# Delegated Monitor completion evidence

Environment for both captures: Linux, real development web client in Chromium,
1280 × 800 CSS viewport, dark theme. Each commit ran its own isolated development
server and fresh database with `T3CODE_DEV_SEED=0`. Both used the same empty,
synthetic workspace and project title, seeded with `seed.py`; no live data copied.

Run `drive.py` with the environment variables documented in its header for each
checkout. It sends the identical coordinator prompt through the real server CLI.
The coordinator and worker both use the Claude provider. Runtime selection is
provided privately through environment variables, never embedded in these files.

Browser sequence on each side:
1. Open the resulting coordinator thread and observe the initial WAITING result.
2. Open its delegated worker row while Monitor is running.
3. Open Monitor Details; allow the command to end without waiting or polling
   from either agent. The worker must answer FINAL.
4. Return using Open parent thread. Inspect the delegated results and answer.

The command is exactly `sleep 60; echo MONITOR_DONE`. Neither agent invoked
`t3_background_wait`; the worker ended its first turn with WAITING. After Monitor
ended, the worker answered FINAL. Base retained only WAITING at the coordinator;
head added a distinct delegated FINAL completion and the coordinator answered FINAL.

Videos retain real-time timing after the initial setup portion. UI loading/setup
is trimmed. Cropping excludes side panels and the composer, which contain private
runtime metadata. Any additional redaction is labeled. `observe.py` produces a
small read-only transcript of the synthetic server's persisted messages and
Monitor invocation; it does not inject events or alter state.

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