continuous-learning

ProductionOS Continuous Learning

Architecture

SESSION ACTIVITY (via PostToolUse hooks)
       ↓
  Observation Log (~/.productionos/analytics/skill-usage.jsonl)
       ↓
  Pattern Extractor (runs at session end via Stop hook)
       ↓
  Instinct Creation (~/.productionos/instincts/project/{hash}/)
       ↓
  Confidence Scoring (0.0 - 1.0, based on repetition + outcome)
       ↓
  Auto-Promotion (confidence > 0.8 → global instinct)
       ↓
  Global Instincts (~/.productionos/instincts/global/)

Instinct Format

Each instinct is a single markdown file:

---
name: {pattern-name}
confidence: 0.65
observations: 3
first_seen: 2026-03-20
last_seen: 2026-03-20
scope: project
promoted_from: null
---

## Pattern
{What was observed}

## Evidence
- Session {date}: {observation 1}
- Session {date}: {observation 2}

## Application
{When to apply this pattern in future sessions}

Confidence Scoring

Confidence	Meaning	Action
0.0 - 0.3	Weak signal	Store but don't act
0.3 - 0.6	Emerging pattern	Suggest when relevant
0.6 - 0.8	Strong pattern	Apply proactively
0.8 - 1.0	Proven pattern	Auto-promote to global

Observation Types

1. Code Pattern — Repeated code structure across sessions
2. Tool Usage — Preferred tool sequences (e.g., always Grep before Edit)
3. Error Recovery — How errors were resolved (reusable fix patterns)
4. Architecture Decision — Repeated structural choices
5. Review Finding — Common issues found in code reviews

Stop Hook: Pattern Extraction

At session end, the Stop hook:

1. Reads ~/.productionos/analytics/skill-usage.jsonl for this session
2. Groups events by type (edit, bash, security_edit, etc.)
3. Identifies patterns (repeated file types, common commands, recurring errors)
4. Creates or updates instinct files
5. Bumps confidence on existing instincts that match

Commands

• pos-instincts list — Show all instincts with confidence
• pos-instincts promote {name} — Manually promote to global
• pos-instincts demote {name} — Reduce confidence
• pos-instincts forget {name} — Delete an instinct

Red Flags — STOP If You See These

• Creating instincts from single observations (need 2+ sessions)
• Promoting instincts below 0.6 confidence
• Instincts that contradict CLAUDE.md or project conventions
• Learning patterns that are project-specific as global instincts
• Extracting patterns from error sessions (failed attempts aren't patterns)