ais-orchestrator

name	ais-orchestrator
description	PRIMARY execution skill. Run, monitor, and orchestrate all agent sessions in tmux via ais CLI. Use ais for ALL task execution — spawning sub-agents, inspecting output, rotating accounts on rate limits, and collecting results. Everything runs in tmux.
metadata	{"openclaw":{"emoji":"🕸️","os":["linux","darwin"],"requires":{"bins":["ais","tmux"]}}}
parameters	{"action":{"type":"string","description":"Action: create, ls, inspect, inject, watch, kill, accounts, status, patrol","required":true},"name":{"type":"string","description":"Session name for create/inspect/inject/kill","required":false},"agent":{"type":"string","description":"Agent type: claude or kimi (default: kimi)","required":false},"account":{"type":"string","description":"Account ID: cc1-cc3/hataricc/nicxxx for claude, 1-N for kimi","required":false},"cmd":{"type":"string","description":"Command to inject after agent loads","required":false}}

AIS Orchestrator — Primary Execution Skill

ais is the primary way to run everything. All tasks, agents, and processes are managed through tmux sessions via ais. Never run long-lived processes outside tmux.

Source: https://github.com/gastown-publish/ais

Core Principle

Everything runs in tmux. Everything is inspected in tmux. Use ais for all execution.

Quick Reference

# === LIFECYCLE ===
ais create <name> -a claude|kimi -A <account> [--yolo] [-c "cmd"] [-d dir]
ais ls                              # list all sessions
ais kill <name>                     # graceful shutdown
ais kill --all                      # kill everything

# === MONITORING ===
ais inspect <name> -n 100           # capture last 100 lines
ais inspect <name> --rate-limit     # check for rate limits
ais watch <name> -i 5               # live tail (every 5s)
ais watch <name> --until "DONE"     # watch until pattern
ais logs <name> -o file.log         # save full scrollback
ais status <name>                   # machine-readable status

# === INTERACTION ===
ais inject <name> "do the thing"    # send command
ais inject <name> "y"               # answer a prompt

# === ACCOUNTS ===
ais accounts                        # list all accounts

Spawning Agents

Claude Code Agent

ais create worker1 -a claude -A cc1 --yolo -c "fix the auth bug in src/auth.py"

Kimi Code Agent (free)

ais create worker2 -a kimi -A 1 --yolo -c "write tests for src/utils.py"

With working directory

ais create task1 -a claude -A cc2 --yolo -d ~/project/services/auth -c "fix all failing tests"

Account Inventory

Check before spawning:

ais accounts

Claude Code Accounts (~/.claude-accounts/)

Account	Model	Best For
cc1	Sonnet	Fast tasks, workers
cc2	Sonnet	Parallel workers
cc3	Opus	Complex reasoning

Env: CLAUDE_CONFIG_DIR=~/.claude-accounts/<id>

Kimi Code Accounts (~/.kimi-accounts/)

Free, no credit cost, has rate limits. Add more with kimi-account add <key>.

Env: KIMI_SHARE_DIR=~/.kimi-accounts/<id>

The 7 Rules of Sub-Agent Prompts

1. One task, one agent. Never give an agent two unrelated goals.
2. Be specific about what, not how. Say "fix the failing test in test_auth.py" not "look at the tests."
3. Name the files. "Edit src/auth/handler.py" not "find the auth code."
4. Set boundaries. "Only modify files in services/order/."
5. Define done. "When done, print TASK COMPLETE: <summary>."
6. Give context, not instructions to find context. Paste the error, not "run the tests to see."
7. Match agent to task. Opus for architecture. Sonnet for fixes. Kimi for free boilerplate.

Prompt Template

[WHAT] One sentence stating the task.
[CONTEXT] The error, test output, function signature.
[SCOPE] Which files to touch. Which to leave alone.
[DONE] How to signal completion.
[CONSTRAINTS] Things to avoid.

Monitoring All Agents

Quick patrol

bash ~/.openclaw/skills/ais-orchestrator/scripts/patrol.sh

Manual check

for name in $(ais ls 2>/dev/null | tail -n +3 | awk '{print $1}'); do
  echo "=== $name ==="
  output=$(ais inspect "$name" -n 10 --rate-limit 2>/dev/null)
  if echo "$output" | grep -q "TASK COMPLETE"; then
    echo "[DONE] $(echo "$output" | grep 'TASK COMPLETE')"
  elif echo "$output" | grep -qiE 'rate.?limit|429'; then
    echo "[RATE LIMITED] — rotate account"
  else
    echo "$output" | tail -3
  fi
done

Rate Limit Recovery

When ais inspect --rate-limit detects a limit:

1. Kill the session: ais kill <name>
2. Pick a different account
3. Respawn: ais create <name> -a <agent> -A <new-account> ...

Rotation order:

• Claude: cc1 → cc2 → cc3 → cc1
• Kimi: 1 → 2 → 3 → 1

Multi-Agent Patterns

Fan-Out: Same task across directories

for svc in order product payment; do
  ais create "fix-${svc}" -a kimi -A $((RANDOM % 3 + 1)) --yolo \
    -d ~/project/services/${svc}-service \
    -c "Fix all failing tests. Print TASK COMPLETE when done."
done

Pipeline: Sequential dependencies

ais create phase1 -a claude -A cc3 --yolo -c "Design the interface. Print TASK COMPLETE."
# Wait for phase 1
while ! ais inspect phase1 -n 20 | grep -q "TASK COMPLETE"; do sleep 30; done
# Fan-out phase 2
ais create impl1 -a claude -A cc1 --yolo -c "Implement using the interface."
ais create impl2 -a kimi -A 1 --yolo -c "Implement the other part."

Specialist Agents

ais create architect -a claude -A cc3 --yolo -c "Review and design. Write PLAN.md."
ais create implement -a claude -A cc1 --yolo -c "Implement plan from PLAN.md."
ais create tester -a kimi -A 1 --yolo -c "Write tests. Free credits."

Log Trail

Every session creates logs at ~/.ais/logs/<name>/:

~/.ais/logs/worker1/
├── events.log      # Timestamped events
├── snapshots.log   # TUI captures
├── status.json     # Machine-readable (for other agents)
└── prompt.txt      # Original task

Other agents can resume work:

ais status worker1                            # read status
cat ~/.ais/logs/worker1/prompt.txt            # read original task
ais create worker1 -a claude -A cc2 -c "$(cat ~/.ais/logs/worker1/prompt.txt)"

Rules

1. Always run ais accounts first — know what's available
2. Spread across accounts — don't overload one account
3. Monitor every 2-5 minutes — catch problems early
4. Rotate immediately on rate limit — don't wait
5. Use Kimi for simple tasks — it's free
6. Save logs before killing — ais kill --save
7. Never send Ctrl-C — always ais kill for graceful shutdown
8. Include TASK COMPLETE signals in every prompt
9. Report to user — what's running, done, failed
10. Everything in tmux — never run agents outside ais