Home

Cool Workflow Wiki

Get a saved report from your AI agent, with every claim tied to a line of code — not a chat message you lose.

Cool Workflow (CW) is a command-line tool that sits between you and your AI coding agent. Point it at a repo or a folder of docs. CW plans the work, hands each task to your agent, records and checks every result, and writes a report where every claim points to a place in the code, like file.ts:42. Everything is saved as plain files under .cw/ on your own disk.

CW never runs the model itself. The model is fuel; CW is the black-box recorder. Your agent spends the tokens; CW keeps the books.

New here? Start with these

Page	Use it for
Getting Started	Install, run the 30-second tamper proof, and make your first report.
User Guide	Your first ten minutes, from a real test: what to type, what you see, what to do when a run stops.
Mental Model	Why CW is built this way — the four commitments, and when it is worth it.
Glossary	Every core term in one place: run, evidence, topology, verifier gate, and more.
Quickstart	The fast command reference, once you know the shape.

Go deeper

Read these in order for the full picture, or jump to the one you need:

Page	Use it for
Architecture	The runtime boundary, state files, verifier gate, and MCP surface.
Repo Map	What every top-level folder and file is for, one line each.
Workflow Apps	Pick between the shipped apps and see how an app is put together.
Commands or API	The stable CLI shapes and MCP entry points.
MCP And Manifests	Generated vendor manifests and CLI ↔ MCP parity.
Trust And Audit	What the trust checks prove — and what they do not.
Recovery And Restore	Resume, export, inspect, import, verify, and rerun saved runs.
Operations	Verify, restore, regenerate manifests, and run release checks.
FAQ	Trust limits, agent setup, reports, and what happens when things fail.

What CW does

• plans a workflow app into phases and tasks,
• sends out workers, each in its own separate space,
• hands each worker to an outside agent — CW has no model SDK inside it and holds no keys,
• takes in each worker's result.md and checks it,
• keeps audit and telemetry ledgers — records that make any later edit visible, signed with ed25519,
• commits only checked state, and
• writes a report from the saved run record, with every claim tied to its source.

These lanes and the commands in the User Guide are the core path; the rest is kept working, not grown.

When evidence, dependencies, or a stated reason are missing, CW does not make up success — it stops at a clear unexplained / blocked state. See the Mental Model for why.

---

This Wiki sums up what is in the repository today; it makes no new claims. The root README is still the official first-run page.