Skip to content

sensor-gate.md

Latest commit

 

History

History
90 lines (72 loc) · 2.65 KB

sensor-gate.md

File metadata and controls

90 lines (72 loc) · 2.65 KB
Raw

Sensor Gate — Intent Classification & Chain Selection

The Sensor Gate is DreamCode's first-stage classification system. It runs on every user turn (before the LLM) to determine intent, select skills, and decide whether to spawn specialist agents.

How It Works

User Prompt
    ↓
Sensor Gate (Python)
  ├─ Intent classification
  ├─ Domain tagging
  ├─ Risk assessment
  ├─ Skill chain selection
  └─ Specialist spawn decision
    ↓
Guardian AI (risk approval)
    ↓
LLM + Skill Chain Execution

Gate Output

The sensor gate Python script produces a structured JSON result:

{
  "intent": "Fix authentication bug in login flow",
  "domain_tags": ["security", "backend", "api"],
  "risk_level": "medium",
  "confidence": 0.85,
  "complexity": "medium",
  "primary_skill": "debugging",
  "support_skills": ["testing", "security"],
  "chain": ["debugging", "testing", "security"],
  "guardian_decision": "APPROVED",
  "guardian_risk": "low",
  "mode": "DREAM_INNOVATION",
  "personas": [
    {"name": "The Detective", "role": "Root Cause Analysis", ...},
    {"name": "The Artisan", "role": "Code Quality", ...}
  ]
}

Spawn Decision

The spawn necessity algorithm (evaluateSpawnNecessity()) determines if specialists are needed:

Factor Weight Condition
Simplicity pattern Skip Prompt starts with fix/update/change/add/remove
Unique domains +3 3+ different domain tags
Unique domains +1 2 different domain tags
High risk +2 risk_level === "high"
Chain length +2-3 effectiveChainLen >= 2 (filtering "always" skills)
DREAM_INNOVATION +2 Innovation mode active
Code + long prompt +1 Code blocks present and prompt > 200 chars

Score → Spawns: ceil(score / 2), capped at 5.

Rate Limiting

  • 5 spawns per 5-minute rolling window per session
  • Enforced in code, not just prompt text
  • When limit hit: rate-limit warning injected, task handled directly

Bypass

The sensor gate is bypassed for:

  • Subagent sessions (they don't re-enter persona spawning)
  • Synthesis responses (synthetic messages don't re-trigger classification)
  • /compact commands (directly trigger context compaction)
  • Greeting detection (is_social_greeting: true skips chain execution)

Configuration

~/.config/dreamcode/config.yaml:

sensor_gate:
  enabled: true
  min_confidence: 0.5
  max_personas: 5

Related