Register intent to edit code BEFORE editing. Returns any collisions with other active agents. Call this first, always. Also pass reads: the declarations you consulted to write this — the functions, methods or types you are calling or implementing against. Tower tells you when one of them changes under you, and shows the old and new signature so you can patch your call sites without re-reading the file. If you are changing a symbol's signature, set declares on it to the new declaration as it will read in the source, up to the body (e.g. function verify(token: string, opts: Opts): boolean): agents whose work reads it are handed that contract immediately and can code against it in parallel instead of waiting for you. If the claim is refused, the response carries alternatives: what to avoid, and advice. Don't stop — work outside avoid; Tower messages you when it frees up.

Check for collisions without registering a claim (a dry run).

Keep an active claim alive; claims auto-expire without heartbeats. Returns invalidations: declarations your work read that have moved, or that another agent has declared it is about to change — with the new signature, so you can adapt before it lands.

Release a claim after committing (optionally record the commit sha). Pass symbols with their final sig/sigText and every agent whose work reads a changed declaration is told what it landed as, and whether it matches what you declared.

Abandon a claim without committing.

List claims, optionally filtered by repo/branch/status.

Record an architecture decision and WHY it was made, for the team's shared memory.

Recall past architecture decisions before acting.

Ask the sequencer for a task whose module is safe to start right now.

Send an async message or task to another agent (toAgentId, or '*' to broadcast). Delivered on their next Tower contact. kind 'task' delegates work; reply with kind 'task_update' (replyTo=the task id) when done.

Read your inbox (marks messages read). Call whenever claim_intent reports unreadMessages > 0.

Read-only count of unread messages + open tasks waiting for you. Marks nothing read — the interactive nudge; if it returns > 0, call fetch_messages / list_tasks.

Claim a delegated task before working on it (first accept wins — prevents two agents doing the same work).

Finish an accepted task: success or failure, with the result and optional commit sha / PR url. Auto-notifies the delegator.

List delegated tasks by repo/status/recipient/assignee (the worker's poll).

Park a task for human approval before running it (remote-approve worker mode) — a person approves it from the board/phone.

Approve or reject a parked task (used by the board; also callable by tools).

Announce that this worker is online and ready to run tasks (call it every poll so the board shows live presence).

BEFORE you research or write anything: say in plain English what you plan to work on. Returns anyone already doing the same work, matched on meaning rather than file paths — so it catches a duplicate even when you would have picked a different filename. One call per task; it is the cheapest check Tower offers and the only one that fires before your tokens are spent.

Record declarations you just read, so your next claim_intent knows what your work is built on. Returns conflicts straight away when another agent is changing what you read — the cheapest moment to find out, before you plan against it. Normally called by the PostToolUse hook, not by you.