Start a Pi worker in the background on a self-contained task and return its id at once. Workers always use the model the user set with pi_model (not selectable here). Give it a complete standalone brief (it has no access to this conversation). Use disjoint dirs or let worktree isolation separate parallel workers. Then use pi_wait / pi_digest to read compact results and judge them. Workers are long-lived RPC processes that keep their context: for a fix or a follow-up in the same area, pi_send an existing worker instead of spawning a new one.

The project dictionary: short term -> definition entries (what a system is called, what it does, where it lives, conventions) that are injected into every worker for that project. Seed it from what you already understand BEFORE the first pi_spawn; after reviewing a worker, add facts you verified (never unverified worker claims). Not for task notes or history. action: show | set (upsert entries) | remove (terms).

The project toolbox: the checks (tests, compile, lint) workers are expected to run THEMSELVES with check <name> before reporting. Shown in every dev/general worker's prompt; required checks are enforced (the worker is reminded, and the digest warns if one was not run or failed). serial checks hold a project-wide lock, so a tool that cannot run twice at once (a Unity editor, a Gradle daemon) is safe with parallel workers: do not forbid workers from running it. A Unity project with no toolbox set gets built-in unity-editmode (required) and unity-playmode checks. action: show | set (upsert checks) | remove (names) | reset (drop all, back to the built-in default).

One line per worker: state, elapsed, last action, files touched.

Compact report of one worker: files changed, commands, errors, final message, git status. detail "full" adds recent raw events. Treat claims as unverified; check with pi_diff.

Wait until the given workers (default: all running) finish or timeoutSec passes (max 60 per call; call again to keep waiting), then return their digests.

Send a follow-up message to a finished worker. Its Pi process (or, after 30 idle minutes, its saved session) keeps everything it already read, so use this for fixes and for the next task in the same area instead of re-spawning (no re-reading, no re-learning the layout). Send only the new instruction; it already has the background. The worker must not be running.

git diff (with stat) of a worker's directory so you can review the real changes.

Abort a running worker and stop its Pi process.

Commit a finished worker's worktree changes and merge its branch (--no-ff) into the repo it was spawned from. Refuses while the worker runs or the main tree has tracked changes; aborts and reports on conflict.

Remove a finished worker's worktree and branch and forget the worker. Refuses if it has unmerged work unless force is true.

Show or change the model every Pi worker uses

Show or change the reasoning effort every Pi worker uses