菜单
Route a task to the right model — a native Claude subagent, Codex, or local MLX — based on task type
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
Use when creating PRs, linking issues, managing PR comments, or creating GitHub issues
Use when creating or editing GitHub Actions workflows that call reusable workflows (uses: OWNER/repo/.github/workflows/...) — org owner references must be the literal current org, and shared-CI homes are under dryvist.
Emit a paste-ready two-part handoff for a fresh session: a `## Goal statement` hard-capped under 4000 characters (measured with wc -c, never estimated) that pastes straight into Claude Code's /goal Stop hook, plus an unbounded `## Full prompt` carrying cwd, ordered reading list, hard rules, pitfalls, and deliverables. Use when forking work to a new session, spinning up an orchestrator, or when wrap-up needs a next-session prompt with a real goal and not just a task list.
Analyzes current session state and repository status without any cleanup. Full mode (default): resolves the active plan file, reads plan checklist + TaskList, gathers unfinished work/issues from conversation history, checks git status, and emits a /handoff-built next-session prompt. Mid-session mode (`/session-status mid`): a fast plain-language 'done vs remaining' snapshot for mid-flight orientation, skipping the history scan, triage, and handoff.
End-of-session handler that first checks whether the current session's plan is actually complete. If complete: refresh repo, run quick retrospective, clean gone branches, and emit a forward-looking follow-up prompt. If incomplete: skip cleanup and emit one or more `cd`-into-worktree blocks paired with ready-to-paste resume prompts so the unfinished work can be picked up cold in a new session.
Automatically finalize pull requests for merge by resolving CodeQL violations, review threads, merge conflicts, and CI failures. Handles single PR (current branch or by number), all open PRs in the repo, or all open PRs across the org. Includes bot-authored PRs in all modes.
| name | delegate-to-ai |
| description | Route a task to the right model — a native Claude subagent, Codex, or local MLX — based on task type |
Picks the right executor when Claude-in-this-session is not the best tool: native Claude subagents for implementation and planning, Codex for an external/adversarial second opinion, and local MLX for private or offline work.
Plan mode or the Plan / general-purpose subagent type). Claude is the
best tool here — keep it in-house.codex MCP
tool or CLI). A genuinely different model catches what a Claude subagent won't.superpowers:dispatching-parallel-agents
skill), optionally adding Codex as one of the voices. Synthesize the results
yourself.Local models are named by capability role, not physical id. Roles resolve to
the resident model via the ai-stack registry
(~/.config/ai-stack/registry.json, written by nix-ai); never hardcode a
physical model id — when the resident model changes, only the registry changes.
Call llama-swap directly (OpenAI-compatible, no gateway hop):
curl -s http://127.0.0.1:11434/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{"model":"default","messages":[{"role":"user","content":"..."}]}'
Roles: default, quickest, coding, tool-calling, large-context,
most-capable, oss. Pick the role that fits the task; the registry maps it to
the physical model.
| Task type | Route | Executor |
|---|---|---|
| Implementation | native subagent | general-purpose subagent |
| Architecture / planning | native subagent | Plan mode / Plan subagent |
| Adversarial review | external model | Codex (codex MCP) |
| Multi-perspective / consensus | parallel subagents | N native subagents (+ Codex) |
| Private / offline / quick local | local MLX | llama-swap :11434, capability role |
curl/Bash to :11434.