start_worktree creates a branch + worktree by policy and moves the session into it.
Settings card: location, branch prefix, fetch-first, setup command, cleanup on archive.
Rail shows the branch; clean worktrees are removed when the session is archived.
Ships surfaced and enabled beside the everyday Coworker (supersedes the
2026-08-21 lineup for Code only); prompt rebuilt around the discipline of
a worktree per feature, explore-first subagents, proof by tests.
shell_task_output returned at once, and run_shell's description told the model to poll it. That is right for a dev server (nothing to wait for) and wrong for a build or test run the next step depends on: every poll is a full model call, and the only alternative, run_shell("sleep N"), is a blind wait capped at 600 s. Measured on a sealed evaluation run: 297 polls on one background job, ~$8.50 of cache reads and ~50 min of a 90-min clock spent asking whether it had finished.
wait_seconds (default 0 = unchanged) blocks until the task exits, prints new output, or the user interrupts, up to the shell's 600 s ceiling; the record gains waited_seconds. The sandbox path carries the wait over the socket and stretches the client's RPC timeout to cover it (it was a hard 30 s). Descriptions reworded, generically: check background processes with shell_task_output; if your next step depends on one finishing, pass wait_seconds rather than polling.
Tests, on both executor variants: returns early on exit (with the last line), returns early on new output, gives up at the deadline while running, wait_seconds 0 / negative / omitted unchanged, cap advertised and applied.
Imported from andrewyng/aisuite@1b4bbf303e
(contents of its platform/ directory, hoisted to the repo root).
Development history prior to this commit lives in that repository.
Co-authored-by: Devika <devikaverma11@gmail.com>