feat(llm): recognize orcarouter/ gateway model refs alongside openrouter

[XiaoHuo888-hue]

This adds a dedicated OrcaRouter model-ref path rather than relying only on the generic OpenAI-compatible base URL, mirroring how this repo already treats the openrouter/ prefix. Named routers such as orcarouter/auto pick an upstream per request. OrcaRouter is an OpenAI-compatible gateway that exposes 150+ models behind one API key, and it also provides gateway-level security controls for AI agents.

What changed

• src/llm-client.ts / src/admission-control.ts — normalizeDirectModelRef / normalizeAdmissionModelRef now recognize the orcarouter/ gateway prefix the same way they already handle openrouter/:
  • orcarouter/anthropic/claude-sonnet-4.6 → anthropic/claude-sonnet-4.6 (the namespaced id a direct OrcaRouter call accepts)
  • orcarouter/auto → stays orcarouter/auto (OrcaRouter rejects the bare auto id with 503 model_not_found, so the auto-router keeps its prefix)
• index.ts — inferProviderFromBaseURL now maps api.orcarouter.ai to the orcarouter provider, so a bare model name plus an OrcaRouter base URL resolves correctly on the reflection path.
• README.md — documented the OpenAI-compatible gateway llm config with an OrcaRouter example (orcarouter/auto auto-router, https://api.orcarouter.ai/v1).

Verification

• npm run build (tsc) passes.
• New + existing unit tests pass: test/admission-model-resolution.test.mjs, test/infer-provider-from-baseurl.test.mjs, test/llm-api-key-client.test.mjs, test/admission-lane-model-affinity.test.mjs (47 tests, 0 failures; llm-clients-and-auth CI group green).
• Live end-to-end against the real gateway using the plugin's own createLlmClient: orcarouter/auto returns a valid JSON completion (200), and anthropic/claude-sonnet-4.6 also returns 200. The triple-prefixed orcarouter/anthropic/claude-sonnet-4.6 form is rejected by OrcaRouter with 400, which is exactly why the normalize step strips the gateway prefix only when a vendor namespace remains.

It also runs gateway-level, zero-trust security for AI agents on the same endpoint — screening every prompt/response and governing every tool call on a default-deny basis, with no application code changes.

I'm an engineer on the OrcaRouter team.

[rwmjhb]

The helper-level normalization is well tested, but the new OrcaRouter behavior is not wired consistently into the production client paths:

1. Primary smart-extraction and CLI clients bypass normalizeDirectModelRef. buildMemoryLlmClient constructs its primary client with raw llmModel, and the CLI path likewise passes raw config.llm.model; createApiKeyClient serializes config.model unchanged. Only admission-client construction calls the normalizer. Therefore orcarouter/anthropic/claude-sonnet-4.6 still reaches those outgoing requests triple-prefixed even though the new helper test says it should become anthropic/claude-sonnet-4.6. Please normalize every API-key direct-client construction path and add request-capture tests for primary and CLI clients.

2. Reflection resolution does not preserve the advertised auto-router semantics. splitProviderModel("orcarouter/auto") produces provider="orcarouter", model="auto", and generateReflectionText passes those values to the embedded runner. The PR itself documents that OrcaRouter rejects bare auto, but no host-level test demonstrates that OpenClaw reconstructs the required gateway model ID or even recognizes this newly inferred provider from the plugin-only llm.baseURL configuration. Please implement or prove the complete runner-to-request behavior with a production-boundary test; helper inference alone is insufficient.

3. A vendor-prefixed reflection model bypasses the new base-URL inference. With llm.baseURL=https://api.orcarouter.ai/v1 and anthropic/claude-sonnet-4.6, split.provider wins over inferProviderFromBaseURL, so reflection selects provider anthropic rather than the configured gateway. Make reflection resolution gateway-aware before generic provider splitting, preserving the model ID accepted by OrcaRouter, and test the complete generateReflectionText runner arguments/request route.

Please centralize the gateway model-resolution rules so direct, admission, CLI, and reflection paths cannot drift. Non-blocking: remove or substantiate the gateway security marketing claims; this diff implements model routing, not those controls.

Local verification: the four related test files pass 42/42, the full orchestrator suite passes, and npm run build leaves the generated tree clean. GitHub currently reports no checks for this PR; the missing integration cases above are not covered by the green local suites.