claude-code-orchestrator（Claude 代码编排器）

Claude Code Orchestrator (tmux-first)

Use tmux-based orchestration for long coding tasks to avoid silent hangs and make progress observable.

Standard workflow

1. Create prompt file (avoid long shell quote issues).
2. Start a dedicated tmux session.
3. Launch claude --dangerously-skip-permissions in interactive mode.
4. Paste prompt into Claude.
5. Require callback command in prompt (via wrapper):

bash {baseDir}/scripts/wake.sh "..." now

1. Share socket/session attach command with user.
2. On completion, review diff + lint/build + risk summary.

Start command

bash {baseDir}/scripts/start-tmux-task.sh \
  --label "gallery-detail-polish" \
  --workdir "/Users/yaxuan/.openclaw/workspace/work/active/02-gallery-ops" \
  --prompt-file "/Users/yaxuan/Downloads/gallery-website-design-system.md" \
  --task "参考这个修改我当前的画廊官网，注意优先打磨细节和质感，对整体结构展示先不用大改。"

Monitor commands

# attach
bash {baseDir}/scripts/monitor-tmux-task.sh --attach --session <session>

# capture last 200 lines
bash {baseDir}/scripts/monitor-tmux-task.sh --session <session> --lines 200

Task overview

List all running cc-* tasks at a glance — useful for "butler-style" summaries.

# Human-readable one-liner per task
bash {baseDir}/scripts/list-tasks.sh

# Structured JSON array (pipe to jq, feed to OpenClaw, etc.)
bash {baseDir}/scripts/list-tasks.sh --json | jq .

Options:

• --lines <n> — number of trailing pane lines to capture per task (default 20).
• --socket <path> — tmux socket path (default $TMPDIR/clawdbot-tmux-sockets/clawdbot.sock).
• --json — emit JSON array instead of human table.
• --target ssh --ssh-host <alias> — list sessions on a remote host.

Each entry contains: label, session, status, sessionAlive, reportExists, reportJsonPath, lastLines, updatedAt.

Combine with OpenClaw to generate a periodic butler summary:

# In an OpenClaw prompt / cron:
bash {baseDir}/scripts/list-tasks.sh --json | \
  openclaw gateway call summarize-tasks --stdin

Rules

• Prefer interactive Claude in tmux for visibility (not long claude -p one-shot for large tasks).
• Always include callback via wrapper bash {baseDir}/scripts/wake.sh "..." now in prompt.
• Startup script now uses robust submit (ready-check + multi-Enter retry + execution-state detection) to avoid "prompt pasted but not submitted".
• If no pane output for >2-3 min, inspect and restart session.
• Kill stale Claude processes before restart.
• Always return: session name + attach command + current status.

Status check (zero-token)

If wake not received within expected time, check task status before consuming tokens:

bash {baseDir}/scripts/status-tmux-task.sh --label <label>

Output: STATUS=running|likely_done|stuck|idle|dead|done_session_ended

• likely_done / done_session_ended → proceed to completion loop
• running → wait
• stuck → inspect (attach or capture-pane)
• dead → session lost, run complete-tmux-task.sh fallback
• idle → Claude may be waiting for input, inspect

Completion loop (mandatory)

When wake event "Claude Code done (...)" arrives, complete this loop immediately:

1. Acknowledge user within 60s: "已收到完成信号，正在评估改动".
2. Preferred path: read completion report generated by Claude Code task:

- /tmp/cc-<label>-completion-report.json

1. If report missing, run local fallback immediately:

- bash {baseDir}/scripts/complete-tmux-task.sh --label <label> --workdir <workdir>

1. Mandatory deep-read: read full JSON/MD report before replying.
2. Read context before replying:

- Read completion report file(s) (/tmp/cc-<label>-completion-report.json/.md) - Read recent tmux transcript (monitor script) to capture what Claude actually did/failed/tried - Incorporate the latest user constraints from current chat

1. Then provide assistant analysis (not a fixed template):

- what was actually completed - what is reliable vs uncertain - key risks/tradeoffs in the user's context - concrete next-step options

1. Ask explicit decision from user if scope drift exists.

Do not stop at wake-only notification. Wake is trigger, not final delivery.

Anti-pattern to avoid

• Forbidden: one-line fixed reply after wake without reading transcript + report.
• Forbidden: only relaying "done + report path" without analysis in user context.
• Forbidden: rigid templated output that ignores current conversation context.

Hard guardrails added

• Prompt now enforces “no wake without report”:

- task must write /tmp/cc-<label>-completion-report.json + .md - final wake must include report=<json_path>

• Recovery command exists for deterministic fallback:

- scripts/complete-tmux-task.sh reproduces evidence and emits structured report

• Delivery SLA remains mandatory:

- wake received -> ack <= 60s -> report