Python: Fix structured output parsing when text contents are not coalesced

[t-anjan]

Summary

ChatResponse.value and AgentResponse.value use self.text to feed into model_validate_json() for structured output parsing. self.text delegates to Message.text, which joins multiple TextContent objects with " ".join().

This is correct for natural language responses, but corrupts JSON when text contents are not fully coalesced into a single content — spaces get injected into JSON keys and values, causing Pydantic
validation failures.

The fix is scoped to the value property, which only executes when response_format is set (i.e., structured output). Message.text and its " ".join behavior are unchanged — natural language responses
are unaffected.

Real-world failures

Observed in production with OpenAI-compatible chat client (OpenRouter → Gemini) using response_format for structured output. Failures were intermittent — retrying the same request succeeded.

Failure 1 — Space in JSON key:
LLM returned valid JSON with "action": "request_evidence", but after Message.text processing, Pydantic received "action ": "request_evidence" (trailing space in key). Error:
Field required [type=missing, input_value={'action ': 'request_evid...}]

Failure 2 — Space in JSON value:
LLM returned "readiness": "not_started", but Pydantic received "readiness": "not_started " (trailing space in value). Error:
Input should be 'not_started', ... [input_value='not_started ']

In both cases, the raw LLM response was valid JSON — the corruption was introduced by " ".join() in Message.text.

Changes

• ChatResponse.value: when parsing structured output via response_format, concatenate text contents directly instead of using self.text (which delegates to Message.text's " ".join)
• AgentResponse.value: same fix
• Message.text is not modified — natural language joining behavior is preserved
• Added tests for both classes demonstrating the bug and fix

Test plan

• New tests: test_chat_response_value_multi_chunk_json and test_agent_response_value_multi_chunk_json
• Each test creates a JSON response split across 3 TextContent objects (simulating uncoalesced streaming chunks), then asserts:
  1. message.text contains injected spaces — confirming " ".join behavior exists: '{"resp onse": "He llo"}'
  2. .value still parses correctly despite that — confirming the fix works
• Without the fix, response.value throws: ValidationError: Field required [type=missing, input_value={'resp onse': 'He llo'}] — the same class of error observed in production
• Existing test_types.py tests continue to pass