Docs menu

System architecture

AGS is a plugin-hosted governance kernel. Engines register MCP tools, gates, and graphs. Your IDE or bot host loads the toolkit; agents call tools in order; hooks and CI re-validate the same rules at commit time.

Control plane

  • CAP — whether the agent may write commits at all
  • Enforcement session — proves the required tools ran this turn
  • Staging — classifies the diff into one Commit-Scope
  • Confidence — blocks low-certainty commits on critical paths
  • Domain engines — schema, pointer, semantic, arrays, canonical, sentinel, …
  • Commit — commit_checkpoint writes only after the chain passes
Before commit text
enforcement_begin
staging_enforce
confidence_enforce
schema_enforce / pointer_enforce / semantic_enforce
performance_validate
testing_validate
sentinel_status
commit_checkpoint

Pipeline diagram

Layers

  1. Authoring — agent edits files; staging classifies the change set into one atomic scope (staging scopes).
  2. Meaning & structure — semantic and pointer domains keep concepts and paths coherent.
  3. Singularity — canonicalization and data arrays pin one SoT and ordered lists.
  4. Reflex — Sentinel converts engine failures into halt / remediate actions.
  5. Ship — versioning and deployment promote tagged work through environments.

Where code runs

Runtime lives in @tmrxjd/agent-governance-system: TypeScript engines, MCP tool factories, CLI scripts, and git hooks. Your host (Cursor MCP, Claude Code, Discord bot, CI job) imports the package and mounts the tool catalog — see API surface and Configuration.

AGS documentation