AMQ CLI
Use the amq CLI for all queue operations. Never edit Maildir files directly.
AMQ transports coordination; the caller or orchestrator still owns task planning,
worktrees, approvals, and delivery.
Start in the correct context
Use the first case that applies:
AM_MEis set (amq coop execoramq launchstarted the session): run bareamqcommands. Do not override--meor reconstruct the root.A Claude Code session with no
AM_ME, such as a chat of the desktop app: runamq coop joinonce; add--me "<handle>"when the user named a handle or the output says the default is taken. Later bareamqcommands in that session use the handle. Follow what the output says about the wake: when the wait hook is not installed, tell the user thatamq coop install-hooksinstalls it once per machine and that it takes effect in a new session. When a later command says the session has no handle (a terminal session after/clear), join again. Do not use theamq envexport here; a shell variable does not last between the shell calls of a Claude Code session. A Codex chat with noAM_ME(CODEX_THREAD_IDis set) joins the same way; its default handle iscodex. When the output says the wait hook is not installed, tell the user thatamq coop install-hooks --harness codexinstalls it, that they approve it in Codex, and that it takes effect in a new chat.Any other shell: resolve and export the complete context before reading or sending:
amq_context="$(amq env --me "<handle>" --export)" && eval "$amq_context" && amq drain --include-body
For a named session, add --session <name> to amq env. Treat the evaluated
context as one terminal bound to one session. Use --session or --project for
deliberate routing; do not point a participating command at another queue with a
raw --root.
Receive under a live wake
If session-start context says that a wake rings this terminal or that a hook
wakes this session, or amq wake check --me <handle> --json reports
live_wake: true with an injection mode other than none, a doorbell is
delivered for you.
Run amq drain --include-body when it fires, then act on the messages.
The default drain reads at most 20 messages in filename order, including held
mail. After a full batch, check amq list --new --limit 1 and continue with bounded drains
as needed; do not assume the inbox is empty. Use --limit 0 only for a known,
manageable batch. Urgent hold bypass does not change drain order.
Do not run amq watch, amq monitor, sleep-poll, or start another inbox
watcher under that live wake or wait hook. A blocking wait holds your turn while the
doorbell queues behind it. When your work is done, finish the turn; do not
keep a tool running or send idle check-ins just to wait for mail.
A wake can hold its first doorbell for routine mail with the hold_normal and
hold_low settings; 5m / 30m is an opt-in example. The configured retry
policy applies after the first attempt.
Urgent priority skips the hold. Ordinary urgent mail still passes through
debounce and input deferral; urgent mail with the configured interrupt label
follows the interrupt path.
Use --priority urgent on send or reply for time-critical verdicts and
unblocking requests. Other review responses remain normal by default.
By default the wake rings a plain doorbell (text and Enter, never held for a
dialog); amq coop exec prints dialog guard: plain or
dialog guard: proper (<harness>). In Claude Code,
amq wake install-dialog-hook makes the wake hold its doorbell while
AskUserQuestion, ExitPlanMode, or a permission dialog is open; the held
doorbell arrives after you answer, or when that turn ends. Codex needs
--harness codex and the entries trusted in /hooks.
Wake settings (hold, debounce, preview, bell, input deferral, interrupt notice)
apply to a running wake without a restart: amq wake config --me <handle>
shows them, --hold-normal 5m --wait sets one and waits for the wake to apply
it, --unset <key> falls back, and --machine sets this user's default.
Injection flags are fixed for a running wake. Exit codes and recovery are in
wake operations.
Without an injecting wake or a wait hook, use the receive methods in the
operations guide.
A notify-only wake (--inject-mode none) paired with a supervisor monitor
service is a separate supported setup.
Send useful handoffs
For one related batch, send one completion handoff with artifact paths, the relevant source or digest, checks, failed or pending parts, and the next action or actor. Send independent blockers promptly; send a partial result when the receiver can act on it before the rest finishes.
Send status when work starts, a state, decision, or result changes, or an agreed escalation is due. Avoid file-by-file completion notices and unchanged check-ins. Automated callers own duplicate suppression by item, revision, and event; AMQ threads correlate messages, but do not deduplicate repeated sends.
A reply to a todo defaults to --kind status; an explicit --kind todo
opens another request. Progress uses --kind status;
completion uses --kind status --labels done; cancellation uses
--kind decision --labels withdrawn. State the result or cancellation
honestly: withdrawing a request does not mean its feature shipped.
The last --refs entry is the direct reply target; a list of refs is lineage,
not a batch-close operation. Close independent requests separately.
The one exception is the requester's own requests to one owner in one thread:
send one reply of --kind decision --labels settled-through against the latest
of them and the pair's older open requests follow. amq thread settle --older-than <dur> composes exactly one such decision per (thread, owner).
amq send --follow-up 30m delivers one notice to the sender's own inbox if the
recipient has not replied to that message when the time passes; a live wake,
the wait hook, or the next drain, read, or monitor checks it, and any
reply from the recipient cancels it.
An unanswered message is a reply obligation, not evidence that work started. Use the host harness's task and turn records when activity matters. Keep declared task state, actual execution, approval waits, and delivery presence separate. A live wake, a recent message, or an assigned tracker item does not prove a task is running. Do not create work items from old coordination messages without confirming an actual requested outcome and remaining gap.
Route the task
| Need | Read or run |
|---|---|
| Send, reply, drain, inspect receipts, or diagnose | Use the operations guide. |
Two-agent research and design, or a received message labeled workflow:spec |
Use the amq-spec skill: do your own research before you read the partner's, then follow the spec/<topic> thread. |
| Co-op roles and phased collaboration | Read co-op mode. |
| Claude Code Agent Teams | Read swarm mode. |
| Cross-project delivery or decision threads | Read cross-project routing. |
| Grok Bot on a registered Mac | Read registered-machine execution. Select the machine before you resolve the project or session. |
Two separate host queues or amq-bridge |
Use companion amq-bridge, never a foreign --root. Read two-host fleets. |
| Symphony or Kanban adapters | Read integrations. |
| Message schema and kinds | Read message format. |
| Multi-round background review | Read review loop. |
Buzz, ACP, or amq-acp |
Companion amq-acp queues to AMQ_ACP_TO; pool workers must not drain. Chat must not pass --root, recipients, or argv, and [Context] is not routing. Read amq-acp. |
Attach to a running harness session (amq-remote up, submit, share) |
Read the amq-remote reference. |
Reach this session from Buzz (/amq-remote) |
Use the amq-remote skill. |
Use amq <command> --help for current flags. The repository README is the
canonical setup path.
Safety and delivery rules
- A wake notification is attention, not consumption proof. Use receipts or
--wait-for drainedwhen delivery proof matters. Adrainedreceipt proves ingestion, not that the requested work is complete. - Preserve the existing thread when replying. Drain again after a doorbell; the newest complete message body is authoritative.
- Cleanup is explicit through
amq cleanup. Do not add automatic deletion. - Before any wake mutation, run
amq wake check --me "<handle>" --json. Act only whenrestart_capability=agent_safe; otherwise preserve state and report the required operator action. - Keep cross-host payloads outside privileged inboxes until the configured bridge has authenticated and applied them. Never copy or remote-drain a foreign Maildir as a substitute for routing.
- Do not translate native Windows injection into Unix
wakeorcoop exec. Read the platform section in the operations guide when Windows is involved. - Dangerous provider bypass flags belong only on an operator-controlled direct
coop execcommand. Do not place them in committed launch configuration.
The full command catalog, environment precedence, setup flow, platform notes, operator gates, and examples are in the operations guide.