dev-task

name	dev-task
description	Allowlisted development (wezdeck (+ optional team roots in local config)) under OpenClaw: wezdeck defaults to primary master; claw worktrees when parallel/isolation needed; 团队仓 still prefers claw-*; human/Claw rails (H1/H2, C1/C2/C3), ACP access layer.

Dev task (allowlisted repos)

When

Write work only under allowlist (see AGENTS.md). Pure Q&A: skip. Other repos: refuse.

Logical	Roots
团队仓	$HOME/work/team-repo, $HOME/work/.worktrees/team-repo
wezdeck	$HOME/github/wezterm-config (primary), optional .worktrees

cwd policy (L0-12/13)

Core: personal projects → default mainline. wezdeck is the local instance.

Repo	Default	Worktree when
wezdeck (personal)	primary master	parallel, long experiment, isolated agent cwd, user asks
团队仓 (team-ish)	claw-* under dirname(primary)/.worktrees/<repo>/	default isolation unless user overrides

Path formula (when used): dirname(realpath(primary))/.worktrees/<basename(primary)>/<slug>/.

Architecture: openclaw/docs/agent-architecture.md.

Checklist

Same steps as AGENTS.md Write-task checklist (wezdeck may skip assess/create). Scripts: dev-task-ledger.sh, claw-worktree.sh, claw-run.sh.

Assess → create (only when isolation/parallel needed)

./openclaw/scripts/claw-worktree.sh assess \
  --title "<subject>" --domain "<area>" --scope "<hint>" [--days N] \
  --cwd "$HOME/github/wezterm-config"
./openclaw/scripts/claw-worktree.sh create \
  --title "…" --lifecycle task|dev|hotfix --domain "…" \
  --cwd "$HOME/github/wezterm-config"

• Prefer-reuse same domain; never human dev-*/task-* as write targets.
• Reclaim never automatic after close (only if a worktree was used).

Rails & modes (user-facing)

轨	方式	旧	Who codes	Main does
人工	H1 人直接	A	User IDE	Ledger/验收 only
人工	H2 原生 Agent	A	Host grok/claude/codex	Assist only
Claw	C1 Main 自写	B	Main (Main-Grok)	Implement + verify
Claw	C2 Handoff	C	Host CLI after handoff	Stop coding that cwd
Claw	C3 ACP 后端	E	ACP → claude | codex	Spawn/close; single writer
—	D	D	—	Forbidden

ACP = access layer only; backends are Claude/Codex. No spawn grok. Do not rewrite host ~/.codex / ~/.grok defaults when fixing ACP (use ~/.openclaw/acpx/codex-home for ACP Codex).

Handoff (C2)

## Handoff
- task_id / cwd / branch / goal / non-goals / acceptance
- 开发方式: C2 本机原生 handoff（C）
- constraints: no force-push; no push main without yes
- after: 本机做完 → 飞书摘要 → main close + reclaim ask
- 本机: cd <cwd> && claude --continue

开发方式推荐卡（必发）

Full names required (never bare "Codex"/"Claude" alone):

Full name	Meaning
Claude-TUI	host claude (H2/C2)
Claude-ACP	C3 agentId=claude
Codex-TUI	host codex + ~/.codex
Codex-ACP	C3 agentId=codex + isolated CODEX_HOME
Codex-Grok-profile	host codex -p grok
Main-Grok	OpenClaw main model
Grok-native	host grok CLI

Before code or ACP, post and wait:

## 开发方式（请抉择）
- 轨: 人工 | Claw
- 推荐: H1 | H2 (Claude-TUI|Codex-TUI|Grok-native) | C1 Main-Grok |
        C2 handoff | C3 (Claude-ACP|Codex-ACP)
- 执行者 / 后端全名: …
- 理由: …（含限制/degraded）
- 备选: …
- 平台约束: 单写者、claw-*、确认前不写码；不改原生默认配置
- 审查建议: review-claude × review-codex-grok | 跳过（理由）
- cwd / task_id: …
请确认。确认前不改代码 / 不 spawn ACP。

Heuristics: C1 small/clear; Claude-ACP multi-file/profile; C2/H2 need TUI; H1 already coding; Codex-ACP explicit Codex stack.

C3 ACP spawn constitution (prepend to task)

[OpenClaw C3 constitution — non-negotiable]
1. Single writer: only you write this cwd; no parallel Main/TUI on same tree.
2. cwd is the path Main gave (wezdeck may be primary master or claw-*).
3. No force-push; wezdeck may push master per owner policy; other repos need explicit yes.
4. Prefer 1–3 logical commits; no secret leakage.
5. On completion report: changed files, summary, blockers; honest fail if blocked.
6. You are Claude-ACP or Codex-ACP (access layer), not a replacement for host TUI config.

Before spawn: reject if cwd missing or outside allowlist. wezdeck may be primary master; 团队仓 usually claw-*. After C3: prefer structured report changed_files / summary / blockers / commits.

Probe: openclaw/scripts/agent-matrix-status.sh.

实现方案块

See AGENTS.md. Always restate mode even if user named it.

落实 / commits

On 落实: review → implement → verify → 1–3 logical commits → push agreed branch → report. wezdeck: default work on primary master; after green checks commit + push master (L0-12/13/20). If a task branch/worktree was used, ff-only into master then push. Other repos: push main/master only with explicit yes. Shell via claw-run when required by exec-risk.

Feishu report default (brevity)

• Progress while working: 1–3 lines, not tool play-by-play.
• Final report: use AGENTS 精简【结果】卡 by default (status, one-liner, commit anchor, decision ask).
• Do not default-paste full acceptance tables, long path lists, full adversarial disclosure, or multi-block bash.
• User says 展开/全文/细节 → then full completion template.
• Failures: 4-line closed-loop first (失败/原因/处置/影响).

Git author & trailer (mandatory)

Rule	Detail
Author	Always repo owner. wezdeck: Yuns <[e-mail]>. Never bot Author names / yuns@local.
Do not	git -c user.name=<bot> or other bot identity overrides.
Trailer	Assisted-by: OpenClaw (backend=…, model=…)
C1 Main	backend=main, model=<short model id> e.g. grok-4.5 (not grok-proxy/… unless debugging)
C2/C3 write	backend=<full name> e.g. Claude-ACP, Codex-ACP + that backend's model
Editorial only	Assisted-by: OpenClaw (editorial-only) or omit
Integrate	wezdeck already on master → push; else rebase → merge --ff-only → push; no default --no-ff

Example message:

feat(scope): subject

- bullet why/what

Assisted-by: OpenClaw (backend=main, model=grok-4.5)

Adversarial review (agent runs it)

Do not ask the human to run run.sh. On review intent or acceptance:

1. Load skills/adversarial-review/SKILL.md
2. Run $REPO/scripts/dev/adversarial-review/run.sh <BASE> --writer <family> --mode strict
3. Report L0-21 disclosure (writer / form / reviewer / refuter / conclusions)

「对抗审查」= multi-role (find+refute minimum). Same agent twice with opposite stance is OK if labeled SINGLE-MODEL. Solo Main monologue = 设计批判 only. Writer-aware selection: --writer main|claude|codex|human (strategy B).

Constitution (all agents)

Usage may differ by rail/limits; criteria do not: L0, skills, scripts, single-writer, honest pass/fail, no fake green, secret hygiene, error closed-loop (Process failed / Exec failed same-turn plain language).