coding-agent（编码剂）

Coding Agent (always backgrounded)

Use bash with background:true for all coding-agent work. Do not use a foreground one-shot path here. Start the agent, get the sessionId, monitor with process, and require the worker to notify the user directly when it finishes.

⚠️ PTY Mode: Codex/Pi/OpenCode yes, Claude Code no

For Codex, Pi, and OpenCode, PTY is required:

# Correct for Codex/Pi/OpenCode
bash pty:true background:true command:"codex exec 'Your prompt'"

For Claude Code (claude CLI), use --print --permission-mode bypassPermissions instead. Do not use PTY for Claude Code here.

# Correct for Claude Code
bash background:true command:"claude --permission-mode bypassPermissions --print 'Your task'"

# Wrong for Claude Code (PTY, wrong flags, no background)
bash pty:true command:"claude --dangerously-skip-permissions 'task'"

Bash Tool Parameters

Process Tool Actions

Mandatory Pattern

Every coding-agent run follows this pattern:

1. Capture the notification route from the current conversation before spawning:

- notifyChannel - notifyTarget - notifyAccount (if applicable) - notifyReplyTo (if replying to a specific message is desired) - notifyThreadId (Telegram topic / Slack thread when applicable)

1. Start the coding CLI with background:true immediately.
2. Include the notification route in the worker prompt and require the worker to call openclaw message send on completion.
3. Monitor with process action:log / poll.
4. If the worker needs input or fails before notifying, handle that explicitly yourself. Do not rely on heartbeat.

If you do not have a trustworthy notification route, say so and do not claim that completion will notify the user automatically.

Notification Route

Do not rely on:

• openclaw system event
• tools.exec.notifyOnExit
• heartbeat delivery
• HEARTBEAT.md

Use a direct outbound completion message instead:

openclaw message send --channel <channel> --target '<target>' --message '<text>'

Add optional routing flags only when they are real and applicable:

• --account <id>
• --reply-to <messageId>
• --thread-id <threadId>

openclaw message send is a direct outbound send. It does not depend on heartbeat being enabled.

Completion Prompt Snippet

Append something like this to every worker prompt:

Notification route for completion:
- channel: <notifyChannel>
- target: <notifyTarget>
- account: <notifyAccount or omit>
- reply_to: <notifyReplyTo or omit>
- thread_id: <notifyThreadId or omit>

When the task is completely finished, send exactly one completion message back to the user with openclaw message send using that route.
If the task fails fatally, send exactly one failure message back to the user with openclaw message send using that route.
Do not use openclaw system event. Do not rely on heartbeat. Do not skip the completion/failure message.

Completion Command Template

openclaw message send \
  --channel <notifyChannel> \
  --target '<notifyTarget>' \
  --message 'Done: <brief summary>'

Optional additions:

  --account <notifyAccount> \
  --reply-to <notifyReplyTo> \
  --thread-id <notifyThreadId>

Quick Start

For scratch Codex work, create a temp git repo first, then start the worker in the background with the completion route injected into the prompt:

SCRATCH=$(mktemp -d)
cd "$SCRATCH" && git init

bash pty:true workdir:$SCRATCH background:true command:"codex exec 'Your prompt here.

Notification route for completion:
- channel: <notifyChannel>
- target: <notifyTarget>
- account: <notifyAccount or omit>
- reply_to: <notifyReplyTo or omit>
- thread_id: <notifyThreadId or omit>

When the task is completely finished, send exactly one completion message back to the user with openclaw message send using that route.
If the task fails fatally, send exactly one failure message back to the user with openclaw message send using that route.
Do not use openclaw system event. Do not rely on heartbeat. Do not skip the completion/failure message.'"

Codex refuses to run outside a trusted git directory. Reuse this same notify-route injection block in every example below; only the task-specific prompt body should change.

Codex CLI

Model: gpt-5.2-codex is the default (set in ~/.codex/config.toml)

Flags

Building/Creating

# Always background immediately
bash pty:true workdir:~/project background:true command:"codex exec --full-auto 'Build a dark mode toggle'"

# More autonomy
bash pty:true workdir:~/project background:true command:"codex --yolo 'Refactor the auth module'"

Reviewing PRs

Never review PRs in OpenClaw's own project folder. Clone to a temp folder or use a worktree.

REVIEW_DIR=$(mktemp -d)
git clone https://github.com/user/repo.git $REVIEW_DIR
cd $REVIEW_DIR && gh pr checkout 130

bash pty:true workdir:$REVIEW_DIR background:true command:"codex review --base origin/main"

Or:

git worktree add /tmp/pr-130-review pr-130-branch
bash pty:true workdir:/tmp/pr-130-review background:true command:"codex review --base main"

Batch PR Reviews

git fetch origin '+refs/pull/*/head:refs/remotes/origin/pr/*'

bash pty:true workdir:~/project background:true command:"codex exec 'Review PR #86. git diff origin/main...origin/pr/86'"
bash pty:true workdir:~/project background:true command:"codex exec 'Review PR #87. git diff origin/main...origin/pr/87'"

process action:list
process action:log sessionId:XXX

Claude Code

bash workdir:~/project background:true command:"claude --permission-mode bypassPermissions --print 'Your task'"

OpenCode

bash pty:true workdir:~/project background:true command:"opencode run 'Your task'"

Pi Coding Agent

# Install: npm install -g @mariozechner/pi-coding-agent
bash pty:true workdir:~/project background:true command:"pi 'Your task'"

# Non-interactive mode
bash pty:true workdir:~/project background:true command:"pi -p 'Summarize src/'"

# Different provider/model
bash pty:true workdir:~/project background:true command:"pi --provider openai --model gpt-4o-mini -p 'Your task'"

Parallel Issue Fixing with git worktrees

git worktree add -b fix/issue-78 /tmp/issue-78 main
git worktree add -b fix/issue-99 /tmp/issue-99 main

bash pty:true workdir:/tmp/issue-78 background:true command:"pnpm install && codex --yolo 'Fix issue #78: <description>. Commit and push after review. Send the completion message with openclaw message send using the provided notify route.'"
bash pty:true workdir:/tmp/issue-99 background:true command:"pnpm install && codex --yolo 'Fix issue #99 from the approved ticket summary. Implement only the in-scope edits. Send the completion message with openclaw message send using the provided notify route.'"

process action:list
process action:log sessionId:XXX

⚠️ Rules

1. Use the right execution mode per agent:

- Codex/Pi/OpenCode: pty:true - Claude Code: --print --permission-mode bypassPermissions (no PTY required)

1. Respect tool choice - if user asks for Codex, use Codex.

- Orchestrator mode: do NOT hand-code patches yourself. - If an agent fails/hangs, respawn it or ask the user for direction, but don't silently take over.

1. Be patient - don't kill sessions because they're "slow"
2. Monitor with process:log - check progress without interfering
3. --full-auto for building - auto-approves changes
4. vanilla for reviewing - no special flags needed
5. Parallel is OK - run many Codex processes at once for batch work
6. NEVER start Codex inside your OpenClaw state directory ($OPENCLAW_STATE_DIR, default ~/.openclaw) - it'll read your soul docs and get weird ideas about the org chart!
7. NEVER checkout branches in ~/Projects/openclaw/ - that's the LIVE OpenClaw instance!
8. Always inject the Completion Prompt Snippet into the worker prompt before spawning. The simplified examples below omit it for brevity — never spawn a worker without it.

Progress Updates (Critical)

When you spawn a coding agent in the background, keep the user in the loop.

• Send 1 short message when you start: what is running and where.
• Update only when something changes:

- a milestone completes - the worker asks a question - you hit an error or need user action - the worker finishes

• If you kill a session, immediately say you killed it and why.
• If you are expecting the worker to self-notify with openclaw message send, say that clearly in your start update.

This prevents the user from seeing only a missing reply and having no idea what happened.

Rules

1. Always background immediately.

- Use background:true for every coding-agent launch. - Do not use the foreground one-shot path in this skill.

1. Use the right execution mode per agent.

- Codex/Pi/OpenCode: pty:true - Claude Code: --print --permission-mode bypassPermissions

1. Respect tool choice.

- If the user asked for Codex, use Codex. - Orchestrator mode: do not hand-code the patch yourself instead of using the requested coding agent.

1. Capture notify routing before spawn.

- Completion messaging must have a real route.

1. Use direct completion messaging.

- Require openclaw message send. - Do not rely on openclaw system event or heartbeat.

1. Do not silently take over.

- If a worker fails or hangs, respawn it or ask for direction. Do not quietly switch to hand-editing.

1. Monitor with process.

- process action:log is the default low-friction check.

1. Be patient.

- Do not kill sessions just because they are slow.

1. Parallel is OK.

- Many background Codex sessions can run at once.

1. Never start Codex in ~/.openclaw/.
2. Never checkout branches in ~/Projects/openclaw/.

Learnings

• PTY is essential for Codex/Pi/OpenCode.
• Git repo required: Codex needs a trusted git directory.
• Use exec under background orchestration: short and long tasks follow the same path now.
• submit vs write: use submit to send input plus Enter.
• Direct message send beats heartbeat for completion notification when the user must be told immediately and heartbeat may be disabled.