| name | hermes-review-loop-validation |
| description | Validate the Hermes supervisory critic review loop: verify Hermes receives correct context, invokes HermesReviewer.evaluate(), and preserves learning signals for later persistence. Use when modifying hermes_feedback.py, the supervisory review path, or the review pipeline. |
Hermes Review Loop Validation
When to Use
- Modifying
src/hermes_feedback.py
- Changing Hermes-supervised context/backstory wiring in
src/swarm_agency.py
- Integrating the real hermes-agent client
- Verifying FTS5 persistence after reviews
Review Loop Contract
Execution output
|
v
Hermes supervisory critic path
|
v
HermesReviewer.evaluate(code, context)
|
v
FTS5 DB (persist learning signals)
Validation Checklist
- Hermes review logic must reference critique and learning signals, not broad execution ownership.
- Hermes review path must not absorb Droid execution responsibilities.
HermesReviewer.evaluate() must receive the actual code output, not a summary.
HermesReviewer.evaluate() must receive repo-grounded context (from ContextBuilder / ProjectContext render path).
- When the real hermes-agent client is integrated:
- Verify it writes to
.hermes/ directory
- Verify
.hermes/ is in IGNORE_DIRS so agents don't ingest the DB as source
- The evaluate response must include structured fields:
ok, notes, code, context at minimum.
Current State
HermesReviewer.evaluate() is implemented with a stable schema and conservative
fallback behavior. Durable FTS5 learning persistence remains deferred.