skillset

JACU

A governance harness for coding agents. It compiles missions, isolates the work, verifies declared scope, and preserves auditable local run state.

Three surfaces over one core: the CLI, the MCP server (serve), and — after SDD-009 — a Go library. MCP is a surface, not the identity.

It has no network, no credential and no deploy. If JACU fails in every way it can, production does not fall. That boundary is the design, not an omission.

Install with one command.

Homebrew (macOS or Linux) — first install adds the tap; after that the name is short:

brew install jacu-dev/jacu/jacu
brew upgrade jacu
brew install jacu

Signed binary via curl: download the script from the same Release, review it, then run it. There is no curl | sh installer.

curl -fsSL -o /tmp/jacu-install.sh \
  https://github.com/jacu-dev/jacu-harness/releases/latest/download/install.sh
bash /tmp/jacu-install.sh

That writes a Sigstore-verified ~/.local/bin/jacu. From source: go install github.com/jacu-dev/jacu-harness/cmd/jacu@latest. Host registration: docs/install.md.

Quick start

jacu doctor
jacu serve

For a host configuration, use jacu init --host cursor (or another supported host) or jacu doctor --emit claude-code.

Core MCP workflow

The names below are MCP tools called by the configured host, not shell commands:

1. jacu_mission_compile turns the objective, scope and verification argv into a governed mission.
2. jacu_workspace_open creates the isolated run. Keep its run_id and make every change only inside the returned worktree_path.
3. jacu_verify runs the mission checks. Only a pass verdict may proceed to review.
4. jacu_diff returns the scoped patch and digest for human review. Any later edit invalidates that review and requires a new diff.
5. After explicit approval, jacu_apply validates and commits the reviewed tree. Use jacu_discard instead when abandoning the run.

Never apply an unreviewed diff. jacu_apply may run the mission's verification commands for up to 10 minutes; MCP hosts must allow at least that long (MCP_TOOL_TIMEOUT in Claude Code).

Diagnostics and local data

Start with jacu status. It lists every run that still owns a worktree, in every project. After installation, run jacu doctor and jacu storage inspect --json.

Runtime state lives under .git/jacu. User-level telemetry and tool caches live under ~/.jacu-harness, with a one-time migration from the previous user-level directory.

The program, its invariants and the definition of done live in docs/sdd/PROGRAM.md. Triggers, refusals and acceptance floors live in docs/decisions/.