Major Overhaul: Episode-Based Goal Emergence

The system replaces its old goal creation pipeline with episode-based emergence. Goals now cluster from episodes using KNN, temporal decay, and LLM synthesis instead of relying on regex intent classification and unmatched signal accumulation.

A new method, _check_goal_emergence(), runs after every episode creation to cluster data and synthesize goals via a background LLM.

Goals now track their origin via a derived_from column, and goal finding services have been updated to append this information.

The previous IntentClassifierService, GoalInferenceService, GoalSignalService, and related signal infrastructure have been deleted as they are superseded by the episode path.

The system now incorporates user timezone support, automatically detecting and persisting the IANA timezone to ensure all time-sensitive features operate in the user’s local context.

Methodology learning V2 implements an “always-create-new-record” design, where every ACT loop execution generates a new knowledge record to prevent embedding drift.

This knowledge compounds by injecting existing_goal_guidance into the reflection process, allowing each new learning record to build upon prior related learnings.

Reliability tracking is updated by replacing CriticService with PostLoopReflectionService, shifting learning focus from tool success to overall goal approach methodology.

• Replaced chat-based goal inference with episode-based goal emergence using KNN and temporal decay.

• Deleted IntentClassifierService and GoalInferenceService.

• Implemented user timezone support, persisting IANA timezone for all time-sensitive features.

• Introduced methodology learning v2: every ACT loop creates a new, compounding knowledge record.

• Replaced CriticService with PostLoopReflectionService for goal-level methodology learning.

• Updated scheduler skill to return structured JSON responses for better LLM interaction.