@goodandready/dsh-key-rotation 0.8.85 → 0.8.87

package/CHANGELOG.md

@@ -1,5 +1,21 @@
 # Changelog
 
+## 0.8.87 (2026-11-03)
+
+### Cache-Affinity Router for Hot Prompt Cache (Block 14 Release)
+- **Cache-Affinity Module (Issue #573)**: Introduced `src/cache-affinity.js` with deterministic FNV-1a prompt prefix hashing (`extractPromptPrefixKey`) and LRU-managed `CacheAffinityRouter`.
+- **Hot Prompt Cache Affinity**: Binds stable system prompts and context prefixes to the provider key that has already warmed the upstream prompt cache, maintaining high hit-rates (>80%) across conversation sessions and multi-step tool calls.
+- **Synergistic Proactive Routing**: Cooperates with `applyPreemptiveYield` to ensure that hot-cache keys are prioritized only when they have safe token/request headroom, preventing cache optimization from triggering 429 rate limit failures.
+- **Config & Schemastery Support**: Added `cacheAffinityEnabled`, `cacheAffinityTtlMs` (default: 5 minutes), and `cacheAffinityPrefixChars` (default: 2048 chars) across Schemastery configuration, pool builders, and TypeScript types.
+
+## 0.8.86 (2026-11-03)
+
+### Pre-emptive Token & Request Yield before 429 (Block 13 Release)
+- **Pre-emptive Yield Module (Issue #571)**: Introduced `src/preemptive-yield.js` to evaluate active token and request headroom across provider keys using upstream `x-ratelimit-remaining-tokens`, `x-ratelimit-remaining-requests`, and reset window headers (`x-ratelimit-reset-tokens`).
+- **Zero-Round-Trip Proactive Rotation**: Keys with low token balance (`< preemptiveSafetyMarginTokens`, default: 4000) or low request count (`<= preemptiveSafetyMarginRequests`, default: 2) yield execution priority to candidate keys with ample quota before any HTTP `429 Too Many Requests` or `insufficient_quota` failure occurs.
+- **Graceful Degradation & Reset Recovery**: When upstream reset windows expire, capacity automatically restores to full priority. If all candidate keys in a pool are running on low quota, requests proceed using the key with highest remaining tokens or soonest reset time, avoiding false exhaustion.
+- **Config & Schemastery Support**: Added `preemptiveRotationEnabled`, `preemptiveSafetyMarginTokens`, and `preemptiveSafetyMarginRequests` across Schemastery configuration, pool builders, and TypeScript types.
+
 ## 0.8.85 (2026-11-03)
 
 ### Inter-Token Stall & Zombie Stream Detector (Block 12 Release)

package/lib/cache-affinity.js

@@ -0,0 +1,237 @@
+// src/cache-affinity.js — Cache-Affinity Router for hot prompt cache (#573).
+// Pure module: zero runtime dependencies, strictly < 600 lines.
+
+export const DEFAULT_CACHE_AFFINITY_TTL_MS = 300000; // 5 minutes
+export const DEFAULT_CACHE_AFFINITY_PREFIX_CHARS = 2048;
+export const MIN_CACHE_PREFIX_CHARS = 32;
+
+/**
+ * 32-bit FNV-1a hash algorithm for fast, deterministic string hashing.
+ * @param {string} str
+ * @returns {string} Hexadecimal hash representation
+ */
+export function fnv1a(str) {
+  if (typeof str !== 'string') return '0';
+  let hash = 0x811c9dc5;
+  for (let i = 0; i < str.length; i++) {
+    hash ^= str.charCodeAt(i);
+    hash = Math.imul(hash, 0x01000193);
+  }
+  return (hash >>> 0).toString(16);
+}
+
+/**
+ * Extracts and hashes a deterministic prompt prefix key from request options.
+ * Inspects system prompt, conversation messages, prompt string, or contents.
+ * @param {object} options - Request options (messages, prompt, system, model)
+ * @param {number} [maxChars=2048]
+ * @returns {string|null} Cache prefix key or null if insufficient content
+ */
+export function extractPromptPrefixKey(options, maxChars = DEFAULT_CACHE_AFFINITY_PREFIX_CHARS) {
+  if (!options || typeof options !== 'object') return null;
+
+  let text = '';
+
+  // 1. Explicit system prompt option
+  if (typeof options.system === 'string' && options.system.trim()) {
+    text += options.system.trim() + '\n';
+  }
+
+  // 2. Messages array (OpenAI / Anthropic format)
+  if (Array.isArray(options.messages) && options.messages.length > 0) {
+    for (const msg of options.messages) {
+      if (!msg) continue;
+      const role = String(msg.role ?? '').toLowerCase();
+      let content = '';
+
+      if (typeof msg.content === 'string') {
+        content = msg.content;
+      } else if (Array.isArray(msg.content)) {
+        // Multipart content (e.g. [{ type: 'text', text: '...' }])
+        content = msg.content
+          .map(part => (typeof part === 'string' ? part : (part?.text ?? '')))
+          .filter(Boolean)
+          .join('\n');
+      }
+
+      if (content.trim()) {
+        text += `[${role}] ${content.trim()}\n`;
+      }
+
+      // Stop once we have reached or exceeded maxChars
+      if (text.length >= maxChars) break;
+    }
+  }
+
+  // 3. Raw prompt string
+  if (!text && typeof options.prompt === 'string' && options.prompt.trim()) {
+    text = options.prompt.trim();
+  }
+
+  // 4. Contents array (Gemini format)
+  if (!text && Array.isArray(options.contents) && options.contents.length > 0) {
+    for (const item of options.contents) {
+      const parts = item?.parts ?? [];
+      for (const p of parts) {
+        if (typeof p?.text === 'string' && p.text.trim()) {
+          text += p.text.trim() + '\n';
+          if (text.length >= maxChars) break;
+        }
+      }
+      if (text.length >= maxChars) break;
+    }
+  }
+
+  text = text.trim();
+  if (text.length < MIN_CACHE_PREFIX_CHARS) {
+    return null;
+  }
+
+  const slice = text.slice(0, maxChars);
+  const hash = fnv1a(slice);
+  const model = options.model ? String(options.model).trim() : '';
+
+  return model ? `${model}:${hash}` : hash;
+}
+
+/**
+ * CacheAffinityRouter tracks which key has warmed the upstream prompt cache.
+ */
+export class CacheAffinityRouter {
+  constructor(options = {}) {
+    this.ttlMs = options.ttlMs ?? DEFAULT_CACHE_AFFINITY_TTL_MS;
+    this.maxEntries = options.maxEntries ?? 1000;
+    this.entries = new Map();
+  }
+
+  resolveAffinity(cacheKey, candidateRefs, now = Date.now()) {
+    if (!cacheKey || !this.entries.has(cacheKey) || !Array.isArray(candidateRefs) || candidateRefs.length === 0) {
+      return null;
+    }
+
+    const entry = this.entries.get(cacheKey);
+    if (!entry) return null;
+
+    if (now >= entry.expiresAt) {
+      this.entries.delete(cacheKey);
+      return null;
+    }
+
+    if (!candidateRefs.includes(entry.ref)) {
+      return null;
+    }
+
+    entry.hits = (entry.hits || 0) + 1;
+    return {
+      ref: entry.ref,
+      hits: entry.hits,
+      warmedAt: entry.warmedAt,
+      expiresAt: entry.expiresAt,
+      ttlRemainingMs: Math.max(0, entry.expiresAt - now),
+    };
+  }
+
+  recordAffinity(cacheKey, ref, options = {}) {
+    if (!cacheKey || !ref) return;
+    const now = options.now ?? Date.now();
+    const ttlMs = options.ttlMs ?? this.ttlMs;
+
+    if (this.entries.size >= this.maxEntries && !this.entries.has(cacheKey)) {
+      const oldestKey = this.entries.keys().next().value;
+      if (oldestKey) this.entries.delete(oldestKey);
+    }
+
+    const existing = this.entries.get(cacheKey);
+    const hits = existing ? (existing.hits || 0) : 0;
+
+    this.entries.set(cacheKey, {
+      ref,
+      warmedAt: existing?.warmedAt ?? now,
+      lastSeenAt: now,
+      expiresAt: now + ttlMs,
+      hits,
+      model: options.model,
+    });
+  }
+
+  deleteAffinity(cacheKey) {
+    if (cacheKey) this.entries.delete(cacheKey);
+  }
+
+  clear() {
+    this.entries.clear();
+  }
+
+  get size() {
+    return this.entries.size;
+  }
+}
+
+/**
+ * Obtains or lazily creates a CacheAffinityRouter on a pool state.
+ * @param {object} pool
+ * @returns {CacheAffinityRouter}
+ */
+export function getOrCreateCacheAffinityRouter(pool) {
+  if (!pool) return new CacheAffinityRouter();
+  if (!pool.state) pool.state = {};
+  if (!pool.state.cacheAffinityRouter || !(pool.state.cacheAffinityRouter instanceof CacheAffinityRouter)) {
+    pool.state.cacheAffinityRouter = new CacheAffinityRouter({
+      ttlMs: pool.cacheAffinityTtlMs ?? DEFAULT_CACHE_AFFINITY_TTL_MS,
+    });
+  }
+  return pool.state.cacheAffinityRouter;
+}
+
+/**
+ * Applies prompt cache affinity ordering to candidate attempt list.
+ * @param {string[]} attemptList
+ * @param {object} pool
+ * @param {object} [options]
+ * @returns {string[]}
+ */
+export function applyCacheAffinity(attemptList, pool, options = {}) {
+  if (!Array.isArray(attemptList) || attemptList.length <= 1 || !pool) {
+    return (attemptList ?? []).slice();
+  }
+
+  const enabled = options.enabled ?? pool.cacheAffinityEnabled ?? true;
+  if (!enabled) return attemptList.slice();
+
+  const cacheKey = options.cacheKey;
+  if (!cacheKey) return attemptList.slice();
+
+  // If a pinned key is configured and already first, keep pinned priority
+  if (pool.pinnedRef && attemptList[0] === pool.pinnedRef && options.preservePinned !== false) {
+    return attemptList.slice();
+  }
+
+  const router = getOrCreateCacheAffinityRouter(pool);
+  const now = options.now ?? Date.now();
+  const affinity = router.resolveAffinity(cacheKey, attemptList, now);
+
+  if (!affinity || !affinity.ref) {
+    return attemptList.slice();
+  }
+
+  const targetRef = affinity.ref;
+  const list = attemptList.slice();
+  const idx = list.indexOf(targetRef);
+
+  if (idx > 0) {
+    list.splice(idx, 1);
+    list.unshift(targetRef);
+
+    if (options.trace && typeof options.trace.addStep === 'function') {
+      options.trace.addStep(targetRef, 'PRIORITIZED', 'CACHE_AFFINITY', `Prompt cache hit affinity (key ${targetRef}, hits: ${affinity.hits})`);
+    }
+    if (options.pushEvent && typeof options.pushEvent === 'function') {
+      options.pushEvent(pool, targetRef, 'CACHE_AFFINITY_HIT', { cacheKey, hits: affinity.hits });
+    }
+    if (pool.state) {
+      pool.state.cacheAffinityHits = (pool.state.cacheAffinityHits ?? 0) + 1;
+    }
+  }
+
+  return list;
+}

package/lib/index.js

@@ -140,6 +140,12 @@ const SettingsSchema = Schema.object({
   stallDetectionEnabled: Schema.boolean().default(true),
   stallTimeoutMs: Schema.number().default(15000),
   stallInitialTimeoutMs: Schema.number().default(30000),
+  preemptiveRotationEnabled: Schema.boolean().default(true),
+  preemptiveSafetyMarginTokens: Schema.number().default(4000),
+  preemptiveSafetyMarginRequests: Schema.number().default(2),
+  cacheAffinityEnabled: Schema.boolean().default(true),
+  cacheAffinityTtlMs: Schema.number().default(300000),
+  cacheAffinityPrefixChars: Schema.number().default(2048),
   providers: Schema.array(Schema.object({
     provider: Schema.string().required(),
     keys: Schema.array(Schema.string()).default([]),
@@ -172,6 +178,12 @@ const SettingsSchema = Schema.object({
     stallDetectionEnabled: Schema.boolean(),
     stallTimeoutMs: Schema.number(),
     stallInitialTimeoutMs: Schema.number(),
+    preemptiveRotationEnabled: Schema.boolean(),
+    preemptiveSafetyMarginTokens: Schema.number(),
+    preemptiveSafetyMarginRequests: Schema.number(),
+    cacheAffinityEnabled: Schema.boolean(),
+    cacheAffinityTtlMs: Schema.number(),
+    cacheAffinityPrefixChars: Schema.number(),
   })).default([...DEFAULT_PROVIDERS]),
 });
 

package/lib/pool-builder.js

@@ -44,6 +44,12 @@ export function buildPoolItem({
   stallDetectionEnabled = true,
   stallTimeoutMs = 15000,
   stallInitialTimeoutMs = 30000,
+  preemptiveRotationEnabled = true,
+  preemptiveSafetyMarginTokens = 4000,
+  preemptiveSafetyMarginRequests = 2,
+  cacheAffinityEnabled = true,
+  cacheAffinityTtlMs = 300000,
+  cacheAffinityPrefixChars = 2048,
 }) {
   const seen = new Set();
   const indexed = (Array.isArray(keys) ? keys : []).map((key, index) => ({ ref: typeof key === 'string' ? key.trim() : '', index }))
@@ -157,6 +163,15 @@ const weightsMap = {};
     stallDetectionEnabled: stallDetectionEnabled ?? true,
     stallTimeoutMs: typeof stallTimeoutMs === 'number' && Number.isFinite(stallTimeoutMs) && stallTimeoutMs > 0 ? Math.floor(stallTimeoutMs) : 15000,
     stallInitialTimeoutMs: typeof stallInitialTimeoutMs === 'number' && Number.isFinite(stallInitialTimeoutMs) && stallInitialTimeoutMs > 0 ? Math.floor(stallInitialTimeoutMs) : 30000,
+    preemptiveRotationEnabled: preemptiveRotationEnabled ?? true,
+    preemptiveSafetyMarginTokens: typeof preemptiveSafetyMarginTokens === 'number' && Number.isFinite(preemptiveSafetyMarginTokens) && preemptiveSafetyMarginTokens >= 0 ? Math.floor(preemptiveSafetyMarginTokens) : 4000,
+    preemptiveSafetyMarginRequests: typeof preemptiveSafetyMarginRequests === 'number' && Number.isFinite(preemptiveSafetyMarginRequests) && preemptiveSafetyMarginRequests >= 0 ? Math.floor(preemptiveSafetyMarginRequests) : 2,
+    cacheAffinityEnabled: cacheAffinityEnabled ?? true,
+    cacheAffinityTtlMs: typeof cacheAffinityTtlMs === 'number' && Number.isFinite(cacheAffinityTtlMs) && cacheAffinityTtlMs > 0 ? Math.floor(cacheAffinityTtlMs) : 300000,
+    cacheAffinityPrefixChars: typeof cacheAffinityPrefixChars === 'number' && Number.isFinite(cacheAffinityPrefixChars) && cacheAffinityPrefixChars > 0 ? Math.floor(cacheAffinityPrefixChars) : 2048,
+    preemptiveRotationEnabled: preemptiveRotationEnabled ?? true,
+    preemptiveSafetyMarginTokens: typeof preemptiveSafetyMarginTokens === 'number' && Number.isFinite(preemptiveSafetyMarginTokens) && preemptiveSafetyMarginTokens >= 0 ? Math.floor(preemptiveSafetyMarginTokens) : 4000,
+    preemptiveSafetyMarginRequests: typeof preemptiveSafetyMarginRequests === 'number' && Number.isFinite(preemptiveSafetyMarginRequests) && preemptiveSafetyMarginRequests >= 0 ? Math.floor(preemptiveSafetyMarginRequests) : 2,
   };
 }
 
@@ -207,6 +222,15 @@ export function buildPools({ cfg = {}, poolState = new Map(), defaultState = ini
       stallDetectionEnabled: p.stallDetectionEnabled ?? cfg.stallDetectionEnabled ?? true,
       stallTimeoutMs: p.stallTimeoutMs ?? cfg.stallTimeoutMs ?? 15000,
       stallInitialTimeoutMs: p.stallInitialTimeoutMs ?? cfg.stallInitialTimeoutMs ?? 30000,
+      preemptiveRotationEnabled: p.preemptiveRotationEnabled ?? cfg.preemptiveRotationEnabled ?? true,
+      preemptiveSafetyMarginTokens: p.preemptiveSafetyMarginTokens ?? cfg.preemptiveSafetyMarginTokens ?? 4000,
+      preemptiveSafetyMarginRequests: p.preemptiveSafetyMarginRequests ?? cfg.preemptiveSafetyMarginRequests ?? 2,
+      cacheAffinityEnabled: p.cacheAffinityEnabled ?? cfg.cacheAffinityEnabled ?? true,
+      cacheAffinityTtlMs: p.cacheAffinityTtlMs ?? cfg.cacheAffinityTtlMs ?? 300000,
+      cacheAffinityPrefixChars: p.cacheAffinityPrefixChars ?? cfg.cacheAffinityPrefixChars ?? 2048,
+      preemptiveRotationEnabled: p.preemptiveRotationEnabled ?? cfg.preemptiveRotationEnabled ?? true,
+      preemptiveSafetyMarginTokens: p.preemptiveSafetyMarginTokens ?? cfg.preemptiveSafetyMarginTokens ?? 4000,
+      preemptiveSafetyMarginRequests: p.preemptiveSafetyMarginRequests ?? cfg.preemptiveSafetyMarginRequests ?? 2,
     }, provRpm, provTpm, provConcurrency);
     if (pool) {
       for (const ref of pool.refs) {
@@ -242,6 +266,15 @@ export function buildPools({ cfg = {}, poolState = new Map(), defaultState = ini
         stallDetectionEnabled: mp.stallDetectionEnabled ?? p.stallDetectionEnabled ?? cfg.stallDetectionEnabled ?? true,
         stallTimeoutMs: mp.stallTimeoutMs ?? p.stallTimeoutMs ?? cfg.stallTimeoutMs ?? 15000,
         stallInitialTimeoutMs: mp.stallInitialTimeoutMs ?? p.stallInitialTimeoutMs ?? cfg.stallInitialTimeoutMs ?? 30000,
+        preemptiveRotationEnabled: mp.preemptiveRotationEnabled ?? p.preemptiveRotationEnabled ?? cfg.preemptiveRotationEnabled ?? true,
+        preemptiveSafetyMarginTokens: mp.preemptiveSafetyMarginTokens ?? p.preemptiveSafetyMarginTokens ?? cfg.preemptiveSafetyMarginTokens ?? 4000,
+        preemptiveSafetyMarginRequests: mp.preemptiveSafetyMarginRequests ?? p.preemptiveSafetyMarginRequests ?? cfg.preemptiveSafetyMarginRequests ?? 2,
+        cacheAffinityEnabled: mp.cacheAffinityEnabled ?? p.cacheAffinityEnabled ?? cfg.cacheAffinityEnabled ?? true,
+        cacheAffinityTtlMs: mp.cacheAffinityTtlMs ?? p.cacheAffinityTtlMs ?? cfg.cacheAffinityTtlMs ?? 300000,
+        cacheAffinityPrefixChars: mp.cacheAffinityPrefixChars ?? p.cacheAffinityPrefixChars ?? cfg.cacheAffinityPrefixChars ?? 2048,
+        preemptiveRotationEnabled: mp.preemptiveRotationEnabled ?? p.preemptiveRotationEnabled ?? cfg.preemptiveRotationEnabled ?? true,
+        preemptiveSafetyMarginTokens: mp.preemptiveSafetyMarginTokens ?? p.preemptiveSafetyMarginTokens ?? cfg.preemptiveSafetyMarginTokens ?? 4000,
+        preemptiveSafetyMarginRequests: mp.preemptiveSafetyMarginRequests ?? p.preemptiveSafetyMarginRequests ?? cfg.preemptiveSafetyMarginRequests ?? 2,
       }, modelRpm, modelTpm, modelConcurrency);
       if (!mpool) continue;
       mpool.basePool = pool;

package/lib/preemptive-yield.js

@@ -0,0 +1,129 @@
+// src/preemptive-yield.js — Pre-emptive Token & Request Yield before 429 (#571).
+// Pure module: zero runtime dependencies, strictly < 600 lines.
+
+export const DEFAULT_PREEMPTIVE_SAFETY_MARGIN_TOKENS = 4000;
+export const DEFAULT_PREEMPTIVE_SAFETY_MARGIN_REQUESTS = 2;
+
+/**
+ * Evaluates whether a key has low remaining quota and should yield to other candidates.
+ * @param {object} pool - Pool definition carrying state
+ * @param {string} ref - Key reference
+ * @param {object} [options]
+ * @param {number} [options.now] - Current timestamp in ms
+ * @param {number} [options.safetyMarginTokens]
+ * @param {number} [options.safetyMarginRequests]
+ * @param {boolean} [options.enabled]
+ * @returns {{ yielding: boolean, reason?: string, remainingTokens?: number, remainingRequests?: number, resetInMs?: number }}
+ */
+export function isPreemptiveYielding(pool, ref, options = {}) {
+  const enabled = options.enabled ?? pool?.preemptiveRotationEnabled ?? true;
+  if (!enabled || !pool || !ref) return { yielding: false };
+
+  const limits = pool.state?.detectedLimits?.get?.(ref);
+  if (!limits) return { yielding: false };
+
+  const now = options.now ?? Date.now();
+  const recordedAt = limits.at || now;
+
+  // Window reset calculation
+  const resetMs = limits.tpmResetMs || limits.rpmResetMs || (typeof limits.resetTokens === 'number' ? limits.resetTokens * 1000 : (typeof limits.resetRequests === 'number' ? limits.resetRequests * 1000 : 0));
+
+  if (resetMs > 0 && now >= recordedAt + resetMs) {
+    // Window has already passed and refreshed upstream
+    return { yielding: false };
+  }
+
+  const remainingTokens = limits.remainingTokens ?? limits.tpmRemaining;
+  const remainingRequests = limits.remainingRequests ?? limits.rpmRemaining;
+
+  const marginTokens = options.safetyMarginTokens ?? pool.preemptiveSafetyMarginTokens ?? DEFAULT_PREEMPTIVE_SAFETY_MARGIN_TOKENS;
+  const marginRequests = options.safetyMarginRequests ?? pool.preemptiveSafetyMarginRequests ?? DEFAULT_PREEMPTIVE_SAFETY_MARGIN_REQUESTS;
+
+  const lowTokens = typeof remainingTokens === 'number' && remainingTokens < marginTokens;
+  const lowRequests = typeof remainingRequests === 'number' && remainingRequests <= marginRequests;
+
+  if (lowTokens || lowRequests) {
+    const resetInMs = resetMs > 0 ? Math.max(0, (recordedAt + resetMs) - now) : 0;
+    return {
+      yielding: true,
+      reason: lowTokens ? `LOW_REMAINING_TOKENS (${remainingTokens} < ${marginTokens})` : `LOW_REMAINING_REQUESTS (${remainingRequests} <= ${marginRequests})`,
+      remainingTokens,
+      remainingRequests,
+      resetInMs,
+    };
+  }
+
+  return { yielding: false, remainingTokens, remainingRequests };
+}
+
+/**
+ * Reorders candidate attempts list so that keys with ample token/request headroom
+ * are attempted first, while yielding keys yield priority.
+ * Gracefully degrades if all candidate keys are low on headroom.
+ * @param {string[]} attemptList
+ * @param {object} pool
+ * @param {object} [options]
+ * @returns {string[]}
+ */
+export function applyPreemptiveYield(attemptList, pool, options = {}) {
+  if (!Array.isArray(attemptList) || attemptList.length <= 1 || !pool) {
+    return (attemptList ?? []).slice();
+  }
+
+  const enabled = options.enabled ?? pool.preemptiveRotationEnabled ?? true;
+  if (!enabled) return attemptList.slice();
+
+  const now = options.now ?? Date.now();
+  const ample = [];
+  const yielding = [];
+
+  for (const ref of attemptList) {
+    // Pinned keys stay first if explicitly configured and not overridden
+    if (ref === pool.pinnedRef && options.preservePinned !== false) {
+      ample.push({ ref, info: { yielding: false } });
+      continue;
+    }
+
+    const info = isPreemptiveYielding(pool, ref, { ...options, now });
+    if (info.yielding) {
+      yielding.push({ ref, info });
+      if (options.trace && typeof options.trace.addStep === 'function') {
+        options.trace.addStep(ref, 'DEPRIORITIZED', 'PREEMPTIVE_YIELD', info.reason);
+      }
+      if (options.pushEvent && typeof options.pushEvent === 'function') {
+        options.pushEvent(pool, ref, 'PREEMPTIVE_YIELD', info);
+      }
+    } else {
+      ample.push({ ref, info });
+    }
+  }
+
+  if (yielding.length === 0) {
+    return attemptList.slice();
+  }
+
+  // Record yield counter in pool state
+  if (pool.state) {
+    pool.state.preemptiveYields = (pool.state.preemptiveYields ?? 0) + yielding.length;
+  }
+
+  // Sort yielding keys among themselves: highest remaining tokens first, then soonest reset
+  yielding.sort((a, b) => {
+    const ta = a.info.remainingTokens ?? -1;
+    const tb = b.info.remainingTokens ?? -1;
+    if (ta !== tb) return tb - ta; // Descending tokens
+    const ra = a.info.resetInMs ?? Infinity;
+    const rb = b.info.resetInMs ?? Infinity;
+    return ra - rb; // Ascending reset duration (soonest reset first)
+  });
+
+  if (ample.length === 0) {
+    // Graceful degradation: all keys are yielding; pick the best yielding key first
+    if (options.logWarn && typeof options.logWarn === 'function') {
+      options.logWarn(`[dsh-key-rotation] ${pool.provider ?? '-'}: all candidate keys are yielding (low quota) — proceeding with highest remaining headroom`);
+    }
+    return yielding.map(y => y.ref);
+  }
+
+  return ample.map(a => a.ref).concat(yielding.map(y => y.ref));
+}

package/lib/resolver.js

@@ -11,6 +11,8 @@
 import { bucketAllow, bucketRetryMs, tpmAllow, tpmRetryMs } from './bucket.js';
 import { envValue, sortAttemptList, isKeyPaused, isKeyRevoked, isKeyBroken, costForDay, costForWeek, budgetVerdict, getSessionSticky, setSessionSticky } from './pool.js';
 import { decryptSecret } from './crypto-storage.js';
+import { applyPreemptiveYield } from './preemptive-yield.js';
+import { extractPromptPrefixKey, applyCacheAffinity } from './cache-affinity.js';
 import { resolveRefPool, isQuotaManagedRequest, providerOfBase, basePoolForBase } from './pool-index.js';
 import {
   isModelQuotaAvailable,
@@ -132,6 +134,13 @@ export function createResolver({
         lastUsed: pool.state.lastUsed,
       });
     }
+    // Issue #573: Cache-Affinity Router
+    const cacheKey = extractPromptPrefixKey(resolveOpts, pool.cacheAffinityPrefixChars);
+    if (cacheKey) {
+      list = applyCacheAffinity(list, pool, { cacheKey, now: Date.now() });
+    }
+    // Issue #571: Pre-emptive Token & Request Yield before 429
+    list = applyPreemptiveYield(list, pool, { now: Date.now() });
     // Fail closed: when this request is managed by a model pool that has local
     // token budgets, an exhausted candidate may not be dispatched and the
     // original credential must never be used as a bypass. Decided once, before

package/lib/rotate.js

@@ -27,6 +27,8 @@ import {
 import { LOCAL_POOL_EXHAUSTED_CODE, resolveKeyDirectly } from './resolver.js';
 import { recordStreamSuccessMetrics, preserveDispatchContext } from './stream-metrics.js';
 import { withStreamWatchdog, StreamStallError } from './stream-watchdog.js';
+import { applyPreemptiveYield } from './preemptive-yield.js';
+import { extractPromptPrefixKey, applyCacheAffinity } from './cache-affinity.js';
 import { penalizeKey } from './rotate-penalty.js';
 import { recordPromptCache } from './prompt-cache.js';
 import { recordRequestOutcome, recordFailureCategory } from './ops-analytics.js';
@@ -108,7 +110,8 @@ export function createRotate(deps) {
       const breaker = runtime0.circuitBreakerEnabled === false ? null : circuitBreaker;
       let activePermit = null;
       const sessionId = options?.sessionId ?? options?.session ?? options?.metadata?.sessionId ?? options?.chatId ?? options?.conversationId;
-      const reqStore = { pool, pickedRef: undefined, concurrencyRef: undefined, startMs: now(), sessionId, model: options.model };
+      const cacheKey = extractPromptPrefixKey(options, pool.cacheAffinityPrefixChars);
+      const reqStore = { pool, pickedRef: undefined, concurrencyRef: undefined, startMs: now(), sessionId, model: options.model, cacheKey };
       setRotateStartMs(reqStore.startMs);
       const wallNow = Date.now();
       const quotaManaged = hasModelQuotaConfig(pool);
@@ -203,6 +206,19 @@ export function createRotate(deps) {
         }
       }
 
+      // Issue #573: Cache-Affinity Router
+      if (reqStore.cacheKey) {
+        attemptList = applyCacheAffinity(attemptList, pool, { cacheKey: reqStore.cacheKey, now: wallNow, trace, pushEvent });
+      }
+
+      // Issue #571: Pre-emptive Token & Request Yield before 429
+      attemptList = applyPreemptiveYield(attemptList, pool, {
+        now: wallNow,
+        trace,
+        pushEvent,
+        logWarn,
+      });
+
       const penalizeRef = (targetRef, errCode, errMsg, rawHeaders) => {
         penalizeKey(targetRef, errCode, errMsg, rawHeaders, {
           pool,
@@ -439,6 +455,7 @@ export function createRotate(deps) {
                       rawHdrs,
                       pushEvent,
                       quotaStore,
+                      cacheKey: reqStore?.cacheKey,
                     });
                   }
                   recordLatency(pool, reqStore);

package/lib/stream-metrics.js

@@ -3,6 +3,7 @@
 
 import { recordSuccess, setSessionSticky, recordFailure, extractRateLimit, isRateLimited } from './pool.js';
 import { recordPromptCache } from './prompt-cache.js';
+import { getOrCreateCacheAffinityRouter } from './cache-affinity.js';
 import { recordRequestOutcome } from './ops-analytics.js';
 import { tpmRecord } from './bucket.js';
 import { extractUsageTokens, consumeModelTokens } from './model-quota.js';
@@ -49,6 +50,7 @@ export function recordStreamSuccessMetrics(pool, activeRef, options, chunk, deps
     rawHdrs,
     pushEvent,
     quotaStore,
+    cacheKey,
   } = deps;
 
   const nowMs = typeof now === 'function' ? now() : Date.now();
@@ -61,6 +63,10 @@ export function recordStreamSuccessMetrics(pool, activeRef, options, chunk, deps
   // 2. Prompt cache and analytics outcome
   const usagePayload = chunk?.usage ?? chunk?.metadata?.usage;
   recordPromptCache(pool, activeRef, options.model, usagePayload, wallNow);
+  if (cacheKey) {
+    const affinityRouter = getOrCreateCacheAffinityRouter(pool);
+    affinityRouter.recordAffinity(cacheKey, activeRef, { now: wallNow, model: options.model });
+  }
   recordRequestOutcome(pool, {
     ref: activeRef,
     success: true,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@goodandready/dsh-key-rotation",
-  "version": "0.8.85",
+  "version": "0.8.87",
   "packageManager": "pnpm@10.33.2",
   "description": "Per-provider API key rotation for DeepSeek Harness: a key pool per provider, auto-created clone routes, and switching to the next key on quota/rate-limit errors. Includes a Settings section (Key Rotation) to edit the key pools, cooldown and switch codes.",
   "keywords": [