Backtestsopenai-agents-pythonrun-20261024-184640#5275
gpt-4o-transcribe is being retired (config.py)
Our patch was judged better than the merged fix.
In this judgment, Patch A = our agent, Patch B = the merged fix.
Both patches make the same core change, switching DEFAULT_STT_MODEL in src/agents/voice/models/openai_model_provider.py from gpt-4o-transcribe to gpt-transcribe, and all four independent tests go from FAIL-ASSERT on base to PASS on both, so correctness on the issue's ask is equal. The difference is in scope and risk. Patch A adds a focused test that asserts the default STT model name is gpt-transcribe and that an explicit model still overrides it; it would fail before and pass after, and nothing else changes. Patch B bundles two things the issue never asks for: (1) an unrelated pyproject.toml change to the sqlalchemy extra (SQLAlchemy[asyncio]), which is pure dependency churn in a model-rename fix, and (2) a behavioural change in openai_stt.py that, via a hard-coded string compare on self.model == 'gpt-transcribe', drops the language kwarg and re-sends it as extra_body={'languages': [...]}. Nothing in the issue supports that the replacement model rejects the language parameter, and the independent test with language set passes identically on A and B, so this remap earns no credit and adds risk: it is a brittle special case that would not apply to any other model name, and it is applied only to the non-streamed path while the streamed session code is untouched (yet B's streamed test asserts a 'languages' key, so that test's validity depends on unchanged code already behaving that way). B's wire-level tests are more thorough than A's, but thoroughness in tests does not offset unrelated dependency edits and speculative API-shape changes mixed into a one-line default swap. A resolves the issue at its root with no extraneous risk, so A is the better fix.
Four judges ruled on this ticket in parallel, and the verdict comes from counting their votes. Each saw the issue and the two patches, labelled A and B in random order (two of them saw ours as A), with issue links and version numbers removed. Where tests could check the fix, they also saw the results of tests a separate agent wrote for the issue, run on both patches. None was told which patch was ours. The reasoning above is one judge's, shown verbatim.
Solve
3m
Wall-clock time in the container
Full test suite
No new failures
852 passed · 3 skipped · 855 total
Contamination audit
Clean
No sign the agent saw the real fix
Patch size
13vs 99 merged
Changed lines, ours vs merged