fix(gemini): preserve tool calls through opengateway

[kevincodex1]

preserve tool calls for gemini

[Vasanthdev2004]

Blockers

None found.

Non-Blocking

• PR description is minimal — "preserve tool calls for gemini" could be more detailed.
• Removes GLM-5.1 from Opengateway — not clear why in the description.

Looks Good

• Preserves Gemini tool call signatures through OpenGateway
• Adds google/gemini-3.1-flash-lite-preview to brand and model catalogs
• Removes GLM-5.1 from Opengateway (likely not available anymore)
• Good test coverage — 2 new tests for tool call preservation (non-streaming and streaming)
• 452 additions, 49 deletions — focused changes
• Updates README and docs to reflect GLM removal

---

Verdict: Approve — clean Gemini tool call fix.

[Vasanthdev2004]

Clean Gemini tool call fix. Good test coverage.

[anandh8x]

Non-blocking streaming robustness note:

The current implementation preserves Gemini thought_signature when extra_content is on the tool call itself or on the same streaming delta as the tool call.

One edge case to consider: OpenAI-compatible streams are delta-based, so a gateway could theoretically emit delta.extra_content.google.thought_signature in an earlier assistant metadata chunk, then emit delta.tool_calls in a later chunk. In that shape the signature would be dropped because the shim doesn’t currently retain pending extra_content across chunks.

Not asking for changes unless this shape is observed in practice, but a small regression test or per-choice pending signature state would make this more future-proof.