Evidence environment: Linux, two detached worktrees at the commits shown in the recordings, each running its own real T3 development server and fresh database. Automatic developer-data seeding was disabled with T3CODE_DEV_SEED=0. No database fixtures, mocks, or tests were used. The same configured real agent runtime and identical prompts were used on both sides; private runtime identifiers and credentials are intentionally absent.

The videos show live MCP response fields printed by driver.py in a white 120-column by 46-row terminal. The full display was recorded at 5 fps and cropped identically to the terminal. Elapsed seconds are relative to the start of each driver, with no time compression. A run ending is distinct from its delegated task finishing; task_status.workState is the waiting-state evidence.

Sequence:
1. A delegates B asynchronously; B replies READY. A receives the completion wake and replies CHILD_READY.
2. A calls t3_thread_send on the same finished B. B delegates C asynchronously. C runs sleep 100 in a foreground shell, then replies GRANDCHILD_DONE. B ends its current turn with BETA and replies FINAL when C completion wakes it.
3. A runs sleep 20, calls task_status once on its follow-up, reports workState and summary, and ends the turn. The outside driver observes server projections without sending additional turns.
4. On the follow-up completion wake, A uses the delivered summary literally in PARENT_RESULT:<summary>.

Run session-bootstrap.mjs from each worktree with the private token-file path as its first argument. It uses T3's existing local auth service to issue a full-access MCP client session for that isolated server. Start both servers with T3CODE_DEV_SEED=0 vp run dev first.

Run driver.py with --origin, --token-file, --provider, --model, --workspace, --output, --side, and --commit. Both sides must use the same runtime selection. Use fresh state for each run. It creates the synthetic project and threads through real MCP calls, and writes each run into a new timestamped directory. Files named private-* contain the complete local protocol data and are not public artifacts. The public transcripts contain selected task fields and synthetic replies only.

record.sh is the recording wrapper, with its scratch location parameterized for sharing. Set EVIDENCE_SCRATCH, EVIDENCE_PROVIDER, and EVIDENCE_MODEL privately, then run it through headless-desktop. Its ports correspond to these two worktrees; adjust only the ports if T3 chooses different ones.

AI-assisted; verified with the tools listed above.