[AndresDiagoM]

⚠️ base-red inherited: #17219

Summary

• Problem. OpenAI-compatible upstreams called with stream_options.include_usage send usage in a separate final chunk with choices: [], after the finish chunk (fix(sse): estimate usage in passthrough stream even with include_usage #12151). open-sse/translator/response/openai-to-antigravity.ts handles a chunk without choices[0] by stashing state._usage and returning null (L27-33 on the tip). The stash is only read when a later chunk with a choice builds a response, and after the finish chunk none comes. The stream-end flush does not rescue it either: stream.ts (L3174, L3218) calls translateResponse(..., null, state), and the translator returns null for a null chunk. So the usage never reaches the client: no usageMetadata, no prompt/candidate/total counts, no thoughts or cached tokens.

• Scope. The same function is registered for OPENAI → GEMINI too (openai-to-gemini.ts:14), so both Antigravity and Gemini client formats lose usage this way. fix(gemini): forward OpenAI usage-only chunks and cached/thought token counts to /v1beta clients #17842 fixed the same gap in the /v1beta SSE converter (openai-to-gemini-sse.ts); this is the sibling path.

• Fix (same approach as fix(gemini): forward OpenAI usage-only chunks and cached/thought token counts to /v1beta clients #17842). A small helper, usageOnlyChunkResponse, handles the no-choices case:

  • Usage before the finish chunk: stashed for the finish chunk, as before.
  • Usage after the finish chunk: emitted as its own chunk (empty text part, no finishReason, usageMetadata, and the stream's modelVersion / responseId).
  • Usage already sent: dropped, so it is never counted twice.

  The inline usage mapping is replaced by the shared buildGeminiUsageMetadata from geminiUsageMetadata.ts (fix(gemini): forward OpenAI usage-only chunks and cached/thought token counts to /v1beta clients #17842), which produces the same fields.

• Not changed: tool-call accumulation, how the finish chunk carries usage, the stash-before-finish behavior, the non-streaming path. A trailing chunk with neither choices nor usage is still dropped.

Proof Of Fix

• TDD in tests/unit/translator-resp-openai-to-antigravity.test.ts, 5 new tests:
  1. A trailing usage-only chunk after the finish chunk is forwarded with all five counts, no finishReason, and the stream's modelVersion / responseId.
  2. Usage already sent on the finish chunk is not repeated.
  3. Usage stashed before the finish chunk is not repeated by a second usage-only chunk.
  4. A trailing chunk without usage stays dropped.
  5. End to end through translateResponse(OPENAI, ANTIGRAVITY): text, finish, usage-only chunk, then the flush null. Exactly one chunk carries usageMetadata, and it is the last one.

# tip (b2a64106c9), new tests against the unfixed translator
ℹ tests 10  pass 8  fail 2   (tests 1 and 5: usage-only chunk after finish → null)
# with the fix (plus translator-resp-openai-to-gemini.test.ts)
ℹ tests 12  pass 12  fail 0

Validation

• check:open-sse-typecheck (0 errors), check:complexity-ratchets --base-ref origin/release/v3.8.53 (complexity violations in the touched file 2 → 1), check:file-size: OK
• prettier + eslint (with config/quality/eslint-suppressions.json) clean; commit hooks incl. check:ai-attribution pass
• Not verified against a live Antigravity / Gemini client; the end-to-end test goes through the real translateResponse registry, not through stream.ts's transform stream.