amq-cli

name	amq-cli
version	0.52.3
description	Coordinate agents via the AMQ CLI for file-based inter-agent messaging. Use this skill whenever you need to send messages to another agent (codex, claude, or any named handle), check your inbox, drain queued messages, set up co-op mode between agents, join a swarm team, route messages across projects, or diagnose delivery issues. Also use it when you receive a message and need to know how to reply, inspect receipts, or handle priority. Covers any multi-agent coordination task where agents need to talk to each other — review requests, questions, status updates, decision threads, wake notifications, and orchestrator integration (Symphony, Kanban). For collaborative spec/design workflows specifically, prefer the /amq-spec skill which provides structured phase-by-phase guidance. Not intended for distributed systems design (RabbitMQ, Kafka), CI/CD pipelines, or single-agent tasks with no partner.
metadata	{"short-description":"Inter-agent messaging via AMQ CLI","compatibility":"claude-code, codex-cli, grok-cli"}

AMQ CLI Skill

File-based message queue for agent-to-agent coordination.

AMQ manages the conversation, not the task plan. Use it for messaging, routing, replies, and adapter-emitted lifecycle events; keep work decomposition and execution in the orchestrator above it.

Prerequisites

Requires amq binary in PATH. Install:

curl -fsSL https://raw.githubusercontent.com/avivsinai/agent-message-queue/main/scripts/install.sh | bash

Environment Rules

AMQ primarily uses AM_ROOT (which mailbox tree) and AM_ME (which agent). Pinned terminals also carry AM_BASE_ROOT plus an independent AM_SESSION identity; sessionless pins use the exact root as AM_BASE_ROOT and an empty AM_SESSION. Getting these wrong means messages go to the wrong place or silently disappear, so let the CLI handle them rather than guessing.

Inside coop exec — everything is pre-configured. Just run bare commands:

amq send --to codex --body "hello"     # correct
amq send --me claude --to codex ...    # wrong — --me overrides the env
./amq send ...                         # wrong — use amq from PATH

The reason: coop exec sets AM_ROOT, AM_ME, AM_BASE_ROOT, and AM_SESSION precisely for the session. Passing --me overrides the identity; for read-side sibling access, use --session <name> instead of overriding the raw root.

Outside coop exec — resolve the root from config, don't hardcode it:

amq_context="$(amq env --me claude)" && eval "$amq_context"  # reads .amqrc chain, replaces the full context
amq_context="$(amq env --session auth --me claude --export)" && eval "$amq_context"  # pin one session

# Or use an isolated subshell without polluting the parent shell:
(
  amq_context="$(amq env --me claude)" &&
  eval "$amq_context" &&
  amq send --to codex --body "hello"
)

Why not hardcode? The root path depends on project and explicit configuration, then context-sensitive implicit fallbacks. Hardcoding skips this and breaks when the project moves or config changes. Every shell-mode amq env invocation replaces the complete context. It emits AM_SESSION unconditionally (empty for a sessionless root), exports AM_BASE_ROOT as the authorized parent for named sessions or the exact root for a sessionless context. --export additionally prints a stderr pin note. Treat the evaluated output as one terminal, one session.

Global fallback: Orchestrator-spawned agents often start outside an AMQ-enabled repo where no project .amqrc or repo-local .agent-mail exists. Set AMQ_GLOBAL_ROOT or ~/.amqrc so amq env and amq doctor still resolve the correct queue. AMQ_GLOBAL_ROOT is explicit authority and therefore precedes repo-local auto-detection. The implicit home config is ineligible inside a Git worktree or bare repository. A Git worktree or bare repository with no eligible root refuses implicit ~/.amqrc fallback because it can silently select another project's mailbox. Participating commands keep that refusal. coop exec honors root precedence, then bootstraps a worktree-local queue at the Git top when no eligible root exists; coop exec --no-init refuses. coop init explicitly targets that local Git top. Bare repositories require a worktree or an explicit --root.

Session pitfall: coop exec defaults to --session collab (i.e., .agent-mail/collab). Outside coop exec, the base root is .agent-mail (no session suffix). These are different mailbox trees — don't mix them up.

Root Resolution Truth-Table

Context	Command	AM_ROOT resolves to
Outside coop exec	amq env --me claude	resolved base root from project .amqrc, AMQ_GLOBAL_ROOT, or an eligible implicit fallback
Git worktree or bare repository, no project .amqrc	amq env --me claude	AMQ_GLOBAL_ROOT when set, otherwise repo-local detected .agent-mail
Git worktree or bare repository, no eligible root	amq env --session auth --me claude	refuses implicit ~/.amqrc; requires a local or explicit root
Git worktree, no eligible root	amq coop exec claude	bootstraps <git-top>/.agent-mail/collab; never consults ~/.amqrc
Git worktree, no eligible root	amq coop exec --session auth claude	bootstraps <git-top>/.agent-mail/auth
Git worktree, no eligible root	amq coop exec --no-init claude	refuses and names amq coop init as the remedy
Bare repository, no eligible root	amq coop exec claude	refuses; use a worktree or explicit --root
Outside coop exec, isolated session	amq env --session auth --me claude	<resolved-base-root>/auth
Inside coop exec (no flags)	automatic	.agent-mail/collab (default session)
Inside coop exec --session X	automatic	.agent-mail/X

Canonical root precedence is:

explicit --root > AM_ROOT > project-local .amqrc > AMQ_GLOBAL_ROOT > implicit fallbacks

Inside a Git worktree or bare repository, the remaining eligible fallback is repo-local detected .agent-mail; outside Git, ~/.amqrc precedes detected .agent-mail.

An initialized cwd-local queue is also a routing safety signal. If an active pin points to another root, implicit participating commands refuse instead of silently following that pin. Repin to the cwd-local queue, route deliberately with --session/--project, or pass an explicit --root to confirm the active queue; ordinary pin checks still apply.

Git worktrees

A relative project root such as {"root":".agent-mail"} and auto-detected roots are intentionally per-worktree. Two terminals in different git worktrees can therefore use the same session name while reading different mailboxes. If a delivery receipt times out, run amq doctor --ops; it can warn when a peer has fresher presence in the same session under another worktree.

To share one mailbox across worktrees, use the same absolute root in each worktree's machine-local .amqrc, or remove the project-relative .amqrc and set AMQ_GLOBAL_ROOT to one absolute base. Keep the relative default when per-worktree isolation is intended. A Git worktree with neither local configuration nor a local queue fails closed instead of inheriting ~/.amqrc; this prevents accidental cross-project delivery.

Task Routing

Before diving in, match the task to the right workflow — this avoids wasted effort:

Your task	What to do
"spec", "design with", "collaborative spec"	Use /amq-spec instead — it has structured phase-by-phase guidance for parallel-research workflows.
Send a message, review request, question	Use amq send (see Messaging below)
Swarm / agent teams	Read references/swarm-mode.md, then use amq swarm
Received message with labels workflow:spec	Follow the spec skill protocol: do independent research first, then engage on the spec/<topic> thread — don't skip straight to implementation.

Quick Start

# One-time project setup
amq coop init

# Per-session (one command per terminal — defaults to --session collab)
amq coop exec claude -- --dangerously-skip-permissions  # Terminal 1
amq coop exec codex -- --dangerously-bypass-approvals-and-sandbox  # Terminal 2
amq coop exec grok  # Terminal 3 (optional peer) — caller flags forwarded unchanged, no baked-in bypass

Without --session or --root, coop exec defaults to --session collab.

Add --no-gitignore when coop exec should auto-initialize the project without changing .gitignore.

Standalone wake interrupt safety

Standalone wake keeps urgent interrupt notices and the bell without injecting Ctrl+C by default:

amq wake --me claude --interrupt-cmd none &

Swarm bridge events are hardcoded priority=normal plus label swarm, so do not bind that combination to Ctrl+C. Use ordinary non-destructive wake:

amq wake --me codex --interrupt-cmd none &

--interrupt-cmd ctrl-c sends a real SIGINT to the foreground process group and can interrupt or crash the agent. Use it only with a separate, operator-controlled label/priority when process-level interruption is intentional; the interrupt label alone never enables Ctrl+C.

Statusline (Claude Code)

To show the current AMQ session in your Claude Code status bar, add this snippet to your statusline script (e.g., ~/.claude/statusline.sh):

# AMQ session segment — try CLI first, fall back to env vars for older amq versions
amq_session=""
if _amq_out=$(amq env --session-name 2>/dev/null) && [ -n "$_amq_out" ]; then
    amq_session="$_amq_out"
elif [ -n "$AM_ROOT" ] && [ -n "$AM_BASE_ROOT" ] && [ "$AM_ROOT" != "$AM_BASE_ROOT" ]; then
    amq_session=$(basename "$AM_ROOT")
fi
if [ -n "$amq_session" ]; then
    output+=$(printf " | \033[33mamq:%s\033[0m" "$amq_session")
fi

amq env --session-name (v0.27+) prints the session name and exits 0 (empty when not in a session). The env-var fallback covers older versions. amq env --json also includes session_name.

To also set the terminal tab title (works in Ghostty, iTerm2, Terminal.app):

# Set tab title to "repo | amq:session" — re-asserts on each statusline refresh.
# Manual titles (e.g. Ghostty's prompt_tab_title) take priority and won't be overwritten.
tab_title="$repo_name"
[ -n "$amq_session" ] && tab_title+=" | amq:${amq_session}"
printf '\033]0;%s\007' "$tab_title" > /dev/tty 2>/dev/null

Integration & Ops Quick Reference

# Global fallback for orchestrator-spawned agents
export AMQ_GLOBAL_ROOT="$HOME/.agent-mail"

# Symphony hooks
amq integration symphony init --me codex
amq integration symphony emit --event after_run --me codex

# Cline Kanban bridge
amq integration kanban bridge --me codex
amq integration kanban bridge --me codex --workspace-id my-workspace

# Runtime diagnostics
amq doctor --ops
amq doctor --ops --json
amq doctor --root <exact-root> --ops
amq wake check --me <agent>
amq wake check --me <agent> --json

# Base-config-only session repair outside the current pin
amq doctor --root <session-root> --base-root <base-root> \
  --ignore-session-pin --fix-mailboxes

Exit Codes

Treat AMQ's process exit code as the stable machine contract:

Code	Meaning
0	Success. The command completed normally.
1	General error. The failure has no more specific exit-code classification.
2	Usage error. Arguments, flags, or command input are invalid.
3	Not found. A requested resource such as a mailbox, message, session, agent, or configuration does not exist.
4	Timeout. A watch, monitor, receipt wait, or delivery wait reached its deadline.
5	Context mismatch. A syntactically valid route was refused, including a pin conflict or an ineligible implicit root inside Git.

Do not parse stderr prose as a stable discriminator. --json preserves the same process exit codes. A read-only list on a mismatched session pin warns and continues; commands that consume or mutate mailbox state fail with code 5.

Delivery Receipts

AMQ records delivery outcomes in consumer-local receipt files. The main stages are:

• drained — a consumer successfully ingested the message
• dlq — the message was moved to the dead letter queue during ingest

Use these when you need confirmation rather than just fire-and-forget messaging:

# Block on delivery for a single-recipient send
amq send --to codex --body "please review" --wait-for drained --wait-timeout 60s

# Query receipt history later
amq receipts list --me codex --msg-id <msg_id>
amq receipts wait --me codex --msg-id <msg_id> --stage drained --timeout 60s

amq read, amq drain, and amq monitor all apply the same strict header validation. Messages in inbox/new that are corrupt or have malformed headers are moved to DLQ and produce a dlq receipt.

DLQ retries use four durable states: ready, pending, delivered, and indeterminate. A successful retry retains a terminal audit in dlq/cur until purge. delivered is idempotent and reports already_delivered plus audit_finalized; --force cannot redeliver it. A pending or legacy indeterminate envelope without a visible inbox destination refuses retry, including with --force; that flag bypasses only the maximum retry count. Bulk JSON separates retried, already_delivered, and skipped, and its count includes only newly retried messages.

amq who and amq doctor --ops report notifier_live only when the wake-lock inspector verifies a live amq wake process identity. That proves prompt notification, not message consumption. recent_activity means only that last_seen is fresh. Use drain or monitor when consumption is required; run long-lived wake/monitor commands under launchd, systemd, or another supervisor rather than treating AMQ itself as a daemon.

Before replacing a wake, run amq wake check --me <agent> --json. It is read-only and reports the running/current image path and version plus an exact next_action. An automated agent may act only when restart_capability=agent_safe. For operator_only, leave the live wake running and hand off to its owning terminal or supervisor. For unavailable, preserve the state and diagnose it. Never kill a live raw wake from a non-TTY process, and never accept an attention-only fallback as a replacement for full-strength input delivery.

Those consuming commands, watch, and all DLQ commands refuse a raw target that conflicts with a complete AM_BASE_ROOT/AM_SESSION pin before touching mailbox state. send and reply apply the same check to their source context. Use --session <name> for deliberate sibling access. The raw-root escape hatch, --ignore-session-pin, requires a non-empty explicit --root; it never blesses an inherited AM_ROOT. list warns and remains available for non-destructive inspection. With no session/tree evidence, scripts and CI remain fail-open. A missing mailbox is an error, not an empty inbox. Empty drain and list --new results may print a stderr note when the same handle has pending messages in a sibling session; follow the exact amq list --session <name> --me <handle> --new command in that note. This is an operational safety check, not an authorization boundary; a local process can deliberately repin or override it.

For doctor, --root selects the exact target but does not waive the active pin. Read-only inspection continues and reports a mismatch warning. --fix-mailboxes and --ops --fix-wake-locks require a matching pin unless an explicit non-empty --root is paired with --ignore-session-pin. --base-root requires --root, supplies retained config authority for the target or one direct child, and never waives the pin.

Session Layout

By default, the root is .agent-mail (from .amqrc or auto-detect). Use --session to create isolated subdirectories:

.agent-mail/              ← default root (configurable in `.amqrc`)
.agent-mail/auth/         ← isolated session (via --session auth)
.agent-mail/api/          ← isolated session (via --session api)

• amq coop exec claude → AM_ROOT=.agent-mail/collab (default session)
• amq coop exec --session auth claude → AM_ROOT=.agent-mail/auth

The main env vars are AM_ROOT (where) + AM_ME (who). coop exec also sets AM_BASE_ROOT for cross-session resolution and AM_SESSION as the independent session identity used by consuming-command guards. The CLI enforces correct routing — run bare commands for the current session or use --session for a named sibling. Default .agent-mail/<session> layouts are recognized even without .amqrc; custom root names still need config or explicit flags/env.

Cross-Project Routing

Send messages to agents in other projects via --project or inline @project:session syntax. Requires peer configuration in .amqrc.

When to use --session vs --project: --session = same project, different session. --project = different project. Change one dimension at a time.

Peer setup

Add project and peers to your .amqrc:

{
  "root": ".agent-mail",
  "project": "my-project",
  "peers": {
    "infra-lib": "/Users/me/projects/infra-lib/.agent-mail"
  }
}

Both projects must register each other as peers for round-trip messaging.

Use --project/--session to route, not a raw --root. A direct --root selects which tree to operate on; it carries no sender-origin metadata, so the recipient can't reply (a naive reply loops back into their own tree). amq send therefore refuses an explicit --root that crosses into a different base tree than your active session (AM_ROOT/AM_BASE_ROOT) when no --project/--session/--from-session is given. To message another project replyably, register the peer and use --project (or inline @project). If a send is genuinely local, set the target as your AM_ROOT instead of passing --root.

Sending cross-project

# Flag syntax
amq send --to codex --project infra-lib --body "hello from here"

# Inline syntax (terser)
amq send --to codex@infra-lib:collab --body "inline syntax"

# Same session name as source (default when --session omitted)
amq send --to codex --project infra-lib --body "delivers to same session"

Replies route automatically

When you receive a cross-project message, reply_project is set in the header. amq reply routes back automatically — no --project flag needed:

amq reply --id <msg_id> --body "got it"  # routes back via reply_project

Thread naming

• Same project P2P: p2p/claude__codex
• Cross-project P2P: p2p/projA:collab:claude__projB:collab:codex
• Topical (cross-project): use same thread ID across projects, e.g., decision/release-v0.24

For full details, see references/cross-project.md.

Cross-project identity (IMPORTANT)

When you receive a message where from matches your own handle (e.g., from: "claude" and you are claude), check from_project and reply_project. If either is present and names a different project, this is NOT an echo — it is a legitimate cross-project message from a different agent instance with the same handle. Process it normally.

AM_ROOT scoping after cross-project sends

After sending a cross-project message (via --project), your AM_ROOT still points to YOUR project. To send to your own partner (same project), use plain amq send --to codex — do NOT use --project. The --project flag is ONLY for sending to agents in OTHER projects.

Decision Threads

Decentralized decision protocol using existing AMQ primitives (no new CLI commands).

• Thread: decision/<topic>
• Kind: decision for all messages
• Labels: decision:proposal, decision:objection, decision:support, decision:final; plus project:<name> for cross-project decisions
• Context on proposals: {"proposal_id": "...", "question": "...", "options": [...], "required_projects": [...], "deadline": "..."}

Process: Propose → Review/Object → Resolve objections → Close when all required projects responded and no unresolved blocking objections.

amq send --to codex --project infra-lib --kind decision \
  --labels "decision:proposal,project:my-project,project:infra-lib" \
  --thread "decision/api-v2" \
  --context '{"proposal_id":"api-v2","question":"Adopt new API?","required_projects":["my-project","infra-lib"]}' \
  --body "Proposal: migrate to API v2. All tests green."

Session-Aware Routing

Users refer to sessions using many words: "session", "stream", "squad", "team", "workspace", "channel", or just a bare name. When the user mentions sending to or talking to an agent in a named context (e.g., "ask codex on stream1", "send to the auth team", "talk to codex in squad-api"), you must discover sessions before routing.

Important: Do not confuse sessions with projects. "Project" in AMQ means a different repo/codebase (cross-project routing via --project). Sessions are isolated mailbox trees within the same project (via --session). If the user says "the infra project", that likely means --project infra, not --session infra.

# Step 1: Discover active sessions and agents
amq who --json
# Returns: [{"name":"collab","agents":[...]},{"name":"stream1","agents":[...]},{"name":"auth","agents":[...]}]

# Step 2: Match the user's name against session names in the output, then send
amq send --to codex --session stream1 --body "Message for stream1"

Recognition patterns — any of these mean "route to a specific session":

• Explicit: "on stream1", "via auth", "in the api session", "the infra squad"
• Bare name: user just says "stream1" or "auth" — could be a session or an agent handle
• Colloquial: "team", "squad", "stream", "workspace", "channel" followed by a name

Note: The agent@name inline syntax (e.g., codex@infra) is for cross-project routing, not cross-session. For same-project session routing, always use --session <name> explicitly.

Rules:

1. When the user names something that could be a session, always run amq who --json first to check if it matches a known session name
2. If the name matches a session, use --session <name> on the send command
3. If it matches both a session and an agent handle, prefer the session interpretation when the user's phrasing implies a group/context ("on X", "in X", "the X team"), and the agent interpretation when it implies a person ("ask X", "tell X")
4. If the target session differs from your current session ($AM_ROOT basename), use --session <name>
5. Never guess — if the name doesn't appear in amq who --json output, tell the user (it may need coop exec --session <name> to initialize)
6. For cross-project routing (different repo), use --project instead — see Cross-Project Routing section

Messaging

amq send --to codex --body "Message"              # Send (uses AM_ROOT/AM_ME from env)
amq drain --include-body                          # Receive (one-shot, silent when empty)
amq drain --session auth --include-body           # Deliberate sibling-session receive
amq reply --id <msg_id> --body "Response"          # Reply in thread
amq watch --timeout 60s                           # Block until message arrives
amq list --new                                    # Peek without side effects
amq send --to grok --body "hello"                 # Grok is a normal peer handle, like codex or claude

Send with metadata

amq send --to codex --subject "Review" --kind review_request --body @file.md
amq send --to codex --priority urgent --kind question --body "Blocked on API"
amq send --to codex --labels "bug,parser" --context '{"paths": ["src/"]}' --body "Found issue"
echo "evidence: tests green" | amq send --to codex --subject "done" --body -   # - reads stdin

Body is fail-closed. --body - (or --body @-, or omitting --body) reads stdin; a literal string or @file is used as-is. A send whose resolved body is empty/whitespace is rejected with a usage error instead of delivering a blank message — so --body - with nothing piped fails loudly rather than shipping an empty body. Pass --allow-empty only when you truly want a blank body (subject carries everything).

Unrouted self-addressing is fail-closed. When --to resolves to your own handle and no --project, --session, or --from-session routing dimension is present, amq send refuses the ambiguous same-root send. Use routing to reach another instance of the same handle. Pass --allow-self only to confirm an intentional same-root self-send; it does not bypass cross-tree or session-pin guards.

Send file paths, not file contents. When attaching source code, configs, or large text for review, send the file path in the message body, not the contents inline. The receiver can open the file with their local tools. If the receiver cannot access that worktree, send a short diff instead of the full source.

Filter

amq list --new --priority urgent
amq list --new --from codex --kind review_request
amq list --new --label bug

Operator Gates

Almost all coordination is agent-to-agent. Occasionally the next required actor is a human: an approval, a manual test, a deploy only a person can run, or sign-off that a goal is complete. AMQ has no separate "gate" feature. You represent this structurally: address a message to the human's mailbox instead of describing the wait in prose to another agent.

The single invariant AMQ relies on here is recipient-as-next-actor: a message addressed to the human handle means a human is who must act next. Everything else below (the gate/<topic> thread name and the APPROVAL: / DONE: subject prefixes) is a naming convention that downstream tools like amq-noc watch for. AMQ routing and message classification do not special-case thread names or subject text; those are plain strings, useful only because humans and tooling agree to read them. They are conventions, not core AMQ semantics.

The human handle is user

By convention the human/operator mailbox is user. AMQ reserves this handle for validation in configured projects, so --to user is accepted wherever the project has a configured agent list. New co-op projects include user in the default agent set. For explicit amq init --agents ... projects or older roots, initialize the mailbox layout before relying on human drain/receipt/DLQ ergonomics:

# Seed the human mailbox alongside the agents (one-time, per project)
amq init --root .agent-mail --agents claude,codex,user
# or, for a coop project:
amq coop init --agents claude,codex,user

Throughout this section, user means "the conventional human handle." In configured projects it is warning-free for strict handle validation; in older or explicitly seeded roots, make sure the agents/user/ mailbox exists before expecting a human to drain and reply from it.

Raising a gate

Use a stable gate/<topic> thread so a gate and its resolution stay together:

# Approval / choice / manual test a human must perform
amq send --to user --thread gate/<topic> --kind question \
  --subject "APPROVAL: <decision>" \
  --body "<what you need a human to approve or run, and why>"

# Human closeout of a completed goal (sign-off that the goal is done)
amq send --to user --thread gate/<topic> --kind decision \
  --subject "DONE: <goal>" \
  --body "<what was completed; what the human should confirm or close>"

The human answers on the same thread from their own terminal or client, e.g. amq send --me user --to <agent> --thread gate/<topic> --kind answer --subject "APPROVED: <decision>" --body "<approval / answer text>" (use DENIED: or ANSWER: for a rejection or a plain answer). Reusing the gate/<topic> thread is what lets a watcher pair the answer with the open gate and clear it.

When NOT to raise a gate

Keep ordinary coordination agent-to-agent. Do not send to user for:

• FYIs and status updates -> status on a normal thread
• Acknowledgements
• Routine code review between agents -> review_request / review_response
• Agent-owned blockers (waiting on another agent, a build, or a flaky test)

If the escalation owner is a lead/CTO agent, they decide whether to escalate, and that decision is agent-to-agent. But once a human action is actually required, it still becomes a to:user gate so tooling can observe it.

Anti-pattern

Prose like operator-held, pending operator, or manual approval inside an agent-to-agent message is not a gate. It is body text a human or tool has to guess at. If a human must act, address the human.

What a gate is, and is not

• It is an observability / handoff signal, not authorization or security. AMQ sender identity is local convention, not authenticated approval. A to:user gate records that a human is the next actor; it does not grant permission or prove a human approved anything. Do not treat it as an access-control boundary.
• Cross-session / cross-project gates must be intentional. Default examples target the human mailbox in the current session/project. Routing a gate to another session's or project's user is a deliberate act (--session / --project), never the default. Gate-clearing is a consumer/orchestrator convention unless AMQ later adds explicit gate state.

Priority Handling

Priority	Action
urgent	Interrupt current work, respond now
normal	Add to TODOs, respond after current task
low	Batch for session end

Message Kinds

Kind	Reply Kind	Default Priority
review_request	review_response	normal
question	answer	normal
decision	—	normal
todo	—	normal
status	—	low
brainstorm	—	low

References

For detailed protocols, read the reference file FIRST, then follow its instructions:

• references/coop-mode.md — Co-op protocol: roles, phased flow, collaboration modes
• references/swarm-mode.md — Swarm mode: agent teams, bridge, task workflow
• references/integrations.md — Symphony + Kanban integration commands, global root fallback, ops checks
• references/message-format.md — Message format: frontmatter schema, field reference
• references/cross-project.md — Cross-project routing: peer config, addressing, decision threads
• references/review-loop.md — Token-efficient review cycles: delegate multi-round reviews to background agents