What is an AI Control Pack?

AI Control Pack is the enterprise-facing name for an OAP Policy Pack. Each pack defines rules for a capability, such as finance.payment.refund.v1, including limits, assurance requirements, and context validation.

How do I add a policy pack to my endpoint?

Install the Express or FastAPI middleware, configure it with your policy pack ID, and wrap your sensitive routes. The middleware calls the verify API before each request.

Can I customize policy pack limits?

Yes. Organizations can customize limits (e.g., max refund amount, export row count) per policy pack. Limits are enforced at verification time.

Back to Policy Packs

Task Completion Gate

Pre-action governance for an agent marking a task complete. Enforces that the agent has provided required deliverable evidence — summary, acceptance criteria attestations, test status, reviewer identity — before done is authorized.

v1.0.0

Requirements

Required Capabilities

deliverable.task.complete

Minimum Assurance Level

Required Limits

Enforcement Rules

These rules are automatically enforced when using this policy pack

criteria_attestation_enforced

summary_enforced

tests_enforced

reviewer_enforced

pattern_scan_enforced

Best Practices

Define acceptance_criteria in passport limits.deliverable.task.complete.acceptance_criteria — this is a new pattern, unique to this capability
acceptance_criteria is an array of {id, description} objects — no other capability uses this structure
Set require_summary: true and min_summary_words to enforce written summaries
Set require_tests_passing: true for code agents to block completion without passing tests
Set require_different_reviewer: true in multi-agent pipelines to prevent self-approval — also submit author_agent_id in context
Set scan_output: true with blocked_patterns to catch TODO, FIXME, placeholder, console.log (max 100 patterns enforced)
The deny_code tells the agent exactly what is missing — it can self-correct and retry
criteria_attestations must include non-empty evidence strings, not just met: true
Use task_id to correlate completion receipts with your task management system