Domain adapters

OpenGantry is a domain-agnostic governance layer for autonomous agents. The mission/verify loop, trace mapping, failure envelope, and gate_command shell gates are domain-neutral. Domain adapters plug deterministic discovery, blueprint, and perimeter checks into that loop.

All docs: index.md · Why features exist: FEATURES.md

Architectural Boundaries vs. Full-Repo AST

Note: OpenGantry enforces architectural perimeters (such as import layers and regex content rules) rather than acting as a general, full-repo AST scanner. While Code Surgeon is utilized for TypeScript AST registration, the primary verification gate relies on strict architectural boundary compliance.

Universal framing

Phase	OpenGantry surface	What it does
Context ingestion	`gantry init --discover --domain <key>`	Walk corpus files; emit evidence-anchored conventions/anomalies
Rules of engagement	`gantry blueprint --domain <key>`	Human interview → `TARGET_ARCHITECTURE.yaml` + `verification_plan.json`
Standardized audit API	`gantry verify` + failure envelope `findings[]`	Deterministic gates + machine-readable verdicts for external executors

Built-in adapters

Key	Corpus	Enforcement rules
`code`	`.ts`, `.js`, …	Import layers (`forbid_import_layer`, `forbid_specifier_substring`, …)
`content`	`.md`, `.html`, `.txt`, `.json`	Regex perimeter (`forbid_pattern`, `require_pattern`)

List adapters: gantry domains

Determinism mandate

Everything in the trust path is binary and reproducible:

Enforcement: regex rules either match or they do not; reruns on an unchanged tree are byte-identical.
Content discovery: verbatim boilerplate blocks (≥2 files, exact byte match) and YAML frontmatter keys only — no statistical “dominant terminology.”
Code discovery: coverage percentages are advisory proposal metadata; humans confirm before rules become law.

Perimeter schema 0.3.0

TARGET_ARCHITECTURE.yaml gains optional domain and pattern rules:

schema_version: "0.3.0"
domain: content
scan_roots:
  - content
layers:
  - id: content
    globs:
      - content/**
rules:
  - id: forbid-claim
    from_layer: content
    applies_to: ["content/**"]
    forbid_pattern: "(?i)cures cancer"
  - id: require-disclaimer
    from_layer: content
    applies_to: ["content/**"]
    require_pattern: "These statements have not been evaluated by the FDA"

Check: gantry perimeter check (domain-neutral alias of gantry arch check).

Schemas 0.1.0 and 0.2.0 remain supported for code-only adopters.

Mission loop is domain-agnostic

Regardless of domain:

Planner legislates mission YAML (gate_command can be any shell command).
Executor works within TMVC roots / forbidden zones.
Trace rows land in EXECUTOR_LOG.md.
gantry verify runs gates and emits findings[] on failure.

See AGENT-LOOP.md for Hermes-style executor integration.

Recipes without built-in adapters

Accounting, legal, custom corpora: use gate_command scripts and TMVC path globs. Document patterns in your repo’s ARCHITECTURE.md; custom adapter plugins are a future extension point.

Example

Full content loop fixture: examples/content-governance/

Roadmap (not yet shipped)

Runtime-loaded adapter plugins (npm packages registered in config)

See CHANGELOG.md for shipped releases (including advisory performance judge #62 in v3.1.0).