fix(openai-adapters): read Gemini embeddings from correct response field

[Pluviobyte]

Description

The Gemini embeddings adapter reads the wrong field from the :batchEmbedContents response. In packages/openai-adapters/src/apis/Gemini.ts, embed() did:

data: data.batchEmbedContents.map((embedding: any) => embedding.values),

but the Gemini API returns the vectors under embeddings, not batchEmbedContents. That field is always undefined, so every embeddings request through the Gemini adapter throws:

TypeError: Cannot read properties of undefined (reading 'map')

This breaks embeddings (e.g. @codebase / @docs indexing) for anyone configuring Gemini as their embed model. The path is reachable: core/llm/llms/Gemini.ts sets useOpenAIAdapterFor = ["chat", "embed", ...], so embeddings go through GeminiApi.embed().

The fix mirrors the core implementation, which already reads the correct field:

// core/llm/llms/Gemini.ts
return data.embeddings.map((embedding: any) => embedding.values);

I also removed the usage block, which read two more non-existent fields (data.total_tokens / data.prompt_tokens). The batchEmbedContents response carries no token usage, and embedding() already defaults usage to zeros — matching core, which sets no usage.

Checklist

• I've read the contributing guide
• The relevant docs, if any, have been updated or created
• The relevant tests, if any, have been updated or created

Screen recording or screenshot

N/A — backend crash fix; covered by the regression test below.

Tests

Added an embed test in packages/openai-adapters/src/apis/Gemini.test.ts that mocks the real :batchEmbedContents response shape ({ embeddings: [{ values }] }) and asserts the parsed vectors. Verified it fails on the old code with Cannot read properties of undefined (reading 'map') and passes with the fix. npx tsc --noEmit and prettier --check both pass locally.

Made with Cursor

---

Summary by cubic

Fixes a crash in the Gemini embeddings adapter by reading vectors from the correct response field. Restores embeddings for Gemini embed models and aligns adapter behavior with core.

• Bug Fixes

  • packages/openai-adapters/src/apis/Gemini.ts: embed() now reads data.embeddings instead of data.batchEmbedContents.
  • Removed the invalid usage block (nonexistent token fields).

• Tests

  • Added a regression test in packages/openai-adapters/src/apis/Gemini.test.ts that mocks the :batchEmbedContents response and asserts parsed vectors.

^{Written for commit e5f1548. Summary will update on new commits.}

Review in cubic

[Pluviobyte]

I have read the CLA Document and I hereby sign the CLA

[github-actions]

All contributors have signed the CLA ✍️ ✅
_{Posted by the CLA Assistant Lite bot.}

[cubic-dev-ai]

No issues found across 2 files

_{Re-trigger cubic}

[Pluviobyte]

I checked the remaining failing statuses. The GitHub Actions checks are green except the old macOS VSIX job showing cancelled, and the Continuous AI failure target I opened returns Cannot GET /api/agent-check-summary/..., so there is no actionable failure detail exposed from the external check.

Could a maintainer rerun/waive the Continuous AI statuses or share the specific task output? I can address any concrete finding once the external check exposes it.