Summary

Gemini 2.5 models (Flash and Pro) return reasoning_content as a dict in their OpenAI-compatible response, not a string. Other thinking/reasoning models via OpenAI-compatible proxies may exhibit the same behavior.

The existing truthy check (if reasoning:) treats a non-empty dict as valid, storing it as ThinkingPart.content. On the second API call, _into_message_param() calls '\n\n'.join(contents) which raises:

TypeError: sequence item 0: expected str instance, dict found

This crashes any multi-step / ReAct agent using Gemini 2.5 after the first tool call.

Fix

Add an isinstance(reasoning, str) guard in both the non-streamed and streamed _process_thinking paths so non-string values are silently ignored:

- if reasoning:  # pragma: no branch
+ if isinstance(reasoning, str) and reasoning:  # guard against non-string values (e.g. Gemini 2.5 returns a dict)

Reproduction

import types

msg = types.SimpleNamespace(
    reasoning_content={"reasoningContent": {"reasoningText": {"text": "", "signature": "abc"}}},
    reasoning=None,
)

reasoning = getattr(msg, "reasoning_content", None)

try:
    "\n\n".join([reasoning])
except TypeError as e:
    print(f"BUG: {e}")  # sequence item 0: expected str instance, dict found

if isinstance(reasoning, str) and reasoning:
    pass
else:
    print("Fix: non-string reasoning_content ignored")

• Closes TypeError when Gemini 2.5 returns reasoning_content as a dict (multi-step agents crash on second LLM call) #5157

Checklist

• Any AI generated code has been reviewed line-by-line by the human PR author, who stands by it.
• No breaking changes in accordance with the version policy.
• PR title is fit for the release changelog.