feat(model): add OrcaRouter as an OpenAI-compatible model provider

[zhenjunchen-png]

What type of PR is this?

feat

Check the PR title.

• This PR title match the format: <type>(optional scope): <description>
• The description of this PR title is user-oriented and clear enough for others to understand.
• Add documentation if the current PR requires user awareness at the usage level.

(Optional) More detailed description for this PR.

This PR adds OrcaRouter as a model provider. OrcaRouter is an OpenAI-compatible router/aggregator: one API key reaches many upstream models across vendors, and orcarouter/auto picks an upstream per request by policy (cost / quality / latency / adaptive). Disclosure: I'm an engineer on the OrcaRouter team.

Scope: additive only. No existing provider's code, config, or tests are modified.

What changed

• Protocol / provider registration
  • idl/app/developer_api.thrift + generated developer_api.go / developer_api.ts: add ModelClass.OrcaRouter.
  • bizpkg/config/modelmgr/deprecate_model_get.go: map protocol string orcarouter to ModelClass_OrcaRouter.
  • bizpkg/config/modelmgr/mode_provider.go: add the OrcaRouter entry to the provider list (icon / name / description), consistent with the other providers.
• Model builder (bizpkg/llm/modelbuilder/orcarouter.go)
  • Reuses the existing OpenAI chat-model construction (OrcaRouter speaks the OpenAI wire protocol), so request/response handling stays in lockstep with the openai builder.
  • Adds optional attribution headers (HTTP-Referer / X-Title) via a custom http.Client, without overriding caller-set headers. Upstreams that don't understand them ignore them.
  • Defaults the base URL to https://api.orcarouter.ai/v1 when none is supplied, so the protocol can't accidentally fall back to the upstream OpenAI default.
  • For orcarouter/auto and known reasoning models, sampling params (temperature / top_p / penalties) are withheld, because the adaptive router's upstream is unknown per request and reasoning models reject temperature (HTTP 400). Pinned non-reasoning models keep full parameter control. (A general fix belongs in the gateway; this is a client-side guard.)
  • Registered in model_builder.go's dispatch map.
• Presets / metadata (following existing conventions)
  • conf/model/template/model_template_orcarouter.yaml: template defaulting to orcarouter/auto.
  • conf/model/model_meta.json (+ helm copy): an OrcaRouter entry with default plus a small set of flagship presets. Reasoning presets omit temperature so the bot config UI does not render a slider that would be ignored.
  • docker/volumes/minio/default_icon/orcarouter.png: provider icon.

Tests: unit tests in orcarouter_test.go cover attribution-header injection (added when absent, never overrides, never mutates the caller's request), the sampling-param suppression decision table, and builder registration. go build ./... passes.

How to verify

1. docker compose -f docker/docker-compose.yml --env-file docker/.env up -d (build the server image from this branch).
2. Open http://localhost:8888/admin/#model-management. The OrcaRouter provider card appears.
3. Add a model: API key (sk-orca-...), model orcarouter/auto (base URL is pre-filled).
4. Create a bot using that model and send a message. The reply is served through OrcaRouter's adaptive routing.

Full model catalog: https://www.orcarouter.ai/models.

(Optional) Which issue(s) this PR fixes:

Fixes #2697

[CLAassistant]

All committers have signed the CLA.