Agent Decision Context Normalizer

Agent Decision Context Normalizer

Normalize a redacted agent decision into a stable evidence contract before any policy review.

Scenario and prerequisites

• - Use for agent actions that need a later explanation of the input, policy snapshot, permissions and result.
• Python 3.10 or newer is required. The script uses only the standard library.
• Input must be authorized, redacted and free of secrets, customer records and production credentials.

Input and output

Input JSON must contain run_id, agent_id, event_time, policy_version, policy_snapshot, decision_input, permissions, action, result_state and evidence_refs.

Output JSON contains READY or BLOCKED, decision_id, context_hash, policy_hash, normalized fields and explicit reasons.

Installation and execution

1. 1. Copy the skill directory into the local skills directory.
2. Run python scripts/normalize_agent_decision.py --input examples/input.json --output examples/output.json.
3. Pass only a READY output to the policy binding step.

Recovery and boundaries

Missing fields, secret-like values, invalid JSON or empty evidence references produce BLOCKED. Fix the redacted source event and rerun; do not bypass the block or infer missing context.

The script never calls an agent, policy engine or external service. It hashes the supplied snapshot; it does not prove that the snapshot was authentic.

Acceptance

The included sample must produce status=READY, a stable decision_id, a non-empty context_hash and policy_hash. Changing the policy snapshot must change policy_hash.