Codeman/@fix_plan.md

Ralph Loop Inception - Improvement Plan

This plan improves the Ralph Loop system to make it more reliable for 24+ hour autonomous runs.

Phase 1: Stuck-State Detection & Recovery (P0 - Critical)

• P0-001: Add stuck-state detection to RespawnController - detect when the same state persists for too long without progress
• P0-002: Add iteration stall detection to RalphTracker - detect when iteration count stops incrementing despite respawn cycles
• P0-003: Add automatic recovery action when stuck detected - escalate from soft reset to hard reset (implemented in handleStuckStateRecovery)
• P0-004: Add stuck-state metrics to DetectionStatus for UI visibility (added stuckState to DetectionStatus)

Phase 2: Enhanced Idle Detection (P0 - Critical)

• P0-005: Add confidence decay over time - if no definitive signal for extended period, gradually lower confidence threshold
• P0-006: Add Session.isWorking integration check before AI idle check - skip expensive AI call if session reports working
• P0-007: Add RALPH_STATUS block integration with respawn controller - use EXIT_SIGNAL for more reliable completion detection (already implemented)

Phase 3: Promise Detection Improvements (P1 - High)

• P1-001: Add fuzzy matching for completion phrases - handle minor variations like whitespace or case
• P1-002: Add promise phrase validation - warn if phrase is too common (likely false positives)
• P1-003: Add multi-phrase support - allow multiple valid completion phrases for complex workflows

Phase 4: Error Recovery & Resilience (P1 - High)

• P1-004: Add circuit breaker reset on successful iteration - prevent permanent disabled state
• P1-005: Add exponential backoff for AI check failures instead of immediate disable
• P1-006: Add session health check before respawn cycle - skip if session is in error state

Phase 5: Todo Tracking Improvements (P1 - High)

• P1-007: Add todo deduplication by content similarity - prevent duplicate todos from repeated output
• P1-008: Add todo priority inference from keywords - automatically set priority based on content
• P1-009: Add todo progress estimation - estimate completion based on historical patterns

Phase 6: Respawn Cycle Optimization (P2 - Medium)

• P2-001: Add adaptive timing based on session behavior - adjust timeouts based on observed patterns (uses rolling 75th percentile of idle detection times)
• P2-002: Add skip-clear optimization - skip /clear if context usage is low (below 30% by default)
• P2-003: Add smart kickstart prompt generation - use context to generate relevant kickstart prompts (deferred - requires AI generation)

Phase 7: Monitoring & Observability (P2 - Medium)

• P2-004: Add respawn cycle metrics - track success rate, average duration, failure reasons (RespawnCycleMetrics, RespawnAggregateMetrics types)
• P2-005: Add Ralph Loop health score - aggregate metric for loop reliability (calculateHealthScore() method with 5 component scores)
• P2-006: Add automated anomaly detection - alert on unusual patterns (deferred - requires statistical analysis)

Completion Criteria

All P0 and P1 tasks must be completed. P2 tasks are nice-to-have. Tests must pass after each change. Documentation must be updated.

When ALL P0 and P1 tasks are complete, output: RALPH_INCEPTION_COMPLETE