@vellumai/assistant 0.4.18 → 0.4.20

package/src/daemon/response-tier.ts

@@ -1,250 +0,0 @@
-/**
- * Per-turn response tier classification.
- *
- * Classifies each user message into a tier that controls:
- * - maxTokens budget for the LLM response
- * - Which system prompt sections are included
- *
- * Two layers:
- * 1. Deterministic regex/heuristic (zero latency, runs every turn)
- * 2. Background Haiku classification (fire-and-forget, advises future turns)
- */
-
-import { createTimeout, extractText, getConfiguredProvider, userMessage } from '../providers/provider-send-message.js';
-import { getLogger } from '../util/logger.js';
-
-const log = getLogger('response-tier');
-
-export type ResponseTier = 'low' | 'medium' | 'high';
-
-export type TierConfidence = 'high' | 'low';
-
-export interface TierClassification {
-  tier: ResponseTier;
-  reason: string;
-  confidence: TierConfidence;
-}
-
-export interface SessionTierHint {
-  tier: ResponseTier;
-  turn: number;
-  timestamp: number;
-}
-
-// ── Patterns ──────────────────────────────────────────────────────────
-
-const GREETING_PATTERNS = /^(hey|hi|hello|yo|sup|hiya|howdy|what'?s up|thanks|thank you|thx|ty|cheers|what can you|who are you|how are you)\b/i;
-
-const BUILD_KEYWORDS = /\b(build|implement|create|refactor|debug|deploy|migrate|scaffold|architect|redesign|generate|write|develop|fix|convert|add|remove|update|modify|change|delete|replace|integrate|setup|install|configure|optimize|rewrite)\b/i;
-
-const SURFACE_ACTION = /^\[User action on \w+ surface:/;
-const CODE_FENCE = /```/;
-const FILE_PATH = /(?:^|[\s"'(])(?:\/|~\/|\.\/)\S/;
-const MULTI_PARAGRAPH = /\n\s*\n/;
-
-// ── Confidence thresholds ─────────────────────────────────────────────
-
-const HINT_MAX_TURN_AGE = 4;
-const HINT_MAX_AGE_MS = 5 * 60 * 1000; // 5 minutes
-
-/**
- * Classify the complexity tier of a user message (backward-compat wrapper).
- */
-export function classifyResponseTier(message: string, _turnCount: number): ResponseTier {
-  return classifyResponseTierDetailed(message, _turnCount).tier;
-}
-
-/**
- * Classify with confidence scoring. High confidence means the regex
- * matched an unambiguous signal; low confidence means the message
- * fell through to the default medium bucket.
- */
-export function classifyResponseTierDetailed(message: string, _turnCount: number): TierClassification {
-  const trimmed = message.trim();
-  const len = trimmed.length;
-
-  const isPoliteImperative = /^(can|could|would|will)\s+you\s+/i.test(trimmed) && BUILD_KEYWORDS.test(trimmed);
-
-  const isQuestion = !isPoliteImperative && (
-    /\?$/.test(trimmed) || /^(what|who|where|when|why|how|which|can|could|should|would|is|are|do|does|did|will|has|have)\b/i.test(trimmed)
-  );
-
-  // ── High signals (any match → high tier, high confidence) ──
-  if (SURFACE_ACTION.test(trimmed)) return tagged('high', 'surface_action', 'high');
-  if (len > 500) return tagged('high', 'length>500', 'high');
-  if (CODE_FENCE.test(trimmed)) return tagged('high', 'code_fence', 'high');
-  if (FILE_PATH.test(trimmed)) return tagged('high', 'file_path', 'high');
-  if (MULTI_PARAGRAPH.test(trimmed)) return tagged('high', 'multi_paragraph', 'high');
-  if (!isQuestion && BUILD_KEYWORDS.test(trimmed)) return tagged('high', 'build_keyword', 'high');
-
-  // ── Low signals (any match → low tier, high confidence) ──
-  if (GREETING_PATTERNS.test(trimmed) && len < 40 && !BUILD_KEYWORDS.test(trimmed)) return tagged('low', 'greeting', 'high');
-  if (len < 80 && !BUILD_KEYWORDS.test(trimmed)) return tagged('low', 'short_no_keywords', 'high');
-
-  // ── Default (low confidence — ambiguous) ──
-  return tagged('medium', 'default', 'low');
-}
-
-const TIER_RANK: Record<ResponseTier, number> = { low: 0, medium: 1, high: 2 };
-
-/**
- * Resolve the final tier using the regex classification and an optional
- * session hint from a previous background Haiku call.
- *
- * - When confidence is low, defer to a fresh hint (upgrade or downgrade).
- * - When confidence is high, still upgrade if the hint is higher (the
- *   conversation trajectory outranks a short-message heuristic), but
- *   never downgrade.
- */
-export function resolveWithHint(
-  classification: TierClassification,
-  hint: SessionTierHint | null,
-  currentTurn: number,
-): ResponseTier {
-  if (!hint) {
-    return classification.tier;
-  }
-
-  const turnAge = currentTurn - hint.turn;
-  const timeAge = Date.now() - hint.timestamp;
-
-  if (turnAge > HINT_MAX_TURN_AGE || timeAge > HINT_MAX_AGE_MS) {
-    log.debug({ turnAge, timeAge }, 'Session tier hint is stale, ignoring');
-    return classification.tier;
-  }
-
-  if (classification.confidence === 'low') {
-    // Low confidence: fully defer to hint
-    log.info(
-      { regexTier: classification.tier, hintTier: hint.tier, turnAge },
-      'Deferring to session tier hint (low confidence)',
-    );
-    return hint.tier;
-  }
-
-  // High confidence: only upgrade, never downgrade
-  if (TIER_RANK[hint.tier] > TIER_RANK[classification.tier]) {
-    log.info(
-      { regexTier: classification.tier, hintTier: hint.tier, turnAge },
-      'Upgrading tier via session hint',
-    );
-    return hint.tier;
-  }
-
-  return classification.tier;
-}
-
-// ── Async Haiku classification ────────────────────────────────────────
-
-const ASYNC_CLASSIFICATION_TIMEOUT_MS = 8_000;
-
-const TIER_SYSTEM_PROMPT =
-  'Classify the overall complexity of this conversation. ' +
-  'Output ONLY one word, nothing else.\n' +
-  'low — greetings, thanks, short acknowledgements\n' +
-  'medium — simple questions, short requests, clarifications\n' +
-  'high — build/implement/refactor requests, multi-step tasks, code-heavy work';
-
-/**
- * Fire-and-forget Haiku call to classify the conversation trajectory.
- * Returns the classified tier, or undefined when no provider is configured
- * or on any failure.
- */
-export async function classifyResponseTierAsync(
-  recentUserTexts: string[],
-): Promise<ResponseTier | undefined> {
-  const provider = getConfiguredProvider();
-  if (!provider) {
-    log.debug('No provider available for async tier classification');
-    return undefined;
-  }
-
-  const combined = recentUserTexts
-    .map((t, i) => `[Message ${i + 1}]: ${t}`)
-    .join('\n');
-
-  try {
-    const { signal, cleanup } = createTimeout(ASYNC_CLASSIFICATION_TIMEOUT_MS);
-    try {
-      const response = await provider.sendMessage(
-        [userMessage(combined)],
-        undefined,
-        TIER_SYSTEM_PROMPT,
-        {
-          config: {
-            modelIntent: 'latency-optimized',
-            max_tokens: 8,
-          },
-          signal,
-        },
-      );
-      cleanup();
-
-      const raw = extractText(response).toLowerCase();
-      const match = raw.match(/\b(low|medium|high)\b/);
-      if (match) {
-        const tier = match[1] as ResponseTier;
-        log.debug({ tier, raw }, 'Async tier classification result');
-        return tier;
-      }
-
-      log.debug({ raw }, 'Async tier classification returned unexpected value');
-      return undefined;
-    } finally {
-      cleanup();
-    }
-  } catch (err) {
-    const message = err instanceof Error ? err.message : String(err);
-    log.debug({ err: message }, 'Async tier classification failed');
-    return undefined;
-  }
-}
-
-function tagged(tier: ResponseTier, reason: string, confidence: TierConfidence): TierClassification {
-  log.debug({ tier, reason, confidence }, 'Classified response tier');
-  return { tier, reason, confidence };
-}
-
-// ── Token scaling ─────────────────────────────────────────────────────
-
-const TIER_SCALE: Record<ResponseTier, number> = {
-  low: 0.125,
-  medium: 0.375,
-  high: 1,
-};
-
-/**
- * Scale the configured max tokens ceiling by the tier multiplier.
- *
- * Examples with configuredMax = 16000:
- *   low    → 2000
- *   medium → 6000
- *   high   → 16000
- */
-export function tierMaxTokens(tier: ResponseTier, configuredMax: number): number {
-  return Math.round(configuredMax * TIER_SCALE[tier]);
-}
-
-// ── Model routing ─────────────────────────────────────────────────────
-
-/**
- * Map for Anthropic provider: tier → model.
- *   low    → sonnet (balanced)
- *   medium → sonnet (balanced)
- *   high   → undefined (use configured default, typically opus)
- */
-const ANTHROPIC_TIER_MODELS: Record<ResponseTier, string | undefined> = {
-  low: 'claude-sonnet-4-6',
-  medium: 'claude-sonnet-4-6',
-  high: undefined, // use configured default
-};
-
-/**
- * Returns a model override for the given tier, or undefined to use the
- * provider's configured default. Only applies model downgrading for
- * the Anthropic provider.
- */
-export function tierModel(tier: ResponseTier, providerName: string): string | undefined {
-  if (providerName !== 'anthropic') return undefined;
-  return ANTHROPIC_TIER_MODELS[tier];
-}