fix(extensions): raise on truncated empty Chat Completions responses

[CoralGarden52]

Summary

This pull request fixes AnyLLM and LiteLLM non-streaming Chat Completions adapters returning successful empty responses when a completion terminates with finish_reason="length" before emitting assistant text, refusal, or tool calls.

Both adapters now raise ModelBehaviorError after usage and tracing data are recorded, matching the built-in OpenAI Chat Completions adapter. Responses containing partial text, refusals, or tool calls remain supported.

Test plan

• Added regression tests for empty content=None and content="".
• Added coverage for partial text, refusals, tool calls, and missing LiteLLM finish_reason.
• Added tracing coverage to verify usage is recorded before ModelBehaviorError is raised.
• Focused adapter tests: 101 passed.
• Full parallel test suite: 9360 passed, 29 skipped.
• Serial test suite: 77 passed, 4 skipped.
• Formatting, lint, mypy, Pyright, and the final code-change verification script passed.

Live verification:

• Model: gpt-5.6-luna
• temperature=0
• max_tokens=256

The raw Chat Completions response returned finish_reason="length" with content=null. The built-in adapter raised ModelBehaviorError, while AnyLLM and LiteLLM previously returned successful empty outputs.

Issue number

Closes #4885

Checks

• I've added new tests, if relevant
• I've run .agents/skills/code-change-verification/scripts/run.sh
• I've confirmed all verification steps pass
• If using Codex, I've run /review before submitting this PR

[seratch]

Thanks for the fix. The new condition also rejects length-truncated responses containing reasoning but no assistant text. The existing adapters can return that reasoning as a ResponseReasoningItem, so this may change more than empty-response handling. Please provide a real-provider comparison for a reasoning-only truncated response through the affected adapters, showing the base and head outputs. We need that evidence to assess compatibility before merging.

[fscfede-beep]

I independently reproduced the review concern in a local fork: a finish_reason=length guard that only checks assistant text/refusal/tool calls also rejects reasoning-only completions that the adapters can convert to a ResponseReasoningItem.

I tested a reasoning-aware variant at fscfede-beep/openai-agents-python@504dcf9:

• AnyLLM also requires _extract_any_llm_reasoning_text(message) to be empty before raising.
• LiteLLM also requires both reasoning_content and thinking_blocks to be empty before raising.
• Added reasoning-only finish_reason=length regression coverage for both adapters.
• Local affected-suite verification: 96 passed; Ruff, Pyright, mypy, and git diff --check pass.

Branch: https://github.com/fscfede-beep/openai-agents-python/tree/fix/empty-length-chat-completions-4885

I have not run the requested live paid-provider base/head comparison, so this is only code-level compatibility evidence, not a substitute for that live verification.

[seratch]

Without confirming real-world impact, we won't move forward with this possibly breaking change to existing use cases.

[fscfede-beep]

I now have a real-provider base/head comparison for the reasoning-only truncation case.

Provider/model used:

• Ollama 0.32.14, isolated local server on loopback
• rumbo-qwen3-4b-q4km:latest (Qwen3 4B)
• temperature = 0
• max output tokens = 8
• prompt held constant across runs

Raw provider behavior is the case from the review concern: the OpenAI-compatible endpoint returns finish_reason="length", empty assistant content, and non-empty reasoning (reasoning_len=18). I am intentionally not posting the reasoning text itself.

Pinned code states:

• base: 1d471a4775bf2f40179f411824da383deb4c3fca
• PR fix(extensions): raise on truncated empty Chat Completions responses #4886 head: 7a437cea5bf7a2e10194ee7239a25d87f59f31ad
• reasoning-aware variant: 504dcf93a52fe86732509a83ebb24f986b9f9133

AnyLLM, end-to-end against the real Ollama provider:

• base -> returns one ResponseReasoningItem (output_types=["reasoning"], output_tokens=8)
• PR fix(extensions): raise on truncated empty Chat Completions responses #4886 -> raises ModelBehaviorError: Chat Completions response terminated with finish_reason='length' but produced no assistant text, tool call, or refusal.
• reasoning-aware variant -> returns one ResponseReasoningItem (output_types=["reasoning"], output_tokens=8)

LiteLLM, end-to-end via Ollama's real OpenAI-compatible /v1/chat/completions endpoint:

• base -> returns one ResponseReasoningItem (output_types=["reasoning"], output_tokens=8)
• PR fix(extensions): raise on truncated empty Chat Completions responses #4886 -> raises the same ModelBehaviorError
• reasoning-aware variant -> returns one ResponseReasoningItem (output_types=["reasoning"], output_tokens=8)

I also independently called litellm.acompletion() against that endpoint; it returned in ~6.65s with finish_reason="length", content_len=0, and reasoning_content_len=18, confirming the LiteLLM provider response itself carries the reasoning before the Agents SDK adapter processes it.

So this is a concrete compatibility break in an existing supported path: before #4886, a length-truncated reasoning-only completion is surfaced as a reasoning item; at the current PR head it becomes an exception. The reasoning-aware guard at 504dcf93... preserves the previous behavior while still rejecting truly empty length-truncated responses.

This is local real-provider evidence, not a hosted/paid-provider claim. The branch remains: https://github.com/fscfede-beep/openai-agents-python/tree/fix/empty-length-chat-completions-4885