Environment for both recordings: Linux, real T3 development server from the stated detached commit, Codex CLI 0.167.0, Code Mode enabled with --enable code_mode. Independent worktrees and databases, T3CODE_DEV_SEED=0; identical synthetic project workspace. No live database copied. Native runtime identifiers are redacted.

Run vp i, then T3CODE_DEV_SEED=0 vp run dev in each checkout. Seed with seed.py CHECKOUT SYNTHETIC_WORKSPACE. Set EVIDENCE_RUNTIME to an available runtime identifier in your environment, then run driver.py CHECKOUT before|after --workspace SYNTHETIC_WORKSPACE. The same runtime and prompt were used on both sides.

The recordings are real-time captures of the real server's parent provider-rollout events, collected live by driver.py. session.cast preserves the timed terminal output. recording.mp4 renders that same captured stream at its original elapsed timing using render.py; it is a terminal transcript recording, not a recording of the web UI. Long quiet intervals on head reflect its blocking tool call. Event inputs are preserved with private runtime identifiers redacted; large tool results are summarized. observations.txt independently lists server projection timestamps, the actual child's shell execution and reply, and native parent tool-call timestamps. inspect.py produces it from only the run's own threads.

Before: delegation ran inside exec, yielded after 31.0 seconds, then required six parent wait calls. After: one direct delegate_task call blocked for 102.1 seconds until DONE_42, with zero parent cell wait calls. Both children executed sleep 90 successfully and returned DONE_42.

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