fix: coerce chat summary to str for RunCompletionEvent

[genisis0x]

Summary

Closes #1811.

RunCompletionEvent.summary is annotated str, but the path that fills it accepted whatever the configured summary_method (or the LLM reflection response) returned. Some providers — Ollama, Cohere, and any model emitting a structured assistant message — surface the completion as a nested dict like:

{"content": "...", "tool_calls": None}

The single-level if isinstance(summary, dict): summary = str(summary.get("content", "")) guard in _summarize_chat only unwraps one level. When the inner "content" is itself a dict (which is what the reporter in #1811 hit with Ollama llama3.1:8b and again with Cohere command-r), the dict propagates straight to RunCompletionEvent, and pydantic rejects it:

ValidationError: 1 validation error for RunCompletionEvent
summary
  Input should be a valid string [type=string_type,
  input_value={'content': '...', 'tool_calls': None}, input_type=dict]

Following the run_and_event_processing two-agent comedians notebook with the LLM swapped to Ollama makes the chat unusable from the first turn.

Change

Add a small _coerce_summary_to_str helper:

• recursively unwraps a "content" key from each nested dict,
• treats None as "",
• stringifies any other non-str value (e.g. a callable returning an int).

Route both _summarize_chat (the gate every ChatResult.summary passes through) and the new dict branch of _last_msg_as_summary through it.

Behaviour for the existing single-level dict and plain-string return values is unchanged, so the existing test_summarize_chat_with_dict_summary still passes.

Test plan

New regression coverage in test/agentchat/test_conversable_agent.py:

• test_summarize_chat_coerces_arbitrary_returns_to_str — parametrised across single-level dict, nested-dict-content, dict-with-None-content, plain None, and a non-string scalar
• test_last_msg_summary_unwraps_dict_content — _last_msg_as_summary follows dict-shaped content and still strips TERMINATE on that path

$ AUTOGEN_USE_DOCKER=0 pytest test/agentchat/test_conversable_agent.py \
    -k "summary or summarize or last_msg" -q
... 7 passed, 1 skipped

(AUTOGEN_USE_DOCKER=0 is only needed locally because UserProxyAgent in the existing test is constructed with the default code_execution_config and tries to run docker; CI sets this differently. The new tests pass code_execution_config=False directly.)

[CLAassistant]

All committers have signed the CLA.

[marklysze]

Clean fix. The existing single-level isinstance(value, dict) check was insufficient for nested dict responses from Ollama/Cohere (e.g. {"content": {"content": "..."}}). The recursive _coerce_summary_to_str helper handles the nesting correctly while still falling back to str() for unexpected types. Tests cover the parametrized return-type matrix and the _last_msg_as_summary dict-content unwrap. Approved.