OMC/OMX 编码任务编排器

PassAudited by VirusTotal on Apr 30, 2026.

Overview

Type: OpenClaw Skill Name: omc-omx-orchestrator Version: 4.3.0 The bundle is a sophisticated orchestration system for managing asynchronous coding tasks using the Claude (OMC) and Codex (OMX) CLI tools. It features a robust task lifecycle management system, including pre-flight checks, structured task metadata (task.json), and a dedicated recovery script (task-recovery.py) to handle agent restarts. While the skill requires high-privilege access to modify local files and explicitly uses flags to bypass tool sandboxes (e.g., --dangerously-bypass-approvals-and-sandbox), these behaviors are transparently documented and essential for its stated purpose of automated code refactoring and auditing. No evidence of malicious intent, data exfiltration, or unauthorized persistence was found.

Findings (0)

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

A dispatched OMX task may edit files or run local commands without the usual sandbox or approval boundary, so an incorrect task or model action could damage the project or local environment.

Why it was flagged

The standard OMX workflow tells the agent to disable both approval prompts and sandbox protections for local coding tasks.

Skill content
始终使用 `--dangerously-bypass-approvals-and-sandbox`(别名 `--madmax`)关闭沙箱,因为任务由用户明确派发,在用户本机执行。
Recommendation

Do not make the dangerous bypass mode the default. Require explicit per-task confirmation, prefer sandboxed or least-privilege modes where possible, run inside a disposable checkout/container, and ensure version control/backups are available.

ConcernHigh Confidence
ASI08: Cascading Failures
What this means

This could terminate other OMX detached tmux sessions that happen to match the pattern, affecting unrelated tasks outside the current dispatch.

Why it was flagged

The OMX command template performs a broad tmux cleanup by pattern, not by the specific task ID or process it launched.

Skill content
tmux ls 2>/dev/null | grep 'omx-.*-detached-' | cut -d: -f1 | xargs -r tmux kill-session -t
Recommendation

Track and clean up only the tmux session or process created for the specific task, and ask before killing unrelated sessions.

NoteHigh Confidence
ASI10: Rogue Agents
What this means

Long-running coding jobs may continue to operate on local files even if the user closes or restarts the main agent.

Why it was flagged

The skill intentionally creates detached background processes that can continue running after the main agent stops.

Skill content
`os.setsid` 创建独立进程组,主 Agent 死亡不影响子进程
Recommendation

Use the task list and recovery tools to monitor running jobs, provide an explicit cancel/kill workflow, and avoid dispatching tasks unless you are comfortable with them continuing in the background.

What this means

Using this skill may consume API quota and grants the local Claude CLI access under the configured Anthropic account.

Why it was flagged

The skill requires an Anthropic API key for one of its core execution modes; this is expected for the Claude CLI but is still sensitive account authority.

Skill content
`claude -p` 需要 `ANTHROPIC_API_KEY`:Claude Code 的 `-p/--print` 管道模式依赖 API Key 认证
Recommendation

Use a scoped key if available, store it securely outside task files/logs, and revoke or rotate it if you no longer use this workflow.

What this means

Task descriptions, logs, and model outputs may remain on disk and could contain project details or secrets if the user includes them in prompts.

Why it was flagged

Task specs, stdout, stderr, and results are persisted under the OpenClaw task directory for recovery and later inspection.

Skill content
所有输出写文件:`claude -p` 和 `omx exec` 的 stdout/stderr 全部重定向到任务目录下的文件
Recommendation

Avoid putting secrets in task descriptions, review task logs before sharing, restrict filesystem permissions on the task directory, and periodically clean up old tasks.