# ACP cancel ordering: real-app reproduction

Environment on both sides: Linux web dev client, Chromium, 1280×800 CSS viewport, light theme; fresh per-checkout T3 home (T3CODE_DEV_SEED=0), synthetic project, OpenCode 1.19.5 via local acpRegistry command `opencode acp`, model `openai/gpt-6-luna`, full access. No replay agent and no application source modifications.

1. Configure provider instance `proofAcp` with driver `acpRegistry`, enabled true, config `{ "source": "local", "commandPath": "/usr/local/bin/opencode", "commandArgs": ["acp"] }` in the isolated home's settings.json before starting the app.
2. Start `T3CODE_DEV_SEED=0 T3CODE_AUTO_BOOTSTRAP_PROJECT_FROM_CWD=0 vp run dev --no-seed` in each checkout. Pair its web client.
3. Add a synthetic git project using `node apps/server/src/bin.ts project add <workspace> --title 'ACP cancel ordering' --base-dir .t3`.
4. Start A and B with `node apps/server/src/bin.ts thread start <workspace> 'Reply exactly READY' --title 'Thread A ready' --provider proofAcp --model openai/gpt-6-luna --base-dir .t3 --wait --json`; use BRAVO and Thread B ready for B.
5. Open A by its environment/thread URL. Start a browser recording.
6. In the checkout, run `python3 fifo-driver.py hold <thread-B-id>`. It replaces B's own provider log with a FIFO, sends B a message, and waits for its flush to block. B may dispatch its own prompt before the batch flush takes the shared log-store lock; A's later prompt is held.
7. Fill A's composer with `Use your shell tool to run exactly this command: sleep 120 && echo SLEEP_DONE > SLEEP_DONE.txt` and click Submit.
8. Confirm OpenCode's log has no new prompt loop for A while held. Click Stop during the blocked dispatch, then immediately run `python3 fifo-driver.py release <thread-B-id>` in the same checkout. Release occurs within the 15-second cancel wait bound.
9. Observe A: base starts Running sleep and shows Agent not responding to Stop / Force stop; head finishes Run interrupted / Stopped by user without executing sleep. Base requires Force stop to clean up.
10. On head, send a follow-up `Reply exactly READY_AFTER_STOP` to the same A and observe whether it completes. In the captured head follow-up, OpenCode first resumed the earlier shell instruction (Running sleep); this is recorded separately from the interrupted turn.

Browser actions were driven through T3's preview automation, using the actual composer and buttons. The fill expression was:

```js
(() => {
  const editor = document.querySelector('[data-testid="composer-editor"]');
  editor.focus();
  document.execCommand('insertText', false,
    'Use your shell tool to run exactly this command: sleep 120 && echo SLEEP_DONE > SLEEP_DONE.txt');
})();
```

In a separate browser action (so React processes the input), submit:

```js
document.querySelector('button[aria-label="Submit message"]').click();
```

Then Stop:

```js
document.querySelector('button[aria-label="Stop generation"]').click();
```

The recordings are continuous captures of these interactions, downscaled from device pixels to 1280×800. A rectangular redaction covers the private environment label at the top right. PNGs are labelled video frames, not separate static-layout proof. Additional failed Haiku setup threads are visible in the base sidebar; they were not used for the reproduction. The head capture was repeated after a panel-maximization action hid the first run's result; its new A had the same READY warm-up and the same FIFO driver and prompt.

## Observed timestamps

Base: Stop reached OpenCode at 00:35:24.175Z; FIFO released 00:35:24.222645Z; A prompt loop began 00:35:24.242Z; sleep permission check 00:35:28.640Z. The app showed Running sleep and Agent not responding to Stop, requiring Force stop.

Head (final capture): FIFO released 00:40:19.552367Z; OpenCode logged cancel 00:40:19.568Z; the server persisted the turn as interrupted. It did not start a prompt loop or sleep before a later explicit follow-up.

Head follow-up: a new user message at 00:40:53 started a new provider loop; at 00:40:57.852Z it chose to run the previous sleep instruction. This continuation behavior is retained as a limitation rather than hidden.

The follow-up completed with READY_AFTER_STOP at 00:43:02Z after running the previous shell command; SLEEP_DONE.txt was created only in that new turn. The earlier turn remained interrupted. This verifies a sequential follow-up was not cancelled, but does not prove overlapping prompt dispatch.
