Overview

Agent Governance System (AGS) is a governance runtime for AI-assisted development. It turns agent work into a repeatable loop: classify the change, prove confidence, enforce repo rules, then commit with a structured message. The package is @tmrxjd/agent-governance-system.

What AGS does

Agent discipline — CAP (Commit Authorization Protocol) controls when agents may commit; staging scopes force one atomic category per commit; commit-msg validation blocks fake “finished” language.
Repo integrity — schema, pointer, semantic, canonical, and data-array engines keep configs, docs, and declared lists aligned with the files that actually exist.
Cost & quality signals — efficiency and impact engines track token burn, loops, and whether the agent is actually improving outcomes over time.
Delivery control — versioning bump rules and deployment environment gates promote work through local → bench → staging → production with explicit checks.

The governed loop

When CAP is on, a typical agent turn ends like this — every step is an MCP tool or package API your host can call:

MCP pipeline (ordered) text

enforcement_begin
staging_enforce
confidence_enforce
schema_enforce / pointer_enforce / semantic_enforce
performance_validate
testing_validate
sentinel_status
commit_checkpoint

Commits themselves look like this — status line + trailers the hook and MCP validator understand:

Agent commit message text

status/checkpoint(auth-middleware): add session refresh helper

Status: checkpoint
Scope: auth-middleware
Commit-Scope: function-create
Not-done: wire helper into request pipeline
Next: module-wire for refresh on 401
Evidence: vitest packages/auth/src/session.test.ts
Token: mid
Graph-ids: none (infra)
User-approved: no
Enforcement: mcp

Where to go next

You are…	Start with
Installing for the first time	Install → Tutorials
Wiring Cursor / MCP	Configuration → API surface
Designing repo policy	Staging scopes → Canonicalization
Choosing a plan	Licensing · Pricing

Package entry points

@tmrxjd/agent-governance-system — plugins, TOC, coordination
…/commit-msg — validate agent commit messages in hooks and CI
…/commit-cap — CAP state (off | on | close-only)
…/gates, …/confidence, …/product-licensing — enforcement helpers

Interactive walkthroughs of each engine live under Showcase. Full copy-paste configs are on Examples.