agent-contracts/dsl_base/guardrails.yaml

Declarative YAML DSL toolkit for defining, validating, and rendering multi-agent development workflows

dsl-readonly-enforcement:
  description: >-
    Prevent audit-purpose agents from modifying DSL source files.
    dsl-auditor is mode=read-write (to produce dsl-audit-report) but
    must not write to dsl-source. This guardrail enforces that boundary.
  scope:
    agents:
      - dsl-auditor
    artifacts:
      - dsl-source
  rationale: >-
    dsl-auditor uses mode=read-write because it produces dsl-audit-report,
    but its constraints prohibit modifying DSL definitions. This guardrail
    blocks writes to dsl-source from audit-purpose agents.

dsl-validate-before-render:
  description: >-
    Ensure DSL validation passes before rendering is attempted.
    Prevents generation of prompts from invalid DSL definitions.
  scope:
    workflows:
      - dsl-update
    tasks:
      - render-dsl-outputs
    artifacts:
      - dsl-generated-output
  rationale: >-
    render-dsl-outputs must not be invoked while validation_result
    of update-dsl-definitions is not 'pass'.

dsl-no-hallucinated-permissions:
  description: >-
    Rendered prompts must not grant permissions not declared in DSL.
    Detected via audit-generated-prompts; any hallucinated-permission
    finding of severity critical must be escalated.
  scope:
    artifacts:
      - dsl-generated-output
    tasks:
      - audit-generated-prompts
    tools:
      - agent-contracts-cli
  rationale: >-
    audit-generated-prompts detects permissions or tool references
    in generated prompts that have no corresponding declaration
    in the agent's DSL definition.