fix(extensions): reject truly empty length-truncated Chat Completions

[fscfede-beep]

Summary

Successor to closed #4886 / addresses closed #4885.

The original change correctly rejected truly empty non-streaming Chat Completions responses ending with finish_reason="length", but the first implementation also rejected reasoning-only responses. This PR preserves the intended empty-response protection while retaining provider reasoning that is already surfaced by AnyLLM/LiteLLM.

Behavior

Raise ModelBehaviorError only when all of these hold:

• finish_reason == "length"
• no assistant text
• no refusal
• no tool calls
• no provider reasoning (reasoning_content / thinking_blocks, or AnyLLM normalized reasoning)

Preserve partial text, refusals, tool calls, and reasoning-only responses.

Verification

Local regression suite for the affected adapters: 96 passed.

Static verification:

• Ruff: PASS
• Pyright: 0 errors, 0 warnings, 0 informations
• mypy: PASS
• git diff --check: PASS

Real-provider compatibility test, same provider/model/prompt and max_tokens=8 across pinned code states:

• provider: Ollama 0.32.14 on isolated loopback server
• model: rumbo-qwen3-4b-q4km:latest
• base: 1d471a4775bf2f40179f411824da383deb4c3fca -> one ResponseReasoningItem
• previous PR fix(extensions): raise on truncated empty Chat Completions responses #4886 head: 7a437cea5bf7a2e10194ee7239a25d87f59f31ad -> ModelBehaviorError
• this fix: 504dcf93a52fe86732509a83ebb24f986b9f9133 -> one ResponseReasoningItem

Both AnyLLM and LiteLLM were tested end-to-end through the real Ollama OpenAI-compatible endpoint. The raw provider returned finish_reason="length", empty content, and non-empty reasoning.

The original PR was closed after the maintainer requested real-world compatibility evidence; that evidence is now available and the incompatible behavior has been corrected.

[fscfede-beep]

Live-provider compatibility evidence

Following the review concern on closed #4886, I ran the same non-streaming request end-to-end through both affected adapters against the same real local OpenAI-compatible provider.

Provider: Ollama 0.32.14, model rumbo-qwen3-4b-q4km:latest, isolated loopback 127.0.0.1:11435, temperature 0, max_tokens=8.

Raw provider result: finish_reason="length", empty visible content, non-empty reasoning (18 chars). No provider mock was used.

Pinned code-state comparison:

• base 1d471a4775bf2f40179f411824da383deb4c3fca: AnyLLM -> one ResponseReasoningItem; LiteLLM -> one ResponseReasoningItem.
• prior PR fix(extensions): raise on truncated empty Chat Completions responses #4886 head 7a437cea5bf7a2e10194ee7239a25d87f59f31ad: AnyLLM -> ModelBehaviorError; LiteLLM -> ModelBehaviorError.
• this fix 504dcf93a52fe86732509a83ebb24f986b9f9133: AnyLLM -> one ResponseReasoningItem; LiteLLM -> one ResponseReasoningItem.

Additional verification on the fix: 96 affected tests passed; Ruff PASS; Pyright 0 errors/0 warnings/0 informations; mypy PASS; git diff --check PASS.

The experiment demonstrates the compatibility regression and that the reasoning-aware guard restores the prior provider behavior while still rejecting truly empty length-terminated turns.

Receipt: OPENAI_AGENTS_4886_LIVE_REASONING_EVIDENCE_20260907.json, SHA-256 546FDA775E97BCB123F6BFC64756194DDDA903AE802D6ABBDB0A80A15015B822.

[fscfede-beep]

@codex review

Please review the current head 504dcf93a52fe86732509a83ebb24f986b9f9133. Focus on whether the new finish_reason="length" guard is correctly scoped so truly empty responses raise while reasoning-only, refusal, tool-call, and partial-text responses remain valid. Also audit the AnyLLM/LiteLLM adapter normalization and provider-reasoning preservation for P0/P1/P2 correctness issues.

[chatgpt-codex-connector]

Note

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.

[fscfede-beep]

Independent review gate check on the current head 504dcf93a52fe86732509a83ebb24f986b9f9133: the proposed condition is appropriately narrow—raise only for truly empty non-streaming length-truncated responses, while preserving reasoning-only, refusal, tool-call, and partial-text responses. The PR documents 96 focused tests and real-provider AnyLLM/LiteLLM + Ollama compatibility evidence. No new review threads are currently present. I am not treating contributor-side validation as upstream CI PASS; the current workflow is action_required with no jobs. Please evaluate the provider compatibility evidence and semantics on the upstream CI/review gate.

[seratch]

Thanks for adding the real-provider compatibility evidence. The Ollama comparison addresses the reasoning-only concern from #4886: both the base and this PR return a reasoning item.

However, that experiment does not exercise the truly empty response targeted by the new exception. It demonstrates compatibility with existing behavior, rather than a user-visible problem that this change fixes. The original reproduction in #4885 supplies a configured response, so the real-world impact needed for this behavior change remains unestablished.

I'm going to close this PR for now. We can reconsider with a real-provider example containing no text, refusal, tool calls, or reasoning, together with the concrete application consequence that requires truncation-specific handling.

[fscfede-beep]

Targeted maintainer-review note: this revision is intentionally narrower than #4886. The guard raises only for finish_reason="length" with no text, refusal, tool calls, or provider reasoning; reasoning-only responses remain valid. Live-provider compatibility was exercised against Ollama 0.32.14 through both AnyLLM and LiteLLM (no mock): base and this head preserve ResponseReasoningItem, while #4886 raised ModelBehaviorError. 96 affected tests + Ruff/Pyright/mypy/diff-check pass locally. Please evaluate the semantic boundary and whether the live-provider evidence is sufficient for merge.