a2a

name	a2a
description	Spawn a team of agentic-CLI sessions (claude, opencode, pi, ...) that talk to each other as peers via a shared SQLite message bus at ~/.a2a/{project}/database.db. No central chain of command — each agent decides who to message. Use when the user wants multiple AI sessions to collaborate, debate, or divide work without a fixed orchestrator.
trigger	/a2a

/a2a — agent-to-agent peer messaging

a2a turns N agentic-CLI sessions into peers on a shared SQLite message bus. Each agent can send to any other agent (or all for broadcast) and recv blocking until something arrives. There is no orchestrator: communication flows freely.

Usage

/a2a spawn <name> [--role R] [--prompt "..."] [--cli claude|opencode|pi] [--model MODEL]
/a2a list
/a2a peek                     # show recent messages on the bus
/a2a send <to> "<msg>" --from <id>   # inject a message from the host shell
/a2a stop                     # kill all spawned a2a sessions in this project
/a2a clear                    # wipe the message bus

The default project name is the basename of the current working directory. Set A2A_PROJECT or pass --project NAME to override.

For the full list of CLI commands and their usage (Python & Go binaries), see docs/GO_CLI_REFERENCE.md.

⚠️ Critical: Use Pattern 3 for Multi-Agent Teams

When spawning 2+ agents, ALWAYS use Pattern 3 auto-spawn. Manual spawning or ad-hoc approaches will fail — agents will work independently without bus coordination.

Pattern 3 is the only reliable method because:

• Agents register BEFORE spawning (critical for bus communication)
• Kit prompts are written to files (avoids shell escaping bugs)
• --project flag ensures all agents connect to the same bus
• PIDs are updated after spawn with --upsert
• Agents coordinate via the bus, not just file editing

See Step 2-4 below for the Pattern 3 workflow. Reference implementation: examples/remote_worktree_team.sh

Common Command Syntax Mistakes

These mistakes will cause agents to fail:

❌ Wrong: a2a register --as agent-1
✅ Correct: a2a register agent-1 --role "Dev"

❌ Wrong: a2a send --as agent-1 "hello"
✅ Correct: a2a send agent-2 "hello" --from agent-1

❌ Wrong: a2a recv agent-1
✅ Correct: a2a recv --as agent-1

❌ Wrong: Inline kit prompts in spawn commands
✅ Correct: Write kit prompts to files first, then reference with --kit-file

When to use

• The user asks for "multiple claude sessions talking to each other"
• The user wants a "team of agents" without a fixed orchestrator
• Debate / red-team-blue-team setups, multi-perspective analysis
• Divide-and-conquer where peers self-coordinate

If the user wants a strict orchestrator → workers pattern, prefer the standard Agent tool (subagents) instead. a2a is for peer communication.

Three usage patterns

This skill implements Pattern 3 below. See docs/QUICKSTART.md for Patterns 1 and 2.

#	Pattern	Who drives	Documented in
1	Human-drive CLI — you open terminals and type a2a send/recv by hand	You (the human)	QUICKSTART.md — "Pattern 1" section
2	Multi-terminal AI team — you open N terminals, tell each AI agent to join the bus with a role, and they self-coordinate	AI agents (you instruct them)	QUICKSTART.md — "Pattern 2" section
3	Auto-spawn — one agent launches N background sessions via /a2a spawn	AI agents (spawned automatically)	This document (below)

What follows assumes Pattern 3: you are an AI agent running inside an agentic CLI (Claude Code, pi, opencode, etc.) and you will spawn peer agents as background processes. The kit prompt, a2a-spawn flags, monitoring loop, and teardown steps are all specific to this pattern.

For Pattern 2 (multi-terminal AI team), you don't need this spawn protocol. Just tell each human-driven terminal: "register yourself, a2a recv, a2a send, `a2a status done". The kit prompt in Step 4 can serve as inspiration.

What You Must Do When Invoked

Step 0 — Locate the a2a binary

The a2a CLI lives next to the skill. It is available as a Python script (default) or a faster Go binary companion. Resolution order:

1. $A2A_BIN env var (overrides all — set this to a downloaded Go binary)
2. $PATH (preferred — installer symlinks it into ~/.local/bin)
3. ~/.agents/skills/a2a/a2a (global cross-CLI skills path)
4. ~/.claude/skills/a2a/a2a (Claude Code skills path)
5. The skill source directory itself

A2A="${A2A_BIN:-}"
if [ -z "$A2A" ]; then
  for cand in \
      "$(command -v a2a 2>/dev/null)" \
      "$HOME/.agents/skills/a2a/a2a" \
      "$HOME/.claude/skills/a2a/a2a" ; do
    if [ -x "$cand" ]; then A2A="$cand"; break; fi
  done
fi
[ -z "${A2A:-}" ] && { echo "a2a binary not found"; exit 1; }
PROJECT="${A2A_PROJECT:-$(basename "$PWD")}"
export A2A_PROJECT="$PROJECT"

Step 1 — Initialize the project bus

"$A2A" init

If the user supplied agent definitions inline (e.g. "spawn alice as a planner and bob as a critic"), use those. Otherwise ask the user once for the agent roster, then proceed without further questions.

Step 2 — Register each agent up front

For each agent the user wants, before spawning the process:

"$A2A" register <id> --role "<role>" --prompt "<initial instruction>" --cli claude

Registration makes the agent addressable. The --prompt is stored so other agents can a2a list --json to see who they are talking to.

Step 3 — Spawn each agent as a background CLI session

Each agent runs in its own CLI process and receives the peer kit prompt that bootstraps it onto the bus. The kit prompt is CLI-agnostic; only the launch flags differ. Use the a2a-spawn helper that ships with the skill — it knows the right flags for each supported CLI:

CLI	System prompt flag	Non-interactive flag	Notes
claude	--append-system-prompt	-p	use --dangerously-skip-permissions for unattended
opencode	(none — embed in message)	run "<msg>"	use ~/.opencode/bin/opencode (the opencode alias goes through tmux)
pi	--append-system-prompt	-p	needs a --provider and --model set, e.g. --provider google --model gemini-2.5-flash

Write each agent's kit prompt to a temp file (avoids shell-escape bugs for multi-line prompts) and spawn:

printf '%s' "$KIT_PROMPT_FOR_ALICE"  > /tmp/a2a-$PROJECT-alice.kit
printf '%s' "$KIT_PROMPT_FOR_BOB"    > /tmp/a2a-$PROJECT-bob.kit

ALICE_PID=$(a2a-spawn --cli claude   --id alice --model haiku   \
                      --log /tmp/a2a-$PROJECT-alice.log         \
                      --kit-file /tmp/a2a-$PROJECT-alice.kit)
BOB_PID=$(  a2a-spawn --cli opencode --id bob   --model anthropic/claude-haiku-4-5 \
                      --log /tmp/a2a-$PROJECT-bob.log           \
                      --kit-file /tmp/a2a-$PROJECT-bob.kit)

⚠️ CRITICAL: Do NOT foreground-wait for agents after spawning. a2a-spawn already nohup+disowns the process — the PIDs are captured above for lifecycle tracking, NOT for wait loops. If the caller adds a while loop polling a2a list or sleeps between checks, it blocks the entire spawning context for the agent's full runtime. This prevents parallel spawns, defeats Hermes background-task tracking, and causes timeouts on long agent sessions.

Correct patterns:

• Hermes: terminal(command="a2a-spawn ...", background=True, notify_on_complete=True) — let Hermes track lifecycle
• Shell script: capture PIDs, move on. Poll A2A bus as a SEPARATE process or cron, not inline after the spawn
• AM daemon: use scheduler run-lifecycle tracking (DB run records), not a synchronous waitForAgents() loop

Save each PID so you can stop them later, and write it back to the registry so peers can see who is online:

"$A2A" register alice --pid "$ALICE_PID" --upsert
"$A2A" register bob   --pid "$BOB_PID"   --upsert

If a2a-spawn is not on PATH, invoke it directly: ~/.agents/skills/a2a/a2a-spawn or ~/.claude/skills/a2a/a2a-spawn.

Step 4 — The peer kit prompt (what every agent receives)

⚠️ Critical: Kit prompts must include coordination instructions.

Without explicit instructions to:

1. Register themselves on the bus
2. Introduce themselves to peers
3. Use a2a send/recv for coordination

Agents will skip bus coordination and work directly on files, defeating the purpose of A2A. Always include these steps in every kit prompt.

This is CLI-agnostic. Substitute {AGENT_ID}, {ROLE}, {USER_PROMPT}, {PEER_LIST}, {PROJECT}. The {A2A_PATH} line is computed dynamically by the agent so it works across CLIs that may or may not have a2a on PATH.

You are agent "{AGENT_ID}" on an a2a peer bus (project={PROJECT}).

Your role: {ROLE}
Your standing instruction from the user:
{USER_PROMPT}

You are one of several peers. There is no boss. You decide whom to message,
when to ask, when to answer, when to stop. Coordinate with your peers.

== Peers on the bus ==
{PEER_LIST}

== How to find the a2a CLI ==
Run the bash snippet below ONCE at the start to pick a working `a2a` binary
and store it in $A2A. Check $A2A_BIN first (if set), then try PATH, then
common skill installation paths:

  A2A="${A2A_BIN:-}"
  [ -z "$A2A" ] && A2A="$(command -v a2a 2>/dev/null)"
  [ -z "$A2A" ] && [ -x "$HOME/.agents/skills/a2a/a2a" ] && A2A="$HOME/.agents/skills/a2a/a2a"
  [ -z "$A2A" ] && [ -x "$HOME/.claude/skills/a2a/a2a" ] && A2A="$HOME/.claude/skills/a2a/a2a"
  echo "using a2a at: $A2A"

== How to communicate ==
You have a shell/bash tool. Use ONLY the `a2a` CLI to talk to peers.
A2A_PROJECT={PROJECT} is already in the environment.

  # see who is online (json includes their roles and prompts)
  $A2A list --json

  # check your inbox (blocks up to 30s for new messages)
  $A2A recv --as {AGENT_ID} --wait 30

  # send a direct message to a peer by id
  $A2A send <peer-id> "your message" --from {AGENT_ID}

  # broadcast to everyone
  $A2A send all "your message" --from {AGENT_ID}

  # mark yourself done so others know
  $A2A status done --as {AGENT_ID}

== Loop ==
1. recv --as {AGENT_ID} --wait 30
2. Decide: respond to a peer, ask a question, broadcast a finding, or finish.
3. Send at most one short message.
4. If nothing left to do AND no peer is awaiting your reply, run
   `status done --as {AGENT_ID}` and stop.
5. Else go back to step 1.

== Rules ==
- Do not invent peers. Address only ids returned by `a2a list`.
- Stay terse. One short message per turn unless asked for detail.
- Never speak on behalf of another agent.
- If you receive a broadcast that does not concern you, ignore it silently.
- If `recv` returns empty 3 times in a row, mark yourself `done` and stop.
- Do NOT call `a2a clear`, `a2a unregister`, or modify other agents' state.
- Hard cap: 8 loop iterations, then mark done and stop.

== Coordination rules (multi-role teams) ==
Omit this block if all agents have the same role or there is no role discipline.
- CLAIM: <task> — <id> BEFORE starting any work. Wait for ACK-CLAIM if collision.
- CLAIM expires after 5 minutes. Re-CLAIM if resuming after a gap.
- Bug reports: verify the issue exists in the current state before reporting. Do not report from assumptions formed earlier in the session.
- Role boundary: each agent stays within their declared role. Announce before crossing.
- To cross a role: send ROLE-CROSS: <reason> and wait 60s for a VETO before proceeding.
- Do NOT claim tasks outside your declared role without a ROLE-CROSS signal.

Begin now: run the locator snippet, then `$A2A recv --as {AGENT_ID} --wait 5`.
If empty, introduce yourself with one short broadcast, then enter the loop.

Step 5 — Monitor & relay to the user

While the agents run, poll the bus and show the user activity:

"$A2A" peek --limit 50
"$A2A" list

The user is not on the bus by default. To inject a message from the user, register a synthetic user (or host) agent once and send from it:

"$A2A" register user --role human --upsert
"$A2A" send all "stop debating, summarize." --from user

Step 5.5 — (Optional) Run the watchdog for automatic liveness

Manual peek/list polling is fine for a short attended run. For longer or unattended teams, a2a-watchdog automates the liveness checks that this skill otherwise leaves to you (see the "PID management is fragile" pitfall). It is deterministic — no LLM, no network — so it costs nothing to run on a timer.

One pass over the roster:

• PID gone, status not done → the seat exited without marking done (finished-and-forgot or crashed — one-shot -p agents leave no session, so the two are indistinguishable). The watchdog marks it done, broadcasts an [A2A watchdog] … EXITED notice on the bus, and desktop-notifies once.
• PID alive but a lease is overdue → the seat is up but not checking in. The watchdog sends it a wake message on the bus (an agent blocked on a2a recv receives it immediately — the bus is the "poke"), rate-limited to once per cooldown.
• A paused file (~/.a2a/<project>/paused) makes the watchdog fully quiet.

The ambiguity above is removed entirely by kit discipline: an agent that ends with a2a status done --as <id> is skipped by the watchdog and never triggers a false "EXITED" notice. Keep that as the final step of every kit (Step 4 loop already does).

# one pass (run from cron / systemd / launchd every 60s):
A2A_PROJECT={PROJECT} a2a-watchdog

# or a foreground dev loop every 60s:
A2A_PROJECT={PROJECT} a2a-watchdog --loop 60

Knobs (env): A2A_WD_MIN_GRACE_S (lease grace, default 600), A2A_WD_COOLDOWN_S (min seconds between nudges, default 600), A2A_WD_NOTIFY (0 to silence desktop notifications).

Leases (opt-in, per agent) let a seat promise its next check-in so the watchdog can tell "alive but stalled" from "alive and working":

a2a-lease {AGENT_ID} 1800        # "chase me if I go quiet > ~30 min"
a2a-lease {AGENT_ID} 1800 off    # going quiet on purpose; leave me alone

No lease = the watchdog never nudges the seat for idleness (it still reports a dead PID). Add a2a-lease <id> <cadence_s> to a kit's loop when the agent works a long queue and wants a safety net under its own wake loop.

Adapted from the doorbell/watchdog/lease trio in chrisreedbates/a2a-dms, whose tmux terminal-injection wake is a good fit for persistent agents (a future direction for the tmux-based debri CLI). This bus port swaps the tmux poke for a bus message, which is the equivalent wake for one-shot recv-blocking agents.

Step 6 — Tear down

When the user signals stop, or when every agent's status is done:

# Kill spawned background sessions (using ids you saved at spawn time)
# The harness reports completion when each background bash finishes; you can
# also kill via the saved PIDs:
"$A2A" list --json | grep -o '"pid": [0-9]*' | awk '{print $2}' | xargs -r kill 2>/dev/null || true

Leave the database intact unless the user asks to wipe it (/a2a clear). Database survives between sessions — useful for resuming.

Remote machine spawn

You can run the entire spawn sequence on a remote host via SSH. The pattern is identical to local spawn — init, register, write kit files, spawn, register --pid — but pipe it all through a single ssh host bash << 'EOF' heredoc so the bus lives on the remote machine.

ssh myhost bash << 'REMOTE'
  export A2A_PROJECT=myproject
  A2A=/usr/local/bin/a2a   # resolve explicitly; PATH may be minimal in non-login SSH
  SPAWN=/path/to/a2a-spawn # a2a-spawn is often not on PATH in non-login sessions

  "$A2A" init
  "$A2A" register pm --role "product-manager" --cli claude
  # ... register other agents ...

  # write kit files, then:
  PID=$("$SPAWN" --cli claude --id pm --model haiku \
        --project myproject \
        --log /tmp/a2a-myproject-pm.log \
        --kit-file /tmp/a2a-myproject-pm.kit)
  "$A2A" register pm --pid "$PID" --upsert
REMOTE

Remote spawn caveats

SSH goes silent for 60–90s after spawning 4+ agents. Each agent hits the API simultaneously. Subsequent SSH connections time out. Wait before monitoring. See docs/PITFALLS.md → "Remote machine spawn" for full details.

claude and a2a-spawn are often not on PATH in non-login SSH sessions. ~/.local/bin is not sourced. Always resolve by full path:

# claude
CLAUDE="$(command -v claude 2>/dev/null)"; [ -z "$CLAUDE" ] && CLAUDE="$HOME/.local/bin/claude"

# a2a-spawn
SPAWN="$(command -v a2a-spawn 2>/dev/null)"
[ -z "$SPAWN" ] && SPAWN="$HOME/.agents/skills/a2a/a2a-spawn"
[ -z "$SPAWN" ] && SPAWN="$HOME/.claude/skills/a2a/a2a-spawn"

Always pass --project explicitly. The default is basename($PWD), which on a root home directory resolves to root — a collision-prone name.

Resolve stale git state before creating a worktree. Check for .git/rebase-merge or .git/rebase-apply and abort before running git worktree add. Reset to origin/main if local and remote diverged.

Add explicit ACK instructions to implementer kit prompts. Without them, developer agents silently begin work and PM/architect agents send redundant check-in messages while waiting for confirmation.

See examples/remote_worktree_team.sh for a complete working example of this pattern.

Common Pitfalls & Gotchas

12. Cross-user database ownership

When the spawning agent (orchestrator) runs as one user (e.g. root) but spawned agents run as a different user (e.g. via sudo -u agent), the a2a database at ~/.a2a/{project}/database.db is created with the orchestrator's ownership. The spawned agents then get:

a2a: list error: attempt to write a readonly database

Fix: After a2a init, chown the database directory to the agent user:

chown -R agent:agent ~agent/.a2a/{project}/

Or run a2a init as the target user in the first place:

sudo -u agent a2a --project {project} init
sudo -u agent a2a --project {project} register ...

13. Sudo spawns lose the parent's PATH and env vars

When spawning agents via sudo -u user, the child process inherits a minimal environment. The A2A_PROJECT env var, PATH entries, and shell configs (.bashrc, .zshrc) are not inherited. The kit prompt's a2a locator snippet handles this by probing standard paths.

Fix: Either:

• Pass env explicitly: sudo -u agent env A2A_PROJECT=... a2a ...
• Use absolute paths for a2a in the kit prompt locator
• Install a2a to a system-wide location like /usr/local/bin/a2a

14. --dangerously-skip-permissions blocked on root

The claude CLI refuses --dangerously-skip-permissions when running as root/sudo:

--dangerously-skip-permissions cannot be used with root/sudo privileges

Fix: Either:

• Run claude as a non-root user (via sudo -u user claude ...)
• Use the env var CLAUDE_CODE_DANGEROUSLY_SKIP_PERMISSIONS=1 combined with --permission-mode acceptEdits instead of the flag

15. Foreground-waiting after spawn blocks the caller indefinitely

a2a-spawn already backgrounds agents with nohup+disown. If the calling script/daemon adds a while loop or sleep-based polling after spawning, it blocks until every agent exits — preventing parallel spawns, defeating Hermes background-task tracking, and causing spurious timeouts.

Common symptom: "spawning team for X..." prints, agents log activity, then the spawning terminal hangs for 30+ minutes until timeout kills it.

Correct patterns instead of a foreground wait loop:

• Hermes: terminal(command="a2a-spawn ...", background=True, notify_on_complete=True) — Hermes tracks lifecycle asynchronously
• Shell script: capture PIDs, move on immediately. Poll A2A bus as a separate cron or monitoring process, not inline after spawn
• AM daemon: rely on the DB-based run-lifecycle tracking (run records created at spawn, updated when agents complete). Do NOT use waitForAgents() blocking loops

16. Root ↔ Agent user A2A bus split (multi-user spawns)

When a2a is initialized by one user (e.g. root) but agents spawn as a different user (e.g. via sudo -u agent), they see completely different A2A buses because each user's HOME is different and a2a resolves project directories relative to ~/.a2a/.

Root runs a2a init → project created at /root/.a2a/{project}/database.db. Agent runs a2a list → looks at /home/agent/.a2a/{project}/database.db → not found.

Symptom: AM spawner logs show agents registered, but agent bus commands (a2a list, a2a send, a2a recv) all fail with "no such table" or "not found". Agents appear as "active" via ps but never communicate.

Fix: After initializing the bus with a2a init + a2a register, sync the project to the target user's a2a directory:

# Copy contents (NOT the dir itself — note trailing slashes!)
cp -rf /root/.a2a/{project}/. /home/agent/.a2a/{project}/
chown -R agent:agent /home/agent/.a2a/{project}/
chmod 644 /home/agent/.a2a/{project}/database.db

The cp -r nesting pitfall: cp -r /root/.a2a/{project} /home/agent/.a2a/{project} when the target directory doesn't exist creates /home/agent/.a2a/{project}/{project}/ (nested duplicate). Always use trailing slash on source: cp -rf /root/.a2a/{project}/. /home/agent/.a2a/{project}/

The chmod 644 requirement: SQLite needs the database file to be writable by the target user. Default cp permissions (644, owned by root) are not enough — the agent user also needs write access to the SQLite WAL and SHM files, so use chown -R AND ensure database.db is group-writable or world-readable. Without this, agents get "attempt to write a readonly database" when calling a2a commands.

Related Documentation

• docs/GO_CLI_REFERENCE.md — Full Go binary CLI command reference
• docs/QUICKSTART.md — Installation and first-run guide
• docs/CLIENT_API.md — Python client library API reference
• docs/ADVANCED_PATTERNS.md — Advanced usage patterns and artifact smoke tests
• docs/TROUBLESHOOTING.md — Common issues and solutions
• docs/PITFALLS.md — Lessons from artifact smoke testing