메뉴
Full error closed-loop for Dex (Main): diagnose, safe self-fix, verify, report; platform Exec failed handling; escalation with options. Load when handling failures or writing failure reports.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
Ledger development tasks to Feishu Base via dev-task-ledger.sh. Write roots from ~/.openclaw/tasks-allowlist.json (per agent). Use on accept/confirm/close of coding/ops work; not for pure Q&A.
Drive the machine Chrome debug browser via OpenClaw-managed MCP (chrome-devtools). Use for page navigation, snapshots, screenshots, clicks, forms, console/network inspection — not for host shell risk.
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.
Host shell MUST go through claw-run (gate then exec). Three layers: rules → Grok re-check → human only if still danger. Prefer claw-run over bare exec.
Multi-role adversarial code review for allowlisted repos. Agent loads this skill and runs host scripts (find → refute → sandbox repro); humans only state intent (e.g. 审一下 / 对抗审查). Use when user asks for adversarial review, task acceptance recommends review, or runtime code changed before merge.
Multi-role adversarial code review for this repo. Agent loads this skill and runs host scripts (find → refute → sandbox repro). Humans only state intent (审一下 / 对抗审查). Use when user asks for adversarial review, acceptance recommends review, or runtime code changed before merge.
| name | error-closed-loop |
| description | Full error closed-loop for Dex (Main): diagnose, safe self-fix, verify, report; platform Exec failed handling; escalation with options. Load when handling failures or writing failure reports. |
Always-on doctrine is short in AGENTS.md L0. This skill expands procedure.
1. Detect — non-zero / Exec failed / empty critical output
2. Diagnose — one factual cause
3. Self-fix — reasonable recovery 1–2 times when safe
4. Verify — smallest check that proves recovery
5. Report — user-facing template (even if recovered)
6. Escalate — only if blocked: situation + ≥2 options + recommendation
| Failure | Fix |
|---|---|
| ETXTBSY | bash path/to/script or wait + retry |
🛠️ Exec failed / SIGTERM batch | Split short execs; re-run missing checks only |
| 429 / flaky network | Backoff once |
| Wrong path | realpath / create only if clearly intended |
| Your leftover lock/process | Stop it, retry |
Never self-fix without explicit yes: force-push, broad rm -rf, prod deploy, disable safeguards.
Exec failedFeishu/runtime 🛠️ Exec failed: run A → run B → … = agent exec batch failed, not noise.
Explain in plain language the same turn; do not leave the arrow-list undecoded.
Prefer short verification batches after writes.
This skill still matters: it covers failures that re-enter the current agent turn as tool results (exec non-zero, readable tool errors, self-heal + verify).
It does not cover all OpenClaw special cases:
| Covered (agent must close loop) | Not covered by agent discipline alone |
|---|---|
| Same-turn toolResult / Exec failed | Runtime fallback error payload when no renderable final text |
| Self-heal + re-verify | Fixed system error strings (model/stream) |
| User-facing 失败/原因/处置/影响 | Delivery failed/partial after run ended (message_sent is observe-only) |
| Task 失败记录 honesty | Trajectory/session truncation diagnostics |
Full table + evidence pointers: openclaw/docs/error-closed-loop-scope.md.
Do not claim this skill eliminates every Feishu tail error banner.
失败：<command / step / Exec failed batch>
原因：<one sentence>
处置：已自愈（…）| 重试中 | 无法自愈
影响：阻塞 | 不阻塞；缺哪项验收；结论是否仍成立
结果/备选：
- 已恢复：验证命令 → pass/fail
- 未恢复：方案 A/B/C + 推荐 + 需拍板点
pass | fail | not run | re-run pass (after …)