feat: Continuous Learning / Instinct System

[dean0x]

Summary

Background agent that analyzes sessions to detect patterns and creates "instincts" — learned behaviors with confidence scores that compound over time. Builds on #99 (Persistent Project Knowledge).

Key Capabilities

• Stop hook captures observations (tool events, corrections, error resolutions)
• Background agent detects patterns: user corrections, repeated workflows, tool preferences
• Creates instincts with confidence 0.0–1.0, domain classification, scope (project vs global)
• /learn — extract pattern from current session
• /learn-eval — same but with quality gate (specificity, actionability, scope fit — all ≥ 3/5)
• /instinct-status — view instincts grouped by domain with confidence
• /evolve — cluster related instincts into higher-order structures
• /promote — move project instinct to global when cross-project pattern detected
• Temporal decay, import/export for team sharing

Context

Our PROJECT-PATTERNS.md is a crude version of this. The instinct system adds: confidence scoring, temporal decay, quality gates on learning, cross-project promotion, and structured import/export for team sharing.

Identified as the single biggest differentiator in competitive analysis (#107).

Effort

Large

Dependencies

• Persistent Project Knowledge (Stateful Agents / Persistent Minds) #99 (Persistent Project Knowledge) — foundation for knowledge storage

[dean0x]

Enrichment from EDD article analysis (session 23)

Two ideas from TDD to EDD worth incorporating:

1. Convergent evidence boosting

When multiple independent signals detect the same pattern (stop hook observation + user correction + repeated workflow), confidence should increase faster than single-signal detection. Analogous to Google's "three LLM judges" pattern — convergent evidence from independent sources is stronger than repeated evidence from one source.

Concrete: If an instinct is reinforced by 3 different signal types, boost confidence more aggressively than 3 reinforcements from the same signal type.

2. Instinct coverage tracking

Periodically measure which domains have instincts vs. which are blank spots. The article's "eval coverage" concept applied to learned behaviors — helps surface areas where the system hasn't learned anything yet (possibly because the user hasn't worked in that area, or because the stop hook isn't capturing relevant signals).

Concrete: /instinct-status should show domain coverage gaps, not just existing instincts.

[dean0x]

Implemented as the self-learning system. Core capabilities delivered:

• session-end-learning hook captures observations after sessions
• background-learning script analyzes batch transcripts (every 3 sessions, 5 at 15+ observations)
• Confidence scoring (0.0–1.0) with temporal decay and daily run caps
• Observations accumulate in .memory/learning-log.jsonl
• Auto-creates artifacts as slash commands (.claude/commands/self-learning/) or skills (.claude/skills/{slug}/)
• devflow learn --enable/--disable/--status/--configure/--reset/--purge CLI
• Toggleable via devflow init --learn/--no-learn

Some originally proposed commands (/evolve, /promote) were not implemented — the system handles observation clustering and artifact creation automatically instead.