Summary

Severity: High/Medium
Area: Gemini adapter
Affected files: agentkit/llm/gemini.py:159-169
Suggested branch: fix/gemini-thought-part-leak
Suggested PR title: fix(gemini): exclude thought parts from the model answer

This is finding AUDIT-06 of 18 from a consolidated audit of agentkit/. It is deliberately self-contained so it can be fixed on its own branch with its own PR.

Related findings: Related: AUDIT-05.

AUDIT-06 — High/Medium — Gemini thought parts are appended to the answer

File — agentkit/llm/gemini.py:159-169

            candidates = getattr(raw_response, "candidates", None)
            if candidates:
                candidate = candidates[0]
                if candidate.content and candidate.content.parts:
                    for part in candidate.content.parts:
                        if getattr(part, "text", None):
                            text_parts.append(part.text)
                        if getattr(part, "function_call", None):

Why this is a bug

types.Part has a thought flag (google-genai 2.25 Part.thought). The loop never checks it, so internal reasoning text is concatenated into LLMResponse.text, which becomes the assistant turn content (agent.py:210-215), the persisted history, and the final_answer returned to the API client.

Reproduced

part = Part(text="SECRET CHAIN OF THOUGHT", thought=True)
adapter LLMResponse.text == "SECRET CHAIN OF THOUGHT"

Pass cases

• A part with thought=True is never included in LLMResponse.text / final_answer / persisted assistant content.
• A part with thought=False and text is still returned normally (no regression).
• A part carrying both thought=True and a function_call still yields the tool call (and its thought_signature) — the call is not lost.
• A regression test asserts the API's answer field contains no thought text for a Gemini response containing a thought part.

Definition of done (per AGENTS.md)

• New/updated unit and integration tests covering the pass cases above, all passing (pytest).
• ruff check ., ruff format --check . and mypy agentkit clean.
• Docstrings updated where behaviour changed; docs/PROGRESS.md updated (status, PR link, commit count, notes).
• A new ADR in docs/DECISIONS.md if an architectural decision was made.
• Branch pushed and PR opened (never commit directly to main).

Baseline at the time of writing: pytest is green on main (all tests pass; tests/real/* skipped unless AGENTKIT_REAL_INFRA=1).