delivery-graph

Delivery Graph Engineering

The neutral ground where any agent's work is proven, not trusted.

---

What is DGE?

Whatever coding agent you use — Claude Code, Copilot CLI, Kimi — it works against one evidence-gated delivery graph, and done is enforced by the engine, not claimed by the agent. A node cannot reach done until its validation evidence exists on disk. The tool decides completion; the agent doesn't get to say "it's done" and be believed.

That gate is the same regardless of which agent produced the work — which is what makes DGE neutral ground: an objective arbiter every agent's output must pass.

DGE also brings the discipline every serious agentic workflow needs — clarify before coding, plan before execution, work in isolated units, validate before completion, review before merge, capture learnings — and models delivery as a graph rather than a linear checklist. But the discipline is table stakes; the enforced, evidence-gated, agent-neutral done is the point.

See the delivery graph

Every graph also has a local, self-contained viewer at delivery-graph/view/index.html. It turns the canonical store into a desktop workspace: choose a demand, follow its dependency graph, see which validation criteria are proven, and inspect a node only when you select it. The viewer stays offline and refreshes at a user-selected rate, so it is a readable projection of the graph rather than another source of truth.

Generate or refresh the viewer whenever you work with the graph:

npx dge status
# Open delivery-graph/view/index.html in your browser

Why it's different

Every coding agent today trusts itself to declare work complete — and agents are documented to over-report "done." In DGE, completion is a state-machine invariant: verified can only be minted by running dge verify against real evidence, and done requires verified. There is no path — not even a "quiet" or automated one — that reaches done without proof.

• Harness-agnostic. The dge CLI is a plain binary and graph.json is a plain file, so Claude Code, Copilot CLI, and others read and write the same delivery graph. DGE rides on top of harnesses; it is not itself a harness.
• Done means proven. Evidence-gated completion is enforced in the engine, not left to the agent, the user, or CI.
• Persistent, machine-readable graph. Demands → requirements → tracks → nodes with real dependency edges and validation contracts, versioned in graph.json — the source of truth harnesses treat as ephemeral.
• Compounds across demands. Completed work leaves behind evidence and learnings that the next design reads, so the toolset gets smarter with each demand.

Built for what no single harness can do. Because the ground is neutral and completion is objective, DGE is designed toward different agents building and verifying each other's work against one source of truth. That parallel, cross-agent validation is on the roadmap — local concurrency and risk-based verifier-policy substrates exist today; production control-plane wiring, isolated worktrees, and durable run history make it fully real.

60-second quickstart

Add DGE to any repo and drive one node from demand to evidence-gated done. Runs fully locally — no Linear, Azure DevOps, or credentials required.

# 1. Install the CLI (npm) + the /dge-* slash commands
npm install --save-dev github:rafaelolsr/delivery-graph
npx dge install-skills                 # auto-detects .claude/ or .github/

# 2. Create the canonical graph store
npx dge init --title "My delivery graph"

# 3. Author one demand -> requirement -> track -> node
npx dge add-demand --title "Safer releases" --source "user" --outcome "Every completed node has proof"
npx dge add-requirement --demand DEM-001 --statement "Nodes require validation evidence" --acceptance "Verify fails without evidence" --evidence "Evidence manifest"
npx dge add-track --title "Validation"
npx dge add-node --title "Add evidence gate" --type test --track TRK-validation --requirements REQ-001 --validation "npm test"

# 4. Capture proof and close the node (fails without passing evidence)
npx dge evidence run NODE-001 --satisfies "npm test" -- npm test
npx dge done NODE-001
npx dge status --save

Then reload skills in Claude Code (restart the session or /reload-plugins). New here? The getting-started walkthrough explains each step.