feat(routing): add routing_rules schema, condition types, and action types (M4-001/002/003)

- Add routing_rules table to packages/db with scope, priority, conditions (jsonb), and action (jsonb) columns; generate migration 0004
- Define RoutingCondition, RoutingAction, RoutingRule, TaskClassification, and RoutingDecision types in apps/gateway/src/agent/routing/routing.types.ts
- Expand @mosaic/types routing/index.ts to export all M4 types (TaskType, Complexity, Domain, CostTier+local, Capability) alongside existing RoutingCriteria/RoutingResult
- Fix pre-existing type errors in routing.service.ts (local CostTier) and default-rules.ts (count optional chaining, unknown cast)
- Fix pre-existing Prettier violations in agent module and provider files

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

packages/types/src/routing/index.ts

@@ -1,10 +1,15 @@
-/** Cost tier for model selection */
-export type CostTier = 'cheap' | 'standard' | 'premium';
+// ─── Legacy simple-routing types (kept for backward compatibility) ────────────
 
-/** Task type hint for routing */
-export type TaskType = 'chat' | 'coding' | 'analysis' | 'summarization' | 'general';
+/** Result of a simple scoring-based routing decision */
+export interface RoutingResult {
+  provider: string;
+  modelId: string;
+  modelName: string;
+  score: number;
+  reasoning: string;
+}
 
-/** Routing criteria for model selection */
+/** Routing criteria for score-based model selection */
 export interface RoutingCriteria {
   taskType?: TaskType;
   costTier?: CostTier;
@@ -15,11 +20,115 @@ export interface RoutingCriteria {
   preferredModel?: string;
 }
 
-/** Result of a routing decision */
-export interface RoutingResult {
-  provider: string;
-  modelId: string;
-  modelName: string;
-  score: number;
-  reasoning: string;
+// ─── Classification primitives (M4-002) ──────────────────────────────────────
+
+/** Category of work the agent is being asked to perform */
+export type TaskType =
+  | 'chat'
+  | 'coding'
+  | 'research'
+  | 'summarization'
+  | 'conversation'
+  | 'analysis'
+  | 'creative'
+  | 'general';
+
+/** Estimated complexity of the task, used to bias toward cheaper or more capable models */
+export type Complexity = 'simple' | 'moderate' | 'complex';
+
+/** Primary knowledge domain of the task */
+export type Domain = 'frontend' | 'backend' | 'devops' | 'docs' | 'general';
+
+/**
+ * Cost tier for model selection.
+ * `local` targets self-hosted/on-premises models.
+ */
+export type CostTier = 'cheap' | 'standard' | 'premium' | 'local';
+
+/** Special model capability required by the task */
+export type Capability = 'tools' | 'vision' | 'long-context' | 'reasoning' | 'embedding';
+
+// ─── Condition types (M4-002) ─────────────────────────────────────────────────
+
+/**
+ * A single predicate that must be satisfied for a routing rule to match.
+ *
+ * - `eq`       — scalar equality: `field === value`
+ * - `in`       — set membership: `value` (array) contains `field`
+ * - `includes` — array containment: `field` (array) includes `value`
+ */
+export interface RoutingCondition {
+  /** The task-classification field to test */
+  field: 'taskType' | 'complexity' | 'domain' | 'costTier' | 'requiredCapabilities';
+  /** Comparison operator */
+  operator: 'eq' | 'in' | 'includes';
+  /** Expected value or set of values */
+  value: string | string[];
+}
+
+// ─── Action types (M4-003) ────────────────────────────────────────────────────
+
+/**
+ * The routing action to execute when all conditions in a rule are satisfied.
+ */
+export interface RoutingAction {
+  /** LLM provider identifier, e.g. `'anthropic'`, `'openai'`, `'ollama'` */
+  provider: string;
+  /** Model identifier, e.g. `'claude-opus-4-6'`, `'gpt-4o'` */
+  model: string;
+  /** Optional: use a specific pre-configured agent config from the agent registry */
+  agentConfigId?: string;
+  /** Optional: override the agent's default system prompt for this route */
+  systemPromptOverride?: string;
+  /** Optional: restrict the tool set available to the agent for this route */
+  toolAllowlist?: string[];
+}
+
+/**
+ * Full routing rule as stored in the database and used at runtime.
+ */
+export interface RoutingRule {
+  /** UUID primary key */
+  id: string;
+  /** Human-readable rule name */
+  name: string;
+  /** Lower number = evaluated first; unique per scope */
+  priority: number;
+  /** `'system'` rules apply globally; `'user'` rules override for a specific user */
+  scope: 'system' | 'user';
+  /** Present only for `'user'`-scoped rules */
+  userId?: string;
+  /** All conditions must match for the rule to fire */
+  conditions: RoutingCondition[];
+  /** Action to take when all conditions are met */
+  action: RoutingAction;
+  /** Whether this rule is active */
+  enabled: boolean;
+}
+
+/**
+ * Structured representation of what an agent has been asked to do,
+ * produced by the task classifier and consumed by the routing engine.
+ */
+export interface TaskClassification {
+  taskType: TaskType;
+  complexity: Complexity;
+  domain: Domain;
+  requiredCapabilities: Capability[];
+}
+
+/**
+ * Output of the routing engine — which model to use and why.
+ */
+export interface RoutingDecision {
+  /** LLM provider identifier */
+  provider: string;
+  /** Model identifier */
+  model: string;
+  /** Optional agent config to apply */
+  agentConfigId?: string;
+  /** Name of the rule that matched, for observability */
+  ruleName: string;
+  /** Human-readable explanation of why this rule was selected */
+  reason: string;
 }