grimoire-intent-routing

name	grimoire-intent-routing
description	Routage d'intention avec scoring de confiance pour sélectionner skill/agent/profil. Use when: intent classification, routing decision, dispatch strategy, choose best skill, ambiguous request, confidence-based routing. Produces deterministic routing decisions with explicit confidence and fallback path.

Intent Routing

Classify user intent and route execution with confidence-aware decisions.

When to Use

• Request is ambiguous or spans multiple domains
• Several skills could match and prioritization is required
• Need deterministic routing justification for auditability
• User asks which path/agent/skill is most appropriate

Intent Model

Classify intent into one primary class and optional secondary classes:

• Implement
• Debug
• Review
• Plan
• Validate
• Explore
• Operate

Procedure

Step 1 — Extract Signals

Use only observable inputs:

• Explicit verbs in user prompt
• Referenced files/symbols
• Error outputs and failing checks
• Session context and current phase

Step 2 — Score Candidate Routes

For each candidate route, assign:

• Relevance score (0-100)
• Risk alignment score (0-100)
• Evidence completeness score (0-100)

Compute route confidence:

$$ R = 0.5Relevance + 0.3RiskAlignment + 0.2Evidence $$

Step 3 — Select Strategy

• R >= 75: direct route to best skill/path
• 50 <= R < 75: route with explicit fallback and one guard check
• R < 50: ask one batched clarification or split into phased execution

Step 4 — Emit Routing Card

## Routing Card

- Primary intent: ...
- Candidate routes: A (XX), B (YY), C (ZZ)
- Selected route: ...
- Confidence: XX/100
- Fallback route: ...
- Reasoning basis: observed signals only

Guardrails

• No hidden routing decisions on low confidence
• Avoid route flapping: keep route stable until a new strong signal appears
• If request becomes safety-critical, force verification path before execution