oh-my-dag

<div align="center"> # oh-my-dag ### The full harness around the model — on the [pi](https://pi.dev) coding-agent runtime. [](LICENSE) [](https://bun.sh) [](https://pi.dev) **English** · [中文](#中文) </div> --- ## What this is In the `oh-my-*` tradition: pi gives you a minimal, extensible agent runtime — this repo gives you **the whole configured outfit** on top of it. **omd** is the bundled terminal agent; every part below mounts as a pi extension and works with any OpenAI-compatible backend. - **A memory that consolidates itself** — runtime signals distill into facts, facts that recur across sessions get promoted, confident facts ground behavior next time. A TTL sweep keeps it bounded; every write passes a validation gate. - **A DAG execution engine** (the flagship — details below) — a conductor turns tasks into typed node graphs that run concurrently, get verified by a cross-model skeptic, and escalate only on failure. - **Gates enforced in code, not prompts** — plan mode blocks writes until the plan is aligned; an edit→verify gate blocks side-effect commands until typecheck/tests pass; dangerous commands hit a fail-closed classifier. - **An MCP router** — every MCP server collapses behind one on-demand menu; large tool outputs offload to a searchable sandbox. - **A 17-skill bundle** — 12 harness skills (session rituals, memory, verification, review) + 5 DAG skills, `.claude/skills` compatible, with a self-improvement flywheel that mines recurring wins into new skill proposals. > The model is the engine. This is the rest of the car. ## The DAG engine A single agent chat does everything serially in one context window; ad-hoc parallel calls lose structure. The engine turns a task into a **typed DAG of small nodes**: ``` task ──▶ conductor (one LLM call) ──▶ plan { nodes, deps } │ ready-set concurrent scheduling ┌───────────┬───────────┼───────────┬───────────┐ command inproc agent map primitive (CLI, no (single-shot (tool-using (runtime (tournament/router/ LLM) LLM call) sub-agent) fan-out) race/escalation/saga) └───────────┴───────────┼───────────┴───────────┘ cross-model verifier │ pass ◀──────────┴──────────▶ fail → escalate & re-plan ``` - **One node, one fault boundary** — a failed node becomes a `[failed]` input downstream; the rest keeps running. Nodes start the moment their own deps settle. - **`map` nodes** — when the work-list is unknown at plan time, a lister discovers it at runtime and a template stamps out one child per item, with stable resumable ids. - **A typed primitive menu** — the conductor *selects* schema-validated control flow (`parallel`, `pipeline`, `tournament`, `router`, `race`, `escalation`, `saga` with reverse compensation), each unit-capped and fail-closed. - **Trust is external** — plans are Zod-validated; a cross-model skeptic judges results (defaults to *fail* when uncertain); escalation to a stronger model happens only on failure, and only if you configured one. - **Recorded and resumable** — plans, per-node results and token usage land in SQLite; artifact-hashed checkpoints let a rerun skip done nodes; `planToMermaid()` draws any run; an ε-greedy bandit learns per node-kind which configured model is worth it. ## Quick start ```bash git clone https://github.com/AbyssCN/oh-my-dag.git oh-my-dag cd oh-my-dag bun install # any supported / OpenAI-compatible backend — no baked default export OMD_RUNTIME_PROVIDER=deepseek export OMD_RUNTIME_MODEL=deepseek-v4-pro export DEEPSEEK_API_KEY=sk-... bun run omd # interactive terminal agent bun run omd -p "..." # one-shot, non-interactive ``` Or `bun run init` for an interactive wizard. `bun link` puts `omd` / `oh-my-dag` on your PATH. Config lives in `.env` — copy [.env.example](.env.example). Legacy `XIHE_*` env names are still accepted. Requires [Bun](https://bun.sh) ≥ 1.3. ### In the TUI | Command | What it runs | |---|---| | `/cg <question>` | Parallel code-retrieval DAG: sync → concurrent query nodes → synthesis | | `/audit` | Multi-lens security/quality audit as an agent-node DAG | | `/sast` | Concurrent semgrep command-nodes → report synthesis | | `/iterate <task>` | Full DAG in a fixpoint loop until a convergence judge passes | | `shift+tab` | Plan mode: read-only deliberation, writes are code-blocked | | `/recall` · `/cost` · `/mcp` | Memory recall · cost ledger · MCP router | ### From the shell (no TUI needed) | Command | What it runs | |---|---| | `bun run dag-research "<question>"` | Search + tiered crawl → multi-lens fanout → judged answer, zero-loss artifact | | `bun run dag-council <goal.json>` | Conductor auto-authors N personas → concurrent candidates → judge + graft | | `bun run dag-fanout <spec.json>` | Hand-written lens spec, straight to fan-out (manual gearbox) | | `bun run dag-review --gate G2` | Adversarial multi-dimension diff review with verify/refute convergence | | `bun run dag-build "<goal>" --oracle-cmd "…"` | Conductor plans → agent leaves build concurrently → oracle gates → heal fixpoint, resumable | Each script has `--help`; the matching skill in [skills/](skills/) documents the discipline. ## Status — what's real today Honesty over marketing. As of v0.2: | Capability | Status | |---|---| | Conductor → typed plan → ready-set concurrent scheduling | ✅ shipped | | Executor kinds: command / inproc / agent / map / primitive | ✅ shipped | | Control-flow primitive menu (tournament, router, race, escalation, saga…) | ✅ shipped (schema-validated, unit-capped, fail-closed) | | Per-node fault boundaries; file-producer guard | ✅ shipped | | Cross-model verifier + escalate-only-on-failure; multi-lens verify / debiased judge | ✅ shipped (needs a second provider key) | | Fixpoint iteration (`/iterate`) + discovery loop | ✅ shipped | | Run recording (SQLite) + node-level checkpoint/resume (artifact-hashed) | ✅ shipped | | Bandit model routing (per node-kind, persisted) | ✅ shipped (no-op with a single model) | | Self-consolidating memory (signals → dream → promoted facts) + recall | ✅ shipped | | Plan mode (read-only gate, decision ledger, best-of-N) | ✅ shipped | | Edit→verify tool gate | ✅ shipped (`OMD_VERIFY_GATE=0` to disable) | | MCP router + output sandbox | ✅ shipped | | 17-skill bundle + skill flywheel (mine → propose) | ✅ shipped | | DAG → Mermaid rendering (`planToMermaid`) | ✅ shipped (API; TUI `/dag` command planned) | | Plan templates / replay a recorded plan as a new run | 🚧 roadmap | | Persistent cross-process workflow store (lease/CAS, node-level refine) | 🚧 roadmap | ## Design rules - **Reliability lives outside the model.** Validated plans, code-enforced gates, verifiers that default to fail, memory writes behind a validation gate. - **Wide over deep.** Parallelism and fault isolation both come from width. - **Pay for quality only on failure.** Cheap model works; escalation is opt-in and triggered by a failed verify, not by vibes. ## License MIT — see [LICENSE](LICENSE). --- --- <div align="center"> # 中文 ### 模型之外的完整装备 — 跑在 [pi](https://pi.dev) coding-agent 运行时上。 </div> ## 这是什么 `oh-my-*` 传统：pi 提供极简可扩展的 agent 运行时，这个仓库给它配上**整套装备**。 **omd** 是内置终端 agent；以下每一层都以 pi extension 挂载，任意 OpenAI 兼容后端可用。 - **自我整理的记忆**——运行时信号蒸馏成事实，跨 session 复现的事实升级为 confident， confident 事实落地为下次的行为；TTL 清扫保持有界，每次写入过校验闸。 - **DAG 执行引擎**（旗舰）——conductor 把任务变成有类型的节点图并发执行：五类执行器 （command / inproc / agent / **map** 运行时展开 / **primitive** 原语菜单含 saga 补偿回滚）、 ready-set 就绪即跑调度、节点级故障边界、跨模型校验失败才升级、checkpoint 断点续跑、 SQLite 落库 + Mermaid 画图、bandit 模型路由。 - **代码级闸门而非 prompt 劝告**——计划模式阻断写操作直到方案对齐；写后必验闸拦住未验证 改动后的副作用命令；危险命令过 fail-closed 分类器。 - **MCP 路由**——所有 MCP server 收敛到一个按需菜单，大输出卸载到可搜索沙箱。 - **17 个技能**——12 个 harness 技能（session 仪式/记忆/验证/审查）+ 5 个 DAG 技能， 兼容 `.claude/skills`，配自进化飞轮（复现的成功模式挖掘成新技能提案）。 > 模型是引擎，这里是整辆车的其余部分。 ## 快速开始 见上方英文 Quick start：`bun install` 后设 `OMD_RUNTIME_PROVIDER/MODEL` + 后端 key， `bun run omd` 启动（旧 `XIHE_*` 环境变量仍兼容）。TUI 内 `/cg` `/audit` `/iterate` 等 命令、shell 侧 `bun run dag-research|dag-council|dag-fanout|dag-review|dag-build`。 ## 现状（诚实版） 上方 Status 表为准：DAG 引擎全链路（五类执行器、原语菜单、校验升级、不动点迭代、 断点续跑、bandit 路由）、记忆层、计划模式、写后必验闸、MCP 路由、17 技能包—— **已交付**；计划模板重放、跨进程持久化工作流存储——**在路线图上**。 ## 设计准则 可靠性在模型之外；宽优于深；只为失败付费。