governed-agent-skills v4.0.0

Instructions with enforcement where it counts.

Six skills tell coding agents how to work. Two guards check the parts that should not depend on judgment.

// the agent that builds a thing never approves it

Two layers

Instruction Layer

Six skills shape the work

They cover reasoning, role separation, planning, implementation, test evidence, and maximum-assurance review. The plugin installs this layer only.

Enforcement Layer

Two guards check mechanical boundaries

The destructive-command guard runs before shell execution. The change-containment guard checks final repository state. Each guard is installed and wired separately.

Not every skill needs a guard. Use enforcement when a rule has a mechanical answer and a miss can cause real damage.

The governed loop

  1. OperatorOwns the outcome and publication.
  2. OrchestratorScopes the work and dispatches it.
  3. Pressure-TesterAttacks the plan before building starts.
  4. BuilderImplements the fixed outcome. Never approves it.
  5. ReviewerChecks the final state in a separate session.

A seat is a role for one session. It is not a person or a vendor. Solo work starts with fewer seats and adds independent review as the stakes rise.

When each skill loads

Any nontrivial task beginsreasoning-doctrine
Work touches a repository, record, or another agentgoverned-operator
Outcome, scope, authority, and evidence are fixedwrite-maintainable-code
The work needs a planportable-adaptive-planning
Tests are written or judgedtest-verification
The operator explicitly asks for maximum assuranceship-it-or-fix-it

The governance dial

G0

Ordinary work

Use proportional self-checks.

G1

Real changes

Use one Builder and one independent Reviewer. Publication stays with the operator.

G2

Maximum assurance

Freeze the acceptance oracle first. Independent judges run it. G2 loads only when the operator says so.

Enforcement Layer guards

Python, pre-execution

Destructive-command guard

Denies a narrow set of catastrophic shell commands before an agent runs them. It has live-session proof for Claude Code, Codex, and Antigravity.

Read the destructive-command guard documentation
Rust, final state

Change-containment guard

Seals allowed change classes, then rejects unclassified repository changes and stale receipts. Version 4 is source only. Live hook dispatch is not claimed.

Read the change-containment guard documentation

Install boundaries

The plugin installs only the six Instruction Layer skills. It does not install, wire, or activate either guard.

Install each Enforcement Layer guard from its own README. The Rust guard has no prebuilt binary or supported native target in version 4.

Read the complete installation guide