oh-my-dag

An agent of its own — and an open execution engine anything can call.

Four ways in: compose one capability, hand off a whole graph, talk to the conductor, or open the TUI.

English · 中文 · Get started →

Four ways in

One engine, four doors. Pick by how much you want to hand over at once.

Every door lands on the same substrate: a typed plan, deterministic passes, oracle gates, a cross-family verifier, per-node checkpoints, model pools, and cost accounting.

Two lanes, one engine

The contract lane — one task, one pass. Research the ground truth, grill the plan until the open questions are named, crystallise it into a written contract, then execute against that contract. The contract is what the executor reads; it never has to guess what the conversation meant.

The map lane — long work across many sessions. A decision map lives in git. Ambiguity becomes typed tickets, tickets get ruled on, a ruled region gets delivered, and delivery flips the ticket to delivered back on the map. The map is the memory the context window does not have.

The two lanes meet at the same engine and the same gates. And they meet at one deliberate bottleneck: deliver is the trigger the owner pulls. A region going quiet only reports; it never starts writing. Automation is free to research, fetch, plan and argue on its own — changing files stays a human decision, every time.

Five problems, five mechanisms

1 · "It said it was done. It wasn't."

An objective gate runs before any model judgement: tsc, the test suite, a scanner, a file that must exist on disk. Zero model in the loop, so it cannot be talked into passing. A node that claims a file it never wrote fails; a "reviewed" screenshot that does not exist fails.

The same discipline applies to the criterion itself. Before an acceptance command is trusted, the engine runs it twice in a throwaway world — once before any work exists (still green means it has nothing to do with this task), and once against a deliberately wrong artifact the classifier had to supply alongside the command (still green means it cannot tell a right answer from a wrong one). Either way the goal drops to exploratory instead of collecting a fake pass. Source: src/harness/goal/acceptance-gate.ts.

2 · "The context ran out and everything started over."

Every finished node lands atomically on disk. Resuming a broken run re-checks input hashes: nodes whose inputs are unchanged stay green and are not re-billed, and only the rest re-runs. solve with detached: true hands the loop to a worker process that outlives your session — close the client, the graph keeps going. Facts you want to keep go into a store with a hybrid recall path (a lexical leg and a deterministic hashed-vector leg, both zero-model), so next week's session can look them up instead of re-deriving them.

3 · "Deep research is expensive, and half of it is made up."

Retrieval has a deterministic floor. omd_web searches and fetches with no model in the loop: full text lands on disk and only an index comes back. Gaps close by re-crawling the missing source, never by a model filling them from memory. The model does synthesis; the engine does recall.

bun run scripts/dag-research.ts "<your question>" --deep

Same question — a mid-2026 MCP ecosystem review — run twice, two configurations of our own:

The cheap configuration independently reproduced 13 of the 15 facts the frontier configuration had verified. Not because small models are secretly frontier-grade — because the part that decides fact coverage is retrieval, and retrieval is the part with no model in it.

→ Deep research guide, seat assignment, full A/B · sample output

4 · "I don't trust a cheap model with anything that matters."

Then don't trust it — check it. Between the plan and execution sit four pure functions (prune dead nodes, merge duplicates by semantic key, enforce the evidence gate, pin a model on every node); after execution sits the oracle gate, and after that a verifier drawn from a different model family than the author, because a verifier that shares the author's family shares its blind spots.

Underneath, work routes to 16 named seats in four classes — decomposer, judge/synthesis, worker, verify. Auto-assign fills them by channel economics: strong where being wrong is expensive and rare, cheap where volume is high and an oracle catches the mistakes, and family diversity spent only where it changes the answer. Pin any seat once in .omd/config.json and every resolver reads that one value. Registry: src/model/seats.ts.

5 · "The method only exists inside one person's prompt."

omd ships 20 methodology skills in the package — adversarial review, root-cause debugging, contract crystallisation, a decision-map workflow, a deletion-only over-engineering audit, and more. They install into ~/.claude/skills/ on first server start, idempotently, and never overwrite a skill you edited.

They are not just for your top-level agent. An agent leaf inside a graph gets the same skill set through the same tool, so a method you wrote once applies whether you invoke it by hand or a node reaches for it forty levels into a fan-out.

Skills: the method ships with the package

Skills are grouped under an umbrella. Your prompt carries the listing — group names and one-line descriptions — not the bodies. A model that wants a method calls read_skill and gets that one body, at that moment. A hundred installed skills therefore cost roughly a hundred lines of prompt, not a hundred documents, and the discovery surface stays the same whether you have three skills or three hundred.

Three roots are scanned, project first: <cwd>/.omd/skills, the package's own set, then ~/.claude/skills. Same name, project wins.

/omd-review for a diff, /omd-debug for a bug, /omd-grill then /omd-contract to lock a plan, /omd-path to open a map — the full list and how to write your own →

What you can call

Six families, 49 tools.

Control flow belongs to the runtime, never to the model: you pick the shape and its parameters — parallel, pipeline, loop-until, verify, judge, discovery, iterate, tournament, router, race, escalation, saga — and the loop, branch, stop and scoring logic is the engine's. A thirteenth, escape-hatch, stays off unless you set OMD_ESCAPE_HATCH=1.

One safety note: for anything unattended, or anything that fetches the open web, run with branchStrategy: 'branch' — an isolated git worktree plus a jail, so the leaf cannot read or write outside it. Details in the engine doc.

→ Full tool reference

Quick start

git clone https://github.com/AbyssCN/oh-my-dag.git && cd oh-my-dag
bun install && bun link      # puts `omd` on your PATH (Bun ≥ 1.3)
omd init                     # wizard: keys, model presets, reachability probe → .env

cd <your-project> && claude mcp add omd -- omd mcp

Then either drive it from your MCP client, or run omd tui for omd's own terminal seat. Prefer to configure by hand? Set OMD_RUNTIME_PROVIDER, OMD_RUNTIME_MODEL and your backend key in .env (copy .env.example). Skill installation opts out with OMD_INSTALL_SKILLS=0.

→ Full walkthrough · command reference

Docs

License

MIT — see LICENSE.