feat(translator): OpenAI SSE → Gemini SSE for /v1beta/models route

[diegosouzapw]

Summary

Ports decolua/9router#225 by @SteelMorgan.

The @google/genai SDK (Gemini CLI) always calls
/v1beta/models/{model}:streamGenerateContent?alt=sse for chat and expects
Gemini SSE chunks. The previous v1beta route forwarded raw OpenAI SSE from
handleChat, so the SDK crashed on the [DONE] sentinel that OpenAI SSE
terminates with (Gemini SSE has no sentinel — the stream just closes):

SyntaxError: Unexpected token 'D', "[DONE]" is not valid JSON

Changes

• New translator open-sse/translator/response/openai-to-gemini-sse.ts

  • transformOpenAISSEToGeminiSSE(upstream, model) — TransformStream that
    re-emits each OpenAI delta as a Gemini-shape candidate, maps
    finish_reason → finishReason (STOP / MAX_TOKENS / SAFETY), attaches
    usageMetadata + modelVersion on the final chunk, and surfaces
    reasoning_content as { thought: true } parts for thinking models.
  • convertOpenAIResponseToGemini(response, model) — sibling JSON
    converter for the non-streaming :generateContent path.
  • OPENAI_TO_GEMINI_FINISH_REASON exported as a shared constant.
  • Chunk-boundary safe: the transformer buffers partial SSE lines across
    TextDecoder boundaries so it never JSON.parses a half-event.

• Route rewrite src/app/api/v1beta/models/[...path]/route.ts

  • Streaming intent now derived from the URL action suffix
    (:streamGenerateContent vs :generateContent), the canonical Gemini
    API convention. generationConfig.stream is not a real Gemini field
    and the SDK never sets it.
  • Streaming branch pipes the upstream OpenAI SSE through
    transformOpenAISSEToGeminiSSE; non-streaming branch hands the JSON
    to convertOpenAIResponseToGemini.
  • Existing Zod validation, buildClientRawRequest plumbing, and
    sanitizeErrorMessage error path are preserved untouched.

• Skipped from upstream: the first hunk of chore(deps): bump actions/cache from 4 to 5 #225 (filter nameless
  Responses-API tools in openai-to-responses.js) is not ported —
  OmniRoute's open-sse/translator/request/openai-responses.ts already has
  a much richer tool-conversion pipeline (web-search passthrough, MCP
  namespace flattening, local_shell mapping, etc.) and the
  empty-name-filter case is unreachable in practice in this codebase.

Test plan

• node --import tsx/esm --test tests/unit/translator-openai-to-gemini-sse.test.ts — 11/11 pass
• node --import tsx/esm --test tests/unit/v1beta-models-route.test.ts — 2/2 pass (route GET path unaffected)
• npm run typecheck:core — clean
• npx eslint on changed files — clean
• node scripts/check/check-file-size.mjs — clean

TDD coverage

11 new tests in tests/unit/translator-openai-to-gemini-sse.test.ts:

• finish-reason map covers all four canonical OpenAI values
• per-chunk mapper: drops role-only deltas, maps content + reasoning,
  attaches usage/modelVersion on the final chunk, maps each
  finish_reason correctly
• full SSE conversion drops the [DONE] sentinel and produces the exact
  Gemini chunk shape the SDK expects
• chunk-boundary buffering test feeds an event split mid-string-literal
  and asserts the transformer re-assembles it
• non-OK upstreams pass through unchanged
• JSON converter: maps choice → candidate, passes through already-Gemini
  bodies, surfaces upstream errors with the right status code

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.