omp-cliproxyapi-provider

CLIProxyAPI dynamic model provider for oh-my-pi (omp).

Port of @router-for-me/pi-cliproxyapi-provider (MIT) to the omp extension API. It discovers the models your CLIProxyAPI (CPA) deployment serves, registers them as a provider, and adds a Fast-mode priority tier, request pausing, catalog refresh, and an elapsed/TPS footer.

Features

| Feature | Command / behavior | | --- | --- | | Dynamic catalog | Models from /v1/models register as cliproxyapi/<slug>; hidden CPA models are skipped, and prices come from models.dev. | | Login flow | /login CLIProxyAPI prompts for the base URL and API key, validates them against the models endpoint, then persists and registers them. | | Fast mode | /cpa-fast [on\|off\|status] injects service_tier: "priority" — only for models whose CPA catalog entry advertises a service tier. | | Pause | /cpa-continue releases omp's process pause gate; /pause (omp built-in) engages it. Both mirror to cliproxyapi.json. | | Catalog refresh | /cpa-refresh re-pulls the catalog, bypassing the cache. | | Elapsed / TPS | The TUI shows a live Elapsed Ns footer per run and a TPS … tok/s. out …, in …, cache r/w …/…, total …, …s summary when it settles. |

Install

omp plugin install omp-cliproxyapi-provider

Or unpack the package into ~/.omp/plugins/node_modules/omp-cliproxyapi-provider/ (the directory omp plugin install targets — note this is ~/.omp/plugins, not under agent/). Pointing extensions: in ~/.omp/agent/config.yml at the package directory works too; package.json's omp.extensions field is the install contract.

Before you install: remove any existing cliproxyapi provider

If ~/.omp/agent/models.yml already defines a provider named cliproxyapi, that static entry collides with this plugin. Both claim the same provider id, so the plugin's registration replaces the static one at runtime — but the static catalog's models can survive on their original transport, leaving you with a mixed provider. Measured on a deployment whose models.yml used api: openai-completions:

| Configuration | Registered cliproxyapi models | | --- | --- | | models.yml only | 41, all openai-completions | | plugin + models.yml | 35 on openai-codex-responses plus 6 stale openai-completions | | plugin, models.yml provider renamed | 35, all openai-codex-responses |

The stale entries are exactly the models the plugin filters out as hidden, so they are both unusable and outside the plugin's Fast gating and catalog refresh.

Remove the provider from models.yml, or rename it to keep it while avoiding the id:

providers:
  cliproxyapi-static:      # renamed: no longer collides with the plugin
    baseUrl: https://cpa.example.com/v1
    api: openai-completions

The plugin's own configuration is independent of models.yml — it reads ~/.omp/agent/cliproxyapi.json and omp's credential store.

Configure

Interactive

omp
/login CLIProxyAPI      # or /login cliproxyapi

Prompts for the CPA base URL (default http://127.0.0.1:8317, or whatever is already configured) and the API key. The key prompt is masked; a host that cannot mask input rejects it rather than echoing the key.

The credentials are stored in omp's credential database at ~/.omp/agent/agent.db, so /logout cliproxyapi removes them (this differs from pi's auth.json).

Non-interactive

~/.omp/agent/cliproxyapi.json (see cliproxyapi.example.json):

{
  "baseUrl": "http://127.0.0.1:8317",
  "apiKey": "12345",
  "fast": false,
  "pause": false
}

Environment overrides, highest precedence first:

| Variable | Purpose | | --- | --- | | CLIPROXYAPI_BASE_URL | CPA base URL. | | CLIPROXYAPI_API_KEY | Ambient request auth. | | CLIPROXYAPI_FAST | true/false/1/0/yes/no/on/off. | | CLIPROXYAPI_PROVIDER_ID / CLIPROXYAPI_PROVIDER_NAME | Override the provider id / display name. | | CLIPROXYAPI_CLIENT_VERSION | client_version sent to /v1/models (default omp), for deployments that reject it. |

baseUrl accepts host:port, http(s)://host:port, …/v1, or …/backend-api; all forms normalize to the …/backend-api/ inference root with the models catalog at …/v1/models.

Commands

All plugin commands live under the cpa- namespace, because omp silently skips an extension command whose name collides with a built-in (there are ~90, including fast, pause, login, logout, model, compact, resume, retry).

| Command | Description | | --- | --- | | /cpa-refresh | Force-refresh models from the remote catalog. | | /cpa-fast [on\|off\|status] | Toggle catalog-gated Fast mode. | | /cpa-continue | Release paused requests. |

Fast mode, /fast, and /cpa-fast

omp has its own /fast, which sets service_tier for models omp's classifier reads as OpenAI. This plugin's /cpa-fast is gated by the CPA catalog instead, which differs in two ways:

• CPA ids omp does not classify as OpenAI (hy4-preview, doubao-*) still get Fast.
• CPA ids omp does classify as OpenAI but that CPA reports with no service tiers (gpt-5.6-*, gpt-6-astra) correctly do not.

Both ultimately send the same service_tier: "priority" on the wire, so either command works where omp's classifier and the CPA catalog agree.

Pause

/pause (omp built-in) freezes every agent loop in the process; the plugin persists that state to cliproxyapi.json and restores it on the next start. /cpa-continue and omp's /resume release it. The gate parks at request boundaries, so an in-flight stream finishes before the run stops.

Behavior notes

• Wire protocol. CPA serves every vendor's models over the codex Responses protocol, so the provider registers the built-in openai-codex-responses API and does not ship a custom transport. omp tolerates non-JWT CPA keys, so no source patching is needed.
• Thinking levels. CPA's low|medium|high|xhigh|max map onto omp's effort ladder. CPA's ultra has no omp equivalent and is dropped.
• Startup. A cached catalog is published immediately and refreshed in the background, so an unreachable CPA never delays omp's startup. The cache is keyed by both the base URL and the inference URL, so changing baseUrl invalidates it.
• Compaction. omp compacts on its own threshold (contextWindow − max(15%, 16384)); this plugin does not add a second one.

Differences from the pi version

• No streamSimple transport patch and no synthetic context_length_exceeded preemption — omp owns compaction, and registering a custom API id that delegates to pi-ai's own dispatcher recurses until the stack overflows.
• Transient-network-error retry is not reimplemented: omp classifies retryable errors inside its own stream pipeline, and extension message_end / turn_end handlers are notification-only (their return values are discarded), so the upstream normalization hook has no equivalent seam. omp already retries stream drops and 5xx/429 responses natively.
• Commands use the cpa- namespace: /cpa-refresh, /cpa-fast, /cpa-continue.
• Secrets live in ~/.omp/agent/agent.db, not auth.json.

Development

bun install
bun run check      # typecheck + unit tests

Verify against a live deployment:

omp --no-extensions -e ./extensions/index.ts -e ./extensions/tps.ts \
  -p "Reply with exactly: OK" --model cliproxyapi/gpt-5.6-luna

License

MIT. Portions derived from @router-for-me/pi-cliproxyapi-provider, also MIT.