[Notaloop763]

Summary

Adds an Opencode provider for the OpenCode gateway. One OpenCode API key reaches both catalogs:

• OpenCode Go (/zen/go/v1), the subscription plan
• OpenCode Zen (/zen/v1), pay-as-you-go from the workspace balance

Model ids follow opencode's own <catalog>/<model> convention, as printed by opencode models. opencode-go/<model> routes to Go and opencode/<model> routes to Zen. A bare id routes to Go, so a missing prefix never spends balance. The API only ever receives the bare id.

API families

The gateway serves each model through one API family. The provider picks the family from the model id, following the Go endpoint table and the routing in opencode's own client and hermes-agent:

Models	API	Endpoint
gpt-*, grok-*, muse-spark*	OpenAI Responses	/v1/responses
claude-*, qwen*, union-alpha*, and on Go minimax-*	Anthropic Messages	/v1/messages
gemini-* (Zen only)	Google generateContent	/v1/models/<model>:generateContent
everything else (DeepSeek, GLM, Kimi, MiMo, LongCat, ...)	chat/completions	/v1/chat/completions

Each family gets its native request body, response parsing (text, finish reason, token usage, output-truncation detection) and auth header (Authorization: Bearer, x-api-key or x-goog-api-key). New models that rotate into the catalogs go to chat/completions unless their id matches a prefix above. Only jev-* on Zen (OpenCode's System One API) is not supported; those models are hidden from the model list and rejected with a clear error.

OpenAICompatibleProvider gains small request-URL, auth-header, payload and response-parsing hooks for this. Its own behavior is unchanged.

Details

• Session header: every request carries a random x-opencode-session id, stable for one translation job, which keeps the gateway's prompt cache warm. Go rejects requests without it. The id is never derived from the machine or the install.
• Model list: both catalogs are fetched concurrently and shown as "OpenCode Go" / "OpenCode Zen" option groups.
• Errors: gateway refusals (ModelProtocolUnsupported, FreeTierError, insufficient balance, "model access is disabled") get actionable messages.
• Wiring: config (OPENCODE_API_KEY, OPENCODE_MODEL, OPENCODE_API_BASE), factory, API routes and settings allow-list, translate/refine/PDF pipelines, CLI, web UI and all 7 locales.
• Docs: new Opencode section in docs/PROVIDERS.md; the provider is also listed in the README, other docs, .env.example and docker-compose.yml.

Shared changes outside the provider

• PDF pipeline: the PDF pipeline now receives the NIM API key as well. It was missing, so PDF jobs on NIM failed.
• Rate limits: in-place sleeps are now capped at MAX_RATE_LIMIT_SLEEP (one hour). A longer Retry-After, such as OpenCode Go's weekly quota, raises RateLimitError, so the pipeline auto-pauses with a checkpoint instead of stalling silently. Shorter throttles still wait in place, as before. This applies to every provider.

Testing

• Live, OpenCode Go: every model in the Go catalog (38) returned a translation through its family: 23 over chat/completions, 9 over Anthropic Messages (Claude Haiku 5.5, MiniMax, Qwen) and 6 over OpenAI Responses (GPT, Grok, Muse Spark).
• Live, OpenCode Zen: Messages and Google requests reach the gateway and get the expected billing refusal (no balance on the test workspace). Successful paid Zen calls are covered by mocked tests only.
• Unit tests: tests/unit/test_opencode_provider.py (routing table, per-family request/response round trips, session header, model list, error hints) and tests/unit/test_opencode_integration.py, plus a rate-limit cap test.

🤖 Generated with Claude Code

[Notaloop763]

@hydropix, friendly ping: this is ready for review, and CI is green. It builds on frank-lar's #278, which has been inactive since mid-September. Their commits are credited as co-author. On top of that, it adds routing for every OpenCode API family (Responses, Anthropic Messages, Google), the Zen catalog, and live tests against all 38 Go models. If it's merged, #278 can be closed.

[hydropix]

Thanks for the ping and for the work on top of #278, and sorry for the wait. I haven't been able to review this one properly yet, and I'd rather not merge a new provider of this size without going through it, so it won't be in the release going out today. It's next on my list, and I'll get back to you here with feedback.