| name | self-healing |
| description | Auto-recovery patterns for marketing automations — token refresh, retry policies, fallback channels, dead-letter queues. |
| canon_references | ["ritson-diagnosis"] |
| canon_source | brain/MARKETING_CANON.md |
| universal | true |
| note | Examples in this skill may reference SunBiz (legacy client); the skill itself is brand-agnostic. Per-brand context lives in brain/clients/<brand>.md. |
| triggers | ["automation is broken fix it","token expired auto-recover","campaign stopped delivering fix","retry failed sends","run self-healing check"] |
SKILL: Self-Healing
5-Dimensional autonomous recovery and maintenance system.
5 Dimensions of Self-Healing
Dimension 1: Memory Self-Healing
Detect: Contradictions between memory files, stale data, bloat
Heal:
- Contradictions → Resolve using most recent verified data
- Stale campaign data → Refresh from API
- Memory bloat → Archive old sessions, compress patterns
Trigger: After memory writes, session start, session end
Dimension 2: Context Self-Healing
Detect: Outdated campaign references, wrong account IDs, stale metrics
Heal:
- Outdated campaign data → Pull fresh from API
- Wrong references → Correct from source of truth (API)
- Stale metrics → Flag as stale, pull fresh
Trigger: Before any campaign operation
Dimension 3: Skill Self-Healing
Detect: Failed operations, declining success rates, unused skills
Heal:
- Repeated failures → Check if API changed, update skill documentation
- Low success rate → Investigate root cause, update approach
- Unused skills → No action (keep for future use)
Trigger: After failures, monthly review
Dimension 4: Infrastructure Self-Healing
Detect: MCP failures, API auth issues, missing env vars, git issues
Heal:
- MCP down → Fall back to Python SDK
- Auth expired → Alert user for token refresh
- Missing env vars → Report which vars are missing
- Git conflicts → Report to user, suggest resolution
Trigger: Session start (health check), after MCP errors
Dimension 5: Campaign Self-Healing
Detect: Rejected ads, learning phase stuck, delivery issues, budget anomalies
Heal:
- Rejected ad → Read rejection reason, suggest compliant alternative
- Learning stuck (>7 days) → Check if sufficient conversions, suggest broadening
- No delivery → Check targeting, budget, bid, ad status
- Budget anomaly → Alert user with recommended action
Trigger: Health check, daily monitoring
Severity Tiers
Tier 1: Auto-Fix (No User Approval Needed)
- Memory file formatting fixes
- Stale reference cleanup
- Session log compression
- Cache refresh
Tier 2: Diagnose & Suggest
- MCP auth failures (suggest fix, user applies)
- Campaign delivery issues (diagnose, suggest fix)
- Performance degradation (analyze, suggest optimization)
Tier 3: Deep Investigation
- Recurring API failures (root cause analysis)
- Systematic performance decline (full audit)
- Memory inconsistencies (cross-reference audit)
Tier 4: Escalate to User
- Token rotation required
- Budget changes >20%
- Campaign structural changes
- Compliance concerns
Self-Healing Checklist (Run at Session End)
