tmpl/multi-agent-mux
1
0 Fork 0
Code Issues 9 Pull Requests Actions Packages Projects Releases Wiki Activity
Files
feceddf49fad8526dc7eaf87a1de006779b408bd
multi-agent-mux/.agents/skills/multi-agent-mux-create/SKILL.md
T

20 KiB
Raw Blame History

name, description, version, author, license, platforms, environments, metadata
name description version author license platforms environments metadata
multi-agent-mux-create Create a new agent session (claude, antigravity/agy) in a dedicated herdr session for context-preserving long-running work. Always creates a herdr session — never backgrounds with nohup/disown. Writes the new session to .mam/agent-sessions.yaml. Use when you want to start a fresh agent (no prior UUID) for a new project workspace. 4.8.4 godopu MIT
linux
macos
terminal
herdr
hermes
tags related_skills prereq_skills
agent
herdr
claude
antigravity
agy
multi-agent
context
session
multi-agent-mux-resume
multi-agent-mux-stop
multi-agent-mux-monitor
claude-code
claude-code

Multi-Agent Create — Start a Fresh Agent in a herdr Session

Companion skills: multi-agent-mux-resume (resume an existing UUID), multi-agent-mux-stop (terminate), multi-agent-mux-monitor (live status). Single source of truth: ./.mam/agent-sessions.yaml (this skill writes to it; never read it ad-hoc — go through this skill).

What this skill does

Spawn a new agent (claude or agy/antigravity-cli) in a dedicated herdr session for context-preserving long-running work. The herdr session is the container; the agent's session ID is data inside the container. This skill creates the container + starts the agent — but does not resume an old conversation (use multi-agent-mux-resume for that).

For all agents: the herdr session name is produced by lib.sh::derive_session_name — the single source of truth shared by create/resume/stop/status/monitor (P0-A). The rule (verbatim from the function):

slug = the two trailing path components of the absolute workspace, _→-, lowercased, joined with -; name = <slug>-creator-<agent>.

So $WORKSPACE_ROOT/landing_page/refer_landing_page + claude → landing-page-refer-landing-page-creator-claude. The workspace basename (refer_landing_page) is included; the hand-written historical entry that dropped it (lab-landing-page-creator-claude) was the bug, not the convention.

Pre-flight checks

create_session.sh checks the environment itself and exits 1 if anything is missing: herdr installed (has_real_herdr), the agent CLI in PATH, the agent's authentication, and the workspace directory. In your own scripts, test for herdr with lib.sh's has_real_herdr, NOT command -v herdr / type -P herdr: once lib.sh is sourced the former matches its herdr() function and the latter matches the .mam/shim wrapper, so both pass on a host with no herdr (B-3).

If a check fails, report the reason (automated path) or tell the user (interactive path). Do not proceed with a half-broken setup.

Standard names

  • herdr session name: derive_session_name <workspace> <agent> (lib.sh)
    • <workspace-slug> = basename $(dirname $WORKSPACE) - basename $WORKSPACE (lowercase, _→-)
    • examples: landing-page-refer-landing-page-creator-claude, paper-pdf2md-creator-agy
    • never re-derive this by hand — source lib.sh and call the function
  • wrapper script (claude only): ~/.local/bin/<workspace-slug>-creator-claude
    • contents: herdr new-session with claude inside; trust/bypass dialogs are handled by the skill for every agent (see --no-auto-trust below)
    • see <workdir>/agent_sessions.md for the canonical wrapper template

Herdr Session Isolation (격리 세션)

When running multiple agent sessions alongside other workflows (e.g., cmux, background workers, manual herdr sessions), sharing the default herdr server can lead to session name conflicts, monitoring clutter, and accidental destruction of user sessions via global commands.

To prevent this, you can run this skill inside an isolated herdr session using the HERDR_SESSION_NAME environment variable or the --herdr-session <name> flag (opt-in; alias: --herdr-server; legacy env alias: HERDR_SERVER_NAME). On success the script also starts the per-workspace monitor daemon through start_watchdog (reconcile.sh --subscribe --idle-timeout 0, one per workspace, see the lifecycle contract in multi-agent-mux-monitor/SKILL.md). Set MAM_NO_SUBSCRIBER_DAEMON=1 to skip that spawn when the monitor is supervised elsewhere or the workspace is ephemeral (Issue #19); the watchdog PID: line is then empty unless a daemon already runs for the workspace, whose PID is still reported.

Additionally, you can specify --herdr-workspace <name> (default: workspace slug without mam- prefix) to name and group the agent panes inside a dedicated Herdr workspace tab in the Herdr runtime (herdr workspace create --label / rename) as well as recording it in .mam/agent-sessions.yaml. Note that --herdr-workspace configures the workspace tab label within the session, whereas --herdr-session selects the daemon socket itself.

Under the hood this maps to a real, separate herdr session (herdr --session <name> — its own socket, its own agent list/workspace list, completely invisible to the default session and vice versa), not just a workspace label inside the same server. lib.sh's shim bootstraps the named session's server headlessly (herdr --session <name> server, backgrounded) the first time it's needed, and scopes every subsequent herdr call to it automatically — this headless bootstrap is what lets it work even when the skill itself is running from inside another herdr-managed pane (a plain interactive herdr --session <name> launch is blocked there by herdr's "nested herdr is disabled" guard; headless server mode isn't).

The bootstrapped server receives a sanitized copy of the caller's environment (Issue #9): per-session agent runtime markers — CLAUDECODE, CLAUDE_CODE_CHILD_SESSION, CLAUDE_CODE_SESSION_ID, the Claude Code messaging/bridge/pid markers, and the ANTIGRAVITY_CONVERSATION_ID / HERMES_SESSION_ID / GROK_SESSION_ID / OPENCODE_SESSION_ID equivalents — are dropped, because herdr hands its environment to every pane it spawns and an inherited CLAUDE_CODE_CHILD_SESSION silently turns transcript saving off in every sub-agent claude (no <uuid>.jsonl, so multi-agent-mux-resume can never find it). User configuration such as CLAUDE_PROJECT_DIR or CLAUDE_CONFIG_DIR is inherited unchanged. Query-only callers set MAM_NO_AUTOSTART=1 to skip the bootstrap entirely (status.sh and reconcile.sh do). A server bootstrapped before this fix keeps its polluted environment until it is restarted (herdr session stop <name>, then create/resume again).

How to use

  1. Via Environment Variable:

    export HERDR_SESSION_NAME=multi-agent-canary
# All subsequent commands (create, status, stop, etc.) will run in the isolated 'multi-agent-canary' herdr session.

  2. Via Option Flag:

    bash scripts/create_session.sh --workspace /path/to/project --agent claude --role developer --herdr-session multi-agent-canary --herdr-workspace my-project

  3. Submit Job Integration: You can automatically register a delegated job with a prompt when creating a session:

    bash scripts/create_session.sh --workspace /path/to/project --agent claude --role developer --submit-job "Task prompt here"

  4. Onboard Integration: You can automatically submit a project alignment/orientation job to the new agent when creating a session:

    bash scripts/create_session.sh --workspace /path/to/project --agent claude --role developer --onboard

    Delivery is judged by the agent's own started event (#23), within the job's ack_timeout_sec (default 180 s) — not by what the pane looked like two seconds after the keys were sent. The script no longer publishes started on the agent's behalf, so a job that shows started really was picked up. A send_keys rc 4 ("Enter not seen accepted", where a still-starting TUI lands) is a warning; rc 1/2/3 still abort at once. No started within the ack timeout marks the job error and warns, but keeps the session: the pane is up and the row is valid, only the handshake never landed, and killing a live agent over a slow onboarding would destroy more than it protects (the same rule this script applies to a TUI that never signalled ready). Re-send with multi-agent-mux-delegate-job submit once the agent is up; the prompt is never re-injected automatically. MAM_ONBOARD_ACK_S overrides how long that wait is — distinct from the wrapper's MAM_INJECT_ACK_GRACE_S.

    Start-up dialogs are answered for every agent, not only claude (#24). The blocking-modal signatures and their accept keys live in each agent adapter (lib_py/agents/adapters/*.py, published as MAM_STARTUP_DIALOGS), and they are answered while waiting for the TUI rather than after it. --no-auto-trust (or MAM_AUTO_TRUST=0 in the environment; an explicit flag wins over the environment) leaves directory-trust dialogs for a human to accept and still answers other blocking notices (for example codex's update prompt, where Skip changes nothing). Trust matters most for codex: trusting a directory makes it load that project's local config, hooks and exec policy, so --no-auto-trust is the switch to reach for when the workspace is not yours. Re-sending is rate-limited by MAM_DIALOG_RESEND_S (default 5 s) because a --source visible snapshot keeps showing a dialog for a while after it was accepted.

    Exit codes: 42 = the TUI never became ready and the process is gone; 43 = the rollback could not kill the herdr session and it is still running. On 43 the message names the session and the manual cleanup command; the YAML row keeps status: running with last_visible_status: rollback-failed and no delegate_job_id, and the onboarding job is closed as cancelled.

Why use --herdr-session?

  • By default, all skills target default herdr session socket — fine for single-workspace use.
  • By using an isolated session via HERDR_SESSION_NAME (or --herdr-session), your agent sessions are completely separated from your default user workspace, ensuring 0% interference — this is backed by a genuinely separate herdr session/socket, not merely a workspace label.
  • To deliberately tear down an entire isolated group at once (all its workspaces and agents), use herdr session stop <HERDR_SESSION_NAME> followed by herdr session delete <HERDR_SESSION_NAME> — this only affects that named session, never the default one.
  • An isolated server MAM starts does not relaunch saved agents when it restarts (#45): it runs with HERDR_CONFIG_PATH=<your herdr config dir>/mam-server-config.toml, your herdr config with [session] resume_agents_on_restore = false. Its saved panes come back as empty shells; resume closes the one of the session it resumes.
  • That file is rebuilt each time MAM starts the server, so a change to your herdr config reaches a MAM server only when it is next started; herdr server reload-config rereads the generated file. A server you start yourself (herdr session attach) does not get it.

Output format

When invoked, the script creates or updates ./.mam/agent-sessions.yaml with:

herdr_sessions:
- name: <workspace>-creator-<agent>           # E.g. landing-page-creator-claude
  status: running                             # Initial status for a freshly created agent
  role: Developer
  herdr_session_created_at: '2026-08-04T12:00:00Z'
  herdr_session_epoch: 1785844800
  herdr_session: <HERDR_SESSION_NAME>         # Isolated session name (default: 'mam-<ws-slug>')
  delegate_job_id: null
  pane:
    index: 0
    pid: 12345
    cmd: claude
    cmd_full: claude --dangerously-skip-permissions
    cwd: /path/to/project
  start_command: "HERDR_SESSION_NAME=<herdr_session> herdr new-session -d -s <SESSION_NAME> -x 140 -y 40 -c <WORKSPACE> <CMD_FULL>"
  attach_command: "HERDR_SESSION_NAME=<herdr_session> herdr agent attach <SESSION_NAME>"
  kill_command: "HERDR_SESSION_NAME=<herdr_session> herdr kill-session -t <SESSION_NAME>"

The three commands require source .agents/skills/lib.sh first — new-session/kill-session are tmux-compat pseudo-commands the shim translates, and HERDR_SESSION_NAME is what the shim reads to route to the right isolated herdr session (real herdr has no env-var-based scoping of its own; herdr_session: default needs no prefix at all).

Workflow

bash .agents/skills/multi-agent-mux-create/scripts/create_session.sh --workspace <path> --agent <agent> --role <role>

create_session.sh does the whole flow: pre-flight, the session name (derive_session_name), the agent command (adapter spawn-spec), spawn, start-up dialogs, waiting for the TUI, pane capture, the YAML row, and the optional job submission.

Option Meaning
--workspace Required. Project directory
--agent Required. claude, agy, hermes, grok, opencode, codex
--role Required. Assigned role
--session (--name) herdr session name (default: derive_session_name)
--wrapper claude only: use the ~/.local/bin/<session> wrapper even if it is not present
--dry-run Print what would be spawned; no spawn and no registry (YAML/DB) update. The checks before it still run, and on macOS they remove the quarantine attribute from the agent binary
--herdr-session (--herdr-server) Isolated herdr server name
--herdr-workspace herdr workspace label recorded in the registry (a label only, never a socket)
--submit-job PROMPT Register a delegated job with this prompt and send it to the new agent
--onboard / --no-onboard Submit the onboarding job (default: on)
--auto-trust / --no-auto-trust Answer a start-up directory-trust dialog or leave it to a human (MAM_AUTO_TRUST; default: answer)
--isolate / --no-isolate Legacy, no-op
Exit Meaning
0 Created; or --dry-run
1 Pre-flight failure, nothing was spawned: workspace not a directory, herdr not installed, agent CLI not in PATH, agent not authenticated. After spawn: the task brief could not be written or the instruction injection failed (the session is rolled back)
2 Bad arguments; control-plane binding refused for a worktree; whitespace in a control-plane path that must be forwarded
3 A herdr session of that name already exists (multi-agent-mux-resume, or multi-agent-mux-stop first)
4 The YAML already has a running row of that name. The YAML write raises it after spawn, and the herdr session is rolled back
42 The TUI never became ready and the process is gone, or it is alive and MAM_KILL_ON_READY_TIMEOUT=1 (rolled back)
43 The rollback could not kill the herdr session; it is still running
other A later step failed and its code is passed through. The EXIT trap (cleanup_herdr_on_error) rolls back the herdr session and closes the onboarding job as cancelled; if that kill fails, the exit is 43 instead

Registering the session in agent-sessions.yaml

create_session.sh writes the herdr_sessions[] row itself (the format is in Output format above); do not append it by hand. pane.cmd_full is the actual command line in the pane, decided by the adapter spawn-spec (lib_py/agents), not the resume command.

Issue worktrees: worktree_prepare.sh (Issue #21 item 4)

Run from the control root's skills tree (lib.sh's WORKSPACE_ROOT = the main checkout that owns .mam):

bash .agents/skills/multi-agent-mux-create/scripts/worktree_prepare.sh \
  --path ../multi-agent-mux-wt/issue-42 --branch fix/issue-42 [--base <ref>] [--submodules] \
  [--creator-agent claude [--session <name>] [--role creator] [--herdr-session <s>] [--herdr-workspace <w>] [--onboard]] [--dry-run]
  • Creates the worktree next to the control root (a path inside it is refused), idempotently; creates the worktree's own .venv from the control root's .venv/bin/python and installs the delegate-job skill's requirements.txt (paho-mqtt, PyYAML - what deploy/install.sh installs) plus requirements.txt / requirements-dev.txt when present (needs network access to the package index; a failed install removes the half-made venv and exits 1 so a re-run starts over; never a link - a leftover link is refused with exit 1; no control-root interpreter is a warning and VENV=skipped); adds /.venv and .pytest_cache/ to the shared info/exclude (a linked worktree's exclude file is $GIT_COMMON_DIR/info/exclude, so it applies to the main checkout too).
  • Submodules are opt-in (--submodules). Initialisation writes submodule.<name>.url to the shared .git/config, and any attempt - even a failed one - leaves a modules/ directory that makes a plain git worktree remove refuse. A failed --submodules undoes only what that attempt created - new submodule checkouts are deinitialised before their git dirs are deleted, so the worktree stays a usable checkout; submodules initialised earlier and pre-existing submodule.* values are kept - and exits 1; re-run without --submodules to finish.
  • --creator-agent creates the creator through this skill with the control root's binding and MAM_NO_SUBSCRIBER_DAEMON=1; the row's pane.cwd is the worktree.
  • Prints BASE=<sha>: record it before delegating the implementation, then review with run_loop.sh --worktree <WT> --base "$BASE" --review-only.
  • A submodule path that is not initialised yet but already holds files refuses --submodules up front (exit 3, nothing changed): the clone would fail on it and a rollback could not tell those files from its own.
  • Exit codes: 0 ok, 1 git / submodule / create failure, 2 usage or path rule, 3 refused to protect existing content (nothing changed).
  • create itself: a control-plane binding that cannot be computed is refused (exit 2) for a --workspace other than the control root and warned about for the control root; the immediate reconcile and the watchdog run the control root's code.

Pitfalls

  • Don't use nohup/disown/setsid for the agent itself — those background the agent outside herdr. The whole point of this skill is the herdr session is the supervisor. nohup is OK only for launching the wrapper (which itself creates the herdr session via herdr new-session -d).
  • Claude Session ID Auto-Assignment: Fresh claude sessions automatically generate a new UUID (mam_gen_uuid) passed via claude --session-id <uuid>. claude_session_id_own is stored in .mam/agent-sessions.yaml with session_id_source: assigned and session_id_verified: false.
  • First Message Materialization: The transcript file .jsonl is created on disk when the first user prompt is sent. The monitor loop (reconcile.sh) verifies the transcript and promotes session_id_verified: true and last_visible_status: pinned.
  • Wrapper script MUST NOT be created via hermes profile alias — that command writes a hermes -p <profile> wrapper that destroys the herdr behavior. Create wrappers manually (see lab-landing-page-creator-claude template).
  • Always use the workspace-relative path in herdr cwd — relative paths break when herdr respawns in a different shell context.

Verification

After spawn + YAML append:

# 1. herdr session is alive (real native command — no lib.sh needed)
herdr agent get "$SESSION_NAME" >/dev/null 2>&1 && echo OK || echo MISSING

# 2. pane has the expected cmd + cwd
herdr agent get "$SESSION_NAME" | python3 -c "
import sys, json
a = json.load(sys.stdin)['result']['agent']
print(f\"cmd={a['agent']} cwd={a['cwd']}\")
"

# 3. agent-sessions.yaml has the new entry
python3 -c "
import yaml
d = yaml.safe_load(open('.mam/agent-sessions.yaml'))
names = [s['name'] for s in d['herdr_sessions']]
assert '$SESSION_NAME' in names, 'session not registered'
print('OK:', names)
"

# 4. Optional: check the TUI status (real native command)
herdr agent read "$SESSION_NAME" --source visible --lines 20   # TUI ready = agent banner visible, no dialog text

herdr has-session / herdr list-panes / herdr capture-pane are tmux-compat pseudo-commands only understood after source .agents/skills/lib.sh — the real herdr binary has no such subcommands. The block above uses only the real herdr agent get/herdr agent read commands, so it works standalone.

When NOT to use this skill

  • Resuming an old conversation → multi-agent-mux-resume
  • Killing an existing session → multi-agent-mux-stop
  • Just attaching to an existing session → herdr agent attach <name> (no skill needed)
  • One-shot print mode (claude -p "...") → no herdr needed; use claude-code skill's print mode
Reference in new issue View Git Blame Copy Permalink