Summary

Decode the final response text without concatenating thought summaries. A thought containing JSON could corrupt the final JSON or be scored as an answer when no final response existed. Vertex uses the same decoder. Thought summaries are opt-in through includeThoughts; the current runner does not request them by default. This fixes decoding of a valid optional provider response, without changing generation configuration.

Evidence and validation

• Reproduced on unchanged upstream main ccf586d43d2a580319b6535c893668904d909eb9 with the regression tests in this PR.
• Focused native tests: python -m pytest -p no:cacheprovider -q tests/test_gemini_thought_parts.py tests/test_imajev_bench_triage.py pass after the patch; pre-patch failures are preserved locally.
• git diff --check passes.
• Unchanged upstream main in this environment: 4 failed, 713 passed, 136 skipped. Two failures import the unavailable optional mlx_vlm dependency; two import the absent test_p3_labeltrain fixture helper.
• Full python -m pytest -p no:cacheprovider -q: Exact PR head 00f2442eed2e: 4 failed, 717 passed, 136 skipped, 35 warnings in 57.80s. Only those same four upstream/environment failures remain; this patch introduces no additional failures.
• This repository currently reports no GitHub status checks; that is not a claim of passing remote CI.
• Tests use local synthetic fixtures; no released model weights, private data, hosted-provider calls, or model quality claims.

Contract references

• https://ai.google.dev/gemini-api/docs/generate-content/thinking — documented part.thought distinguishes thought summaries from final text.

AI assistance

Prepared with AI assistance; the reproduction, patch and tests were reviewed before submission.