Real web app comparison, Codex CLI 0.167.2, Linux, light theme, 1280x800 CSS viewport (captures 2000x1250 device pixels).
Each commit ran in its own detached worktree with its own fresh database and T3CODE_DEV_SEED=0. Synthetic project created through project add. Both use the same scratch Codex configuration and local Responses API stand-in; no Bedrock account was used.

The final videos show the same first turn in a new thread: type "Reply with just: ok", then Enter. The base shows a provider error and the head displays "ok". Auxiliary title-generation requests also reach the stand-in. requests.jsonl records only reasoning and acceptance, never prompts, model identifiers, credentials or filesystem paths. Main-turn requests use Medium effort: base contains summary=detailed and is rejected; head omits summary and succeeds.

For privacy, model-selector and workspace-details elements were hidden with capture-only CSS on both sides, and the sidebar was collapsed. No conversation or error content was changed. bedrock-test is a synthetic stand-in label, not a runtime model identifier. Both final stills were captured twice at least 500 ms apart and their PNG files compared byte-identical. Both stills and recording contact sheets were visually inspected.

The reproduction helper describes the commands and browser steps used; it is supplied for repeatability and was assembled from the executed commands, rather than run as a single script. The stand-in script itself was run for these captures.

AI-assisted; verified with the tests and tools listed above.