Wire global model fallback chain on rate/usage limits + reset-aware cooldown

packages/daemon/src/lib/acp/acp-query-runner.ts

@@ -929,7 +929,7 @@ export class AcpQueryRunner {
 
   private async handleSDKMessage(message: SDKMessage): Promise<void> {
     await this.ctx.onSDKMessage(message);
-    await this.ctx.onMarkApiSuccess();
+    await this.ctx.onMarkApiSuccess(message);
   }
 
   private async ensureRequiredMcpServersForAcp(queryOptions: Options): Promise<Options> {

packages/daemon/src/lib/agent/fallback-recovery.ts

@@ -0,0 +1,350 @@
+/**
+ * Fallback model chain + format-agnostic rate/usage-limit reset handling.
+ *
+ * Pure module (no session/DB deps) consumed by `rate-limit-watchdog.ts`. Kept
+ * separate so the chain resolution, timestamp extraction, and backoff ladder
+ * are fully unit-testable without an AgentSession or database.
+ *
+ * Background: `GlobalSettings.fallbackModels` / `modelFallbackMap`
+ * (`@hyperneo/shared` settings.ts) were editable in the UI but had no runtime
+ * consumers. On 429/usage-cap exhaustion this module resolves the chain, picks
+ * the next untried model, and — when the chain is exhausted — computes a
+ * cooldown from a reset time extracted format-agnostically from the error text
+ * (never by matching vendor-specific phrasing) or an exponential backoff ladder.
+ */
+
+import type { FallbackModelEntry } from '@hyperneo/shared';
+
+export type { FallbackModelEntry };
+
+/** Maximum plausible quota-reset window. Parsed timestamps beyond this are rejected. */
+export const MAX_RESET_HORIZON_MS = 7 * 24 * 60 * 60 * 1000; // 7 days
+
+/** Buffer added to a parsed reset time so we retry just after the window lifts. */
+export const RESET_BUFFER_MS = 30 * 1000; // 30s
+
+/** Backoff ladder (ms), indexed by cooldown step. Used when no reset time is known. */
+export const BACKOFF_LADDER_MS: readonly number[] = [
+  10 * 60 * 1000, // 10m
+  30 * 60 * 1000, // 30m
+  60 * 60 * 1000, // 1h
+  2 * 60 * 60 * 1000, // 2h
+  4 * 60 * 60 * 1000, // 4h
+];
+
+/** Hard cap for any single backoff wait. */
+export const BACKOFF_CAP_MS = 8 * 60 * 60 * 1000; // 8h
+
+/** Jitter fraction: actual wait = capped base * (1 ± BACKOFF_JITTER). */
+export const BACKOFF_JITTER = 0.15;
+
+/** Minimum backoff wait (floor) so jitter can't shrink a step below this. */
+const BACKOFF_FLOOR_MS = 60 * 1000; // 1m
+
+/**
+ * Stable dedup key for a fallback entry. We never fall back to the same
+ * provider+model that just failed, so this key is also the "tried" marker.
+ */
+export function entryKey(entry: FallbackModelEntry): string {
+  return `${entry.provider}/${entry.model}`;
+}
+
+/**
+ * Resolve the fallback chain for a (provider, model) pair.
+ *
+ * Priority: when the `modelFallbackMap` has an entry for
+ * `"${provider}/${model}"` (by key PRESENCE, not length), that override wins —
+ * including an explicitly empty chain, which the settings UI treats as
+ * "disable fallback for this model" (a separate Delete action removes the key
+ * to inherit the global list). Otherwise the global `fallbackModels` list is
+ * used. Returns a defensive copy so callers cannot mutate the live arrays.
+ *
+ * Pure: both settings fields are passed in by the caller.
+ */
+export function resolveFallbackChain(
+  provider: string,
+  model: string,
+  modelFallbackMap: Record<string, FallbackModelEntry[]> | undefined,
+  fallbackModels: FallbackModelEntry[] | undefined
+): FallbackModelEntry[] {
+  const key = entryKey({ provider, model });
+  if (modelFallbackMap && Object.hasOwn(modelFallbackMap, key)) {
+    // Key present — honor it verbatim (empty = disable fallback for this model).
+    return [...(modelFallbackMap[key] ?? [])];
+  }
+  if (fallbackModels && fallbackModels.length > 0) {
+    return [...fallbackModels];
+  }
+  return [];
+}
+
+export type FallbackSkipReason = 'none' | 'tried' | 'unavailable';
+
+export interface FallbackSelection {
+  /** The chosen next entry, or null if the chain is exhausted. */
+  next: FallbackModelEntry | null;
+  /** True when no untried+available entry remains (caller falls through to cooldown). */
+  exhausted: boolean;
+  /** Why the candidate(s) were skipped, for logging. */
+  skipReason: FallbackSkipReason;
+}
+
+/**
+ * Pick the next chain entry to try.
+ *
+ * Iteration order = chain order. The first entry whose key is NOT in `triedKeys`
+ * AND for which `isAvailable` returns true is returned. The caller must already
+ * have added the current (failed) provider/model to `triedKeys` so we never
+ * re-select it.
+ *
+ * `isAvailable` is a synchronous predicate; the watchdog pre-resolves async
+ * provider availability into a Set before calling so this function stays pure.
+ *
+ * `keyFn` defaults to `entryKey` (raw `provider/model`). The watchdog passes a
+ * canonical-ID key function so an alias (e.g. `sonnet`) and its canonical
+ * fallback entry dedupe and the chain isn't re-entered in a loop.
+ */
+export function selectNextFallback(
+  chain: FallbackModelEntry[],
+  triedKeys: ReadonlySet<string>,
+  isAvailable: (entry: FallbackModelEntry) => boolean,
+  keyFn: (entry: FallbackModelEntry) => string = entryKey
+): FallbackSelection {
+  if (chain.length === 0) {
+    return { next: null, exhausted: true, skipReason: 'none' };
+  }
+
+  let lastSkip: FallbackSkipReason = 'none';
+  for (const entry of chain) {
+    if (triedKeys.has(keyFn(entry))) {
+      lastSkip = 'tried';
+      continue;
+    }
+    if (!isAvailable(entry)) {
+      lastSkip = 'unavailable';
+      continue;
+    }
+    return { next: entry, exhausted: false, skipReason: 'none' };
+  }
+  return { next: null, exhausted: true, skipReason: lastSkip };
+}
+
+export type ResetTimestampStrategy = 'iso8601' | 'yyyymmdd-hms' | 'epoch-millis' | 'epoch-seconds';
+
+export interface ParsedReset {
+  /** Epoch-ms of the parsed reset moment. */
+  resetAtMs: number;
+  /** Which strategy matched, for telemetry/logging. */
+  strategy: ResetTimestampStrategy;
+}
+
+// ISO-8601 with an explicit offset or Z (most precise — unambiguous timezone).
+// Global flag so `matchAll` can scan every candidate (a past request timestamp
+// may precede the future quota reset in the same message).
+const ISO_WITH_TZ_RE =
+  /(\d{4})-(\d{2})-(\d{2})[T ](\d{2}):(\d{2}):(\d{2})(?:\.\d+)?(Z|[+-]\d{2}:?\d{2})/g;
+
+// YYYY-MM-DD HH:mm:ss with NO timezone (e.g. the Chinese relay shape). Parsed
+// as daemon-local time. Tried only after ISO_WITH_TZ_RE so an explicit offset
+// always wins. The trailing negative lookahead rejects the date-time PREFIX of
+// an ISO timestamp that carries a `Z` or `[+-]HH(:MM)` offset — otherwise a
+// zoned timestamp the ISO pass rejected (e.g. stale `11:00+08:00`) would be
+// reparsed here as a daemon-local `11:00`, producing a false future reset. The
+// `\.\d` term extends that to fractional seconds: a zoned timestamp like
+// `11:00:00.000+08:00` has a `.` immediately after the seconds, which a bare
+// local datetime never has, so rejecting `.<digit>` stops the local strategy
+// from re-accepting a fractional-zoned timestamp the ISO pass already rejected.
+const LOCAL_DATETIME_RE =
+  /(\d{4})-(\d{2})-(\d{2})[ T](\d{2}):(\d{2}):(\d{2})(?![Zz]|[+-]\d{2}|\.\d)/g;
+
+// 13-digit epoch millis (word-bounded to avoid UUID/request-id fragments).
+const EPOCH_MILLIS_RE = /\b\d{13}\b/g;
+
+// 10-digit epoch seconds (word-bounded; tried last to minimise false positives).
+const EPOCH_SECONDS_RE = /\b\d{10}\b/g;
+
+function isValidReset(ms: number, now: number): boolean {
+  if (!Number.isFinite(ms)) return false;
+  return ms > now && ms < now + MAX_RESET_HORIZON_MS;
+}
+
+function parseLocalGroups(groups: RegExpMatchArray): number {
+  const [, yyyy, mm, dd, hh, mi, ss] = groups;
+  // `new Date('YYYY-MM-DDTHH:mm:ss')` (no Z) parses as LOCAL time per ES spec
+  // (V8/Node). Reconstruct with a T separator from the space-separated capture.
+  return new Date(`${yyyy}-${mm}-${dd}T${hh}:${mi}:${ss}`).getTime();
+}
+
+function parseIsoWithTzGroups(groups: RegExpMatchArray): number {
+  const [, yyyy, mm, dd, hh, mi, ss, tz] = groups;
+  // Normalise offset shape to `±HH:MM` (drop a bare `Z`).
+  let offset = '';
+  if (tz !== 'Z') {
+    const raw = tz.replace(':', '');
+    offset = `${raw.slice(0, 3)}:${raw.slice(3)}`;
+  }
+  const iso = `${yyyy}-${mm}-${dd}T${hh}:${mi}:${ss}${tz === 'Z' ? 'Z' : offset}`;
+  return new Date(iso).getTime();
+}
+
+/**
+ * Extract the first plausible quota-reset timestamp from an error message.
+ *
+ * Locale- and vendor-agnostic: matches digit/separator shapes only, NEVER
+ * Chinese/English phrasing like "将在…重置" / "resets at". Vendors keep
+ * changing phrasing; the digit shape is stable.
+ *
+ * For the relay example
+ *   `Request rejected (429) · [1308][已达到 5 小时的使用上限。您的限额将在 2026-07-22 17:55:10 重置。]`
+ * the `LOCAL_DATETIME_RE` strategy matches `2026-07-22 17:55:10` (the `[1308]`
+ * code is 4 digits and never matches the epoch regexes).
+ *
+ * Accepts only timestamps in the future and within `MAX_RESET_HORIZON_MS`. Past
+ * or far-future matches are rejected (returns null → caller uses backoff).
+ */
+export function extractResetTimestamp(
+  errorMessage: string,
+  now: number = Date.now()
+): ParsedReset | null {
+  if (!errorMessage) return null;
+
+  // Each strategy scans EVERY match (not just the first): an error can contain
+  // a past request timestamp followed by the future quota reset, and the first
+  // token failing isValidReset must not abort the search.
+  const isoMatches = errorMessage.matchAll(ISO_WITH_TZ_RE);
+  for (const m of isoMatches) {
+    const ms = parseIsoWithTzGroups(m);
+    if (isValidReset(ms, now)) return { resetAtMs: ms, strategy: 'iso8601' };
+  }
+
+  const localMatches = errorMessage.matchAll(LOCAL_DATETIME_RE);
+  for (const m of localMatches) {
+    const ms = parseLocalGroups(m);
+    if (isValidReset(ms, now)) return { resetAtMs: ms, strategy: 'yyyymmdd-hms' };
+  }
+
+  for (const m of errorMessage.matchAll(EPOCH_MILLIS_RE)) {
+    const ms = Number.parseInt(m[0], 10);
+    if (isValidReset(ms, now)) return { resetAtMs: ms, strategy: 'epoch-millis' };
+  }
+
+  for (const m of errorMessage.matchAll(EPOCH_SECONDS_RE)) {
+    const ms = Number.parseInt(m[0], 10) * 1000;
+    if (isValidReset(ms, now)) return { resetAtMs: ms, strategy: 'epoch-seconds' };
+  }
+
+  return null;
+}
+
+export type CooldownReason = 'parsed-reset' | 'backoff-ladder';
+
+export interface CooldownDecision {
+  /** ms from now until the retry should fire. */
+  delayMs: number;
+  /** Absolute retryAt epoch-ms (for state serialization / UI). */
+  retryAtMs: number;
+  /** Why this delay was chosen. */
+  reason: CooldownReason;
+  /** Ladder index (0-based) when reason==='backoff-ladder', else -1. */
+  ladderIndex: number;
+  /**
+   * Whether this wait is "free" — does NOT count toward maxAutoRetries.
+   * - parsed-reset: true (we know when the window lifts; waiting isn't a guess).
+   * - backoff-ladder: false (each ladder step is one budgeted retry).
+   */
+  freeWait: boolean;
+  /** Parsed reset (when reason==='parsed-reset'), for surfacing to the UI. */
+  reset: ParsedReset | null;
+}
+
+/**
+ * Decide the next cooldown delay after the fallback chain is exhausted.
+ *
+ * @param errorMessage The 429/usage-limit error text.
+ * @param cooldownRetryCount Number of (non-free) cooldown steps already used.
+ * @param now Injected for testability.
+ * @param jitterFn Injected randomness so tests are deterministic. Defaults to
+ *   Math.random scaled to [-1, 1].
+ */
+export function computeCooldown(
+  errorMessage: string,
+  cooldownRetryCount: number,
+  now: number = Date.now(),
+  jitterFn: () => number = () => Math.random() * 2 - 1
+): CooldownDecision {
+  const parsed = extractResetTimestamp(errorMessage, now);
+  if (parsed) {
+    const retryAtMs = parsed.resetAtMs + RESET_BUFFER_MS;
+    const delayMs = Math.max(0, parsed.resetAtMs - now) + RESET_BUFFER_MS;
+    return {
+      delayMs,
+      retryAtMs,
+      reason: 'parsed-reset',
+      ladderIndex: -1,
+      freeWait: true,
+      reset: parsed,
+    };
+  }
+
+  const lastIndex = BACKOFF_LADDER_MS.length - 1;
+  const ladderIndex = Math.min(cooldownRetryCount, lastIndex);
+  const base = Math.min(BACKOFF_LADDER_MS[ladderIndex], BACKOFF_CAP_MS);
+  const jitter = base * BACKOFF_JITTER * jitterFn();
+  const delayMs = Math.max(BACKOFF_FLOOR_MS, Math.round(base + jitter));
+  return {
+    delayMs,
+    retryAtMs: now + delayMs,
+    reason: 'backoff-ladder',
+    ladderIndex,
+    freeWait: false,
+    reset: null,
+  };
+}
+
+// Keywords (lowercased; ASCII + the Chinese cap/usage characters) that signal a
+// usage CAP rather than a transient rate limit. Used to classify the paused
+// status surfaced to the UI. ASCII-only matching is case-insensitive; the CJK
+// characters are matched literally.
+//
+// Deliberately narrow: generic phrases like "exceeded" / "limit reached" appear
+// in BOTH transient rate-limit messages ("rate limit exceeded, retry in 60s")
+// and usage-cap messages, so they don't discriminate and are excluded. Only
+// cap-specific terms classify as a usage_limit; everything else is a transient
+// rate_limit.
+const USAGE_CAP_KEYWORDS = [
+  'usage',
+  'cap',
+  'quota',
+  'daily',
+  'weekly',
+  '上限',
+  '额度',
+  '小时',
+  '周',
+];
+
+/**
+ * Classify a paused rate-limit episode as a short `rate_limit` or a
+ * daily/weekly `usage_limit`, for surfacing the right task status.
+ *
+ * Heuristic: a known reset time (parsed timestamp) implies a CAP window, so it
+ * is a `usage_limit`. Otherwise, cap/usage keywords in the message override to
+ * `usage_limit`; the default is a transient `rate_limit`.
+ */
+export function classifyLimitKind(
+  errorMessage: string,
+  decision: CooldownDecision
+): 'rate_limit' | 'usage_limit' {
+  if (decision.reason === 'parsed-reset') {
+    return 'usage_limit';
+  }
+  const lower = errorMessage.toLowerCase();
+  if (USAGE_CAP_KEYWORDS.some((kw) => lower.includes(kw.toLowerCase()))) {
+    return 'usage_limit';
+  }
+  // CJK keywords are not lowercased meaningfully; check the raw message too.
+  if (USAGE_CAP_KEYWORDS.some((kw) => errorMessage.includes(kw))) {
+    return 'usage_limit';
+  }
+  return 'rate_limit';
+}