菜单
Use when a coding task is not understood well enough to choose an implementation direction, including broad feature requests, unfamiliar code paths, bugs with multiple plausible causes, and work that benefits from parallel read-only exploration before architecture or planning.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Write or rewrite concise reviewer-first pull request titles and descriptions for Pardes manager publication. Use approved intent, owner reports, conditional verifier reports, and narrow factual details without filler.
Use when a completed or active Codex coding session should be evaluated as a workflow run, especially to identify where orchestration skills, sub-agent delegation, worktree routing, verification handoffs, PR monitoring, or compaction recovery worked well or caused friction.
Use when a user installs the Codex workstreams plugin, asks to set it up, or sees permission prompts for workstream checkpoints, reports, or temporary pull-request body files.
Use when a coding objective spans multiple workflow phases or pull requests and needs a persistent manager to coordinate investigation, architecture, planning, implementation, verification, PR creation, review feedback, and merge gates across a long-running session.
Use when an approved PR-sized coding goal needs execution through an owner worker and optional supporting workers in manually managed Git worktrees.
Use when a verified owner-worker branch is ready to publish, or when an open pull request needs blocking lifecycle management through CI failures, review feedback, conflicts, and merge.
| name | investigate-codebase |
| description | Use when a coding task is not understood well enough to choose an implementation direction, including broad feature requests, unfamiliar code paths, bugs with multiple plausible causes, and work that benefits from parallel read-only exploration before architecture or planning. |
Gather evidence before choosing a change. Keep the manager on synthesis and delegate codebase reading to focused explorers.
explorer agents for the distinct questions. Use
multiple explorers when the questions are independent.wait_agent with all
unresolved explorer IDs and a 900000 ms timeout. Treat an ordinary timeout
as "still working": wait again with a long timeout instead of short-polling
or narrating routine status.drive-workstream for user discussion and scope
approval. Record discovered fixes as candidate follow-ups; do not implement
them during investigation.Give each explorer one bounded question. Ask for file references, relevant commands or logs, verified observations, uncertainties, and a concise report. Do not send multiple explorers over the same surface.
Report: