# Claude steer reproduction

Environment on both sides: Linux web development build; Chromium, 1280 x 800, light theme; Claude Code 2.1.328, Haiku 5.5; fresh isolated T3 home with T3CODE_DEV_SEED=0 and a synthetic empty Git project. No replay provider or test doubles.

Run seed.sh from each isolated checkout. Open the returned thread in the web client, and close the thread details panel.

Start a tab recording. Submit this exact composer prompt:

Run exactly three separate foreground Bash tool calls in parallel. In each call run python3 -c "import time; time.sleep(90); print('MARKER_A')" (use MARKER_B and MARKER_C in the other two). Set timeout to 120000. Do not run in background. Wait for all three calls, then report their markers. Keep running even if I send a progress update.

Once Running python3 appears (all three tool requests were emitted), fill the composer with:

Progress update: keep all three foreground calls running. After they finish, acknowledge STEER_RECEIVED and list the three markers.

Click Steer message. Expand the command group and the individual original commands to inspect their output. Wait for the run to finish. Claude Code executes the three foreground Bash requests sequentially in this runtime, despite the requested parallel invocation; allow at least 270 seconds on the head.

On base, original MARKER_B and MARKER_C are cancelled with: The user doesn't want to take this action right now. STOP what you are doing and wait for the user to tell you how to proceed. Claude then says the calls were declined and retries them. These retries are not evidence of the original calls succeeding.
