Autonomous agent loop integration

OpenGantry is the CLI contract + verify envelope for external autonomous agents (e.g. Hermes). Contract = mission YAML (scope + gate_command); verdict = gantry verify pass/fail with findings[] on failure. The executor agent implements code, builds skills, and retries — it does not parse raw terminal output.

All docs: index.md · Use cases: USE-CASES.md

Workflow

1. gantry init --discover [--domain code|content] — fast-path scanner emits .gitagent/discovery-proposal.json (no baseline writes until confirmed). Use --domain content for brand/compliance corpora (markdown, HTML).
2. gantry blueprint [--domain code|content] — interview produces ARCHITECTURE.md, TARGET_ARCHITECTURE.yaml, and .gitagent/verification_plan.json.
3. Executor reads required_skills in the verification plan and creates missing tooling before coding.
4. Executor legislates mission using gate_commands from the verification plan as gate_command.
5. gantry verify --mission … --json — on failure, ingest findings[]:

{
  "status": "failed",
  "envelope_schema_version": 2,
  "findings": [
    {
      "failed_gate": "gate",
      "offending_file": "",
      "line": 0,
      "severity": "error",
      "resolution_hint": "run gate (npm test); append evidence to EXECUTOR_LOG.md"
    }
  ]
}

6. Retry loop — executor maps each finding to a fix task until verify passes.

Surfaces

See ADR-0032.

Content domain example

Brand/compliance ad copy uses the same loop with --domain content and gantry perimeter check (regex rules in schema 0.3.0). Walkthrough fixture: examples/content-governance/. See DOMAINS.md.