# Native subagent settlement proof

Captured by gpt-6.1-sol through Codex in the web client, Chromium, 1280 × 800, light theme.

Use a checkout of current main with the scenario runner and install dependencies with `vp i`.
Place the downloaded scenario files under `scripts/scenarios/` without committing them.
Claude Code must be installed and authenticated using the operator's existing credentials.
The scenarios enable the native Claude provider on their isolated server and use Haiku 5.5.
They append the same explicit native-background-Agent exception on both commits, because the
normal T3 delegation preference can otherwise make the model refuse this reproduction.
No credentials, copied databases, or real project data are included.

Run the settlement comparison:

```sh
T3CODE_DEV_SEED=0 T3_SCENARIO_INHERIT_ENV='T3CODE_DEV_SEED,ANTHROPIC_*,CLAUDE_*' \
vp run scenarios:prove scripts/scenarios/native-subagent-settlement.e2e.ts \
  --base 82479522cb667e09616a04705648ac4fab68b997 \
  --head 805cd20868269afc295242623769333e0c054be8 \
  --upload-command 'ghdrop upload'
```

The driver seeds one synthetic project and one thread on a fresh database, requires the
literal final reply and two working native agents after the parent run finishes, records
Settle, reopens the same thread, and waits 151 seconds before its final capture and assertion.
The two native agents are instructed to run `sleep 150` in the foreground.
Recording starts after the parent reply so startup and model commentary are excluded.

Run the supplementary release comparison with the same command, replacing the scenario name
with `native-subagent-release-paths.e2e.ts`. The supplementary file supplied with the evidence is the exact version used for that successful comparison; it reopens through `t3.open` after release. It captures archive/unarchive, attempted workspace
and provider/model changes, explicit session detach, and unexpected CLI termination. The latter
uses a wrapper around the real Claude CLI which records its spawned PID and Linux process start
time; only that captured process is terminated. A blocked change is reported as blocked, not
as a tested release. Each completed release is captured after five seconds and the same thread
is reopened. Generic ACP replay cannot create native Claude subagents.

The runner recreates its isolated home separately for each commit, starts each commit's real
server and web client, and stops its app processes. Its `proof.json` contains the comparison,
per-side steps, capture URLs, and recordings. No unit-test result is used as visual evidence.

The primary Settle videos are unedited recordings. The supplementary MP4 videos retain every
frame and their original timing, with only the transient dev pairing-token field masked.
To apply the same redaction, run `python3 redact-pairing.py pairing-heading.png input.webm output.mp4`.
The heading template contains only public UI text. No agent status, reply, sidebar, or timing is masked.
