Desktop runs leave a rollout file with cli_version and the model per turn. App-server companion rounds leave no session artifact, and neither the run output nor the job record states the resolved model or which codex binary served the run — the binary is PATH-decided and therefore silently swappable.
Why it matters: "which model reviewed which change" has to be unambiguous after the fact. We maintain a per-round record by hand purely because of this gap, which means the record is only as good as the discipline keeping it.
Ask: echo the resolved model id and the binary version in the run header (and, if practical, in the job record).
Adjacent data point, offered while filing: gpt-5.6-sol completes the structured-output adversarial path on CLI 0.144.1, so the gpt-5.4 pin in 1.0.3 appears obsolete on current state.
Related: #654 asks for --model to be documented on the review commands; this issue is about recording what was actually resolved.
Desktop runs leave a rollout file with
cli_versionand the model per turn. App-server companion rounds leave no session artifact, and neither the run output nor the job record states the resolved model or whichcodexbinary served the run — the binary is PATH-decided and therefore silently swappable.Why it matters: "which model reviewed which change" has to be unambiguous after the fact. We maintain a per-round record by hand purely because of this gap, which means the record is only as good as the discipline keeping it.
Ask: echo the resolved model id and the binary version in the run header (and, if practical, in the job record).
Adjacent data point, offered while filing:
gpt-5.6-solcompletes the structured-output adversarial path on CLI 0.144.1, so thegpt-5.4pin in 1.0.3 appears obsolete on current state.Related: #654 asks for
--modelto be documented on the review commands; this issue is about recording what was actually resolved.