claw-graceful-recovery

PassAudited by ClawScan on May 5, 2026.

Overview

This instruction-only skill is coherent and safety-oriented, but users should notice that it can automatically abort stuck tasks and write a small local recovery log.

This appears safe for its stated purpose. Before installing, make sure you want Claw to automatically abandon stuck permission-failing tasks, and consider editing the templates/logging guidance to avoid storing or sending sensitive paths or task details.

Findings (3)

Artifact-based informational review of SKILL.md, metadata, install specs, static scan signals, and capability signals. ClawScan does not execute the skill or run runtime probes.

What this means

This can prevent loops, but it may also prematurely stop a task and require the user to resend context.

Why it was flagged

The skill changes the agent's stopping behavior by telling it to abort and clear the current task context when recovery signals are detected.

Skill content
同一操作连续失败 ≥ 2 次;工具调用 > 30 秒无进展...停止当前操作,不重试...清空当前任务的执行上下文
Recommendation

Install only if you want automatic recovery behavior; consider adding a short preserved summary or confirmation step for non-permission-related repeated failures.

What this means

Operation descriptions or error types may remain on disk locally, which could matter if task descriptions contain sensitive information.

Why it was flagged

The skill records failure details internally and may persist recovery metadata to a local log file when WeChat is unavailable.

Skill content
记录三项内容:失败的操作描述(一句话)、错误类型、用户的原始指令...静默写入本地日志:`echo "[$(date)] <操作> | <错误类型>" >> ~/claw_recovery.log`
Recommendation

Keep log entries minimal and redacted, define when the log should be deleted, and avoid including secrets or full sensitive paths in operation descriptions.

What this means

The agent could copy the path-containing examples into WeChat feedback, revealing local path details despite the stated redaction rule.

Why it was flagged

Some feedback templates show full system paths even though the same reference tells the agent not to expose full file paths.

Skill content
Skipped:\n  - /etc/hosts (insufficient permissions)\n  - /var/log/auth.log (access denied) ... Do not expose full file paths
Recommendation

Revise the templates to use generic labels such as 'system log file' or redacted paths consistently.