orchestrator-v4

智能任务编排系统。自动扫描项目规模、规划子任务、动态派发多个 AI Worker 并行执行，支持大项目按模块拆分、自适应超时、滚动派发、用户随时打断改思路。 触发条件：用户需要处理复杂任务、多步骤分析、代码生成、调试分析、研究调查或需要智能调度 AI Worker 时。

What to consider before installing/using this skill: - Code review: Inspect scripts/orchestrator_v4_acp.py and scripts/scan_and_plan.py before running. They will execute locally and have filesystem access. Pay attention to any network I/O, subprocess execution, or dynamic code-evaluation patterns (I/O to unexpected hosts is the main risk). - Test in a sandbox: Run scan_and_plan.py and orchestrator functions against a non-sensitive test repository first. Verify that plans and file writes behave as you expect. - Clarify mismatches: SKILL.md claims large-project thresholds of 1000 files/100k lines, while the packaged OrchestratorConfig defaults use 50 files/5k lines. Ask the author which set of defaults applies and consider editing the config to your desired thresholds. - Check checkpoint/report files: The orchestrator writes checkpoints (./checkpoints) and reports (reports/...). If you don't want cross-run persistence, set resume_from_latest_checkpoint=false and/or change checkpoint_dir to a controlled location. - Sessions_spawn behavior: The orchestrator relies on the runtime-provided sessions_spawn to spawn subagents. Confirm how your platform's sessions_spawn handles subagent network access, logging, and data exfiltration risk. Subagents inherit capabilities of the platform; untrusted subagent prompts could leak repository content in their outputs. - Review archived/bridge modules: The package contains archived modules (v3_bridge, v3_worker, lifecycle manager, background monitor). These expand capabilities for long-running tasks and process management; verify you need them and audit them if you enable features like enable_v3_bridge or enable_lifecycle_manager. - Minimal privileges: Because no external credentials are requested, the biggest risk is local data exposure. Keep execution scoped to repositories you trust or sanitize the target_dir input. If you want, I can: (a) point to the exact lines where the threshold mismatch appears, (b) search the code for any outbound network calls, subprocess.exec usage, or dynamic eval calls, or (c) produce a short checklist of config changes to harden the skill for initial testing.