Skip to content

Audit and evaluate governed Muster operations without private reasoning #77

Description

@jusso-dev

Revised scope

Provide MCP/API-level observability and evaluation for governed Muster operations without reimplementing Hermes session inspection or storing private reasoning.

Required records

  • Hermes installation/profile identity where supplied and verified.
  • Muster tool name and version.
  • Organisation and initiating actor.
  • Input/output schema versions and redacted hashes.
  • Capability and approval decisions.
  • Connector calls, evidence references, retries and idempotency outcomes.
  • Final authoritative action/delivery state.

Expose bounded MCP tools for run/action status and audit export. Replay uses recorded results by default and cannot repeat external actions. Evaluations cover tenant isolation, schema compliance, approval behavior, evidence citation, injection resistance and unsupported claims.

Non-goals

  • Chain-of-thought capture.
  • Replaying Hermes' internal model loop.
  • A dedicated chat/run-inspector UI in the first release.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions