OpenAI Agents Onboarding Playbook

Permanent-form status

The commands below exercise current generated contracts only. They do not exercise the permanent PF-01/PF-02 condition and evidence model, and the resulting artifacts cannot establish initial-profile support or launch readiness. The permanent LLM contract is documented in llm_native_playbook.md; its executable projections remain PF-09 work.

Purpose

This playbook is the correctness-first onboarding path for new teams using OpenAI Agents with Decision Gate.

Onboarding-harness readiness is based on its named deterministic checks and artifact validity, not elapsed calendar time. Passing this playbook does not qualify a production deployment profile or close Decision Gate’s permanent- form distributed-commit, residency, or effect-delivery gaps.

Canonical Inputs

• Protocol contract: Docs/guides/llm_native_playbook.md
• Onboarding fixture pack: examples/agentic/onboarding/basic
• Lifecycle reference: examples/python/basic_lifecycle.py
• System evidence: system-tests/src/suites/monotone_dag.rs

End-to-End Flow

1. Capability discovery:
   • read the generated tool/schema resources;
   • inspect the exact configured local-acquisition authorities;
   • do not infer unavailable network/provider capability from MCP transport.
2. Authoring artifacts:
   • claim_inventory
   • capability_matrix
   • claim_condition_map
3. Fast loop:
   • decision_gate_schemas_register
   • decision_gate_scenario_precheck_stage
4. Live boundary:
   • decision_gate_scenario_define
   • decision_gate_scenario_start
   • decision_gate_scenario_open_stage
   • decision_gate_scenario_evaluate_stage
5. Verification:
   • decision_gate_runpack_export
   • decision_gate_runpack_verify

Runbook

1. Start Decision Gate MCP endpoint.
2. Ensure OpenAI Agents adapter dependencies are installed.
3. Run the onboarding loop:

bash scripts/adapters/adapter_tests.sh --frameworks=openai_agents --validate

4. Inspect output JSON for required artifacts and verification status.

Correctness Acceptance

This bounded onboarding exercise is complete only if all checks pass:

1. Required artifacts are present and structurally valid:
   • claim_inventory
   • capability_matrix
   • claim_condition_map
   • enforcement_verdict
2. The exact bound run reaches accepted status completed.
3. runpack_verify.status is passed.
4. Blocking scenarios emit explicit blocking reasons.

The public runpack_verify result revalidates scenario law and reconstructible retained evidence through hostile constructors before replaying accepted history. It does not by itself prove non-retained source facts, distributed commit, delivery, external binding, or nonrepudiation.

Failure Taxonomy

Classify failures into deterministic buckets:

• capability_mismatch
• schema_mismatch
• comparator_or_type_mismatch
• evidence_policy_unsatisfied
• runpack_verify_failed

Do not use an unclassified unknown bucket for release decisions.