HERO Query Healer and Learned Queries
The healer proposes reusable redacted query templates for frequent failure clusters. It never activates a template automatically. Environment requirements The task requires dedicated DEV values:...
Last updated
The healer proposes reusable redacted query templates for frequent failure clusters. It never activates a template automatically.
Environment requirements
The task requires dedicated DEV values:
HERO_HEALER_SUPABASE_URLHERO_HEALER_SUPABASE_SECRET_KEY
The URL must resolve to the approved DEV Supabase project. Missing values, a different host, or ambient runtime credentials cause the task to abort before database or model work.
Attempt policy
Each cluster has at most five lifetime model attempts:
- Sonnet
- Sonnet
- Sonnet
- Opus, only after all three Sonnet attempts failed
- Opus
Every call reserves a durable attempt slot and conservative cost before provider execution. Outstanding reservations count toward the global UTC-day budget. Successful validation creates a pending_review pair; an admin must approve it.
Safe source requirements
A candidate needs retained trusted source context. Redacted-only clusters without usable source context do not spend model budget. Before save or approval, the system checks:
- canonical redacted question and SQL templates;
- typed
:property_id; - allowed table manifest;
- canonicalization and policy versions;
- permissions and selected table context; and
- SQL validation.
Interpreting outcomes
- Pending review: validated proposal awaiting a platform admin.
- Flagged: rejected or invalidated; requires reason and reapproval.
- Active: explicitly approved and still revalidated at retrieval time.
- Healed cluster: a validated pending proposal was committed with the cluster transition.
Provider failure and validation failure are retained separately with actual recorded cost where available.