pre-response-message-guard

Runtime enforcement extension for pre-response guidance and MRAA (Memory Recall Assurance Architecture).

Enforcement Path

1. message_sending hook calls beforeResponse() from skills/pre-response-hook/index.js.
2. Hook output is validated through requirePreResponseGuidance().
3. Fail-closed mode returns { cancel: true, reason } on any enforcement failure.

MRAA Gates

• Gate A (Retrieval Proof): validates metadata.mraa.retrieval_receipt for trigger-domain requirements.
• Gate B (Usage Proof): validates draft response usage against retrieved memories.
  • Requires memory reference metadata when retrieval is required.
  • Requires lexical overlap between draft response and retrieved snippets.
  • Blocks on critical omissions and contradictions.

Expected Payload Inputs

• User message candidates: event.prompt, event.message, event.content, ctx.prompt, ctx.lastUserMessage.
• Draft response candidates: event.response, event.assistantResponse, event.output, event.content, ctx.draftResponse.
• Draft metadata candidates: event.draftMetadata, event.responseMetadata, event.metadata, ctx.draftMetadata.

Key Reason Codes

• MRAA_REQUIRED_DOMAIN_MISSING_HIT:<domain>
• MRAA_MUST_CONSIDER_RECALL_MISSED:<slug...>
• MRAA_USAGE_REFERENCE_MISSING
• MRAA_USAGE_OVERLAP_TOO_LOW
• MRAA_CRITICAL_MEMORY_OMITTED:<slug...>
• MRAA_MEMORY_CONTRADICTION_DETECTED

Telemetry

• JSONL stream: state/mraa-telemetry.jsonl
• Summary JSON: state/mraa-summary.json
• Learning state: state/mraa-learning-state.json