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。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/p10ns11y/skills --skill split-to-prs命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
Engineering-judgment overlay for control-graph + eva-emptiness: diagnose Odysseus mistakes (Cyclops leak, Sirens rewrite, Helios prod, Circe overbuild, Winds unbounded, Scylla big-bang, ignored prophecy), prescribe professional antidotes, and gate Odysseus spirits (Metis, endurance, curiosity, leadership) so the ship still reaches Ithaca. Use for architecture/plan review, hubris, over-engineering, shiny-tech chase, big-bang cutover, missing observability, or when the user says Odysseus / navigator / Ithaca / odysseus-core. Not domain product logic.
Pre-control-graph feeder: rewrite a dumped prompt, pick tools/plugins/skills, then route light vs control-graph Outer vs EVA Inner. Use when starting from a paste, --prompt-file, --prompt-json, -p/--single, or a messy dump before opening a Control Card. Triggers: control-feeder, /control-feeder, dumped prompt, rewrite wording, feed the graph, pre-control, prompt feed.
Agent control-plane as a graph: outer state machine (or cyclic loop) over named phases, nested inner DAGs or bounded sub-loops, hard budgets, HITL pause gates, and multi-model routing (fast/explore/coding/deep/review). Use for multi-step agent cycles, unbounded-ReAct risk, model routing, structured loops — not domain product logic. Formerly "looper". Composes with agent-orchestrator, subagent-delegation, fusion-sage, concurrent-cli-agents. Triggers: control-graph, looper, state machine, SM+DAG, loop card, model routing, cognitive load, CLT, cognitive strain, load balance.
基于 SOC 职业分类
正在显示 SKILL.md
| name | split-to-prs |
| description | Split current work into small reviewable PRs. Use when the user asks to split a chat, set of changes, branch, or PR. |
CLT: ../rules/clt-dual-load.mdc — plan HITL first (chunked options); after approval, agent may execute split mechanics in parallel where independent.
Turn one pile of work into a few small PRs.
reset --hard, clean -fdx, branch deletion, force-push, history rewrite) without explicit approval.main, so do not assume there is already a safe branch.git add . / git add -A.Compare the current work to the repo's default branch, including committed and uncommitted changes. Summarize the real slices you see, and use the chat history to recover intent.
Before proposing slices, find ownership signals for the touched paths (CODEOWNERS, nested ownership files, tools/ownership/PRODUCTOWNERS, or repo equivalents) and use them to identify natural reviewer boundaries.
Use judgment on detail. Usually PR titles are enough. Add a one-line scope note only when a title is unclear. Show a Mermaid diagram when there are multiple slices.
Optimize for reviewer-aligned PRs with minimal unrelated diff: split independent owners or concerns, keep tightly coupled changes together, and when stacking is necessary, order foundations before consumers.
Default to independent PRs off the default branch. Stack PRs only when the dependency is real.
Ask for approval before starting.
If there is uncommitted work, save a recoverable snapshot without changing the working tree:
SHA=$(git stash create "pre-split")
if [ -n "$SHA" ]; then
git update-ref "refs/backup/pre-split-$(date +%s)" "$SHA"
fi
For each approved slice, create a branch from the right base, stage and commit only the planned files or hunks, then push and open the PR.
Keep it short: PR titles and URLs, plus anything left on the starting branch or working tree. Do not delete the backup ref or original branch unless the user asks.