Closes #11. Cancellation - Every `docker run` gets a unique `--name` (agent-gwt-<agent>-<hex>). - runDocker accepts `signal`, `stdin`, and `containerName`; on abort it force-removes the container synchronously (an async removal was cut short by worker teardown when the timed-out test was the last in its worker), kills the client, and rejects with the abort reason. Exit, SIGTERM and SIGINT handlers remove any still-tracked container, then re-raise. - RunAgentOptions gains `signal`; agent({ timeoutMs }) puts a deadline on the context and executing_the_agent aborts the run when it passes, throwing `agent run exceeded <n> ms`. - Under vitest, executing_the_agent also aborts when the test finishes first (its own timeout, a failure elsewhere), so the agent stops working and billing. vitest is loaded dynamically because it is an optional peer; other runners skip this. Prompt on stdin - Both bindings pass the prompt through `docker run -i` on stdin instead of argv. Linux caps one argument at 128 KB; stdin has no limit, and the prompt no longer shows in `ps` on the host. Verified against both CLIs. - buildClaudeDockerArgs / buildDockerArgs take an optional `containerName` instead of `prompt` (they build argv only). Tests - _runProcess.spec: stdin delivery, abort kills the child and rejects, already-aborted signal rejects at once (uses node as the child). - Specs for both bindings assert the name, `-i`, no prompt on argv, and that the runner receives stdin, containerName, and the signal. - executing_the_agent.spec: timeoutMs cancels and reports. - e2e/cancel.spec: against real docker, an aborted run rejects and its container is gone; a 300 KB stdin reaches the container intact. README: cancellation and timeouts section; stdin noted where the run commands are described.

Commits

Commits on Aug 30, 2026

feat: cancel runs that time out, and send the prompt on stdin