuth-style entry), so it always appears in /login.
2. Interactive setup collects baseUrl + apiKey: /login CLIProxyAPI or /login cliproxyapi. The final step validates credentials against {root}/v1/models?client_version=pi and re-prompts on failure.
3. Maps the CLIProxyAPI catalog into omp models, including Fast service-tier capability, and registers inference against {root}/backend-api/.
4. Provides /fast to toggle OpenAI priority processing for supported models.
5. Caches the model catalog in ~/.omp/agent/cliproxyapi-models.json, refreshes it in the background on startup, and provides /cliproxyapi-refresh to force a refresh.
6. In interactive TUI sessions, shows an elapsed-time status during runs and emits a TPS / token usage notification when the run settles.

# from npm
omp install npm:@router-for-me/omp-cliproxyapi-provider   # when published / supported

# from a local checkout
git clone https://github.com/ngctro/omp-cliproxyapi-provider

For a local checkout, either symlink/copy it into your user extensions directory:

ln -s "$PWD" ~/.omp/agent/extensions/omp-cliproxyapi-provider

or point settings at it (~/.omp/agent/config.yml):

extensions:
  - /absolute/path/to/omp-cliproxyapi-provider

The package manifest (omp.extensions) points at src/index.ts; no build step is required. Run once with -e /path/to/checkout/src/index.ts to test ad hoc.

Login-style setup (recommended)

The provider needs both baseUrl and apiKey.

/login CLIProxyAPI     # or /login cliproxyapi

Then enter:

1. base URL — preferred form is host:port, e.g. http://127.0.0.1:8317
2. API key

Final login validation calls {root}/v1/models?client_version=pi:

• HTTP 200 → login succeeds (empty model list is still OK); models are registered immediately and the catalog cache is rewritten
• non-200 / network error → re-prompt; nothing is persisted until validation passes

On success:

• baseUrl / apiKey are written to ~/.omp/agent/cliproxyapi.json
• credentials are stored by OMP's auth store (agent.db) so inference requests authenticate automatically

Re-run /login CLIProxyAPI anytime to reconfigure. OMP's built-in /logout <provider> removes only the stored credential; cliproxyapi.json is left in place — remove or edit that file if you also need to clear provider configuration.

Non-interactive configuration

~/.omp/agent/cliproxyapi.json:

{
  "baseUrl": "http://127.0.0.1:8317",
  "apiKey": "12345",
  "fast": false,
  "pause": false,
  "providerId": "cliproxyapi",
  "providerName": "CLIProxyAPI"
}

Environment overrides: CLIPROXYAPI_BASE_URL, CLIPROXYAPI_API_KEY, CLIPROXYAPI_PROVIDER_ID, CLIPROXYAPI_PROVIDER_NAME, CLIPROXYAPI_FAST (true/false, 1/0, yes/no, on/off). Resolution order: env > config file > default baseUrl (http://127.0.0.1:8317). Unlike upstream there is no auth.json fallback — after a successful /login the same values are mirrored into cliproxyapi.json.

baseUrl normalization

Preferred form is host:port only. All of these resolve the same way as upstream:

Input	Inference baseUrl
http://127.0.0.1:8317	http://127.0.0.1:8317/backend-api/
http://127.0.0.1:8317/backend-api	http://127.0.0.1:8317/backend-api/
http://127.0.0.1:8317/v1	http://127.0.0.1:8317/backend-api/
127.0.0.1:8317	http://127.0.0.1:8317/backend-api/

Inference traffic posts to {inference}/codex/responses.

Fast mode

Requests the OpenAI priority service tier for supported models (lower latency, higher cost).

Fast is off by default. Toggle with /fast; the preference persists in cliproxyapi.json. When enabled at startup, the session's OpenAI-family service tier is set to priority immediately.

Implementation note (OMP divergence): instead of payload hooks, Fast uses OMP's native per-session service-tier knob for the openai family. Requests carry service_tier: "priority" when the selected model id looks like an OpenAI model (gpt-/o3-/codex-/…); other ids are left unchanged. As upstream, /fast warns that the current model cannot use Fast when enabling it with an unsupported model. When models.dev publishes experimental.modes.fast.cost, re-registration swaps in those rates.

Fast capability stays catalog-driven: a CPA model is Fast-capable when its service_tiers field is a non-empty array. Supported models do not produce separate status text (the OMP status line already surfaces priority tiers).

Pausing provider requests

/pause      # waits before every provider request
/continue   # clears the pause

Both persist pause in cliproxyapi.json. Before every request the setting is reread; while paused, requests poll every 200 ms until cleared.

Model cache

~/.omp/agent/cliproxyapi-models.json stores model metadata plus derived endpoint URLs and a fetchedAt timestamp — never credentials. Remote query timeout: 60 s. Cache scope: current baseUrl (a different base URL ignores it).

Startup behavior matches upstream: cached models register immediately, then a background refresh rewrites the cache on success; without a cache the first query is synchronous and a failure logs a warning. /cliproxyapi-refresh forces a remote refresh, rewrites the cache, and updates registered models. Delete the file to clear the cache manually.

Model mapping

Same rules as upstream: slug→id, display_name→name, context_window, input_modalities→input types, supported_reasoning_levels[].effort→thinking efforts (non-canonical spellings ignored; unsupported levels simply absent from the effort list rather than hidden via a null map), visibility:"hide" skipped. Prices match canonical entries from models.dev (24 h cache at ~/.omp/agent/tmp/models-dev-cache.json); ambiguous reseller prices fall back to zero rather than guessing.

Elapsed time and TPS (TUI)

Active only for primary interactive TUI sessions (ctx.hasUI && ctx.mode === "tui"):

• While the agent runs, an Elapsed … status updates every second.
• On terminal settle (including auto-retry continuations within one run), a notification reports approximate TPS plus token usage (out/in/cache r/w/total) and the final elapsed time.
• Subagent and print-mode sessions do not own the timer, clear the status, or emit toasts.

Divergences from upstream (and why)

Upstream piece	This port	Reason
codex-stream.ts runtime source patching	Removed; uses built-in openai-codex-responses API	OMP's transport derives the account-id header from a JWT and omits it for plain keys, so plain CPA keys work natively
retry.ts transient-error normalization	Removed	OMP TurnRecovery classifies closed/refused connections and similar transient failures as retryable natively
auto-compact.ts proactive compaction	Removed	Context overflow routes into OMP's native compaction-recovery path
fast-footer.ts footer / model-status "fast" chip	Removed (no footer seam in OMP)	OMP exposes setStatus/widgets instead; the service tier itself is visible through the session UI
tps.ts agent_settled event	Finalize on terminal agent_end (willContinue !== true)	Same lifecycle, native OMP events
Files under ~/.pi/agent/	~/.omp/agent/	OMP agent dir
auth.json credential fallback	Removed (env > cliproxyapi.json > default)	Credentials live in OMP's agent store; /login mirrors them into cliproxyapi.json

CLIENT_VERSION stays "pi": CLIProxyAPI filters its catalog by this parameter and the pi vocabulary is the documented contract.

Failure behavior

Equivalent to upstream: unauthenticated providers still appear in /login with no models yet; 401/unreachable during startup keeps an existing matching cache active and only logs when no cache exists; login validation failure never persists anything.

Development

bun install
bun run check   # tsc --noEmit + bun test
bunx tsc --noEmit
bun test        # all tests

Tests stub CLIProxyAPI over local HTTP where needed; no external network required.