samzong

Route the Steward maintainer workflow in Codex. Use when the user invokes steward directly, asks for an ambiguous gate, or wants the umbrella workflow entrypoint.

Dispatch a bug or task report into conflict-aware PR-sized worktrees with `gmc wt add` and write LLM-ready `todo.md` files in each worktree. Use when the user explicitly asks to dispatch bugs from a report, split a bug or task report into PRs, create worktrees from a report, or hand off fixes into separate worktrees. Typical input is the output of `critical-bug-finder` or any structured bug report with identifiable files, problems, and fix goals. Do not use for generic bug finding, generic issue triage, or normal project planning without a report file.

Generate a concise Chinese evaluation prompt for DevOps CLI / command-line tool projects based on clig.onev.dev guidelines, with weights, thresholds, and hard gates. Use when a user asks to create or store a reusable CLI review prompt, a scoring rubric, an evaluation checklist, or wants to evaluate a command-line interface, terminal tool, or CLI best practices against design guidelines.

Generate a deep, code-first project wiki and VitePress documentation site (DeepWiki-style) for a repository. Use when asked to create a complete wiki/manual for a new codebase, build a code-first documentation site, or generate a `codewiki/` folder with structured Markdown, diagrams, and navigation based on the source code (not existing docs).

Read, write, and sync project memory stored as Markdown under `~/.agents/memories/{project-name}/`. Use when starting work and needing project context, when the user asks to read or update memory, or when ending a session and preserving stable learnings. Triggers: "read memory", "load memory", "write memory", "save memory", "update memory", "sync memory", "refresh memory", "读取记忆", "加载记忆", "写入记忆", "保存记忆", "更新记忆", "同步记忆".

Find critical implementation bugs that can crash production, corrupt data, bypass security, deadlock, race, or break core logic. Use for explicit fatal bug hunts such as "find critical bugs", "audit for fatal bugs", "security vulnerability audit", "race condition audit", "find crash or data loss bugs", or reliability incident reviews. Do not use for general code review, refactoring, style feedback, or routine performance analysis.

Review a GitHub Pull Request as a responsible project owner using the `gh` CLI. Use when the user provides a PR URL (e.g. https://github.com/ORG/REPO/pull/N) or a PR number for the current git repo (prefer upstream, else origin) and wants an owner-grade review document `review-N.md` written in Chinese with copy-pastable GitHub comments in English. Scope the review to lines changed by the PR (do not nitpick unrelated pre-existing code), but apply best practices and flag any clear bugs, security issues, or CI failures caused by the change.

Export current session state as a structured HANDOFF.md for another LLM to continue work. Captures: current goal, files changed, what was tried, what worked, dead ends, and next steps. Use when: user says "handoff", "export state", "save progress", "session export", "pass to another LLM", "clone yourself", "create handoff", "wrap up", "done for now", "bye", or at session end when ongoing work needs continuity. Output is a self-contained prompt that another LLM can use to resume work exactly where this session left off.

GitHub issue management toolkit for project maintainers and contributors. Four modes: (1) scan — find contributable issues ranked P0-P5, (2) reply — draft and post concise maintainer responses to specific issues, (3) triage — classify issues and assign/remove labels, (4) analyze — deep root-cause or feature-design analysis saved to markdown. Use when: managing GitHub issues, finding contribution opportunities, replying to issues, triaging issues, labeling issues, analyzing bugs, designing features from issues, "what can I work on", "reply to this issue", "triage issues", "analyze this bug", "label these issues", "find easy issues", "show me contributable issues", issue cleanup, issue management, project maintenance, open source contribution discovery. Do not use for: PR review, code review, issue creation from scratch, or general GitHub operations unrelated to issue management.

Multi-model code review for pull requests and git diffs. Use when asked to "review my code", "review this PR", "review my changes", "find reviewers", or "run a code review". Suggests appropriate reviewers based on which files changed, runs iterative review rounds across multiple AI models, and tracks findings across sessions.

Pre-push bot-adversarial self-review gate. Simulates 7 AI reviewer lenses (coderabbit / copilot / codex / strict-maintainer / gemini / claude-code / greptileai) on staged or just-committed diff, auto-fixes MUST-FIX findings atomically, loops to convergence (hard cap 2 rounds), writes audit trail to commit body and a sentinel file for the Claude Code hook. Self-contained — does NOT call `pr-eval` (which is post-push only). Invoked by `ship` Step 3.5, by `commit` before commit, or manually before any push. Use when: "pre-ship", "pre-push scan", "bot exhaust", "ship-ready check", "push 之前扫一遍", "本地审", or invoked from `ship`.

Generate repo-local coding and review skills from a repository's git history and GitHub PR review activity. Use when asked to create, refresh, or update repository-specific coding/review skills, learn from PRs, analyze commit history, extract coding patterns, or generate coding guidelines for the current repo or a specific module.

Ship staged changes through a narrow release flow: auto-clean staged files, block unexpected non-i18n CJK additions, create or keep a branch, commit staged work, push, and open a pull request — all without confirmation. Use mainly when the user explicitly invokes `/ship` or says "ship it", "push and PR", or "commit and PR". Do not use for commit-only, push-only, PR-only, or existing-PR editing requests. All generated git and GitHub text must be in English.

Ruthless code simplifier: flatten abstractions, inline wrappers, remove unnecessary layers, delete dead code — without changing behavior. Use when: user says "simplify", "simplify this", "flatten", "inline", "too complex", "over-engineered", "remove abstraction", "unwrap", "reduce complexity", "make it simpler", "this is too complicated", or points at code that has unnecessary indirection. Does NOT change behavior, break public APIs, or remove meaningful error handling.

Cross-tool vibe coding profiler. Scans AI coding tool data on the machine (Claude Code sessions, OpenCode DB, Codex sessions), combined with git history and project memory, to build a true user portrait, discover workflow automation opportunities, and update the target tool's instruction file (CLAUDE.md or AGENTS.md) accordingly. Use when: "calibrate", "vibe-calibrate", "分析我的习惯", "profile me", "update my CLAUDE.md based on my history", "我的效率怎么提升", "analyze my patterns", "优化我的配置", or at the start of a long-term engagement with a new user.

撰写微信公众号技术文章的结构化工作流。使用场景：用户要写一篇公众号文章、需要创建文章目录、组织素材、迭代草稿。强制遵循 blogs/wechat/README.md 的命名规范和目录结构。

Challenge the current Steward story and append adversarial evidence without changing status. Use when the user asks for steward adversarial or wants hidden assumptions tested.

Manually archive a Steward case by freezing the final verdict and setting the case status to closed. Use when the user explicitly asks for steward close or needs to clean up abnormal case state.

Decide whether a task should be fixed and whether the proposed approach is acceptable. Use when the user asks for steward feasibility or wants a first maintainer decision.

Decide whether a change is worth merging after source-backed review. Use when the user asks for steward merge-value or wants a maintainer merge-worthiness verdict.

Suggest the next Steward step from the current case file without writing files. Use when the user asks what Steward should do next or asks for steward next.

Run one evaluative Steward review gate for a task, issue, PR, or change. Use for feasibility, adversarial, source-align, or merge-value review requests when the user wants maintainer judgment.

Compare the implementation with original source context, PR or issue discussion, and upstream state. Use when the user asks for steward source-align or wants source-backed maintainer review.

Show the current Steward case status without writing files. Use when the user asks where a Steward review stands, wants the current verdict, or asks for steward status.

Refresh the current Steward case file without changing status. Use before handoff, tool switching, restart, or when the user asks for steward sync.

Draft a paste-ready Steward review reply. Use when the user asks for steward reply or wants a concise maintainer response.