gait-policy-test-rollout

name	gait-policy-test-rollout
description	Evaluate Gait policy changes safely with deterministic fixture tests and rollout simulation. Use when asked to validate allow or block behavior, check reason codes, compare policy outcomes, or plan enforce-by-stage rollout.
disable-model-invocation	true

Policy Test Rollout

Execute this workflow to validate policy behavior before enforcement.

Gait Context

Gait is an offline-first runtime for AI agents that enforces tool-boundary policy, emits signed and verifiable evidence artifacts, and supports deterministic regressions.

Use this skill when:

• incident triage involves policy decision verification
• CI gate failures require deterministic policy verdict checks
• receipt/evidence generation needs policy verdict traceability

Do not use this skill when:

• Gait CLI is unavailable in the environment
• no policy fixture inputs or Gait artifacts are available

Workflow

1. Require both files:
   • <policy.yaml>
   • <intent_fixture.json>
2. Run deterministic policy test (required):
   • gait policy test <policy.yaml> <intent_fixture.json> --json
3. Parse and report fields:
   • ok, policy_digest, intent_digest, verdict, reason_codes, violations, summary
4. If rollout simulation is requested, run:
   • gait gate eval --policy <policy.yaml> --intent <intent_fixture.json> --simulate --json
5. Return structured rollout recommendation:
   • current verdict
   • blocking reasons or required approvals
   • suggested next stage (observe, require_approval, enforce)

Exit Code Contract

• 0: allow
• 3: block
• 4: require approval
• 6: invalid input

Safety Rules

• Never bypass policy test by inferring results from YAML text.
• For replay workflows, prefer gait run replay (stub mode default); require explicit unsafe flags for real tool replay.
• Never claim a policy digest or verdict without command output.
• Keep simulation and enforcement clearly separated in reporting.
• Keep this skill wrapper-only: do not re-implement policy parsing or verdict logic.

Determinism Rules

• Always use --json outputs.
• Report exact reason_codes and violations as emitted.
• Preserve fixture-based evaluation flow for repeatability.