Agmsg

<!-- agmsg:render-root -->

Step 0: First-run bootstrap

agmsg keeps its SQLite database, team registry, and runtime state under ~/.agents/skills/agmsg/. The ./install.sh install path creates that tree; the Claude Code plugin install path does not (the plugin marketplace only copies this repository into ~/.claude/plugins/cache/). Before any other command, bootstrap if needed:

if [ ! -d ~/.agents/skills/agmsg ]; then
  # Newest cached copy of the plugin. Several versions can sit side by side, so
  # pick by version folder name (numeric, portable -- not sort -V, not mtime).
  cache="$HOME/.claude/plugins/cache/fujibee-agmsg/agmsg"
  newest=$(ls "$cache" 2>/dev/null | sort -t. -k1,1n -k2,2n -k3,3n | tail -1)
  installer="$cache/$newest/install.sh"
  if [ -n "$newest" ] && [ -f "$installer" ]; then
    bash "$installer" --cmd agmsg
  else
    echo "agmsg not installed. Either:" >&2
    echo "  - run ./install.sh in the agmsg repo, or" >&2
    echo "  - install via /plugin marketplace add fujibee/agmsg && /plugin install agmsg@fujibee-agmsg" >&2
    exit 1
  fi
fi

Once ~/.agents/skills/agmsg/ exists this step does nothing, so it is safe to run every time.

Agent messaging command. IMPORTANT: Always use the provided scripts. NEVER directly read or edit config files, DB, or team data. There is NO register.sh — use join.sh to join a team.

Use agmsg, not the host agent's own inter-session messaging. Several agent CLIs ship a native way for one session to message another on the same machine (in Claude Code, the SendMessage / ListAgents tools over its peer-session list). While a project is on agmsg, route agent-to-agent messages through agmsg instead. A message sent natively does not exist as far as agmsg is concerned: it is absent from history.sh and the team's export, it never reaches a member on another machine through remote sync, it does not mark read or advance any cursor, and it cannot address a member whose CLI is a different type. Half the conversation living somewhere unrecorded is worse than either channel alone, and the gap is invisible until someone reads the history and finds a decision with no message behind it. The native channel stays fine for anything outside the team — a subagent you spawned for your own task, or a session that has not joined.

Shell requirement: All agmsg scripts are Bash scripts. Always execute them via bash, never via PowerShell or cmd directly. If your default shell is not Bash (e.g. PowerShell on Windows), wrap every command with bash -lc '...'. Example: bash -lc '~/.agents/skills/agmsg/scripts/send.sh myteam alice bob "hello"'. Do NOT construct DB paths manually — the scripts handle path resolution internally. If you need to redirect storage, use AGMSG_STORAGE_PATH (the supported override).

Identity

If you already know your AGENT and TEAMS from a previous /agmsg call in this session, skip to Execute below.

Otherwise, run: ~/.agents/skills/agmsg/scripts/whoami.sh "$(pwd)" claude-code

Four possible outputs:

A) Single identity: agent=<name> teams=<t1,t2,...> type=claude-code project=<path> → Remember AGENT and TEAMS, then go to Execute.

B) Multiple identities: multiple=true agents=<n1,n2,...> teams=<t1,t2,...> type=claude-code project=<path> → Ask the user which agent name to use for this session, then go to Execute.

C) Not in a team: not_joined=true available_teams=<t1,t2,...> (or available_teams=none) → Show the user the available teams from the output, then:

Before first-time setup, inspect the user's request. If they ask to join, import, or bring in a team that already exists on a server, do not call join.sh. Go directly to remote pull under Execute. First run ~/.agents/skills/agmsg/scripts/team-list.sh --json --scope all; if a same-named local team has binding_state none or disconnected, stop and ask the user how to proceed. After pull succeeds, return to Identity setup so the user can register a new local agent in the pulled team.

First-time setup required. Joining a team so this agent can send and receive messages.

• Team name: a group of agents that can message each other (available: <list from output>)
• Agent name: this agent's identity within the team

1. Ask: "Enter a team name (joins existing or creates new)". If that name is not in available_teams and ~/.agents/skills/agmsg/scripts/team-list.sh --json --scope all shows a team whose binding_state is not none, do not create it yet: ask whether to create a new local team or bring in the team of that name from a server (remote pull; after it succeeds, return to Identity setup). With no such team, create it locally without mentioning remote.
2. If the team name given already appears in available_teams, run ~/.agents/skills/agmsg/scripts/team.sh <team> to see the current roster (name, type, project) and note the names already in use. Look for a naming convention already in play (e.g. a shared base name with role and number suffixes (<base>-<role><n>), or names derived from the team name) and, when one exists, propose 2-3 unused names that extend it; otherwise propose 2-3 short, distinctive identity names (not a bare tool-type label like codex/cc). Either way, names must not collide with the roster. Then ask: "Enter a name for this agent (suggestions: <name1>, <name2>, <name3> — or type your own)". For a brand-new team, skip the roster check and just ask: "Enter a name for this agent".
3. You MUST use join.sh — run: ~/.agents/skills/agmsg/scripts/join.sh <team> <agent_name> claude-code "$(pwd)"
4. Show the result and explain:

Joined! You can now use /agmsg to check and send messages.

• /agmsg — check inbox
• /agmsg send <agent> <message> — send a message
• /agmsg team — list team members
• /agmsg history — message history

<!-- agmsg:render-overlay claude-code -->

5. REQUIRED — Do NOT skip this step. Ask the user to pick monitor, turn, both, or off delivery. Empty input means monitor. Run ~/.agents/skills/agmsg/scripts/delivery.sh set <mode> claude-code "$(pwd)" and follow the printed AGMSG-DIRECTIVE block.

6. Then check inbox for the newly joined team.

D) Suggestions for reuse: suggest=true agents=<n1,n2,...> teams=<t1,t2,...> type=claude-code project=<path> available_teams=<t1,t2,...> → No exact registration exists for this project, but there are same-type agent names registered elsewhere.

1. Show the suggested agent names to the user.
2. Ask whether to reuse one of those names or choose a new one.
3. Ask for the team name to join (existing or new). If that name is not in available_teams and ~/.agents/skills/agmsg/scripts/team-list.sh --json --scope all shows a team whose binding_state is not none, do not create it yet: ask whether to create a new local team or bring in the team of that name from a server (remote pull; after it succeeds, return to Identity setup). With no such team, create it locally without mentioning remote. If the team is in available_teams, run ~/.agents/skills/agmsg/scripts/team.sh <team> and check that the agent name you are about to use (reused or new) is not already in its roster.
4. Run: ~/.agents/skills/agmsg/scripts/join.sh <team> <agent_name> claude-code "$(pwd)"
5. Then continue with the normal post-join flow above.

Execute

Only use scripts in ~/.agents/skills/agmsg/scripts/ — do not read or modify files under teams/ or db/ directly. Treat the storage layout as internal: never construct a database path or invoke sqlite3 directly. The scripts resolve the active store, including AGMSG_STORAGE_PATH overrides.

Terminal/pane self-awareness. Asked about this session's own terminal, pane, or driver — or before using arrange, peek, or poke below — run where.sh (see the "where" argument below) first and answer from its terminal=/capabilities= fields. Never infer the driver from environment variables or a grep/ps guess: that is how a session under a real driver ends up reporting a false negative about its own placement, or claiming a capability or a whole driver does not exist when it does (#1171). Each driver's own operational detail lives in ~/.agents/skills/agmsg/scripts/drivers/terminals/<terminal>/README.md, named by where.sh's own terminal= field — never guessed at from a remembered syntax.

Asked about a teammate's placement or status, or what can be done to one, that is team.sh <team>'s question (see the "team" argument below), not something to infer from a stale memory of their last known pane. Act on a teammate with peek.sh/poke.sh/arrange.sh <team> <name> directly rather than guessing reachability first — its exit code says whether it worked and, if not, why (see the "peek"/"poke"/"arrange" arguments below).

If no arguments provided (DEFAULT action — always do this when the command is invoked without arguments):

1. IMMEDIATELY run inbox check for each TEAM: ~/.agents/skills/agmsg/scripts/inbox.sh $TEAM $AGENT
2. Do NOT ask the user what to do — just run the inbox check.
3. If there are messages, read and respond appropriately. To reply: ~/.agents/skills/agmsg/scripts/send.sh $TEAM $AGENT <to_agent> "<message>"

Ensure monitor is running first. In monitor or both mode, keep one persistent watch.sh task for this session. Switch that watcher when actas changes the active role.

If asked, in ordinary language and in either English or Japanese, to re-arm this session's own agmsg monitor (no fixed trigger word — read the request as it is phrased): invoke Monitor with the standard command and description for this seat, and say nothing else.

If asked to re-arm every Claude Code seat in the team together (not just this session's own — no dedicated command for this, do it through poke: #1321): run ~/.agents/skills/agmsg/scripts/team.sh <team> --json, select the rows whose type is claude-code and whose delivery is monitor or both, deduplicated by member name. Whole team — never narrow this to your own project. For each selected seat, in turn with a gap of a few seconds between seats (poking them all at once starts every seat's model turn in the same instant and risks rate limits): write a one-line message asking that seat to re-arm its own agmsg monitor — phrase it in whoever asked's own words, or something equivalent — to a file, then ~/.agents/skills/agmsg/scripts/poke.sh <team> <seat> --retries 5 --retry-delay 2 --backoff exponential --body-file <path>. A seat whose input box was still busy after every retry (poke exits 14) could not be reached this way; name it in your report rather than silently skipping it.

Claude Code commands may need permission and sandbox allowlists for ~/.agents/skills/agmsg/scripts/ and its writable db/, teams/, and run/ directories.

Permission prompts. Every command here runs through the Bash tool, so each call is gated by the permission system until the script directory is allowlisted. Without this the user is asked to confirm essentially every agmsg call. Add to ~/.claude/settings.json (or project-level .claude/settings.local.json):

{
  "permissions": {
    "allow": [
      "Bash(~/.agents/skills/agmsg/scripts/*)",
      "Bash(/Users/<you>/.agents/skills/agmsg/scripts/*)",
      "Bash(bash ~/.agents/skills/agmsg/scripts/*)",
      "Bash(bash /Users/<you>/.agents/skills/agmsg/scripts/*)"
    ]
  }
}

Four entries are needed because a rule matches the command string as written, and these scripts are invoked both as ~/... and as an absolute path, with or without an explicit bash prefix. Replace /Users/<you> with the user's home directory.

Sandbox compatibility. When Claude Code's sandbox is enabled, watch.sh (monitor mode) runs inside the sandbox and needs to write pidfiles and SQLite WAL files under ~/.agents/skills/agmsg/. If monitor mode fails with write/permission errors there, add an allowlist entry to ~/.claude/settings.json (or project-level .claude/settings.local.json):

{
  "sandbox": {
    "filesystem": {
      "allowWrite": [
        "~/.agents/skills/agmsg/"
      ]
    }
  }
}

The allowlist does not enable sandboxing by itself. Use /sandbox in Claude Code to choose a sandbox mode, or add "enabled": true alongside "filesystem" under "sandbox" to configure it in settings. The allowlist has no effect until sandboxing is enabled.

If argument is "history":

1. Run: ~/.agents/skills/agmsg/scripts/history.sh $TEAM $AGENT

If argument starts with "team list" (e.g. "team list", "team list --json", "team list --scope project"):

1. Run: ~/.agents/skills/agmsg/scripts/team-list.sh <the rest of the args after "team list", unchanged>
2. This is a distinct command from bare "team" below — check for "team list" FIRST so "list" is never mistaken for a team name.

If argument is "team" or "team --json":

1. For each TEAM, run: ~/.agents/skills/agmsg/scripts/team.sh $TEAM [--json], preserving the option when present. --json returns every observed field.
2. This is read-only. It reports each member's identity cells and whether they are consistent — including a member that does not answer at all — but writes nothing and pokes no one. A seat that is wrong or unresponsive is not something this command, or any other, repairs from the outside: typing into another seat's session to fix it is exactly the mistake that used to happen here, and it is gone on purpose, not replaced by another form of the same thing. A member repairs its own identity cells by running fix (below), from itself. A seat that cannot or will not do that gets despawned and restarted, or a person takes it — not patched over from another pane.

If argument starts with "send" (e.g. "send misaki check the server"):

1. Parse target agent and message from the arguments
2. Determine which team the target agent belongs to, then run: ~/.agents/skills/agmsg/scripts/send.sh $TEAM $AGENT <to_agent> "<message>"

If argument is "config":

1. Run: ~/.agents/skills/agmsg/scripts/config.sh show
2. Show the output to the user.

If argument starts with "config set" (e.g. "config set hook.check_interval 30"):

1. Parse key and value from the arguments.
2. Run: ~/.agents/skills/agmsg/scripts/config.sh set <key> <value>

If argument is "version":

1. Run: ~/.agents/skills/agmsg/scripts/version.sh
2. Show the output — the installed version (git-describe provenance recorded at install time).

If asked to update or upgrade agmsg:

1. Run: npx agmsg@latest install --update — add --cmd <name> when this machine has more than one agmsg install. It keeps the database and team configs and replaces the scripts.
2. Tell the user to restart running agent sessions so they pick up the new scripts.

If argument is "where" (e.g. asked to report this session's own pane or placement):

1. Run: ~/.agents/skills/agmsg/scripts/where.sh
2. Report exactly what it prints. Do not try to answer this by naming a terminal yourself or running any terminal-specific command directly — this call already asked every driver on this session's behalf.
3. resolved=true placement=<terminal>:<id> is a known pane; resolved=true placement=none is a GENUINE negative (this session's own terminal confirmed it has no addressable pane). resolved=false means placement could NOT be determined — reason names which terminal(s) were asked and why. Never report a resolved=false answer as "no pane" or "not attached to a pane"; those are different answers to different questions, and the difference is the entire point of this command (#1171).
4. where.sh's output also carries capabilities=<list> (#1082) — that resolved terminal's own manifest, space-separated, verbatim. Before using arrange, peek, or poke below, check that the verb is in this list; if it is not, report it as unavailable for this terminal (name the terminal) rather than attempting it and finding out from an exit code. If it IS listed, read that terminal's own file — ~/.agents/skills/agmsg/scripts/drivers/terminals/<terminal>/README.md — before reporting a peek/poke/arrange failure: exit-code meanings differ by driver, and that file, not this one, is where they live.

If argument starts with "actas" followed by an agent name (e.g. "actas alice"):

1. Parse the new role name. If none was given (e.g. bare "actas", or the user asks you to suggest one), run ~/.agents/skills/agmsg/scripts/team.sh <team> for each TEAM to see the current roster. Look for a naming convention already in play (e.g. a shared base name with role and number suffixes (<base>-<role><n>), or names derived from the team name) and, when one exists, propose 2-3 unused names that extend it; otherwise propose 2-3 short, distinctive identity names (not a bare tool-type label). Either way, names must not collide with the roster. Ask the user to pick one or type their own before continuing.
2. Run ~/.agents/skills/agmsg/scripts/identities.sh "$(pwd)" claude-code to see whether the role is already registered for this (project, type).
3. If the name does not appear in the output, join under the existing team. Read TEAMS from the in-session whoami state (it may be a single team or comma-separated). For a single team, run ~/.agents/skills/agmsg/scripts/join.sh <team> <name> claude-code "$(pwd)". For multiple teams, ask the user which team to join the new role into, then run join.sh for that team.
4. Pre-flight claim the actas exclusivity lock so this role isn't already owned by another live session: ~/.agents/skills/agmsg/scripts/actas-claim.sh "$(pwd)" claude-code <name> "$CLAUDE_CODE_SESSION_ID". Read the status= line of the output:
   • status=ok ...: proceed to step 5.
   • status=held team=<team> owner=<sid>: another live session currently owns <name> in <team>. Tell the user: "Cannot actas as <name> — it is held by session <sid> in team <team>. Run /agmsg drop <name> in that session first, then retry." Then abort — do NOT touch the running Monitor.
   • status=not_registered: shouldn't happen if step 3 ran; treat as an error.
5. Switch receive too — exclusive role mode. a. Find this session's agmsg watch Monitor task: a task in TaskList whose description begins with "agmsg inbox stream", or the task_id returned by a Monitor call you made earlier in this conversation. Not every environment lists Monitor tasks in TaskList, so an empty list does not mean none is running. b. If you have such a task_id: TaskStop it. c. If you have none (typical when /agmsg actas runs as the first command of a fresh session — SessionStart hasn't fired the Monitor directive yet, or you're invoking actas before the agent acted on it): skip TaskStop. Do NOT attempt TaskStop with a guessed or empty task_id — it will fail with "Invalid tool parameters" and confuse the flow. Starting a new watcher for this session in step d replaces any previous one. d. Run ~/.agents/skills/agmsg/scripts/delivery.sh status claude-code "$(pwd)" and read its first line.

   • mode: monitor or mode: both: invoke a fresh Monitor, regardless of whether step b or c applied:

     • command: ~/.agents/skills/agmsg/scripts/watch.sh $CLAUDE_CODE_SESSION_ID "$(pwd)" claude-code <name> --max-seconds=1750
     • description: agmsg inbox stream (acting as <name>)
     • persistent: true
     • timeout_ms: 1800000

     This watch renews itself. A little before the 30-minute cap it prints one line on its own and exits: on agmsg watch: re-arm - ..., invoke Monitor again with exactly the command and description that line names (persistent: true, timeout_ms: 1800000), silently — no message to the user, no "re-armed", no acknowledgement, no summary, since announcing it every 30 minutes wastes tokens for no benefit; on agmsg watch: stopping - ..., do not re-arm it. If the watch is instead killed at the cap and no such line arrived (an agmsg install from before this), re-arm it only when the expiry notification says it delivered something.

   • mode: turn: leave it stopped, silently. has_st=1 is the one case delivery.sh can actually confirm was a deliberate choice — someone configured turn-based delivery for this project — so actas starting nothing here needs no explanation.

   • mode: off (no agmsg delivery hooks installed for this project): leave it stopped (actas must not start automatic delivery a project wasn't configured for), but do not treat this as silently deliberate. delivery.sh cannot tell whether someone ran mode off here or this project was simply never configured — both leave the exact same settings file (#687 review round 3). Tell the user — e.g. "agmsg delivery hooks are not installed for this project; automatic delivery remains stopped. Run /agmsg mode <choice> if you want to configure it." Keep it matter-of-fact, not a warning. Do not report actas as complete without saying this.

   • mode: off (unrecognized: ...): leave it stopped too (same rule — do not guess a mode), but this is a stronger case than the no-hooks-installed one above: delivery.sh could not even find or read a settings file for this project, most often because the working directory does not match how the project was actually registered. Tell the user explicitly — e.g. "agmsg could not find a delivery configuration for this project at <path from the message> — delivery is stopped, but this may mean the project isn't registered here rather than that it was deliberately turned off. Check the path, or run /agmsg mode <choice> to configure it explicitly." Do not report actas as complete without saying this — a silent stop here is indistinguishable from the other off cases and is what let this go unnoticed before (#687). The 4th argument to watch.sh restricts the subscription to messages addressed to <name> only — other roles' inbound messages stop reaching this session until another actas or session end.

6. Set the session's active FROM to <name> — use <name> in every send.sh call for the rest of this session.
7. Tell the user: "Now acting as <name>. Sends use <name> as from; receive restricted to <name> only."
8. Only if this session was NOT launched via spawn — check the environment variable AGMSG_SPAWNED (e.g. printenv AGMSG_SPAWNED): spawn exports AGMSG_SPAWNED=1 and already named the session <team>-<agent> via -n, so when it is set, skip this tip entirely. When it is UNSET (a human typed claude then actas'd, so the session has no convention name), additionally suggest to the user: "Tip: rename this session to <team>-<name> with /rename <team>-<name> so it's easy to find in the /resume picker and stays labeled after a restart." /rename is a user-typed slash command — you cannot invoke it yourself, so only suggest it.
9. Confirm the Monitor actually attached — only when step 5d invoked a fresh Monitor (mode: monitor or mode: both): confirm the Monitor call from step 5d started (it returned a task id rather than an error). TaskList may list this task, but not every environment does (the desktop app's Code tab runs the Monitor and delivers its events without listing it), so a task missing from TaskList is not a failure: judge by the Monitor call starting and its events arriving. The background-task footer is not a reliable check either. If the Monitor call failed, retry the invocation from step 5d once. If it still fails after the retry, tell the user actas completed but delivery could not be confirmed as attached, and do not describe delivery as active. If argument starts with "drop" followed by an agent name (e.g. "drop alice"):
10. Parse the role name.
11. Run ~/.agents/skills/agmsg/scripts/reset.sh "$(pwd)" claude-code <name> "$CLAUDE_CODE_SESSION_ID" to remove only that role's registration for this project. If the role has no other registrations left, reset.sh also drops it from the team config. The 4th argument releases any actas exclusivity locks this session held on the role so peers can pick it up immediately (see #62).
12. If the session's active FROM was <name>, clear that state. Then: a. Find this session's agmsg watch Monitor task: a task in TaskList whose description begins with "agmsg inbox stream", or the task_id returned by a Monitor call you made earlier in this conversation. An empty TaskList does not mean none is running. b. If you have such a task_id: TaskStop it. c. If you have none: skip TaskStop. Do NOT attempt TaskStop with a guessed or empty task_id. d. Run ~/.agents/skills/agmsg/scripts/delivery.sh status claude-code "$(pwd)" and read its first line.

    • mode: monitor or mode: both: invoke a fresh Monitor with the default subscription (no actas name filter — receives every (team, agent) pair currently registered for this project that isn't held by another session):

      • command: ~/.agents/skills/agmsg/scripts/watch.sh $CLAUDE_CODE_SESSION_ID "$(pwd)" claude-code --max-seconds=1750
      • description: agmsg inbox stream
      • persistent: true
      • timeout_ms: 1800000

      This watch renews itself. A little before the 30-minute cap it prints one line on its own and exits: on agmsg watch: re-arm - ..., invoke Monitor again with exactly the command and description that line names (persistent: true, timeout_ms: 1800000), silently — no message to the user, no "re-armed", no acknowledgement, no summary, since announcing it every 30 minutes wastes tokens for no benefit; on agmsg watch: stopping - ..., do not re-arm it. If the watch is instead killed at the cap and no such line arrived (an agmsg install from before this), re-arm it only when the expiry notification says it delivered something.

    • mode: turn: leave it stopped, silently — the one case delivery.sh can confirm was deliberate.

    • mode: off (no agmsg delivery hooks installed for this project): leave it stopped, but say so — same reasoning as the actas step this mirrors: this state is indistinguishable from "never configured" (#687 review round 3), so do not report it as deliberate. Do not report the drop as complete without mentioning it.

    • mode: off (unrecognized: ...): leave it stopped, but say so with the stronger diagnostic — same reasoning as the actas step this mirrors (#687). Do not report the drop as complete without mentioning it.

13. Tell the user: "Dropped role <name> from this project." If argument starts with "spawn" (e.g. "spawn codex reviewer", "spawn claude-code alice --window"):
14. Parse <type> (a spawnable agent type), <name>, and any options (--boot-prompt <text>, --project <path>, --team <team>, --window, --split h|v, --terminal <template>, --no-wait, --ready-timeout <secs>, --model <id>, --fresh).
15. Run: ~/.agents/skills/agmsg/scripts/spawn.sh <type> <name> --project "$(pwd)" [options]
    • spawn.sh pre-joins <name>, then opens a new pane or window through the terminal driver and launches the target CLI with /agmsg actas <name> as its initial prompt. --boot-prompt appends a first task to that prompt. Which terminal that is is the driver's decision, never something to name here.
    • By default it blocks until a spawned Claude Code agent's watcher attaches and prints status=ready; --no-wait returns immediately. A spawned Codex agent has no Monitor and skips the readiness wait.
    • It refuses early when <name> is already held by another live session, the target CLI is missing, the project path is invalid, or the terminal driver has nowhere to place it.
16. Show the script's output. Do not TaskStop or relaunch this session's own Monitor; spawn affects a separate agent.