intelligence-learning

Overview

The learning system detects patterns in user behavior and plugin outcomes, then injects learned preferences into future plugin runs. It operates in three tiers shipped sequentially:

1. Taste Learning (Tier 1) — output preference detection
2. Workflow Optimization (Tier 2) — plugin chain detection
3. Confidence-Gated Autonomy (Tier 3) — graduated trust

The Cycle

OBSERVE → RETRIEVE → JUDGE → DISTILL → CONSOLIDATE → APPLY
   ↑                                                    │
   └────────────────── feedback ─────────────────────────┘

Stage	What Happens	When
Observe	Hooks capture events	During command run
Retrieve	Pull relevant past patterns	Before next run
Judge	Compare outcome to patterns	After post_command
Distill	Extract generalizable insight	After 3+ observations
Consolidate	Store with confidence score	After distillation
Apply	Inject as instruction	Before next matching run

Confidence Mechanics

• Formula: confirmations / max(1, confirmations + rejections * 2)
• New patterns start at confidence 0.0 (candidate)
• Applied when confidence >= threshold AND observations >= 3
• Rejections weigh double to prevent bad patterns from persisting

Pattern Lifecycle

candidate (conf < 0.5) → active (conf >= 0.5, auto-applied with notice)
                       → approved (user /founder-os:intel:approve, conf locked at 1.0)
                       → rejected (user rejected or conf < 0.3)

Applying Patterns

When a plugin command runs:

1. Query patterns table for matching plugin + command (or cross-plugin patterns)
2. Filter to status IN ('active', 'approved')
3. For each matching pattern, inject its instruction text into the command context
4. Notify user: [Intel] Applying learned preference: "{description}"

Cross-Plugin Promotion

Patterns referencing a Company or Contact are promoted to the Memory Engine:

• taste + company → Memory Engine category preference
• workflow → Memory Engine category workflow
• Confidence conversion: round(confidence * 100)