@juspay/neurolink 11.1.1 → 11.2.1

package/dist/providers/openaiCompatCatalog.js

@@ -0,0 +1,271 @@
+import { AIProviderName } from "../constants/enums.js";
+import { CloudflareModels, FireworksModels, GroqModels, MistralModels, PerplexityModels, TogetherAIModels, XaiModels, } from "../constants/enums.js";
+import { AuthenticationError, InvalidModelError, ProviderError, } from "../types/index.js";
+import { DEFAULT_ERROR_RULES } from "../utils/errorClassifier.js";
+import { createCloudflareConfig, createFireworksConfig, createGroqConfig, createMistralConfig, createPerplexityConfig, createTogetherAIConfig, createXaiConfig, } from "../utils/providerConfig.js";
+function buildCloudflareBaseURL(accountId) {
+    return `https://api.cloudflare.com/client/v4/accounts/${accountId}/ai/v1`;
+}
+/**
+ * Config-driven catalog of the 7 zero-quirk OpenAI-compatible providers.
+ * Each entry fully replaces what used to be a hand-written
+ * OpenAIChatCompletionsProvider subclass — see ConfiguredOpenAICompatProvider
+ * for the class that reads these entries, and providerRegistry.ts for the
+ * registration loop that consumes this array.
+ *
+ * `errorRules` mirrors each provider's LIVE `formatProviderError` rule array
+ * (post plan-07/wave-2 migration), not the original hand-rolled ladder these
+ * providers had when plan 05 was first drafted: every provider below now
+ * keeps only its bespoke rule(s) — auth, plus Groq's model_decommissioned and
+ * xAI's insufficient_quota — before spreading the SAME exported
+ * `DEFAULT_ERROR_RULES` constant that the live subclasses spread (never an
+ * inlined copy, so this catalog cannot drift from that table independently).
+ * See plan-05/progress.md Ruling R4 for the full rationale.
+ *
+ * To add a new zero-quirk OpenAI-compatible provider: add one entry here.
+ * Do NOT add a provider here if it needs any hook override beyond the 3
+ * mandatory ones (getProviderName/getDefaultModel/formatProviderError) —
+ * write a dedicated subclass instead (see deepseek.ts, azureOpenai.ts, and
+ * Task 14's docs task for the deciding criteria).
+ */
+export const OPENAI_COMPAT_CATALOG = [
+    {
+        providerName: AIProviderName.GROQ,
+        aliases: ["groq"],
+        apiKeyEnvVar: "GROQ_API_KEY",
+        baseURLEnvVar: "GROQ_BASE_URL",
+        defaultBaseURL: "https://api.groq.com/openai/v1",
+        configOptions: createGroqConfig(),
+        modelEnvVar: "GROQ_MODEL",
+        defaultModel: GroqModels.LLAMA_3_3_70B_VERSATILE,
+        registryDefaultModel: GroqModels.LLAMA_3_3_70B_VERSATILE,
+        registryDefaultModelChecksEnvVar: true,
+        fallbackModelName: GroqModels.LLAMA_3_1_8B_INSTANT,
+        fallbackModels: [
+            GroqModels.LLAMA_3_3_70B_VERSATILE,
+            GroqModels.LLAMA_3_1_8B_INSTANT,
+            GroqModels.GEMMA_2_9B_IT,
+            GroqModels.MIXTRAL_8X7B_32768,
+            GroqModels.LLAMA_3_2_90B_VISION_PREVIEW,
+            GroqModels.LLAMA_3_2_11B_VISION_PREVIEW,
+        ],
+        // KNOWN GAP (not reproducible by this data-only array — flagged for PR C,
+        // see plan-05/task-A-report.md): live GroqProvider.formatProviderError()
+        // intercepts TimeoutError and returns a plain ProviderError BEFORE ever
+        // calling classifyProviderError, overriding that function's own
+        // (non-overridable) rule that TimeoutError always maps to NetworkError.
+        // ConfiguredOpenAICompatProvider.formatProviderError() has no such
+        // pre-check hook, so migrating Groq onto it as-is would silently
+        // reclassify Groq timeouts as NetworkError. This entry's errorRules
+        // still mirrors Groq's live rule array faithfully for every other error
+        // shape; the TimeoutError special case is a structural gap in
+        // ConfiguredOpenAICompatProvider, not a data error here.
+        errorRules: [
+            {
+                match: (ctx) => ctx.statusCode === 401 ||
+                    /Invalid API key|Authentication|invalid_api_key/i.test(ctx.message),
+                errorClass: AuthenticationError,
+                message: "Invalid Groq API key. Check GROQ_API_KEY. Get one at https://console.groq.com/keys",
+            },
+            {
+                match: (ctx) => /model_decommissioned/i.test(ctx.message),
+                errorClass: InvalidModelError,
+                message: (ctx) => `Groq model '${ctx.modelName}' was decommissioned. Pick a current model from https://console.groq.com/docs/models.`,
+            },
+            ...DEFAULT_ERROR_RULES,
+        ],
+    },
+    {
+        providerName: AIProviderName.XAI,
+        aliases: ["xai", "grok"],
+        apiKeyEnvVar: "XAI_API_KEY",
+        baseURLEnvVar: "XAI_BASE_URL",
+        defaultBaseURL: "https://api.x.ai/v1",
+        configOptions: createXaiConfig(),
+        modelEnvVar: "XAI_MODEL",
+        defaultModel: XaiModels.GROK_3,
+        registryDefaultModel: XaiModels.GROK_3,
+        registryDefaultModelChecksEnvVar: true,
+        fallbackModelName: XaiModels.GROK_3_MINI,
+        fallbackModels: [
+            XaiModels.GROK_3,
+            XaiModels.GROK_3_MINI,
+            XaiModels.GROK_2_LATEST,
+            XaiModels.GROK_2_VISION_LATEST,
+            XaiModels.GROK_BETA,
+        ],
+        errorRules: [
+            {
+                match: (ctx) => ctx.statusCode === 401 ||
+                    /Invalid API key|Authentication|invalid_api_key/i.test(ctx.message),
+                errorClass: AuthenticationError,
+                message: "Invalid xAI API key. Please check your XAI_API_KEY environment variable. Get one at https://console.x.ai/",
+            },
+            {
+                match: (ctx) => /insufficient_quota|quota exceeded/i.test(ctx.message),
+                errorClass: ProviderError,
+                message: "xAI account has insufficient quota. Top up at https://console.x.ai/",
+            },
+            ...DEFAULT_ERROR_RULES,
+        ],
+    },
+    {
+        providerName: AIProviderName.TOGETHER_AI,
+        aliases: ["together-ai", "together"],
+        apiKeyEnvVar: "TOGETHER_API_KEY",
+        baseURLEnvVar: "TOGETHER_BASE_URL",
+        defaultBaseURL: "https://api.together.xyz/v1",
+        configOptions: createTogetherAIConfig(),
+        modelEnvVar: "TOGETHER_MODEL",
+        defaultModel: TogetherAIModels.LLAMA_3_3_70B_INSTRUCT_TURBO,
+        registryDefaultModel: TogetherAIModels.LLAMA_3_3_70B_INSTRUCT_TURBO,
+        registryDefaultModelChecksEnvVar: true,
+        fallbackModelName: TogetherAIModels.LLAMA_3_1_8B_INSTRUCT_TURBO,
+        fallbackModels: [
+            TogetherAIModels.LLAMA_3_3_70B_INSTRUCT_TURBO,
+            TogetherAIModels.LLAMA_3_1_405B_INSTRUCT_TURBO,
+            TogetherAIModels.LLAMA_3_1_70B_INSTRUCT_TURBO,
+            TogetherAIModels.LLAMA_3_1_8B_INSTRUCT_TURBO,
+            TogetherAIModels.MIXTRAL_8X22B_INSTRUCT,
+            TogetherAIModels.QWEN_2_5_72B_INSTRUCT_TURBO,
+            TogetherAIModels.DEEPSEEK_R1,
+            TogetherAIModels.DEEPSEEK_V3,
+        ],
+        errorRules: [
+            {
+                match: (ctx) => ctx.statusCode === 401 ||
+                    /Invalid API key|Authentication/i.test(ctx.message),
+                errorClass: AuthenticationError,
+                message: "Invalid Together AI API key. Get one at https://api.together.xyz/settings/api-keys",
+            },
+            ...DEFAULT_ERROR_RULES,
+        ],
+    },
+    {
+        providerName: AIProviderName.FIREWORKS,
+        aliases: ["fireworks"],
+        apiKeyEnvVar: "FIREWORKS_API_KEY",
+        baseURLEnvVar: "FIREWORKS_BASE_URL",
+        defaultBaseURL: "https://api.fireworks.ai/inference/v1",
+        configOptions: createFireworksConfig(),
+        modelEnvVar: "FIREWORKS_MODEL",
+        defaultModel: FireworksModels.DEEPSEEK_V4_PRO,
+        registryDefaultModel: FireworksModels.DEEPSEEK_V4_PRO,
+        registryDefaultModelChecksEnvVar: true,
+        fallbackModelName: FireworksModels.DEEPSEEK_V4_PRO,
+        fallbackModels: [
+            FireworksModels.DEEPSEEK_V4_PRO,
+            FireworksModels.GLM_5P1,
+            FireworksModels.GLM_5,
+            FireworksModels.KIMI_K2P6,
+            FireworksModels.KIMI_K2P5,
+            FireworksModels.GPT_OSS_120B,
+        ],
+        errorRules: [
+            {
+                match: (ctx) => ctx.statusCode === 401 ||
+                    /Invalid API key|Authentication/i.test(ctx.message),
+                errorClass: AuthenticationError,
+                message: "Invalid Fireworks API key. Get one at https://fireworks.ai/account/api-keys",
+            },
+            ...DEFAULT_ERROR_RULES,
+        ],
+    },
+    {
+        providerName: AIProviderName.PERPLEXITY,
+        aliases: ["perplexity", "pplx"],
+        apiKeyEnvVar: "PERPLEXITY_API_KEY",
+        baseURLEnvVar: "PERPLEXITY_BASE_URL",
+        defaultBaseURL: "https://api.perplexity.ai",
+        configOptions: createPerplexityConfig(),
+        modelEnvVar: "PERPLEXITY_MODEL",
+        defaultModel: PerplexityModels.SONAR,
+        registryDefaultModel: PerplexityModels.SONAR,
+        registryDefaultModelChecksEnvVar: true,
+        // Perplexity's live class does NOT override getFallbackModelName() — it
+        // inherits the base class default "gpt-3.5-turbo". Preserved here
+        // verbatim, not "fixed" to a Perplexity model — that's a real,
+        // pre-existing quirk this plan is not authorized to change.
+        fallbackModelName: "gpt-3.5-turbo",
+        fallbackModels: [
+            PerplexityModels.SONAR,
+            PerplexityModels.SONAR_PRO,
+            PerplexityModels.SONAR_REASONING,
+            PerplexityModels.SONAR_REASONING_PRO,
+            PerplexityModels.SONAR_DEEP_RESEARCH,
+        ],
+        errorRules: [
+            {
+                match: (ctx) => ctx.statusCode === 401 ||
+                    /Invalid API key|Authentication/i.test(ctx.message),
+                errorClass: AuthenticationError,
+                message: "Invalid Perplexity API key. Get one at https://www.perplexity.ai/settings/api",
+            },
+            ...DEFAULT_ERROR_RULES,
+        ],
+    },
+    {
+        providerName: AIProviderName.MISTRAL,
+        aliases: ["mistral"],
+        apiKeyEnvVar: "MISTRAL_API_KEY",
+        baseURLEnvVar: "MISTRAL_BASE_URL",
+        defaultBaseURL: "https://api.mistral.ai/v1",
+        configOptions: createMistralConfig(),
+        modelEnvVar: "MISTRAL_MODEL",
+        defaultModel: MistralModels.MISTRAL_SMALL_2506,
+        // The one documented registry-vs-class default-model quirk (see this
+        // plan's "Design reference" section): the registry passes the bare
+        // literal MISTRAL_LARGE_LATEST with no env-var check, while
+        // MistralProvider.getDefaultModel() checks MISTRAL_MODEL and defaults to
+        // MISTRAL_SMALL_2506. Preserved exactly, not reconciled.
+        registryDefaultModel: MistralModels.MISTRAL_LARGE_LATEST,
+        registryDefaultModelChecksEnvVar: false,
+        fallbackModelName: MistralModels.MISTRAL_SMALL_2506,
+        fallbackModels: [
+            MistralModels.MISTRAL_SMALL_2506,
+            MistralModels.MISTRAL_LARGE_LATEST,
+        ],
+        errorRules: [
+            {
+                match: (ctx) => ctx.statusCode === 401 ||
+                    /API_KEY_INVALID|Invalid API key|Unauthorized/i.test(ctx.message),
+                errorClass: AuthenticationError,
+                message: "Invalid Mistral API key. Please check your MISTRAL_API_KEY environment variable.",
+            },
+            ...DEFAULT_ERROR_RULES,
+        ],
+    },
+    {
+        providerName: AIProviderName.CLOUDFLARE,
+        aliases: ["cloudflare", "workers-ai", "cf-ai"],
+        apiKeyEnvVar: "CLOUDFLARE_API_KEY",
+        computedBaseURL: {
+            envVar: "CLOUDFLARE_ACCOUNT_ID",
+            missingValueMessage: "CLOUDFLARE_ACCOUNT_ID is required (or pass credentials.cloudflare.accountId). Get the account id from https://dash.cloudflare.com/",
+            build: buildCloudflareBaseURL,
+        },
+        configOptions: createCloudflareConfig(),
+        modelEnvVar: "CLOUDFLARE_MODEL",
+        defaultModel: CloudflareModels.LLAMA_3_3_70B_FAST,
+        registryDefaultModel: CloudflareModels.LLAMA_3_3_70B_FAST,
+        registryDefaultModelChecksEnvVar: true,
+        fallbackModelName: CloudflareModels.LLAMA_3_1_8B_FAST,
+        fallbackModels: [
+            CloudflareModels.LLAMA_3_3_70B_FAST,
+            CloudflareModels.LLAMA_3_1_70B_INSTRUCT,
+            CloudflareModels.LLAMA_3_1_8B_FAST,
+            CloudflareModels.LLAMA_3_2_11B_VISION,
+            CloudflareModels.MISTRAL_7B_INSTRUCT_V0_2,
+            CloudflareModels.QWEN_1P5_14B_CHAT_AWQ,
+        ],
+        errorRules: [
+            {
+                match: (ctx) => ctx.statusCode === 401 ||
+                    /Invalid API key|Authentication/i.test(ctx.message),
+                errorClass: AuthenticationError,
+                message: "Invalid Cloudflare API key. Use a token with Workers AI Read+Write scope. Get one at https://dash.cloudflare.com/profile/api-tokens",
+            },
+            ...DEFAULT_ERROR_RULES,
+        ],
+    },
+];

package/dist/proxy/accountCooldown.js

@@ -2,7 +2,8 @@ import { readFile } from "node:fs/promises";
 import { homedir } from "node:os";
 import { join } from "node:path";
 import { AsyncMutex } from "../utils/asyncMutex.js";
-import { ACCOUNT_COOLING_REASONS } from "./routingEvidence.js";
+import { logger } from "../utils/logger.js";
+import { ACCOUNT_COOLING_REASONS, MAX_COOLDOWN_MS_BY_REASON, } from "./routingEvidence.js";
 import { writeJsonSnapshotAtomically } from "./snapshotPersistence.js";
 const COOLDOWN_FILE = "account-cooldowns.json";
 const VALID_REASONS = new Set(ACCOUNT_COOLING_REASONS);
@@ -32,13 +33,45 @@ function isPersistedCooldown(value) {
         typeof candidate.reason === "string" &&
         VALID_REASONS.has(candidate.reason));
 }
+/**
+ * Cap a persisted cooldown at what its reason can plausibly mean, measured from
+ * when it was written.
+ *
+ * Entries written before per-reason ceilings existed can hold a wildly
+ * out-of-range wait — a "session" cooldown running for days, from a single stale
+ * reset timestamp. Clamping on load heals those without operator action.
+ * Clamping rather than dropping keeps a legitimate long weekly cooldown intact.
+ */
+function sanitizePersistedCooldown(accountKey, entry) {
+    const ceiling = MAX_COOLDOWN_MS_BY_REASON[entry.reason];
+    if (ceiling === undefined) {
+        return entry;
+    }
+    const latest = entry.updatedAt + ceiling;
+    if (entry.coolingUntil <= latest) {
+        return entry;
+    }
+    // Announce it: an account silently parked far beyond what its reason can mean
+    // is exactly the condition that is hard to diagnose from the outside, and this
+    // runs once per process so it cannot become noise.
+    const hours = (ms) => (ms / 3_600_000).toFixed(1);
+    logger.always(`[proxy] cooldown clamp: ${accountKey} ${entry.reason} entry healed from ` +
+        `${hours(entry.coolingUntil - entry.updatedAt)}h to ` +
+        `${hours(ceiling)}h — the stored wait exceeded what "${entry.reason}" can mean`);
+    return { ...entry, coolingUntil: latest };
+}
 async function ensureAccountCooldownsLoaded() {
     if (!cacheLoaded) {
         if (!cacheLoadPromise) {
             cacheLoadPromise = (async () => {
                 try {
                     const parsed = JSON.parse(await readFile(getCooldownFilePath(), "utf8"));
-                    memoryCache = Object.fromEntries(Object.entries(parsed).filter((entry) => isPersistedCooldown(entry[1])));
+                    memoryCache = Object.fromEntries(Object.entries(parsed)
+                        .filter((entry) => isPersistedCooldown(entry[1]))
+                        .map(([key, entry]) => [
+                        key,
+                        sanitizePersistedCooldown(key, entry),
+                    ]));
                 }
                 catch {
                     memoryCache = {};

package/dist/proxy/accountQuota.d.ts

@@ -9,7 +9,14 @@
  * updates an in-memory cache and debounces disk writes so the request/response
  * path is never blocked by file I/O.
  */
-import type { AccountQuota } from "../types/index.js";
+import type { AccountQuota, AccountQuotaWindow } from "../types/index.js";
+/**
+ * Collapse a wire model id to its family by dropping the snapshot date, so
+ * `claude-fable-5-20260115` and `claude-fable-5-20260320` both tag the same
+ * scoped window. Without this a window would stop matching the day Anthropic
+ * ships a new snapshot.
+ */
+export declare function modelFamilyToken(model: string): string;
 /** Read and normalize Anthropic's authoritative top-level unified status. */
 export declare function getUnifiedRateLimitStatus(headers: Headers | Record<string, string>): string | undefined;
 /**
@@ -20,13 +27,32 @@ export declare function getUnifiedRateLimitStatus(headers: Headers | Record<stri
  * fallback percentage together with an allowed overage status, which is the
  * equivalent provider state.
  */
-export declare function isQuotaOverageAvailable(quota: Pick<AccountQuota, "fallbackPercentage" | "fallbackStatus" | "overageStatus" | "overageInUse" | "upgradePaths"> | null | undefined): boolean;
+export declare function isQuotaOverageAvailable(quota: Pick<AccountQuota, "fallbackPercentage" | "fallbackStatus" | "overageStatus" | "overageInUse" | "overageEnabled" | "overageDisabledReason" | "upgradePaths"> | null | undefined): boolean;
 /**
  * Parse Anthropic rate-limit / quota headers into an `AccountQuota`.
  * Returns `null` when key headers are absent.
  * Pure computation — no I/O, no blocking.
  */
-export declare function parseQuotaHeaders(headers: Headers | Record<string, string>): AccountQuota | null;
+export declare function parseQuotaHeaders(headers: Headers | Record<string, string>, opts?: {
+    model?: string;
+    now?: number;
+}): AccountQuota | null;
+/**
+ * Merge dynamic limit windows across snapshots from different sources.
+ *
+ * The two sources see different things and neither is a superset: the usage API
+ * reports every plan bucket but only when explicitly refreshed, while response
+ * headers report only the window(s) touched by the request just served — but do
+ * so continuously. A plain overwrite in either direction loses real data, which
+ * is why a header capture used to erase the model-scoped windows a `/limits`
+ * refresh had just fetched.
+ */
+export declare function mergeQuotaWindows(existing: AccountQuotaWindow[] | undefined, incoming: AccountQuotaWindow[] | undefined): AccountQuotaWindow[] | undefined;
+/**
+ * Fold a freshly observed snapshot onto the previous one for the same account,
+ * preserving dynamic windows the new snapshot does not carry.
+ */
+export declare function mergeQuotaSnapshot(previous: AccountQuota | undefined, incoming: AccountQuota): AccountQuota;
 /**
  * Initialise the quota module with a custom file path.
  * When set, all reads/writes go to this path instead of the default

package/dist/proxy/accountQuota.js

@@ -33,6 +33,82 @@ function getHeader(headers, name) {
     }
     return undefined;
 }
+/** Enumerate header names, working for both `Headers` and a plain record. */
+function forEachHeaderName(headers, visit) {
+    if (typeof headers.forEach === "function") {
+        headers.forEach((_value, name) => visit(name));
+        return;
+    }
+    for (const name of Object.keys(headers)) {
+        visit(name);
+    }
+}
+/**
+ * Collapse a wire model id to its family by dropping the snapshot date, so
+ * `claude-fable-5-20260115` and `claude-fable-5-20260320` both tag the same
+ * scoped window. Without this a window would stop matching the day Anthropic
+ * ships a new snapshot.
+ */
+export function modelFamilyToken(model) {
+    return model
+        .trim()
+        .replace(/[-_](latest)$/i, "")
+        .replace(/-\d{6,8}$/, "");
+}
+/** Unified header window tokens that map to the flat session/weekly fields. */
+const FLAT_UNIFIED_WINDOW_TOKENS = new Set(["5h", "7d"]);
+const UNIFIED_UTILIZATION_HEADER = /^anthropic-ratelimit-unified-([a-z0-9_]+)-utilization$/;
+/**
+ * Discover model-scoped rate-limit windows from response headers.
+ *
+ * Anthropic reports a per-model weekly cap as its own header family — today
+ * `anthropic-ratelimit-unified-7d_oi-*`, sent only on responses for the model
+ * that cap applies to. The token is matched generically rather than hardcoded
+ * so a future `7d_xx` is captured without a code change, mirroring how
+ * `mapUsageLimit` preserves the provider's vocabulary verbatim.
+ *
+ * Requires the request's model: the header states a limit but never says which
+ * model it scopes, and an untagged window cannot be matched to a later request.
+ */
+function parseScopedQuotaWindows(headers, model, now) {
+    if (!model) {
+        return [];
+    }
+    const scopeModel = modelFamilyToken(model);
+    if (!scopeModel) {
+        return [];
+    }
+    const tokens = [];
+    forEachHeaderName(headers, (name) => {
+        const match = UNIFIED_UTILIZATION_HEADER.exec(name.toLowerCase());
+        if (match?.[1] && !FLAT_UNIFIED_WINDOW_TOKENS.has(match[1])) {
+            tokens.push(match[1]);
+        }
+    });
+    const windows = [];
+    for (const token of tokens) {
+        const P = `anthropic-ratelimit-unified-${token}-`;
+        const used = parseFloat(getHeader(headers, `${P}utilization`) ?? "");
+        if (Number.isNaN(used)) {
+            continue;
+        }
+        const resetRaw = getHeader(headers, `${P}reset`);
+        const status = getHeader(headers, `${P}status`)?.trim().toLowerCase();
+        windows.push({
+            kind: token.startsWith("7d") ? "weekly_scoped" : "session_scoped",
+            group: token.startsWith("7d") ? "weekly" : "session",
+            used,
+            status: status ?? "unknown",
+            resetsAt: resetRaw ? parseInt(resetRaw, 10) || 0 : 0,
+            scopeModel,
+            scopeModelId: model,
+            headerWindow: token,
+            source: "headers",
+            updatedAt: now,
+        });
+    }
+    return windows;
+}
 /** Read and normalize Anthropic's authoritative top-level unified status. */
 export function getUnifiedRateLimitStatus(headers) {
     const value = getHeader(headers, "anthropic-ratelimit-unified-status");
@@ -48,7 +124,25 @@ export function getUnifiedRateLimitStatus(headers) {
  * equivalent provider state.
  */
 export function isQuotaOverageAvailable(quota) {
-    if (quota?.overageStatus?.trim().toLowerCase() !== "allowed") {
+    // `extra_usage.is_enabled` from the usage API is the account's own setting and
+    // is reported even for an account that has never served a request, which the
+    // header signals below cannot cover. Positive only: it is refreshed far less
+    // often than headers are, so a stale `false` must not veto live evidence that
+    // overage is actually serving.
+    //
+    // It is also sticky — the merge carries it forward whenever a payload omits
+    // `extra_usage` — so a live header saying overage is switched off must be
+    // able to veto it. Without that veto an org disabling extra usage would leave
+    // the flag true forever, suppressing every cooldown and sending request after
+    // request that is certain to 429.
+    const overageStatus = quota?.overageStatus?.trim().toLowerCase();
+    const providerDisabledOverage = overageStatus === "rejected" || quota?.overageDisabledReason !== undefined;
+    if (quota?.overageEnabled === true && !providerDisabledOverage) {
+        return true;
+    }
+    // Explicit null check: overageStatus is now read before this point, so the
+    // optional-chain no longer narrows `quota` for the accesses below.
+    if (!quota || overageStatus !== "allowed") {
         return false;
     }
     if (quota.overageInUse === true) {
@@ -69,7 +163,7 @@ export function isQuotaOverageAvailable(quota) {
  * Returns `null` when key headers are absent.
  * Pure computation — no I/O, no blocking.
  */
-export function parseQuotaHeaders(headers) {
+export function parseQuotaHeaders(headers, opts) {
     // Anthropic prefixes all quota headers with "anthropic-ratelimit-"
     const P = "anthropic-ratelimit-";
     const sessionUtilRaw = getHeader(headers, `${P}unified-5h-utilization`);
@@ -85,6 +179,10 @@ export function parseQuotaHeaders(headers) {
     const sessionResetRaw = getHeader(headers, `${P}unified-5h-reset`);
     const weeklyResetRaw = getHeader(headers, `${P}unified-7d-reset`);
     const fallbackRaw = getHeader(headers, `${P}unified-fallback-percentage`);
+    const now = opts?.now ?? Date.now();
+    const scopedWindows = parseScopedQuotaWindows(headers, opts?.model, now);
+    const overageDisabledReason = getHeader(headers, `${P}unified-overage-disabled-reason`);
+    const representativeClaim = getHeader(headers, `${P}unified-representative-claim`);
     return {
         unifiedStatus: getUnifiedRateLimitStatus(headers),
         sessionUsed,
@@ -94,15 +192,114 @@ export function parseQuotaHeaders(headers) {
         weeklyStatus: getHeader(headers, `${P}unified-7d-status`) ?? "unknown",
         weeklyResetAt: weeklyResetRaw ? parseInt(weeklyResetRaw, 10) || 0 : 0,
         fallbackPercentage: fallbackRaw ? parseFloat(fallbackRaw) || 0 : 0,
+        // Anthropic does not send `unified-fallback` on the current wire, so this is
+        // always "unknown" in practice, which keeps the legacy back-compat branch of
+        // isQuotaOverageAvailable inert. Left as-is deliberately: making that branch
+        // reachable would stop cooling accounts that today park correctly, and the
+        // authoritative extra-usage signal now comes from `overageEnabled` instead.
         fallbackStatus: getHeader(headers, `${P}unified-fallback`) ?? "unknown",
         upgradePaths: getHeader(headers, `${P}unified-upgrade-paths`),
         overageStatus: getHeader(headers, `${P}unified-overage-status`) ?? "unknown",
         overageInUse: getHeader(headers, `${P}unified-overage-in-use`)?.trim().toLowerCase() ===
             "true",
-        lastUpdated: Date.now(),
+        ...(overageDisabledReason ? { overageDisabledReason } : {}),
+        ...(representativeClaim ? { representativeClaim } : {}),
+        lastUpdated: now,
         source: "headers",
+        ...(scopedWindows.length > 0 ? { windows: scopedWindows } : {}),
     };
 }
+/**
+ * Identity of a window across refreshes.
+ *
+ * Scope identity uses `scopeModel` — the model *family* on header-derived
+ * windows — ahead of the dated wire id, so a new model snapshot updates the
+ * existing window instead of appending a second one for the same cap and
+ * growing the array on every release. `source` keeps the two providers' views
+ * of the same cap distinct, since they name it differently and are reconciled
+ * by freshness rather than merged.
+ */
+function quotaWindowKey(window) {
+    return [
+        window.kind,
+        window.source ?? "usage-api",
+        window.headerWindow ?? "",
+        window.scopeModel ?? window.scopeModelId ?? "",
+        window.scopeSurface ?? "",
+    ].join("|");
+}
+/**
+ * Merge dynamic limit windows across snapshots from different sources.
+ *
+ * The two sources see different things and neither is a superset: the usage API
+ * reports every plan bucket but only when explicitly refreshed, while response
+ * headers report only the window(s) touched by the request just served — but do
+ * so continuously. A plain overwrite in either direction loses real data, which
+ * is why a header capture used to erase the model-scoped windows a `/limits`
+ * refresh had just fetched.
+ */
+export function mergeQuotaWindows(existing, incoming) {
+    if (!incoming?.length) {
+        return existing;
+    }
+    if (!existing?.length) {
+        return incoming;
+    }
+    const merged = new Map();
+    const incomingFromUsageApi = incoming.some((window) => (window.source ?? "usage-api") === "usage-api");
+    for (const window of existing) {
+        // A usage-API sweep is authoritative for every bucket it reports, but it
+        // never reports the header-only scoped windows — so those are carried over.
+        if (incomingFromUsageApi && (window.source ?? "usage-api") !== "headers") {
+            continue;
+        }
+        merged.set(quotaWindowKey(window), window);
+    }
+    for (const window of incoming) {
+        merged.set(quotaWindowKey(window), window);
+    }
+    return [...merged.values()];
+}
+/**
+ * Fold a freshly observed snapshot onto the previous one for the same account,
+ * preserving dynamic windows the new snapshot does not carry.
+ */
+export function mergeQuotaSnapshot(previous, incoming) {
+    if (!previous) {
+        return incoming;
+    }
+    const windows = mergeQuotaWindows(previous.windows, incoming.windows);
+    const next = { ...incoming };
+    if (windows !== undefined) {
+        next.windows = windows;
+    }
+    // Account configuration, not per-response state: each source reports only
+    // some of these, so a plain overwrite makes the value flicker in and out
+    // depending on which source wrote last. `overageEnabled` comes only from the
+    // usage API and `overageDisabledReason` only from response headers, so
+    // whichever wrote last would otherwise erase the other's field.
+    if (next.overageEnabled === undefined &&
+        previous.overageEnabled !== undefined) {
+        next.overageEnabled = previous.overageEnabled;
+    }
+    if (next.overageDisabledReason === undefined &&
+        previous.overageDisabledReason !== undefined) {
+        next.overageDisabledReason = previous.overageDisabledReason;
+    }
+    if (next.representativeClaim === undefined &&
+        previous.representativeClaim !== undefined) {
+        next.representativeClaim = previous.representativeClaim;
+    }
+    // windowsUpdatedAt tracks the last full usage-API sweep; a header capture
+    // adds one window and must not claim to have refreshed all of them.
+    const windowsUpdatedAt = incoming.source === "usage-api"
+        ? incoming.windowsUpdatedAt
+        : (incoming.windowsUpdatedAt ?? previous.windowsUpdatedAt);
+    if (windowsUpdatedAt !== undefined) {
+        next.windowsUpdatedAt = windowsUpdatedAt;
+    }
+    return next;
+}
 // ---------------------------------------------------------------------------
 // In-memory cache + debounced async persistence
 // ---------------------------------------------------------------------------
@@ -236,15 +433,9 @@ export async function loadAccountQuota(accountKey) {
 export async function saveAccountQuota(accountKey, quota) {
     await stateMutex.runExclusive(async () => {
         await ensureAccountQuotasLoaded();
-        const next = { ...quota };
-        // Header-sourced saves carry no dynamic windows; a passive capture right
-        // after a usage-API refresh must not erase the refreshed buckets.
-        const existing = memoryCache[accountKey];
-        if (next.windows === undefined && existing?.windows !== undefined) {
-            next.windows = existing.windows;
-            next.windowsUpdatedAt = existing.windowsUpdatedAt;
-        }
-        memoryCache[accountKey] = next;
+        // A header capture reports only the windows the served request touched, so
+        // it must fold onto the existing snapshot rather than replace it.
+        memoryCache[accountKey] = mergeQuotaSnapshot(memoryCache[accountKey], quota);
         dirty = true;
         cacheVersion += 1;
     });