Environment: Linux web dev servers, separate worktrees and fresh T3 homes, T3CODE_DEV_SEED=0, Chromium preview browser, 1280x800 CSS viewport, 2560x1600 source captures, light theme, Pi 1.5.2 with real authenticated API. Base: a61291e3e2c401852454b701e82278847c9c4ab1 Head: ad573ca3460cb9d70105855173121361ef58f201 Both use the same synthetic README and math.ts, and saved writer selection: {instanceId:pi, model:anthropic/claude-sonnet-4-6, options:[{id:thinking,value:low}]}. Pi config is isolated per side, initialized with defaultThinkingLevel=medium. The observer forwards commands/output without modifying them. It retains only command metadata and assistant message_end thinking/model/stopReason fields. Observation JSON excludes unrelated health-check sessions and process IDs. Both first runs generated and created 'Add math utility with add function'. Base sent no set_thinking_level; assistant message_end reported medium. Head sent set_thinking_level(low), Pi accepted it, then message_end reported low. Both config files remained exactly {"defaultThinkingLevel": "medium"}. The existing chat is a Replay seed, unrelated to the real Pi commit generation. The chat composer's model selection does not control the source control writer. Videos are unsped excerpts of the first Commit interaction: base seconds 28–48, head seconds 8–28. Only the machine label is masked, and output is scaled to 1280x800. Commit stills are frames at base 42s and head 22s. Settings stills are later captures of the same unchanged saved selection; two captures 500ms apart were identical. Annotated before/after images combine a crop of these real Settings captures with the observed RPC transcript. A later attempt to rerecord after soft-resetting the synthetic commits captured stale UI state; it is excluded. All behavioral evidence here is the first run. No repository tests are offered as evidence.