oh-my-dag

列出断掉/失败的 DAG run(带 goal)让 owner 挑一个续跑——从磁盘 checkpoint 重载 plan、跳过已绿节点接着跑。Trigger:/omd-resume、续跑、接着跑、断了的图、resume、跑一半挂了、429 断了。Skip:全新任务(/omd-execute 或 dag_run)、查状态不续跑(dag_status)。

锁 plan/SDD 前的对抗式审问:沿决策树走、先给推荐答案、事实自查·技术 Decision 自裁·真 owner 岔口才阻塞问、对标外部实现逼问「为何偏离」;宽解岔口就地开 council;产决策记录表喂 /omd-sdd。审议纪律:只讨论不动手。Trigger:/omd-grill、审问、盘问这个方案、把这事讨论清楚、压测计划、stress-test。

打开/新建/列出 omd pathfinder 决策地图(经 omd MCP server)。大而模糊的多 session 工作先开图:把待决问题变成票,渐进散雾。Trigger:/omd-path、pathfinder、开图、决策地图、这活儿太大先规划。

把已定的 SDD/规划交给 omd DAG 引擎执行(经 MCP dag_run 三段式),完成后按交叉验证 checklist 逐条判、四选一验收。Trigger:/omd-execute、开始执行、按 SDD 干、执行计划。

视频→逐段结构化笔记 (MiMo-v2.5 原生吃画面+音频, 非 whisper 转写; 可重入管线)。讲解/课程视频里 PPT 框架图/代码/提示词是画面独有、音频拿不到的信息。产 ALL-NOTES.md 交 /omd-council 或 dag_research 做综合。Trigger:/omd-video、抖音/B站/YouTube 讲解视频、课程系列、把这些视频学一遍/提炼、画面里有代码/图表/PPT。Skip:文字原文综合→/omd-council;网页内容→dag_research(检索版)。

One table of every routed oh-my-dag skill: name → when to use. The one name to remember. Trigger: /omd / list skills / which skill / 有哪些技能 / 该用哪个技能.

执行 omd pathfinder 已散尽区域(编译 slice → 跑 DAG 真改文件 → 票翻 delivered),标 delivered 前逐票核对真身(delivered✅≠东西真在)。deliberate/build 的权力闸,仅 owner 明令。Trigger:/omd-deliver、交付、开始执行、动手吧。

裁决 omd pathfinder 前沿票(记录 owner 决策到真相文件),写 ruling 前过终裁判定树(真源三层/灰态三画法/敏感清单)。仅 owner 明确裁决时用。Trigger:/omd-rule、裁决、就这么定、按方案X来。

查看 omd pathfinder 当前地图前沿票,并拉取已落地的 AFK research 结果(pull 回流 + 预算内自续)。Trigger:/omd-tickets、前沿、看票、研究结果回来了吗。

对一批 diff 做对抗式审查:默认多维并行召回 + 确定性跨模型证伪,gate 阶梯控深度,deep 档单 agent 读全仓深审;finding≠真理,按误报裁决程序证伪后再定性。Trigger:/omd-review、审查这批改动、代码审查、review diff、release 闸、深审。

安全专项审计:经 omd dag_run 派多视角安全审查 DAG,按信任边界清单扫注入/认证/fail-open + untrusted 入口校验覆盖率,产按严重度排序的结构化报告。Trigger:/omd-audit、安全审计、查漏洞、信任边界、security review。

多视角专家议会:宽解空间的设计/决策问题,经 omd dag_research(council 模式)并行多 persona 出方案 + 多 lens 择优;领域岔口走接地档(反 happy-path)。Trigger:/omd-council、开会审议、多方案对比、给我几个选项、拿不准选哪个。

系统化根因调试:历史搜索→复现→scope lock→模式匹配→假设验证→修复→验证→报告 + 修完扫同类。铁律:无根因不修。Trigger:/omd-debug、debug、调查、排查、为什么挂了、这个bug、根因分析、investigate、stack trace、昨天还好的。

Session wrap-up ritual: update the active plan / next-state file + write a session log + capture memory. Skipping = the next session loses context. Trigger: end / wrap up / handoff / hand over / save progress / call it for today / session end / 结束 / 收工 / 交接 / 保存进度 / 今天到这 / session结束. Skip: commit code (/commit) / start a new session (/start).

git 热点架构加深扫描:经 omd dag_deepen 确定性挑高摩擦模块,并发找浅模块(删除测试/深接口草图),跨热点归并重复逻辑,按 leverage 排序产 HTML 报告。Trigger:/omd-deepen、架构加深扫描、热点扫描、找浅模块、找重构机会、deepening。

记忆巩固(同步):经 omd dream_consolidate 把近期事件窗压缩进 L0–L6 记忆层,带节制三闸 + 层级红线 + 健康比自检,只留 derived 洞察不吞业务真相。Trigger:/omd-dream、记忆整理、巩固记忆、consolidate memory、跑一次 dream。

外层 fixpoint 迭代执行:跑 omd DAG → 评结果 → 带失败原因重画,直到收敛或到轮数上限。Trigger:/omd-iterate、迭代收敛、没收敛再来一轮、定点修。

把定型的决策/参考资料轻量记进持久台账 docs/plan/NOTES.md,供 /omd-sdd 收割。记 why 不复制内容、遮敏感信息、追加不覆盖。含 ref 职能。Trigger:/omd-note、记下来、这个决定记一下、留个引用。

主动召回 omd 自记忆:推理/写作/决策卡住时主动查既有事实,不等被动触发。经 memory_recall 语义+词法混合检索,回带 confidence + source。Trigger:/omd-recall、查历史、以前怎么做的、翻一下记忆、有没有先例。

确定性 semgrep 静态扫描(SAST)+ 结构化报告。零 LLM 零成本,规则命中即事实。Trigger:/omd-sast、静态扫描、semgrep、扫一下代码。

把审议结论结晶成正式 SDD 文档落盘 docs/plan/,当 /omd-execute 的执行契约。承接 /omd-grill 的决策记录表,写给没有对话上下文的执行器看。含 crystallize/crystals 职能。Trigger:/omd-sdd、写成 SDD、结晶、方案定了记下来、列结晶。

精益审计:先扫 ponytail: 债务台账记已知欠账,再经 omd dag_slim 派车队按 global-first 两遍法找过度工程(镀金/过度抽象/死复杂度),只删不加。Trigger:/omd-slim、找过度工程、瘦身、YAGNI 检查、债务台账。

Ultra-compressed communication mode. Cuts token usage ~75% by dropping filler, articles, and pleasantries while keeping full technical accuracy. Use when user says "caveman mode", "talk like caveman", "use caveman", "less tokens", "be brief", or invokes /caveman.

Shared vocabulary for designing deep modules — module / interface / depth / seam / adapter / leverage / locality, the deletion test, deep-vs-shallow. Use when designing or improving a module's interface, deciding where a seam goes, making code testable or AI-navigable, or when another skill (dag-deepen / dag-slim) needs the design vocabulary. Trigger: design this module / where should the seam go / is this module too shallow / deep module / deletion test / codebase-design. Skip: finding what to refactor (/dag-deepen) / finding what to delete (/dag-slim) / reviewing a diff (/dag-review).

Smart git commit: analyze changes + zone checks (tsc/test/build) + conventional message (Chinese) + git commit. --ship mode: merge base + full tests + PR. Trigger: /commit / save changes / git commit / ship / ship it / create PR / 提交 / 保存变更 / 提交代码 / 帮我提交 / 发布 / 创建PR. Skip: verify only without committing (/verify) / code quality review (/review).

Multi-perspective parallel generation + judge-and-select (a panel of experts 'meets' to deliberate a winner): facing a wide solution space for a design/decision, dispatch N different persona+angle in parallel to produce candidates → multi-lens judge → select-and-synthesize (graft the runners-up's highlights), instead of giving one averaged answer in a single shot. Trigger: multiple options / compare options / best of n / bestof / fanout / council / parallel exploration / multiple perspectives / give me a few options / which option is better / option comparison / explore options / not sure which to pick / domain fork / grounded / anti-happy-path decision / 多个方案 / 对比方案 / 并行探索 / 多视角 / 给我几个选项 / 哪种方案好 / 方案对比 / 拿不准选哪个. Skip: a single clear solution (just do it) / pure execution needing no selection / code review (/review) / root-cause debugging (/investigate).

Decomposable coding task as a DAG: conductor plans, agent leaves build files concurrently, a deterministic oracle command gates, a heal fixpoint retries with failure context, halt-judge stops unproductive loops, checkpoints make runs resumable. Trigger: build this module across N files / 实装这个可分解任务 / parallel implementation with a test gate / dag-build. Skip: small single-file change (edit directly — dispatch overhead loses) / root-cause unknown (investigate first) / frontend visual work (needs a human eye).

Auto-authored expert council as a DAG: give it a goal + ground truth, the conductor authors N distinct personas, they produce candidates concurrently, a multi-lens judge selects and grafts a winner. Grounded tier for domain-correctness forks. Trigger: multiple options / which approach / best of n / council / 多个方案 / 哪种方案好 / 拿不准选哪个 / domain fork / grounded / anti-happy-path decision / dag-council. Skip: single clear solution (just do it) / you want to hand-write the lens spec yourself (/dag-fanout) / needs web evidence first (/dag-research feeds this).

Architecture-deepening scan as a DAG: deterministic git-hotspot discovery picks the highest-friction modules, one agent leaf per hotspot concurrently hunts shallow modules (deletion test, deep-interface sketches), a synthesis leaf dedups cross-hotspot repetition and ranks by leverage, HTML report of candidates. Trigger: improve architecture / 找重构机会 / deepen modules / architecture report / dag-deepen. Skip: what to delete (/dag-slim) / diff review (/dag-review) / design vocabulary only (/codebase-design).

Hand-written lens-spec fanout: you author the lenses (persona + sub-angles), the engine runs them concurrently and judges. Manual gearbox of dag-council. Trigger: fanout with my own lenses / 手写 lens / 我已明确要哪几个视角 / dag-fanout. Skip: want the conductor to author lenses automatically (/dag-council, the default) / web research (/dag-research).

Web research as a DAG: retrieve (multi-query search + tiered crawl) → multi-lens fanout → judged synthesis, one command, zero-loss corpus artifact. Trigger: research this / what is X and which is better / grounded answer from the web / 调研 / 查一下X怎么做 / 综合网上说法 / dag-research. Skip: you only need raw page content (fetch directly) / question answerable from the codebase (read code) / wide design decision without web grounding (/dag-council).

Adversarial multi-dimension diff review as a DAG: correctness/security/boundary/contract lenses review the diff concurrently, then a verify/refute layer converges findings (kills plausible-but-wrong ones). Trigger: review this diff / PR review / 审查这组改动 / 提交前对抗审查 / find P0s before merge / dag-review. Skip: root-cause debugging of a known bug (investigate directly) / building code (/dag-build embeds its own gated review).

Over-engineering-only review as a DAG: two passes, GLOBAL first (one strong reviewer hunts cross-file dedup/collapse/layer/couple — the -50% win), then LOCAL per-file hunks concurrently (delete/stdlib/native/yagni). One line per finding: what to cut, what replaces it. Trigger: what can we delete / over-engineered? / simplify review / 精简审查 / dag-slim. Skip: correctness/security bugs (/dag-review) / deepening candidates (/dag-deepen) / behavioral lazy mode (/ponytail).

Manually trigger Dream consolidation: distill an agent's raw events into durable memory facts, with safeguards. Runs all layers, budget cap ~10k tokens. Trigger: /dream / consolidate memory / dream run / 做梦 / 整理记忆 / 跑一次 dream / 记忆整理. Skip: routine verification (/verify) / commit (/commit) / history recall (/recall).

Forces the laziest solution that actually works — simplest, shortest, most minimal, optimized for the WHOLE system, not the local line. Channels a senior dev who has seen every over-engineered codebase: question whether the task needs to exist (YAGNI), reuse before rewriting, stdlib before custom code, platform before dependencies, one line before fifty. Intensity: lite / full (default) / ultra. Trigger: ponytail / be lazy / yagni / do less / simplest solution / 最简方案. Skip: review-for-deletion of a diff (/dag-slim) / harvesting the shortcut ledger (omd-debt).

Engineering retrospective from git history: analyzes commit patterns, type mix (feat/fix/refactor), focus score, test discipline, and productivity trends. Use for Sprint/weekly reviews. Users may say: "retro", "retrospective", "sprint summary", "what did I do this week", "work analysis", "复盘", "回顾", "Sprint总结", "这周做了什么", "工作分析". Do NOT use for session wrap-up (use /handoff) or code review (use /review).

On-demand audit: security / coverage / tech-debt / full Gate / PR code review (dispatch specialist lenses + optional cross-model review). For Phase/Gate milestones, after a security incident, pre-merge PR review. Trigger: audit / security scan / full check / Gate check / review --security / review --gate / tech-debt inventory / security audit / review PR / code review / 审计 / 安全扫描 / 全量检查 / Gate检查 / 技术债盘点 / 代码审查 / PR审查. Skip: routine verification (/verify) / commit (/commit).

Create new skills, modify and improve existing skills, and measure skill performance. Use when users want to create a skill from scratch, update or optimize an existing skill, run evals to test a skill, benchmark skill performance with variance analysis, or optimize a skill's description for better triggering accuracy.

Session initialization ritual: read the project's next-steps file + git status in parallel to restore the previous context, output a Briefing with the current task + suggested next step. The first command of every new session to prevent context loss. Trigger: start / 开始 / 恢复进度 / 继续上次的 / 新session / 从哪里继续. Skip: save progress (/handoff) / commit (/commit).

Systematic root-cause debugging: 8 phases — history search / reproduce / scope lock / pattern match / hypothesis test / fix / verify / report. Iron Law: no fix without a root cause. Trigger: investigate / debug / 调查 / 排查 / 为什么挂了 / 这个bug / 根因分析 / 错误报错 / 500 / stack trace / 昨天还好的. Skip: general code review (/review) / verification (/verify) / planned refactor (/implement).