orchestrator-v4
staleAudited by VirusTotal on Apr 5, 2026.
Overview
No VirusTotal analysis has been recorded yet. File reputation checks will appear here once the artifact hash has been scanned.
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.
A moderately sized project could trigger multi-agent dispatch much earlier than the documentation suggests.
The executable config shown in scripts/orchestrator_v4_acp.py sets large_project_file_threshold to 50 and large_project_line_threshold to 5000. That materially understates when automatic large-project splitting and worker fan-out will occur.
大项目检测：超过 1000 文件或 10 万行代码时自动切换大项目模式 ... `large_project_file_threshold` | 1000 ... `large_project_line_threshold` | 100000
Align the documentation and code defaults, and show the actual thresholds in the plan before spawning workers.
Multiple agents could read and modify project files in parallel, making unexpected or hard-to-review changes.
The skill directs automatic rolling subagent dispatch and, for creation/fix tasks, encourages workers to write files before summarizing. The artifacts do not require an explicit user approval step, diff review, or rollback plan before those changes.
按 max_parallel（默认 5）取前 N 个 subtasks，逐个调用 sessions_spawn ... 前一批全部完成后，立即派发下一批 ... 子代理 prompt 中要求"先写文件，最后再输出总结"
Require user confirmation after plan generation, show affected files, cap batch size, and require diff review before applying fix-mode changes.
Any spawned worker may have the same access to the project environment as the main session.
Subagents are configured to use the main agent identity and inherit the sandbox, so delegated workers can operate with the same effective workspace permissions.
subagent_agent_id: str = "main" ... subagent_sandbox: str = "inherit"
Use this only in workspaces where same-privilege subagents are acceptable, and prefer a restricted sandbox if available.
Task context or orchestration state may remain on disk and be reused later.
The orchestrator defines default checkpoint and resume behavior, which can persist task state or context locally when the full orchestrator API is used.
checkpoint_dir: str = "./checkpoints" ... auto_checkpoint: bool = True ... resume_from_latest_checkpoint: bool = True
Review or clear the checkpoint directory, and disable auto-checkpoint/resume if persistent local state is not desired.
Long-running or retried workers may continue consuming resources until stopped or timed out.
The skill includes long-worker, retry, lifecycle, and background-monitor concepts. They are purpose-aligned for orchestration, but users should be aware of autonomous worker management.
三级 Worker 路由（Fast / Slow / Long） ... 并发限流、失败自动重试（指数退避） ... 后台监控（心跳、超时）
Monitor active sessions, keep long-running modes disabled unless needed, and set conservative timeouts and restart limits.
Users have less provenance information for the bundled scripts they may run locally.
The skill ships executable Python files, but the registry metadata does not provide an external source or install provenance.
Source: unknown ... Homepage: none ... No install spec — this is an instruction-only skill ... 12 code file(s)
Review the bundled scripts before use and prefer pinned, documented sources for future releases.