Parallel Agent Planner

Split one requirement into task packs you can hand to several coding agents at once - by file ownership, with frozen contracts, write-conflict and missing-dependency detection, a per-agent task brief, and a computed merge order. Use when the user asks "can I split this refactor across agents", "how do I run Claude Code and Codex on the same repo without conflicts", "how should I divide work between agents", "how do I merge parallel agent work", "my agents keep colliding", "turn this PRD into parallel tasks", or "should I even go multi-agent". Capabilities: (1) a go/no-go gate - single-file changes, trial-and-error work, hard sequential chains and unfrozen contracts are told to stay serial (2) four ways to cut work ranked by safety: new files vs existing files, by directory, by layer (the one that burns people), rename/refactor as a separate pass (3) a list of always-contended files - dependency manifests, migration dirs, type/contract layers, wiring points - pulled into a serial prerequisite task (4) scripts/split_check.py validates the plan: write conflicts (E003), reading a file whose owner you do not depend on (E004), dependency cycles (E005), unowned contracts (W101), missing acceptance commands (W102), with --repo comparing real files instead of guessing at globs (5) topological merge order where each level is a parallel batch (6) one-command task briefs per agent: files you may change, read-only files, frozen contracts, acceptance commands, explicit non-goals (7) split_check.py suggest scans a repo for candidate ownership blocks with file and line counts. Do not trigger for: running one agent sequentially (just do the work); writing the code itself (use a coding skill); scheduling containers, dashboards or session monitoring (that is Conductor / Orca / vibe-kanban territory); backlog prioritisation (use a project tool).

存档时间线

版本索引永久保留；内容副本只保留最新 1 份，更早版本仅留索引与哈希（存档时间线的证据链不会因此断裂）。

下载存档内容副本