2.1 KiB
Root Cause Analysis: Deterministic Engine Bouncer & Authorization Gate
- Executive Summary
- 1. Issue: Automated High-Risk Execution
- 2. Design Decision: Org-native Approval
- 3. Permanent Learnings
Executive Summary
Implemented the "Planning Mode" Bouncer to intercept high-risk Probabilistic Engine proposals (e.g., shell commands, Lisp evaluation). The system now forces these actions into an asynchronous "Flight Plan" Org node for manual Sovereign approval, fulfilling the "everything is a node" and high-integrity mandates.
1. Issue: Automated High-Risk Execution
Symptoms
Probabilistic Engine proposals involving `shell` or `eval` were executed immediately upon passing the `decide` gate's safety harness. This lacked human-in-the-loop oversight for irreversible or complex operations.
Root Cause
Architecture gap. The system lacked an authorization state between "Safe" and "Executed".
Resolution
- Interceptor: Added `bouncer-check` to `deterministic.lisp`. It flags high-risk actions that lack the `:approved t` property.
- Asynchronous Event: If flagged, the harness emits an `:approval-required` event.
-
Flight Plan Skill: Created `org-skill-bouncer.org` to:
- Catch the event and create a serialized Org node with state `PLAN`.
- Monitor the Memory for `APPROVED` states.
- Re-inject approved actions with the `:approved t` bypass flag.
2. Design Decision: Org-native Approval
Requirement
Align with "Homoiconic Memory" and "Lisp Machine Sovereignty".
Selected Path
State-Based Approval (Org-native).
- Pros: Auditable, asynchronous, utilizes existing Org-mode workflows.
- Cons: Slightly more latency than an interactive prompt.
Alignment
Ensures that the agent's "Flight Plans" are first-class citizens in the Memex, allowing the Sovereign to review and approve them using standard GTD tools.
3. Permanent Learnings
- Serial Bypass: Always include a specific bypass flag (e.g., `:approved t`) when re-injecting intercepted actions to prevent infinite interception loops.
- Heartbeat Listeners: Periodic scanning of the Memory for state transitions is an effective way to implement asynchronous authorization gates without blocking the harness.