claude-flow
Version:
Ruflo - Enterprise AI agent orchestration for Claude Code. Deploy 60+ specialized agents in coordinated swarms with self-learning, fault-tolerant consensus, vector memory, and MCP integration
55 lines (50 loc) • 2.26 kB
text/typescript
/**
* @claude-flow/browser - Production-Aware UCT (ADR-122 Phase 7)
*
* Extends Phase 4's plain UCB1 with the substrate's production-aware formula:
*
* score = Q + C·√(ln(parent_visits) / child_visits)
* + λ_R · replayability
* − λ_risk · risk
* − μ_cost · cost_usd
* − α_auth · auth_fragility
*
* Q = mean trajectory value (success / similarity)
* R = replayability bonus (encourages reusable winning paths)
* λ_R = replay weight
*
* The penalties keep MCTS from chasing high-Q paths that are expensive,
* irreversible, or auth-fragile in production.
*/
import type { ProductionUctSignals, ProductionUctWeights } from '../domain/workflow.js';
import { DEFAULT_PRODUCTION_UCT_WEIGHTS } from '../domain/workflow.js';
export interface ProductionUctInput {
/** Visits for the candidate branch. */
visits: number;
/** Total visits across all siblings (for the ln(N) term). */
parentVisits: number;
signals: ProductionUctSignals;
}
/** Compute the production-aware UCT score. Unvisited branches return +Infinity. */
export function productionUct(
input: ProductionUctInput,
weights: ProductionUctWeights = DEFAULT_PRODUCTION_UCT_WEIGHTS,
): number {
if (input.visits === 0) return Number.POSITIVE_INFINITY;
const exploitation = input.signals.qValue;
const exploration = weights.c * Math.sqrt(Math.log(Math.max(1, input.parentVisits)) / input.visits);
const replayBonus = weights.replayBonus * input.signals.replayability;
const riskPenalty = weights.riskPenalty * input.signals.risk;
const costPenalty = weights.costPenalty * input.signals.costUsd;
const authPenalty = weights.authPenalty * input.signals.authFragility;
return exploitation + exploration + replayBonus - riskPenalty - costPenalty - authPenalty;
}
/**
* Compose two scoring sources — useful when a branch carries both a raw
* scorer value (HNSW similarity, Phase 4) and observed signals (cost, auth).
*/
export function blendQ(scorerValue: number, replaySuccessRate?: number): number {
if (replaySuccessRate === undefined) return scorerValue;
// Linear blend so both sources matter — replay carries 30% weight by default.
return 0.7 * scorerValue + 0.3 * replaySuccessRate;
}