# Stop while a Claude steer is queued

Environment: real web app in Chromium, 1280×800 CSS pixels, light theme, separate fresh databases at each commit, real Claude CLI. Synthetic project and messages only. The replay scenario runner was not used because generic ACP replay cannot exercise Claude prompt UUID acknowledgments.

1. Start the included hold-endpoint.py with Python 3. It holds each response for 240 seconds on loopback port 18787.
2. In each isolated app, add an empty git project titled Steer delivery. Start a Claude thread titled Stop during foreground command and send: `Reply READY only. Do not inspect any files.` Wait for READY. Project and initial thread were created through the app CLI.
3. Settings → Chat → Follow-ups → Follow-up behavior: Steer (already selected on both). The supplied reproduction's Settings → General location is now Chat.
4. Send exactly: `Run this exact command with Bash in the foreground, not background: curl --max-time 300 http://127.0.0.1:18787/hold . The endpoint intentionally holds its response for four minutes. Set the Bash timeout to 300000 ms. Do not read files or run any other command. Wait for curl to finish before replying.`
5. Wait for Running curl, a request at the endpoint, and the provider's system/task_started with is_backgrounded=false and task_type=local_bash.
6. Send: `Codeword: PELICAN. Remember it, I will ask for it later.`
7. Press Stop before curl completes. Base leaves the steer visible with Run interrupted and no delivery warning. Head displays an orange unconfirmed-delivery warning; expand it to read the original steer and resend guidance.
8. Ask: `What codeword did I ask you to remember? Reply with the codeword only; if no codeword reached you, reply NO_CODEWORD. Do not use tools.`

Before: 3a8c250cb0c8056f2f563091fd412fe5db809fa1
After: d3471b1d693972f7d3b273ffc3dcd41f059dbf68

Captures hide runtime identifiers. The head recording has one opaque redaction over a model-transition label; no reproduction content is replaced. Before/after stills were checked pixel-identical against another screenshot taken more than 500 ms apart.

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