claw-dispatch

name	claw-dispatch
description	Use when dispatching work through the Hermes gateway with /claw, especially when the task may resolve slash commands or hand off into AO worker orchestration.

Claw Dispatch

Default behavior — AO workers first

/claw defaults to spawning AO workers directly. Hermes gateway is the fallback, not the default.

Input	Default action
PR number (#633, PR 633, 633)	Expand to draft-first readiness followed by /green, then post to Slack → Hermes (5 attempt cap)
General task description	Post to Slack → Hermes
--max-attempts N	Override attempt cap (default 5 for PR tasks)
--bidi prefix	Hermes interactive session (streaming)
--hermes prefix	Force through Hermes gateway
Slash command resolution (e.g. /green)	Resolve skill then pass to Hermes via Slack

Project auto-detection for PR tasks: resolve from current git remote (git remote get-url origin).

When to use Hermes instead of AO: only when the user explicitly says --hermes, uses --bidi for interactive output, or the task is explicitly a Hermes-native operation (routing config, gateway status, etc.).

Overview

/claw routes work to AO workers by default. Use:

• ~/.claude/commands/ao.md
• ~/.claude/skills/ao-operator-discipline/SKILL.md

User-specified AO parameters remain mandatory.

Requirements

• For AO tasks: ao session ls must not show ≥30 active sessions (spawn cap)
• For Hermes fallback: gateway must be healthy via hermes gateway status
• Slash command resolution must search .claude/commands first, then .claude/skills

Execution

When invoked with a task description:

TASK_DESCRIPTION="$ARGUMENTS"
set -euo pipefail

LOGDIR="/tmp/hermes"
mkdir -p "$LOGDIR"
chmod 700 "$LOGDIR" 2>/dev/null || true
STATUS_LOG="$(mktemp "$LOGDIR/.claw-status-XXXXXXXX")"
LAUNCHD_PLIST="$HOME/Library/LaunchAgents/ai.hermes.prod.plist"

if [ -f "$LAUNCHD_PLIST" ]; then
  while IFS='=' read -r key value; do
    [ -n "$key" ] || continue
    export "$key=$value"
  done < <(
    python3 - "$LAUNCHD_PLIST" <<'PY'
import plistlib
import sys

path = sys.argv[1]
try:
    with open(path, "rb") as fh:
        data = plistlib.load(fh)
except FileNotFoundError:
    raise SystemExit(0)

env = data.get("EnvironmentVariables") or {}
for key in ("HERMES_HOME", "HERMES_LOG_LEVEL"):
    value = env.get(key)
    if value:
        print(f"{key}={value}")
PY
  )
fi

export HERMES_HOME="${HERMES_HOME:-/.hermes}"
HERMES_CFG=

 [ ! -f  ]; 
   
   1

 ! python3 -c  2>/dev/null; 
   
   1















HERMES_PORT= 2>/dev/null)}$( )http://127.0.0.1:/healthHermes gateway is not reachable on :.(Override via CLAW_HERMES_PORT=<port> or  gateway.port  )statusokHermes gateway health check did not  ok.6976PR 6976#6976https://github.com//pull/PR 
  
  TASK_WITH_RESOLVED=
  TASK_DESCRIPTION=




PR_URL=$(   | python3 -c  2>/dev/null)
 [ -n  ]; 
  TASK_WITH_RESOLVED=


SLASH_CMD=$(   | python3 -c  2>/dev/null)

 [ -n  ]; 
  RESOLVED_CONTENT=
  RESOLVED_SOURCE=
   search_dir   ; 
     [ -f  ]; 
      RESOLVED_CONTENT=$(  2>/dev/null)
      RESOLVED_SOURCE=
      
    
  
   [ -z  ]; 
     search_dir   ; 
       [ -f  ]; 
        RESOLVED_CONTENT=$(  2>/dev/null)
        RESOLVED_SOURCE=
        
       [ -f  ]; 
        RESOLVED_CONTENT=$(  2>/dev/null)
        RESOLVED_SOURCE=
        
      
    
  
   [ -n  ]; 
     
    TASK_WITH_RESOLVED=
  



    | grep -q -- ; 
  CLAW_MAX_ATTEMPTS=$(   | grep -oE --  | awk  |  -1)
  TASK_WITH_RESOLVED=$(   | sed  | sed )
   CLAW_MAX_ATTEMPTS



BIDI_MODE=
CONTINUE_SESSION=

    | grep -q ; 
  BIDI_MODE=
  TASK_WITH_RESOLVED=$(   | sed )


    | grep -q ; 
  CONTINUE_SESSION=$(   | sed  | awk )



FORCE_HERMES=
    | grep -q ; 
  FORCE_HERMES=
  TASK_WITH_RESOLVED=$(   | sed )








 [  !=  ] && [  !=  ] && [ -z  ]; 
  TASK_WITH_RESOLVED=


 [  =  ]; 
   
  hermes chat \
    -q  \
    --yolo \
    --max-turns 30 \
    -- tool \
    --pass-session-id
   
   
   0


 [ -n  ]; 
   
  hermes chat \
    --  \
    --yolo \
    -- tool
   0



    | grep -q ; 
  STATUS_LOG=$(   | sed  | awk )
   [ ! -f  ]; 
     
     1
  
  LOG_LINES=$( -l <)
  LAST_LINE=$( -1  2>/dev/null)
  PID_RUNNING=$(pgrep -f  |  -1)
   
   
   
   
   
   -20 
   0








CLAW_CHANNEL=   
CLAW_BOT_ID=     
CLAW_TOKEN=
 [ -z  ]; 
   
   
   1




TASK_TEXT=$(   | sed -E )
MESSAGE=
MESSAGE=



POST_LOG=
curl -sS -m 15 -X POST \
  -H  \
  -H  \
  --data-binary  \
   > 2>&1


OK=$(python3 -c  2>/dev/null)
 [  !=  ]; 
  ERR=$(python3 -c  2>/dev/null)
   
   
   1



THREAD_URL=$(python3 -c )
MSG_TS=$(python3 -c )

 
 
 
 





 
ACKED=0
 i  $( 1 30); 
   1
  
  REACT=$(curl -sS -m 3 -H  \
     2>/dev/null)
  REACT_HIT=$(   | python3 -c  2>/dev/null)
   [  =  ]; 
     
    ACKED=1
    
  
  
  REPLIES=$(curl -sS -m 3 -H  \
     2>/dev/null)
  REPLY_HIT=$(   | python3 -c  2>/dev/null)
   [  =  ]; 
     
    ACKED=1
    
  

 [  !=  ]; 
   
   
   0  


 
 
 
 
 
 
 

Post-dispatch output (MANDATORY)

After every /claw dispatch — AO or Hermes — always print these lines in your reply:

AO spawn path

✅ AO worker spawned: <session-name>
   Attach:    ao attach <session-name>
   Status:    ao status <session-name>
   tmux:      tmux attach-session -t <container-id>-<session-name>
   Dashboard: http://localhost:3030

Get the session name from ao session ls --project <project> immediately after spawn. The dashboard URL is the live AO web UI (next.js on :3030 in prod, :3020 in dev).

Hermes chat path (fallback)

✅ Hermes worker dispatched (PID: <pid>)
   Log:     <logfile>
   Monitor: tail -f <logfile>
   Kill:    kill <pid>

Always emit the exact logfile path so the user can monitor or kill.

Slack path

✅ Dispatched to Hermes via Slack
   Channel:  #claw-dispatch
   Thread:   <thread_url>
   Hermes:   U0AEZC7RX1Q (@hermes)
   Monitor:  open the thread URL above; Hermes acks within ~30s, then works the task in thread
   Re-check: /claw --status <logfile>

Never omit these lines. The user must always know how to attach, monitor, or kill the worker without asking.

Notes

• Hermes replaced OpenClaw as the live gateway agent
• slash commands are resolved before dispatch
• /claw sessions are hidden from normal Hermes session lists via --source tool
• --yolo bypasses tool approvals for autonomous runs
• when /claw leads to AO dispatch, spawned AO sessions must still be verified post-spawn

Modes

Prefix	Behavior
(none)	Async fire-and-forget with ack detection (✅ ACK / ❌ RATE_LIMITED / ❌ FAILED)
--bidi	Synchronous streaming — blocks until Hermes finishes, shows session ID for follow-up
--continue <name>	Resume a previous --bidi session interactively
--status <logfile>	Check tail + completion status of any async claw session