@earendil-works/pi-ai 1.3.0 → 1.4.0

package/dist/utils/identity.js

@@ -0,0 +1,141 @@
+import { PI_AI_VERSION } from "./version.js";
+/** pi's attribution, used by pi-agent-core, pi-durable, and the coding agent. */
+export const PI_ATTRIBUTION = { name: "pi", url: "https://pi.dev", categories: ["cli-agent"] };
+const NVIDIA_NIM_HOST = "integrate.api.nvidia.com";
+const OPENCODE_HOST = "opencode.ai";
+/** Merge identities field by field; later sources win and undefined fields are skipped. */
+export function mergeClientIdentity(...sources) {
+    let merged;
+    for (const source of sources) {
+        if (!source)
+            continue;
+        merged = { ...merged };
+        if (source.userAgent !== undefined)
+            merged.userAgent = source.userAgent;
+        if (source.attribution !== undefined)
+            merged.attribution = source.attribution;
+        if (source.version !== undefined)
+            merged.version = source.version;
+    }
+    return merged;
+}
+/** The version in the detailed `User-Agent`. */
+export function getClientVersion(identity) {
+    return identity?.version ?? PI_AI_VERSION;
+}
+function runtimeDescription() {
+    if (typeof process === "undefined")
+        return undefined;
+    const versions = process.versions;
+    if (versions?.bun)
+        return `bun/${versions.bun}`;
+    if (versions?.node)
+        return `node/${process.version}`;
+    return undefined;
+}
+/** The `User-Agent` value of an identity. */
+export function formatUserAgent(identity) {
+    const mode = identity?.userAgent ?? "detailed";
+    if (typeof mode === "object")
+        return mode.custom;
+    if (mode === "short")
+        return "pi";
+    const product = `pi/${getClientVersion(identity)}`;
+    const runtime = runtimeDescription();
+    return runtime ? `${product} (${process.platform}; ${runtime}; ${process.arch})` : `${product} (browser)`;
+}
+function matchesHost(baseUrl, expectedHost) {
+    try {
+        return new URL(baseUrl).hostname === expectedHost;
+    }
+    catch {
+        return false;
+    }
+}
+function isOpenRouter(target) {
+    return target.provider === "openrouter" || target.baseUrl.includes("openrouter.ai");
+}
+function isNvidiaNim(target) {
+    return target.provider === "nvidia" || matchesHost(target.baseUrl, NVIDIA_NIM_HOST);
+}
+function isOpenCode(target) {
+    return (target.provider === "opencode" || target.provider === "opencode-go" || matchesHost(target.baseUrl, OPENCODE_HOST));
+}
+/** Provider attribution headers of an identity for a request target. Empty when attribution is off. */
+export function getAttributionHeaders(target, identity) {
+    const attribution = identity?.attribution;
+    if (!attribution)
+        return {};
+    const { name, url, categories } = attribution;
+    if (isOpenRouter(target)) {
+        return {
+            "HTTP-Referer": url,
+            "X-OpenRouter-Title": name,
+            ...(categories === undefined || categories.length === 0
+                ? {}
+                : { "X-OpenRouter-Categories": categories.join(",") }),
+        };
+    }
+    // NVIDIA expects a capitalized origin, such as `Pi`.
+    if (isNvidiaNim(target))
+        return { "X-BILLING-INVOKE-ORIGIN": name.charAt(0).toUpperCase() + name.slice(1) };
+    if (isOpenCode(target))
+        return { "x-opencode-client": name };
+    return {};
+}
+/**
+ * OpenCode's per-conversation routing header for requests to OpenCode, including custom providers pointing at it.
+ * Undefined without a session or for other targets.
+ */
+export function getOpenCodeSessionHeaders(target, sessionId) {
+    return sessionId && isOpenCode(target) ? { "x-opencode-session": sessionId } : undefined;
+}
+/**
+ * Default headers an identity adds to a request: `User-Agent` and provider attribution. APIs apply them below model,
+ * provider, and request headers, so configured headers override them and `null` request headers remove them.
+ */
+export function getIdentityHeaders(target, identity) {
+    return { "User-Agent": formatUserAgent(identity), ...getAttributionHeaders(target, identity) };
+}
+function isMutableModels(models) {
+    return "setProvider" in models && "deleteProvider" in models && "clearProviders" in models;
+}
+export function withClientIdentity(models, identity) {
+    const withIdentity = (options) => ({ ...options, identity: mergeClientIdentity(identity, options?.identity) });
+    const wrapped = {
+        getProviders: () => models.getProviders(),
+        getProvider: (id) => models.getProvider(id),
+        getModels: (provider) => models.getModels(provider),
+        getModel: (provider, id) => models.getModel(provider, id),
+        getModelsOfType: (type, provider) => models.getModelsOfType(type, provider),
+        getModelOfType: (type, provider, id) => models.getModelOfType(type, provider, id),
+        getAllModels: (provider) => models.getAllModels(provider),
+        refresh: (options) => models.refresh(options),
+        checkAuth: (providerId, options) => models.checkAuth(providerId, options),
+        getAvailable: (providerId, options) => models.getAvailable(providerId, options),
+        getAvailableOfType: (type, providerId, options) => models.getAvailableOfType(type, providerId, options),
+        getAllAvailable: (providerId, options) => models.getAllAvailable(providerId, options),
+        getAuth: models.getAuth.bind(models),
+        login: (providerId, type, interaction, options) => models.login(providerId, type, interaction, options),
+        logout: (providerId, options) => models.logout(providerId, options),
+        stream: (model, context, options) => models.stream(model, context, withIdentity(options)),
+        complete: (model, context, options) => models.complete(model, context, withIdentity(options)),
+        streamSimple: (model, context, options) => models.streamSimple(model, context, withIdentity(options)),
+        completeSimple: (model, context, options) => models.completeSimple(model, context, withIdentity(options)),
+        streamDeferred: (model, handle, options) => models.streamDeferred(model, handle, withIdentity(options)),
+        fetchDeferred: (model, handle, options) => models.fetchDeferred(model, handle, withIdentity(options)),
+        cancelDeferred: (model, handle, options) => models.cancelDeferred(model, handle, withIdentity(options)),
+        generateImages: (model, context, options) => models.generateImages(model, context, withIdentity(options)),
+        classify: (model, context, options) => models.classify(model, context, withIdentity(options)),
+    };
+    if (!isMutableModels(models))
+        return wrapped;
+    const mutable = {
+        ...wrapped,
+        setProvider: (provider) => models.setProvider(provider),
+        deleteProvider: (id) => models.deleteProvider(id),
+        clearProviders: () => models.clearProviders(),
+    };
+    return mutable;
+}
+//# sourceMappingURL=identity.js.map

package/dist/utils/identity.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"identity.js","sourceRoot":"","sources":["../../src/utils/identity.ts"],"names":[],"mappings":"AACA,OAAO,EAAE,aAAa,EAAE,MAAM,cAAc,CAAC;AAqB7C,iFAAiF;AACjF,MAAM,CAAC,MAAM,cAAc,GAAsB,EAAE,IAAI,EAAE,IAAI,EAAE,GAAG,EAAE,gBAAgB,EAAE,UAAU,EAAE,CAAC,WAAW,CAAC,EAAE,CAAC;AAqBlH,MAAM,eAAe,GAAG,0BAA0B,CAAC;AACnD,MAAM,aAAa,GAAG,aAAa,CAAC;AAEpC,2FAA2F;AAC3F,MAAM,UAAU,mBAAmB,CAAC,GAAG,OAAuC;IAC7E,IAAI,MAAkC,CAAC;IACvC,KAAK,MAAM,MAAM,IAAI,OAAO,EAAE,CAAC;QAC9B,IAAI,CAAC,MAAM;YAAE,SAAS;QACtB,MAAM,GAAG,EAAE,GAAG,MAAM,EAAE,CAAC;QACvB,IAAI,MAAM,CAAC,SAAS,KAAK,SAAS;YAAE,MAAM,CAAC,SAAS,GAAG,MAAM,CAAC,SAAS,CAAC;QACxE,IAAI,MAAM,CAAC,WAAW,KAAK,SAAS;YAAE,MAAM,CAAC,WAAW,GAAG,MAAM,CAAC,WAAW,CAAC;QAC9E,IAAI,MAAM,CAAC,OAAO,KAAK,SAAS;YAAE,MAAM,CAAC,OAAO,GAAG,MAAM,CAAC,OAAO,CAAC;IACnE,CAAC;IACD,OAAO,MAAM,CAAC;AACf,CAAC;AAED,gDAAgD;AAChD,MAAM,UAAU,gBAAgB,CAAC,QAAyB;IACzD,OAAO,QAAQ,EAAE,OAAO,IAAI,aAAa,CAAC;AAC3C,CAAC;AAED,SAAS,kBAAkB;IAC1B,IAAI,OAAO,OAAO,KAAK,WAAW;QAAE,OAAO,SAAS,CAAC;IACrD,MAAM,QAAQ,GAAG,OAAO,CAAC,QAAQ,CAAC;IAClC,IAAI,QAAQ,EAAE,GAAG;QAAE,OAAO,OAAO,QAAQ,CAAC,GAAG,EAAE,CAAC;IAChD,IAAI,QAAQ,EAAE,IAAI;QAAE,OAAO,QAAQ,OAAO,CAAC,OAAO,EAAE,CAAC;IACrD,OAAO,SAAS,CAAC;AAClB,CAAC;AAED,6CAA6C;AAC7C,MAAM,UAAU,eAAe,CAAC,QAAyB;IACxD,MAAM,IAAI,GAAG,QAAQ,EAAE,SAAS,IAAI,UAAU,CAAC;IAC/C,IAAI,OAAO,IAAI,KAAK,QAAQ;QAAE,OAAO,IAAI,CAAC,MAAM,CAAC;IACjD,IAAI,IAAI,KAAK,OAAO;QAAE,OAAO,IAAI,CAAC;IAClC,MAAM,OAAO,GAAG,MAAM,gBAAgB,CAAC,QAAQ,CAAC,EAAE,CAAC;IACnD,MAAM,OAAO,GAAG,kBAAkB,EAAE,CAAC;IACrC,OAAO,OAAO,CAAC,CAAC,CAAC,GAAG,OAAO,KAAK,OAAO,CAAC,QAAQ,KAAK,OAAO,KAAK,OAAO,CAAC,IAAI,GAAG,CAAC,CAAC,CAAC,GAAG,OAAO,YAAY,CAAC;AAC3G,CAAC;AAED,SAAS,WAAW,CAAC,OAAe,EAAE,YAAoB;IACzD,IAAI,CAAC;QACJ,OAAO,IAAI,GAAG,CAAC,OAAO,CAAC,CAAC,QAAQ,KAAK,YAAY,CAAC;IACnD,CAAC;IAAC,MAAM,CAAC;QACR,OAAO,KAAK,CAAC;IACd,CAAC;AACF,CAAC;AAED,SAAS,YAAY,CAAC,MAA4B;IACjD,OAAO,MAAM,CAAC,QAAQ,KAAK,YAAY,IAAI,MAAM,CAAC,OAAO,CAAC,QAAQ,CAAC,eAAe,CAAC,CAAC;AACrF,CAAC;AAED,SAAS,WAAW,CAAC,MAA4B;IAChD,OAAO,MAAM,CAAC,QAAQ,KAAK,QAAQ,IAAI,WAAW,CAAC,MAAM,CAAC,OAAO,EAAE,eAAe,CAAC,CAAC;AACrF,CAAC;AAED,SAAS,UAAU,CAAC,MAA4B;IAC/C,OAAO,CACN,MAAM,CAAC,QAAQ,KAAK,UAAU,IAAI,MAAM,CAAC,QAAQ,KAAK,aAAa,IAAI,WAAW,CAAC,MAAM,CAAC,OAAO,EAAE,aAAa,CAAC,CACjH,CAAC;AACH,CAAC;AAED,uGAAuG;AACvG,MAAM,UAAU,qBAAqB,CAAC,MAA4B,EAAE,QAAyB;IAC5F,MAAM,WAAW,GAAG,QAAQ,EAAE,WAAW,CAAC;IAC1C,IAAI,CAAC,WAAW;QAAE,OAAO,EAAE,CAAC;IAC5B,MAAM,EAAE,IAAI,EAAE,GAAG,EAAE,UAAU,EAAE,GAAG,WAAW,CAAC;IAC9C,IAAI,YAAY,CAAC,MAAM,CAAC,EAAE,CAAC;QAC1B,OAAO;YACN,cAAc,EAAE,GAAG;YACnB,oBAAoB,EAAE,IAAI;YAC1B,GAAG,CAAC,UAAU,KAAK,SAAS,IAAI,UAAU,CAAC,MAAM,KAAK,CAAC;gBACtD,CAAC,CAAC,EAAE;gBACJ,CAAC,CAAC,EAAE,yBAAyB,EAAE,UAAU,CAAC,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC;SACvD,CAAC;IACH,CAAC;IACD,qDAAqD;IACrD,IAAI,WAAW,CAAC,MAAM,CAAC;QAAE,OAAO,EAAE,yBAAyB,EAAE,IAAI,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC,WAAW,EAAE,GAAG,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC,EAAE,CAAC;IAC5G,IAAI,UAAU,CAAC,MAAM,CAAC;QAAE,OAAO,EAAE,mBAAmB,EAAE,IAAI,EAAE,CAAC;IAC7D,OAAO,EAAE,CAAC;AACX,CAAC;AAED;;;GAGG;AACH,MAAM,UAAU,yBAAyB,CACxC,MAA4B,EAC5B,SAA6B;IAE7B,OAAO,SAAS,IAAI,UAAU,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC,EAAE,oBAAoB,EAAE,SAAS,EAAE,CAAC,CAAC,CAAC,SAAS,CAAC;AAC1F,CAAC;AAED;;;GAGG;AACH,MAAM,UAAU,kBAAkB,CAAC,MAA4B,EAAE,QAAyB;IACzF,OAAO,EAAE,YAAY,EAAE,eAAe,CAAC,QAAQ,CAAC,EAAE,GAAG,qBAAqB,CAAC,MAAM,EAAE,QAAQ,CAAC,EAAE,CAAC;AAChG,CAAC;AAED,SAAS,eAAe,CAAC,MAAc;IACtC,OAAO,aAAa,IAAI,MAAM,IAAI,gBAAgB,IAAI,MAAM,IAAI,gBAAgB,IAAI,MAAM,CAAC;AAC5F,CAAC;AASD,MAAM,UAAU,kBAAkB,CAAC,MAAc,EAAE,QAAwB;IAC1E,MAAM,YAAY,GAAG,CAA0C,OAAsB,EAAK,EAAE,CAC3F,CAAC,EAAE,GAAG,OAAO,EAAE,QAAQ,EAAE,mBAAmB,CAAC,QAAQ,EAAE,OAAO,EAAE,QAAQ,CAAC,EAAE,CAAM,CAAC;IACnF,MAAM,OAAO,GAAW;QACvB,YAAY,EAAE,GAAG,EAAE,CAAC,MAAM,CAAC,YAAY,EAAE;QACzC,WAAW,EAAE,CAAC,EAAE,EAAE,EAAE,CAAC,MAAM,CAAC,WAAW,CAAC,EAAE,CAAC;QAC3C,SAAS,EAAE,CAAC,QAAQ,EAAE,EAAE,CAAC,MAAM,CAAC,SAAS,CAAC,QAAQ,CAAC;QACnD,QAAQ,EAAE,CAAC,QAAQ,EAAE,EAAE,EAAE,EAAE,CAAC,MAAM,CAAC,QAAQ,CAAC,QAAQ,EAAE,EAAE,CAAC;QACzD,eAAe,EAAE,CAAC,IAAI,EAAE,QAAQ,EAAE,EAAE,CAAC,MAAM,CAAC,eAAe,CAAC,IAAI,EAAE,QAAQ,CAAC;QAC3E,cAAc,EAAE,CAAC,IAAI,EAAE,QAAQ,EAAE,EAAE,EAAE,EAAE,CAAC,MAAM,CAAC,cAAc,CAAC,IAAI,EAAE,QAAQ,EAAE,EAAE,CAAC;QACjF,YAAY,EAAE,CAAC,QAAQ,EAAE,EAAE,CAAC,MAAM,CAAC,YAAY,CAAC,QAAQ,CAAC;QACzD,OAAO,EAAE,CAAC,OAAO,EAAE,EAAE,CAAC,MAAM,CAAC,OAAO,CAAC,OAAO,CAAC;QAC7C,SAAS,EAAE,CAAC,UAAU,EAAE,OAAO,EAAE,EAAE,CAAC,MAAM,CAAC,SAAS,CAAC,UAAU,EAAE,OAAO,CAAC;QACzE,YAAY,EAAE,CAAC,UAAU,EAAE,OAAO,EAAE,EAAE,CAAC,MAAM,CAAC,YAAY,CAAC,UAAU,EAAE,OAAO,CAAC;QAC/E,kBAAkB,EAAE,CAAC,IAAI,EAAE,UAAU,EAAE,OAAO,EAAE,EAAE,CAAC,MAAM,CAAC,kBAAkB,CAAC,IAAI,EAAE,UAAU,EAAE,OAAO,CAAC;QACvG,eAAe,EAAE,CAAC,UAAU,EAAE,OAAO,EAAE,EAAE,CAAC,MAAM,CAAC,eAAe,CAAC,UAAU,EAAE,OAAO,CAAC;QACrF,OAAO,EAAE,MAAM,CAAC,OAAO,CAAC,IAAI,CAAC,MAAM,CAAC;QACpC,KAAK,EAAE,CAAC,UAAU,EAAE,IAAI,EAAE,WAAW,EAAE,OAAO,EAAE,EAAE,CAAC,MAAM,CAAC,KAAK,CAAC,UAAU,EAAE,IAAI,EAAE,WAAW,EAAE,OAAO,CAAC;QACvG,MAAM,EAAE,CAAC,UAAU,EAAE,OAAO,EAAE,EAAE,CAAC,MAAM,CAAC,MAAM,CAAC,UAAU,EAAE,OAAO,CAAC;QACnE,MAAM,EAAE,CAAC,KAAK,EAAE,OAAO,EAAE,OAAO,EAAE,EAAE,CAAC,MAAM,CAAC,MAAM,CAAC,KAAK,EAAE,OAAO,EAAE,YAAY,CAAC,OAAO,CAAC,CAAC;QACzF,QAAQ,EAAE,CAAC,KAAK,EAAE,OAAO,EAAE,OAAO,EAAE,EAAE,CAAC,MAAM,CAAC,QAAQ,CAAC,KAAK,EAAE,OAAO,EAAE,YAAY,CAAC,OAAO,CAAC,CAAC;QAC7F,YAAY,EAAE,CAAC,KAAK,EAAE,OAAO,EAAE,OAAO,EAAE,EAAE,CAAC,MAAM,CAAC,YAAY,CAAC,KAAK,EAAE,OAAO,EAAE,YAAY,CAAC,OAAO,CAAC,CAAC;QACrG,cAAc,EAAE,CAAC,KAAK,EAAE,OAAO,EAAE,OAAO,EAAE,EAAE,CAAC,MAAM,CAAC,cAAc,CAAC,KAAK,EAAE,OAAO,EAAE,YAAY,CAAC,OAAO,CAAC,CAAC;QACzG,cAAc,EAAE,CAAC,KAAK,EAAE,MAAM,EAAE,OAAO,EAAE,EAAE,CAAC,MAAM,CAAC,cAAc,CAAC,KAAK,EAAE,MAAM,EAAE,YAAY,CAAC,OAAO,CAAC,CAAC;QACvG,aAAa,EAAE,CAAC,KAAK,EAAE,MAAM,EAAE,OAAO,EAAE,EAAE,CAAC,MAAM,CAAC,aAAa,CAAC,KAAK,EAAE,MAAM,EAAE,YAAY,CAAC,OAAO,CAAC,CAAC;QACrG,cAAc,EAAE,CAAC,KAAK,EAAE,MAAM,EAAE,OAAO,EAAE,EAAE,CAAC,MAAM,CAAC,cAAc,CAAC,KAAK,EAAE,MAAM,EAAE,YAAY,CAAC,OAAO,CAAC,CAAC;QACvG,cAAc,EAAE,CAAC,KAAK,EAAE,OAAO,EAAE,OAAO,EAAE,EAAE,CAAC,MAAM,CAAC,cAAc,CAAC,KAAK,EAAE,OAAO,EAAE,YAAY,CAAC,OAAO,CAAC,CAAC;QACzG,QAAQ,EAAE,CAAC,KAAK,EAAE,OAAO,EAAE,OAAO,EAAE,EAAE,CAAC,MAAM,CAAC,QAAQ,CAAC,KAAK,EAAE,OAAO,EAAE,YAAY,CAAC,OAAO,CAAC,CAAC;KAC7F,CAAC;IACF,IAAI,CAAC,eAAe,CAAC,MAAM,CAAC;QAAE,OAAO,OAAO,CAAC;IAC7C,MAAM,OAAO,GAAkB;QAC9B,GAAG,OAAO;QACV,WAAW,EAAE,CAAC,QAAQ,EAAE,EAAE,CAAC,MAAM,CAAC,WAAW,CAAC,QAAQ,CAAC;QACvD,cAAc,EAAE,CAAC,EAAE,EAAE,EAAE,CAAC,MAAM,CAAC,cAAc,CAAC,EAAE,CAAC;QACjD,cAAc,EAAE,GAAG,EAAE,CAAC,MAAM,CAAC,cAAc,EAAE;KAC7C,CAAC;IACF,OAAO,OAAO,CAAC;AAChB,CAAC","sourcesContent":["import type { Models, MutableModels } from \"../models.ts\";\nimport { PI_AI_VERSION } from \"./version.ts\";\n\n/**\n * The `User-Agent` sent to providers:\n * - `\"detailed\"`: `pi/<version> (<platform>; <runtime>/<runtime version>; <arch>)`, or `pi/<version> (browser)`\n *   outside Node.js and Bun.\n * - `\"short\"`: `pi`.\n * - `{ custom }`: the given value.\n */\nexport type UserAgentMode = \"short\" | \"detailed\" | { custom: string };\n\n/** Who requests are attributed to. `getAttributionHeaders()` maps the fields to each provider's headers. */\nexport interface ClientAttribution {\n\t/** App name: OpenRouter app title, NVIDIA invoke origin (first letter upper-cased), OpenCode client. */\n\tname: string;\n\t/** App URL: OpenRouter `HTTP-Referer`, the app's identity in OpenRouter's rankings, grouped by origin. */\n\turl: string;\n\t/** App categories, e.g. `[\"cli-agent\"]`: OpenRouter `X-OpenRouter-Categories`, comma-separated. */\n\tcategories?: string[];\n}\n\n/** pi's attribution, used by pi-agent-core, pi-durable, and the coding agent. */\nexport const PI_ATTRIBUTION: ClientAttribution = { name: \"pi\", url: \"https://pi.dev\", categories: [\"cli-agent\"] };\n\n/**\n * How pi-ai identifies the calling app to providers. Every field is optional; absent fields use pi-ai's defaults:\n * a detailed `User-Agent` with pi-ai's version and no attribution headers. Plain JSON, so it can be persisted.\n */\nexport interface ClientIdentity {\n\t/** Default: `\"detailed\"`. */\n\tuserAgent?: UserAgentMode;\n\t/** Provider attribution headers. Absent or `null`: none; `null` also turns off attribution set by a merged identity. */\n\tattribution?: ClientAttribution | null;\n\t/** Version in the detailed `User-Agent`. Default: the pi-ai version. */\n\tversion?: string;\n}\n\n/** The provider and endpoint a request goes to. */\nexport interface ClientIdentityTarget {\n\treadonly provider: string;\n\treadonly baseUrl: string;\n}\n\nconst NVIDIA_NIM_HOST = \"integrate.api.nvidia.com\";\nconst OPENCODE_HOST = \"opencode.ai\";\n\n/** Merge identities field by field; later sources win and undefined fields are skipped. */\nexport function mergeClientIdentity(...sources: (ClientIdentity | undefined)[]): ClientIdentity | undefined {\n\tlet merged: ClientIdentity | undefined;\n\tfor (const source of sources) {\n\t\tif (!source) continue;\n\t\tmerged = { ...merged };\n\t\tif (source.userAgent !== undefined) merged.userAgent = source.userAgent;\n\t\tif (source.attribution !== undefined) merged.attribution = source.attribution;\n\t\tif (source.version !== undefined) merged.version = source.version;\n\t}\n\treturn merged;\n}\n\n/** The version in the detailed `User-Agent`. */\nexport function getClientVersion(identity?: ClientIdentity): string {\n\treturn identity?.version ?? PI_AI_VERSION;\n}\n\nfunction runtimeDescription(): string | undefined {\n\tif (typeof process === \"undefined\") return undefined;\n\tconst versions = process.versions;\n\tif (versions?.bun) return `bun/${versions.bun}`;\n\tif (versions?.node) return `node/${process.version}`;\n\treturn undefined;\n}\n\n/** The `User-Agent` value of an identity. */\nexport function formatUserAgent(identity?: ClientIdentity): string {\n\tconst mode = identity?.userAgent ?? \"detailed\";\n\tif (typeof mode === \"object\") return mode.custom;\n\tif (mode === \"short\") return \"pi\";\n\tconst product = `pi/${getClientVersion(identity)}`;\n\tconst runtime = runtimeDescription();\n\treturn runtime ? `${product} (${process.platform}; ${runtime}; ${process.arch})` : `${product} (browser)`;\n}\n\nfunction matchesHost(baseUrl: string, expectedHost: string): boolean {\n\ttry {\n\t\treturn new URL(baseUrl).hostname === expectedHost;\n\t} catch {\n\t\treturn false;\n\t}\n}\n\nfunction isOpenRouter(target: ClientIdentityTarget): boolean {\n\treturn target.provider === \"openrouter\" || target.baseUrl.includes(\"openrouter.ai\");\n}\n\nfunction isNvidiaNim(target: ClientIdentityTarget): boolean {\n\treturn target.provider === \"nvidia\" || matchesHost(target.baseUrl, NVIDIA_NIM_HOST);\n}\n\nfunction isOpenCode(target: ClientIdentityTarget): boolean {\n\treturn (\n\t\ttarget.provider === \"opencode\" || target.provider === \"opencode-go\" || matchesHost(target.baseUrl, OPENCODE_HOST)\n\t);\n}\n\n/** Provider attribution headers of an identity for a request target. Empty when attribution is off. */\nexport function getAttributionHeaders(target: ClientIdentityTarget, identity?: ClientIdentity): Record<string, string> {\n\tconst attribution = identity?.attribution;\n\tif (!attribution) return {};\n\tconst { name, url, categories } = attribution;\n\tif (isOpenRouter(target)) {\n\t\treturn {\n\t\t\t\"HTTP-Referer\": url,\n\t\t\t\"X-OpenRouter-Title\": name,\n\t\t\t...(categories === undefined || categories.length === 0\n\t\t\t\t? {}\n\t\t\t\t: { \"X-OpenRouter-Categories\": categories.join(\",\") }),\n\t\t};\n\t}\n\t// NVIDIA expects a capitalized origin, such as `Pi`.\n\tif (isNvidiaNim(target)) return { \"X-BILLING-INVOKE-ORIGIN\": name.charAt(0).toUpperCase() + name.slice(1) };\n\tif (isOpenCode(target)) return { \"x-opencode-client\": name };\n\treturn {};\n}\n\n/**\n * OpenCode's per-conversation routing header for requests to OpenCode, including custom providers pointing at it.\n * Undefined without a session or for other targets.\n */\nexport function getOpenCodeSessionHeaders(\n\ttarget: ClientIdentityTarget,\n\tsessionId: string | undefined,\n): Record<string, string> | undefined {\n\treturn sessionId && isOpenCode(target) ? { \"x-opencode-session\": sessionId } : undefined;\n}\n\n/**\n * Default headers an identity adds to a request: `User-Agent` and provider attribution. APIs apply them below model,\n * provider, and request headers, so configured headers override them and `null` request headers remove them.\n */\nexport function getIdentityHeaders(target: ClientIdentityTarget, identity?: ClientIdentity): Record<string, string> {\n\treturn { \"User-Agent\": formatUserAgent(identity), ...getAttributionHeaders(target, identity) };\n}\n\nfunction isMutableModels(models: Models): models is MutableModels {\n\treturn \"setProvider\" in models && \"deleteProvider\" in models && \"clearProviders\" in models;\n}\n\n/**\n * `models` with `identity` as the default identity of its requests: a request's `identity` overrides it field by\n * field, and it overrides the identity `models` was created with. Everything else is delegated unchanged; a\n * `MutableModels` stays mutable.\n */\nexport function withClientIdentity(models: MutableModels, identity: ClientIdentity): MutableModels;\nexport function withClientIdentity(models: Models, identity: ClientIdentity): Models;\nexport function withClientIdentity(models: Models, identity: ClientIdentity): Models {\n\tconst withIdentity = <T extends { identity?: ClientIdentity }>(options: T | undefined): T =>\n\t\t({ ...options, identity: mergeClientIdentity(identity, options?.identity) }) as T;\n\tconst wrapped: Models = {\n\t\tgetProviders: () => models.getProviders(),\n\t\tgetProvider: (id) => models.getProvider(id),\n\t\tgetModels: (provider) => models.getModels(provider),\n\t\tgetModel: (provider, id) => models.getModel(provider, id),\n\t\tgetModelsOfType: (type, provider) => models.getModelsOfType(type, provider),\n\t\tgetModelOfType: (type, provider, id) => models.getModelOfType(type, provider, id),\n\t\tgetAllModels: (provider) => models.getAllModels(provider),\n\t\trefresh: (options) => models.refresh(options),\n\t\tcheckAuth: (providerId, options) => models.checkAuth(providerId, options),\n\t\tgetAvailable: (providerId, options) => models.getAvailable(providerId, options),\n\t\tgetAvailableOfType: (type, providerId, options) => models.getAvailableOfType(type, providerId, options),\n\t\tgetAllAvailable: (providerId, options) => models.getAllAvailable(providerId, options),\n\t\tgetAuth: models.getAuth.bind(models),\n\t\tlogin: (providerId, type, interaction, options) => models.login(providerId, type, interaction, options),\n\t\tlogout: (providerId, options) => models.logout(providerId, options),\n\t\tstream: (model, context, options) => models.stream(model, context, withIdentity(options)),\n\t\tcomplete: (model, context, options) => models.complete(model, context, withIdentity(options)),\n\t\tstreamSimple: (model, context, options) => models.streamSimple(model, context, withIdentity(options)),\n\t\tcompleteSimple: (model, context, options) => models.completeSimple(model, context, withIdentity(options)),\n\t\tstreamDeferred: (model, handle, options) => models.streamDeferred(model, handle, withIdentity(options)),\n\t\tfetchDeferred: (model, handle, options) => models.fetchDeferred(model, handle, withIdentity(options)),\n\t\tcancelDeferred: (model, handle, options) => models.cancelDeferred(model, handle, withIdentity(options)),\n\t\tgenerateImages: (model, context, options) => models.generateImages(model, context, withIdentity(options)),\n\t\tclassify: (model, context, options) => models.classify(model, context, withIdentity(options)),\n\t};\n\tif (!isMutableModels(models)) return wrapped;\n\tconst mutable: MutableModels = {\n\t\t...wrapped,\n\t\tsetProvider: (provider) => models.setProvider(provider),\n\t\tdeleteProvider: (id) => models.deleteProvider(id),\n\t\tclearProviders: () => models.clearProviders(),\n\t};\n\treturn mutable;\n}\n"]}

package/dist/utils/overflow.d.ts

@@ -18,6 +18,7 @@ import type { AssistantMessage } from "../types.ts";
  * - Google Gemini: "input token count exceeds the maximum"
  * - xAI (Grok): "maximum prompt length is X but request contains Y"
  * - Groq: "reduce the length of the messages"
+ * - GitHub Copilot: HTTP 413 "failed to parse request" (request byte-size overflow)
  * - Cerebras: 400/413 status code (no body)
  * - Mistral: "Prompt contains X tokens ... too large for model with Y maximum context length"
  * - OpenRouter (most backends): "maximum context length is X tokens"

package/dist/utils/overflow.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"overflow.d.ts","sourceRoot":"","sources":["../../src/utils/overflow.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,gBAAgB,EAAE,MAAM,aAAa,CAAC;AAiFpD;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAsDG;AACH,wBAAgB,iBAAiB,CAAC,OAAO,EAAE,gBAAgB,EAAE,aAAa,CAAC,EAAE,MAAM,GAAG,OAAO,CAkC5F;AAED;;;;;GAKG;AACH,wBAAgB,mBAAmB,CAAC,OAAO,EAAE,gBAAgB,EAAE,gBAAgB,EAAE,MAAM,GAAG,OAAO,CAEhG;AAED;;GAEG;AACH,wBAAgB,mBAAmB,IAAI,MAAM,EAAE,CAE9C"}
+{"version":3,"file":"overflow.d.ts","sourceRoot":"","sources":["../../src/utils/overflow.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,gBAAgB,EAAE,MAAM,aAAa,CAAC;AAqFpD;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAuDG;AACH,wBAAgB,iBAAiB,CAAC,OAAO,EAAE,gBAAgB,EAAE,aAAa,CAAC,EAAE,MAAM,GAAG,OAAO,CAkC5F;AAED;;;;;GAKG;AACH,wBAAgB,mBAAmB,CAAC,OAAO,EAAE,gBAAgB,EAAE,gBAAgB,EAAE,MAAM,GAAG,OAAO,CAEhG;AAED;;GAEG;AACH,wBAAgB,mBAAmB,IAAI,MAAM,EAAE,CAE9C"}

package/dist/utils/overflow.js

@@ -2,7 +2,7 @@
  * Regex patterns to detect context overflow errors from different providers.
  *
  * These patterns match error messages returned when the input exceeds
- * the model's context window.
+ * the model's context window or the provider's request byte-size limit.
  *
  * Provider-specific patterns (with example error messages):
  *
@@ -20,6 +20,7 @@
  * - llama.cpp: "the request exceeds the available context size, try increasing it"
  * - LM Studio: "tokens to keep from the initial prompt is greater than the context length"
  * - GitHub Copilot: "prompt token count of X exceeds the limit of Y"
+ * - GitHub Copilot: 'github-copilot API error (413): {"message":"failed to parse request","code":""}'
  * - MiniMax: "invalid params, context window exceeds limit"
  * - Kimi For Coding: "Your request exceeded model token limit: X (requested: Y)"
  * - DS4: "Prompt has X tokens, but the configured context size is Y tokens"
@@ -58,6 +59,9 @@ const OVERFLOW_PATTERNS = [
     /context[_ ]length[_ ]exceeded/i, // Generic fallback
     /too many tokens/i, // Generic fallback
     /token limit exceeded/i, // Generic fallback
+    // HTTP 413 means request-size overflow regardless of provider or body text.
+    // Match raw SDK errors and formatProviderError's prefix, not a status mentioned in the body.
+    /^(?:[\w -]+ API error \(413\):|413(?:\s|:|$))/i,
 ];
 const CEREBRAS_BODYLESS_OVERFLOW_PATTERN = /^4(?:00|13)\s*(?:status code)?\s*\(no body\)/i;
 /**
@@ -93,6 +97,7 @@ const NON_OVERFLOW_PATTERNS = [
  * - Google Gemini: "input token count exceeds the maximum"
  * - xAI (Grok): "maximum prompt length is X but request contains Y"
  * - Groq: "reduce the length of the messages"
+ * - GitHub Copilot: HTTP 413 "failed to parse request" (request byte-size overflow)
  * - Cerebras: 400/413 status code (no body)
  * - Mistral: "Prompt contains X tokens ... too large for model with Y maximum context length"
  * - OpenRouter (most backends): "maximum context length is X tokens"

package/dist/utils/overflow.js.map

@@ -1 +1 @@
-{"version":3,"file":"overflow.js","sourceRoot":"","sources":["../../src/utils/overflow.ts"],"names":[],"mappings":"AAEA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAiCG;AACH,MAAM,iBAAiB,GAAG;IACzB,0BAA0B,EAAE,oCAAoC;IAChE,4BAA4B,EAAE,kCAAkC;IAChE,oBAAoB,EAAE,kDAAkD;IACxE,wCAAwC,EAAE,iBAAiB;IAC3D,6BAA6B,EAAE,uCAAuC;IACtE,4FAA4F,EAAE,sCAAsC;IACpI,yCAAyC,EAAE,kBAAkB;IAC7D,+BAA+B,EAAE,aAAa;IAC9C,oCAAoC,EAAE,OAAO;IAC7C,uCAAuC,EAAE,6BAA6B;IACtE,kEAAkE,EAAE,sBAAsB;IAC1F,iFAAiF,EAAE,cAAc;IACjG,2BAA2B,EAAE,iBAAiB;IAC9C,qCAAqC,EAAE,mBAAmB;IAC1D,kCAAkC,EAAE,YAAY;IAChD,+BAA+B,EAAE,UAAU;IAC3C,6BAA6B,EAAE,kBAAkB;IACjD,sDAAsD,EAAE,UAAU;IAClE,+EAA+E,EAAE,aAAa;IAC9F,gCAAgC,EAAE,yDAAyD;IAC3F,oDAAoD,EAAE,iCAAiC;IACvF,kCAAkC,EAAE,8BAA8B;IAClE,gCAAgC,EAAE,mBAAmB;IACrD,kBAAkB,EAAE,mBAAmB;IACvC,uBAAuB,EAAE,mBAAmB;CAC5C,CAAC;AAEF,MAAM,kCAAkC,GAAG,+CAA+C,CAAC;AAE3F;;;;;;;;GAQG;AACH,MAAM,qBAAqB,GAAG;IAC7B,2CAA2C,EAAE,oFAAoF;IACjI,aAAa,EAAE,wBAAwB;IACvC,oBAAoB,EAAE,yBAAyB;CAC/C,CAAC;AAEF;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAsDG;AACH,MAAM,UAAU,iBAAiB,CAAC,OAAyB,EAAE,aAAsB;IAClF,uCAAuC;IACvC,IAAI,OAAO,CAAC,UAAU,KAAK,OAAO,IAAI,OAAO,CAAC,YAAY,EAAE,CAAC;QAC5D,oFAAoF;QACpF,MAAM,aAAa,GAAG,qBAAqB,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC,OAAO,CAAC,YAAa,CAAC,CAAC,CAAC;QACvF,IAAI,CAAC,aAAa,EAAE,CAAC;YACpB,IAAI,iBAAiB,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC,OAAO,CAAC,YAAa,CAAC,CAAC,EAAE,CAAC;gBAClE,OAAO,IAAI,CAAC;YACb,CAAC;YACD,IAAI,OAAO,CAAC,QAAQ,KAAK,UAAU,IAAI,kCAAkC,CAAC,IAAI,CAAC,OAAO,CAAC,YAAY,CAAC,EAAE,CAAC;gBACtG,OAAO,IAAI,CAAC;YACb,CAAC;QACF,CAAC;IACF,CAAC;IAED,8EAA8E;IAC9E,IAAI,aAAa,IAAI,OAAO,CAAC,UAAU,KAAK,MAAM,EAAE,CAAC;QACpD,MAAM,WAAW,GAAG,OAAO,CAAC,KAAK,CAAC,KAAK,GAAG,OAAO,CAAC,KAAK,CAAC,SAAS,CAAC;QAClE,IAAI,WAAW,GAAG,aAAa,EAAE,CAAC;YACjC,OAAO,IAAI,CAAC;QACb,CAAC;IACF,CAAC;IAED,sFAAsF;IACtF,qFAAqF;IACrF,gEAAgE;IAChE,IAAI,aAAa,IAAI,OAAO,CAAC,UAAU,KAAK,QAAQ,IAAI,OAAO,CAAC,KAAK,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;QACpF,MAAM,WAAW,GAAG,OAAO,CAAC,KAAK,CAAC,KAAK,GAAG,OAAO,CAAC,KAAK,CAAC,SAAS,CAAC;QAClE,IAAI,WAAW,IAAI,aAAa,GAAG,IAAI,EAAE,CAAC;YACzC,OAAO,IAAI,CAAC;QACb,CAAC;IACF,CAAC;IAED,OAAO,KAAK,CAAC;AACd,CAAC;AAED;;;;;GAKG;AACH,MAAM,UAAU,mBAAmB,CAAC,OAAyB,EAAE,gBAAwB;IACtF,OAAO,OAAO,CAAC,UAAU,KAAK,QAAQ,IAAI,gBAAgB,GAAG,CAAC,IAAI,OAAO,CAAC,KAAK,CAAC,MAAM,GAAG,gBAAgB,CAAC;AAC3G,CAAC;AAED;;GAEG;AACH,MAAM,UAAU,mBAAmB;IAClC,OAAO,CAAC,GAAG,iBAAiB,CAAC,CAAC;AAC/B,CAAC","sourcesContent":["import type { AssistantMessage } from \"../types.ts\";\n\n/**\n * Regex patterns to detect context overflow errors from different providers.\n *\n * These patterns match error messages returned when the input exceeds\n * the model's context window.\n *\n * Provider-specific patterns (with example error messages):\n *\n * - Anthropic: \"prompt is too long: 213462 tokens > 200000 maximum\"\n * - Anthropic: \"413 {\\\"error\\\":{\\\"type\\\":\\\"request_too_large\\\",\\\"message\\\":\\\"Request exceeds the maximum size\\\"}}\"\n * - OpenAI: \"Your input exceeds the context window of this model\"\n * - OpenAI/LiteLLM: \"Requested token count exceeds the model's maximum context length of 131072 tokens\"\n * - OpenAI-compatible: \"Input length (265330) exceeds model's maximum context length (262144).\"\n * - Google: \"The input token count (1196265) exceeds the maximum number of tokens allowed (1048575)\"\n * - xAI: \"This model's maximum prompt length is 131072 but the request contains 537812 tokens\"\n * - Groq: \"Please reduce the length of the messages or completion\"\n * - OpenRouter: \"This endpoint's maximum context length is X tokens. However, you requested about Y tokens\"\n * - OpenRouter/Poolside: \"Input length X exceeds the maximum allowed input length of Y tokens.\"\n * - Together AI: \"The input (X tokens) is longer than the model's context length (Y tokens).\"\n * - llama.cpp: \"the request exceeds the available context size, try increasing it\"\n * - LM Studio: \"tokens to keep from the initial prompt is greater than the context length\"\n * - GitHub Copilot: \"prompt token count of X exceeds the limit of Y\"\n * - MiniMax: \"invalid params, context window exceeds limit\"\n * - Kimi For Coding: \"Your request exceeded model token limit: X (requested: Y)\"\n * - DS4: \"Prompt has X tokens, but the configured context size is Y tokens\"\n * - Cerebras: \"400/413 status code (no body)\"\n * - Mistral: \"Prompt contains X tokens ... too large for model with Y maximum context length\"\n * - z.ai: `{\"code\":\"1261\",\"message\":\"Prompt too long\"}`, `{\"code\":\"1261\",\"message\":\"Prompt exceeds max length\"}` (CN endpoint), or silent overflow via usage.input > contextWindow\n * - Xiaomi MiMo: Truncates input to fill contextWindow exactly, then returns finish_reason \"length\"\n *   with output=0 (no room left to generate). Detected via stopReason \"length\" + zero output +\n *   input filling the context window.\n * - DashScope/Qwen: \"Range of input length should be [1, X]\" (HTTP 400 invalid_parameter_error)\n * - Ollama: Some deployments truncate silently, others return errors like \"prompt too long; exceeded max context length by X tokens\"\n */\nconst OVERFLOW_PATTERNS = [\n\t/prompt (?:is )?too long/i, // Anthropic and z.ai token overflow\n\t/prompt exceeds max length/i, // z.ai CN endpoint token overflow\n\t/request_too_large/i, // Anthropic request byte-size overflow (HTTP 413)\n\t/input is too long for requested model/i, // Amazon Bedrock\n\t/exceeds the context window/i, // OpenAI (Completions & Responses API)\n\t/exceeds (?:the )?(?:model'?s )?maximum context length(?: of [\\d,]+ tokens?|\\s*\\([\\d,]+\\))/i, // OpenAI-compatible proxies (LiteLLM)\n\t/input token count.*exceeds the maximum/i, // Google (Gemini)\n\t/maximum prompt length is \\d+/i, // xAI (Grok)\n\t/reduce the length of the messages/i, // Groq\n\t/maximum context length is \\d+ tokens/i, // OpenRouter (most backends)\n\t/exceeds (?:the )?maximum allowed input length of [\\d,]+ tokens?/i, // OpenRouter/Poolside\n\t/input \\(\\d+ tokens\\) is longer than the model'?s context length \\(\\d+ tokens\\)/i, // Together AI\n\t/exceeds the limit of \\d+/i, // GitHub Copilot\n\t/exceeds the available context size/i, // llama.cpp server\n\t/greater than the context length/i, // LM Studio\n\t/context window exceeds limit/i, // MiniMax\n\t/exceeded model token limit/i, // Kimi For Coding\n\t/too large for model with \\d+ maximum context length/i, // Mistral\n\t/prompt has [\\d,]+ tokens?, but the configured context size is [\\d,]+ tokens?/i, // DS4 server\n\t/model_context_window_exceeded/i, // z.ai non-standard finish_reason surfaced as error text\n\t/prompt too long; exceeded (?:max )?context length/i, // Ollama explicit overflow error\n\t/range of input length should be/i, // DashScope / Qwen Token Plan\n\t/context[_ ]length[_ ]exceeded/i, // Generic fallback\n\t/too many tokens/i, // Generic fallback\n\t/token limit exceeded/i, // Generic fallback\n];\n\nconst CEREBRAS_BODYLESS_OVERFLOW_PATTERN = /^4(?:00|13)\\s*(?:status code)?\\s*\\(no body\\)/i;\n\n/**\n * Patterns that indicate non-overflow errors (e.g. rate limiting, server errors).\n * Error messages matching any of these are excluded from overflow detection\n * even if they also match an OVERFLOW_PATTERN.\n *\n * Example: Bedrock formats throttling errors as \"ThrottlingException: Too many tokens,\n * please wait before trying again.\" which would match the /too many tokens/i overflow\n * pattern without this exclusion.\n */\nconst NON_OVERFLOW_PATTERNS = [\n\t/^(Throttling error|Service unavailable):/i, // AWS Bedrock non-overflow errors (human-readable prefixes from formatBedrockError)\n\t/rate limit/i, // Generic rate limiting\n\t/too many requests/i, // Generic HTTP 429 style\n];\n\n/**\n * Check if an assistant message represents a context overflow error.\n *\n * This handles three cases:\n * 1. Error-based overflow: Most providers return stopReason \"error\" with a\n *    specific error message pattern.\n * 2. Silent overflow: Some providers accept overflow requests and return\n *    successfully. For these, we check if usage.input exceeds the context window.\n * 3. Length-stop overflow: Xiaomi MiMo can return \"length\" with zero output when\n *    the input fills the context window.\n *\n * ## Reliability by Provider\n *\n * **Reliable detection (returns error with detectable message):**\n * - Anthropic: \"prompt is too long: X tokens > Y maximum\" or \"request_too_large\"\n * - OpenAI (Completions & Responses): \"exceeds the context window\", \"exceeds the model's maximum context length of X tokens\", or \"exceeds model's maximum context length (X)\"\n * - Google Gemini: \"input token count exceeds the maximum\"\n * - xAI (Grok): \"maximum prompt length is X but request contains Y\"\n * - Groq: \"reduce the length of the messages\"\n * - Cerebras: 400/413 status code (no body)\n * - Mistral: \"Prompt contains X tokens ... too large for model with Y maximum context length\"\n * - OpenRouter (most backends): \"maximum context length is X tokens\"\n * - OpenRouter/Poolside: \"Input length X exceeds the maximum allowed input length of Y tokens.\"\n * - Together AI: \"The input (X tokens) is longer than the model's context length (Y tokens).\"\n * - llama.cpp: \"exceeds the available context size\"\n * - LM Studio: \"greater than the context length\"\n * - Kimi For Coding: \"exceeded model token limit: X (requested: Y)\"\n * - DS4: \"Prompt has X tokens, but the configured context size is Y tokens\"\n * - DashScope/Qwen: \"Range of input length should be [1, X]\"\n * - z.ai: \"Prompt too long\"\n *\n * **Unreliable detection:**\n * - z.ai: Sometimes accepts overflow silently (detectable via usage.input > contextWindow),\n *   sometimes returns rate limit errors. Pass contextWindow param to detect silent overflow.\n * - Xiaomi MiMo: Truncates input to fit contextWindow then returns stopReason \"length\" with\n *   output=0. Pass contextWindow param to detect via the \"filled context + zero output\" signal.\n * - Ollama: May truncate input silently for some setups, but may also return explicit\n *   overflow errors that match the patterns above. Silent truncation still cannot be\n *   detected here because we do not know the expected token count.\n *\n * ## Custom Providers\n *\n * If you've added custom models via settings.json, this function may not detect\n * overflow errors from those providers. To add support:\n *\n * 1. Send a request that exceeds the model's context window\n * 2. Check the errorMessage in the response\n * 3. Create a regex pattern that matches the error\n * 4. The pattern should be added to OVERFLOW_PATTERNS in this file, or\n *    check the errorMessage yourself before calling this function\n *\n * @param message - The assistant message to check\n * @param contextWindow - Optional context window size for detecting silent overflow (z.ai)\n * @returns true if the message indicates a context overflow\n */\nexport function isContextOverflow(message: AssistantMessage, contextWindow?: number): boolean {\n\t// Case 1: Check error message patterns\n\tif (message.stopReason === \"error\" && message.errorMessage) {\n\t\t// Skip messages matching known non-overflow patterns (e.g. throttling / rate-limit)\n\t\tconst isNonOverflow = NON_OVERFLOW_PATTERNS.some((p) => p.test(message.errorMessage!));\n\t\tif (!isNonOverflow) {\n\t\t\tif (OVERFLOW_PATTERNS.some((p) => p.test(message.errorMessage!))) {\n\t\t\t\treturn true;\n\t\t\t}\n\t\t\tif (message.provider === \"cerebras\" && CEREBRAS_BODYLESS_OVERFLOW_PATTERN.test(message.errorMessage)) {\n\t\t\t\treturn true;\n\t\t\t}\n\t\t}\n\t}\n\n\t// Case 2: Silent overflow (z.ai style) - successful but usage exceeds context\n\tif (contextWindow && message.stopReason === \"stop\") {\n\t\tconst inputTokens = message.usage.input + message.usage.cacheRead;\n\t\tif (inputTokens > contextWindow) {\n\t\t\treturn true;\n\t\t}\n\t}\n\n\t// Case 3: Length-stop overflow (Xiaomi MiMo style) - server truncates oversized input\n\t// to fit the context window, leaving no room for output. Returns stopReason \"length\"\n\t// with output=0 and input+cacheRead filling the context window.\n\tif (contextWindow && message.stopReason === \"length\" && message.usage.output === 0) {\n\t\tconst inputTokens = message.usage.input + message.usage.cacheRead;\n\t\tif (inputTokens >= contextWindow * 0.99) {\n\t\t\treturn true;\n\t\t}\n\t}\n\n\treturn false;\n}\n\n/**\n * Check whether a length stop ended below the caller or model's intended output limit.\n * Such responses may be caused by context pressure or provider-side truncation, so callers\n * can make one bounded compact-and-retry attempt. `desiredMaxOutput` must be the original\n * limit before any context-based clamping.\n */\nexport function isRecoverableLength(message: AssistantMessage, desiredMaxOutput: number): boolean {\n\treturn message.stopReason === \"length\" && desiredMaxOutput > 0 && message.usage.output < desiredMaxOutput;\n}\n\n/**\n * Get the overflow patterns for testing purposes.\n */\nexport function getOverflowPatterns(): RegExp[] {\n\treturn [...OVERFLOW_PATTERNS];\n}\n"]}
+{"version":3,"file":"overflow.js","sourceRoot":"","sources":["../../src/utils/overflow.ts"],"names":[],"mappings":"AAEA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAkCG;AACH,MAAM,iBAAiB,GAAG;IACzB,0BAA0B,EAAE,oCAAoC;IAChE,4BAA4B,EAAE,kCAAkC;IAChE,oBAAoB,EAAE,kDAAkD;IACxE,wCAAwC,EAAE,iBAAiB;IAC3D,6BAA6B,EAAE,uCAAuC;IACtE,4FAA4F,EAAE,sCAAsC;IACpI,yCAAyC,EAAE,kBAAkB;IAC7D,+BAA+B,EAAE,aAAa;IAC9C,oCAAoC,EAAE,OAAO;IAC7C,uCAAuC,EAAE,6BAA6B;IACtE,kEAAkE,EAAE,sBAAsB;IAC1F,iFAAiF,EAAE,cAAc;IACjG,2BAA2B,EAAE,iBAAiB;IAC9C,qCAAqC,EAAE,mBAAmB;IAC1D,kCAAkC,EAAE,YAAY;IAChD,+BAA+B,EAAE,UAAU;IAC3C,6BAA6B,EAAE,kBAAkB;IACjD,sDAAsD,EAAE,UAAU;IAClE,+EAA+E,EAAE,aAAa;IAC9F,gCAAgC,EAAE,yDAAyD;IAC3F,oDAAoD,EAAE,iCAAiC;IACvF,kCAAkC,EAAE,8BAA8B;IAClE,gCAAgC,EAAE,mBAAmB;IACrD,kBAAkB,EAAE,mBAAmB;IACvC,uBAAuB,EAAE,mBAAmB;IAC5C,4EAA4E;IAC5E,6FAA6F;IAC7F,gDAAgD;CAChD,CAAC;AAEF,MAAM,kCAAkC,GAAG,+CAA+C,CAAC;AAE3F;;;;;;;;GAQG;AACH,MAAM,qBAAqB,GAAG;IAC7B,2CAA2C,EAAE,oFAAoF;IACjI,aAAa,EAAE,wBAAwB;IACvC,oBAAoB,EAAE,yBAAyB;CAC/C,CAAC;AAEF;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAuDG;AACH,MAAM,UAAU,iBAAiB,CAAC,OAAyB,EAAE,aAAsB;IAClF,uCAAuC;IACvC,IAAI,OAAO,CAAC,UAAU,KAAK,OAAO,IAAI,OAAO,CAAC,YAAY,EAAE,CAAC;QAC5D,oFAAoF;QACpF,MAAM,aAAa,GAAG,qBAAqB,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC,OAAO,CAAC,YAAa,CAAC,CAAC,CAAC;QACvF,IAAI,CAAC,aAAa,EAAE,CAAC;YACpB,IAAI,iBAAiB,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC,OAAO,CAAC,YAAa,CAAC,CAAC,EAAE,CAAC;gBAClE,OAAO,IAAI,CAAC;YACb,CAAC;YACD,IAAI,OAAO,CAAC,QAAQ,KAAK,UAAU,IAAI,kCAAkC,CAAC,IAAI,CAAC,OAAO,CAAC,YAAY,CAAC,EAAE,CAAC;gBACtG,OAAO,IAAI,CAAC;YACb,CAAC;QACF,CAAC;IACF,CAAC;IAED,8EAA8E;IAC9E,IAAI,aAAa,IAAI,OAAO,CAAC,UAAU,KAAK,MAAM,EAAE,CAAC;QACpD,MAAM,WAAW,GAAG,OAAO,CAAC,KAAK,CAAC,KAAK,GAAG,OAAO,CAAC,KAAK,CAAC,SAAS,CAAC;QAClE,IAAI,WAAW,GAAG,aAAa,EAAE,CAAC;YACjC,OAAO,IAAI,CAAC;QACb,CAAC;IACF,CAAC;IAED,sFAAsF;IACtF,qFAAqF;IACrF,gEAAgE;IAChE,IAAI,aAAa,IAAI,OAAO,CAAC,UAAU,KAAK,QAAQ,IAAI,OAAO,CAAC,KAAK,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;QACpF,MAAM,WAAW,GAAG,OAAO,CAAC,KAAK,CAAC,KAAK,GAAG,OAAO,CAAC,KAAK,CAAC,SAAS,CAAC;QAClE,IAAI,WAAW,IAAI,aAAa,GAAG,IAAI,EAAE,CAAC;YACzC,OAAO,IAAI,CAAC;QACb,CAAC;IACF,CAAC;IAED,OAAO,KAAK,CAAC;AACd,CAAC;AAED;;;;;GAKG;AACH,MAAM,UAAU,mBAAmB,CAAC,OAAyB,EAAE,gBAAwB;IACtF,OAAO,OAAO,CAAC,UAAU,KAAK,QAAQ,IAAI,gBAAgB,GAAG,CAAC,IAAI,OAAO,CAAC,KAAK,CAAC,MAAM,GAAG,gBAAgB,CAAC;AAC3G,CAAC;AAED;;GAEG;AACH,MAAM,UAAU,mBAAmB;IAClC,OAAO,CAAC,GAAG,iBAAiB,CAAC,CAAC;AAC/B,CAAC","sourcesContent":["import type { AssistantMessage } from \"../types.ts\";\n\n/**\n * Regex patterns to detect context overflow errors from different providers.\n *\n * These patterns match error messages returned when the input exceeds\n * the model's context window or the provider's request byte-size limit.\n *\n * Provider-specific patterns (with example error messages):\n *\n * - Anthropic: \"prompt is too long: 213462 tokens > 200000 maximum\"\n * - Anthropic: \"413 {\\\"error\\\":{\\\"type\\\":\\\"request_too_large\\\",\\\"message\\\":\\\"Request exceeds the maximum size\\\"}}\"\n * - OpenAI: \"Your input exceeds the context window of this model\"\n * - OpenAI/LiteLLM: \"Requested token count exceeds the model's maximum context length of 131072 tokens\"\n * - OpenAI-compatible: \"Input length (265330) exceeds model's maximum context length (262144).\"\n * - Google: \"The input token count (1196265) exceeds the maximum number of tokens allowed (1048575)\"\n * - xAI: \"This model's maximum prompt length is 131072 but the request contains 537812 tokens\"\n * - Groq: \"Please reduce the length of the messages or completion\"\n * - OpenRouter: \"This endpoint's maximum context length is X tokens. However, you requested about Y tokens\"\n * - OpenRouter/Poolside: \"Input length X exceeds the maximum allowed input length of Y tokens.\"\n * - Together AI: \"The input (X tokens) is longer than the model's context length (Y tokens).\"\n * - llama.cpp: \"the request exceeds the available context size, try increasing it\"\n * - LM Studio: \"tokens to keep from the initial prompt is greater than the context length\"\n * - GitHub Copilot: \"prompt token count of X exceeds the limit of Y\"\n * - GitHub Copilot: 'github-copilot API error (413): {\"message\":\"failed to parse request\",\"code\":\"\"}'\n * - MiniMax: \"invalid params, context window exceeds limit\"\n * - Kimi For Coding: \"Your request exceeded model token limit: X (requested: Y)\"\n * - DS4: \"Prompt has X tokens, but the configured context size is Y tokens\"\n * - Cerebras: \"400/413 status code (no body)\"\n * - Mistral: \"Prompt contains X tokens ... too large for model with Y maximum context length\"\n * - z.ai: `{\"code\":\"1261\",\"message\":\"Prompt too long\"}`, `{\"code\":\"1261\",\"message\":\"Prompt exceeds max length\"}` (CN endpoint), or silent overflow via usage.input > contextWindow\n * - Xiaomi MiMo: Truncates input to fill contextWindow exactly, then returns finish_reason \"length\"\n *   with output=0 (no room left to generate). Detected via stopReason \"length\" + zero output +\n *   input filling the context window.\n * - DashScope/Qwen: \"Range of input length should be [1, X]\" (HTTP 400 invalid_parameter_error)\n * - Ollama: Some deployments truncate silently, others return errors like \"prompt too long; exceeded max context length by X tokens\"\n */\nconst OVERFLOW_PATTERNS = [\n\t/prompt (?:is )?too long/i, // Anthropic and z.ai token overflow\n\t/prompt exceeds max length/i, // z.ai CN endpoint token overflow\n\t/request_too_large/i, // Anthropic request byte-size overflow (HTTP 413)\n\t/input is too long for requested model/i, // Amazon Bedrock\n\t/exceeds the context window/i, // OpenAI (Completions & Responses API)\n\t/exceeds (?:the )?(?:model'?s )?maximum context length(?: of [\\d,]+ tokens?|\\s*\\([\\d,]+\\))/i, // OpenAI-compatible proxies (LiteLLM)\n\t/input token count.*exceeds the maximum/i, // Google (Gemini)\n\t/maximum prompt length is \\d+/i, // xAI (Grok)\n\t/reduce the length of the messages/i, // Groq\n\t/maximum context length is \\d+ tokens/i, // OpenRouter (most backends)\n\t/exceeds (?:the )?maximum allowed input length of [\\d,]+ tokens?/i, // OpenRouter/Poolside\n\t/input \\(\\d+ tokens\\) is longer than the model'?s context length \\(\\d+ tokens\\)/i, // Together AI\n\t/exceeds the limit of \\d+/i, // GitHub Copilot\n\t/exceeds the available context size/i, // llama.cpp server\n\t/greater than the context length/i, // LM Studio\n\t/context window exceeds limit/i, // MiniMax\n\t/exceeded model token limit/i, // Kimi For Coding\n\t/too large for model with \\d+ maximum context length/i, // Mistral\n\t/prompt has [\\d,]+ tokens?, but the configured context size is [\\d,]+ tokens?/i, // DS4 server\n\t/model_context_window_exceeded/i, // z.ai non-standard finish_reason surfaced as error text\n\t/prompt too long; exceeded (?:max )?context length/i, // Ollama explicit overflow error\n\t/range of input length should be/i, // DashScope / Qwen Token Plan\n\t/context[_ ]length[_ ]exceeded/i, // Generic fallback\n\t/too many tokens/i, // Generic fallback\n\t/token limit exceeded/i, // Generic fallback\n\t// HTTP 413 means request-size overflow regardless of provider or body text.\n\t// Match raw SDK errors and formatProviderError's prefix, not a status mentioned in the body.\n\t/^(?:[\\w -]+ API error \\(413\\):|413(?:\\s|:|$))/i,\n];\n\nconst CEREBRAS_BODYLESS_OVERFLOW_PATTERN = /^4(?:00|13)\\s*(?:status code)?\\s*\\(no body\\)/i;\n\n/**\n * Patterns that indicate non-overflow errors (e.g. rate limiting, server errors).\n * Error messages matching any of these are excluded from overflow detection\n * even if they also match an OVERFLOW_PATTERN.\n *\n * Example: Bedrock formats throttling errors as \"ThrottlingException: Too many tokens,\n * please wait before trying again.\" which would match the /too many tokens/i overflow\n * pattern without this exclusion.\n */\nconst NON_OVERFLOW_PATTERNS = [\n\t/^(Throttling error|Service unavailable):/i, // AWS Bedrock non-overflow errors (human-readable prefixes from formatBedrockError)\n\t/rate limit/i, // Generic rate limiting\n\t/too many requests/i, // Generic HTTP 429 style\n];\n\n/**\n * Check if an assistant message represents a context overflow error.\n *\n * This handles three cases:\n * 1. Error-based overflow: Most providers return stopReason \"error\" with a\n *    specific error message pattern.\n * 2. Silent overflow: Some providers accept overflow requests and return\n *    successfully. For these, we check if usage.input exceeds the context window.\n * 3. Length-stop overflow: Xiaomi MiMo can return \"length\" with zero output when\n *    the input fills the context window.\n *\n * ## Reliability by Provider\n *\n * **Reliable detection (returns error with detectable message):**\n * - Anthropic: \"prompt is too long: X tokens > Y maximum\" or \"request_too_large\"\n * - OpenAI (Completions & Responses): \"exceeds the context window\", \"exceeds the model's maximum context length of X tokens\", or \"exceeds model's maximum context length (X)\"\n * - Google Gemini: \"input token count exceeds the maximum\"\n * - xAI (Grok): \"maximum prompt length is X but request contains Y\"\n * - Groq: \"reduce the length of the messages\"\n * - GitHub Copilot: HTTP 413 \"failed to parse request\" (request byte-size overflow)\n * - Cerebras: 400/413 status code (no body)\n * - Mistral: \"Prompt contains X tokens ... too large for model with Y maximum context length\"\n * - OpenRouter (most backends): \"maximum context length is X tokens\"\n * - OpenRouter/Poolside: \"Input length X exceeds the maximum allowed input length of Y tokens.\"\n * - Together AI: \"The input (X tokens) is longer than the model's context length (Y tokens).\"\n * - llama.cpp: \"exceeds the available context size\"\n * - LM Studio: \"greater than the context length\"\n * - Kimi For Coding: \"exceeded model token limit: X (requested: Y)\"\n * - DS4: \"Prompt has X tokens, but the configured context size is Y tokens\"\n * - DashScope/Qwen: \"Range of input length should be [1, X]\"\n * - z.ai: \"Prompt too long\"\n *\n * **Unreliable detection:**\n * - z.ai: Sometimes accepts overflow silently (detectable via usage.input > contextWindow),\n *   sometimes returns rate limit errors. Pass contextWindow param to detect silent overflow.\n * - Xiaomi MiMo: Truncates input to fit contextWindow then returns stopReason \"length\" with\n *   output=0. Pass contextWindow param to detect via the \"filled context + zero output\" signal.\n * - Ollama: May truncate input silently for some setups, but may also return explicit\n *   overflow errors that match the patterns above. Silent truncation still cannot be\n *   detected here because we do not know the expected token count.\n *\n * ## Custom Providers\n *\n * If you've added custom models via settings.json, this function may not detect\n * overflow errors from those providers. To add support:\n *\n * 1. Send a request that exceeds the model's context window\n * 2. Check the errorMessage in the response\n * 3. Create a regex pattern that matches the error\n * 4. The pattern should be added to OVERFLOW_PATTERNS in this file, or\n *    check the errorMessage yourself before calling this function\n *\n * @param message - The assistant message to check\n * @param contextWindow - Optional context window size for detecting silent overflow (z.ai)\n * @returns true if the message indicates a context overflow\n */\nexport function isContextOverflow(message: AssistantMessage, contextWindow?: number): boolean {\n\t// Case 1: Check error message patterns\n\tif (message.stopReason === \"error\" && message.errorMessage) {\n\t\t// Skip messages matching known non-overflow patterns (e.g. throttling / rate-limit)\n\t\tconst isNonOverflow = NON_OVERFLOW_PATTERNS.some((p) => p.test(message.errorMessage!));\n\t\tif (!isNonOverflow) {\n\t\t\tif (OVERFLOW_PATTERNS.some((p) => p.test(message.errorMessage!))) {\n\t\t\t\treturn true;\n\t\t\t}\n\t\t\tif (message.provider === \"cerebras\" && CEREBRAS_BODYLESS_OVERFLOW_PATTERN.test(message.errorMessage)) {\n\t\t\t\treturn true;\n\t\t\t}\n\t\t}\n\t}\n\n\t// Case 2: Silent overflow (z.ai style) - successful but usage exceeds context\n\tif (contextWindow && message.stopReason === \"stop\") {\n\t\tconst inputTokens = message.usage.input + message.usage.cacheRead;\n\t\tif (inputTokens > contextWindow) {\n\t\t\treturn true;\n\t\t}\n\t}\n\n\t// Case 3: Length-stop overflow (Xiaomi MiMo style) - server truncates oversized input\n\t// to fit the context window, leaving no room for output. Returns stopReason \"length\"\n\t// with output=0 and input+cacheRead filling the context window.\n\tif (contextWindow && message.stopReason === \"length\" && message.usage.output === 0) {\n\t\tconst inputTokens = message.usage.input + message.usage.cacheRead;\n\t\tif (inputTokens >= contextWindow * 0.99) {\n\t\t\treturn true;\n\t\t}\n\t}\n\n\treturn false;\n}\n\n/**\n * Check whether a length stop ended below the caller or model's intended output limit.\n * Such responses may be caused by context pressure or provider-side truncation, so callers\n * can make one bounded compact-and-retry attempt. `desiredMaxOutput` must be the original\n * limit before any context-based clamping.\n */\nexport function isRecoverableLength(message: AssistantMessage, desiredMaxOutput: number): boolean {\n\treturn message.stopReason === \"length\" && desiredMaxOutput > 0 && message.usage.output < desiredMaxOutput;\n}\n\n/**\n * Get the overflow patterns for testing purposes.\n */\nexport function getOverflowPatterns(): RegExp[] {\n\treturn [...OVERFLOW_PATTERNS];\n}\n"]}

package/dist/utils/retry.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"retry.d.ts","sourceRoot":"","sources":["../../src/utils/retry.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,gBAAgB,EAAE,MAAM,aAAa,CAAC;AAiHpD;;;;;;GAMG;AACH,MAAM,WAAW,WAAW;IAC3B,OAAO,EAAE,OAAO,CAAC;IACjB,qFAAqF;IACrF,UAAU,EAAE,MAAM,CAAC;IACnB,0FAA0F;IAC1F,WAAW,EAAE,MAAM,CAAC;IACpB,+EAA+E;IAC/E,eAAe,CAAC,EAAE,MAAM,CAAC;CACzB;AAED,eAAO,MAAM,gCAAgC,QAAS,CAAC;AAEvD,wBAAgB,YAAY,CAAC,MAAM,EAAE,IAAI,CAAC,WAAW,EAAE,aAAa,GAAG,iBAAiB,CAAC,EAAE,OAAO,EAAE,MAAM,GAAG,MAAM,CAIlH;AAED,kFAAkF;AAClF,MAAM,WAAW,cAAc;IAC9B,0EAA0E;IAC1E,gBAAgB,CAAC,EAAE,CAClB,OAAO,EAAE,MAAM,EACf,WAAW,EAAE,MAAM,EACnB,OAAO,EAAE,MAAM,EACf,YAAY,EAAE,MAAM,KAChB,IAAI,GAAG,OAAO,CAAC,IAAI,CAAC,CAAC;IAC1B,mFAAmF;IACnF,mBAAmB,CAAC,EAAE,MAAM,IAAI,GAAG,OAAO,CAAC,IAAI,CAAC,CAAC;IACjD,mFAAmF;IACnF,eAAe,CAAC,EAAE,CAAC,OAAO,EAAE,OAAO,EAAE,OAAO,EAAE,MAAM,EAAE,UAAU,CAAC,EAAE,MAAM,KAAK,IAAI,GAAG,OAAO,CAAC,IAAI,CAAC,CAAC;CACnG;AA0BD;;;;;;;;;;;;;;;;;GAiBG;AACH,wBAAsB,kBAAkB,CACvC,OAAO,EAAE,MAAM,OAAO,CAAC,gBAAgB,CAAC,EACxC,MAAM,EAAE,WAAW,GAAG,SAAS,EAC/B,MAAM,EAAE,WAAW,GAAG,SAAS,EAC/B,SAAS,CAAC,EAAE,cAAc,GACxB,OAAO,CAAC,gBAAgB,CAAC,CA6C3B;AAED;;;;;;;;GAQG;AACH,wBAAgB,yBAAyB,CAAC,OAAO,EAAE,gBAAgB,GAAG,OAAO,CAK5E"}
+{"version":3,"file":"retry.d.ts","sourceRoot":"","sources":["../../src/utils/retry.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,gBAAgB,EAAE,MAAM,aAAa,CAAC;AA0HpD;;;;;;GAMG;AACH,MAAM,WAAW,WAAW;IAC3B,OAAO,EAAE,OAAO,CAAC;IACjB,qFAAqF;IACrF,UAAU,EAAE,MAAM,CAAC;IACnB,0FAA0F;IAC1F,WAAW,EAAE,MAAM,CAAC;IACpB,+EAA+E;IAC/E,eAAe,CAAC,EAAE,MAAM,CAAC;CACzB;AAED,eAAO,MAAM,gCAAgC,QAAS,CAAC;AAEvD,wBAAgB,YAAY,CAAC,MAAM,EAAE,IAAI,CAAC,WAAW,EAAE,aAAa,GAAG,iBAAiB,CAAC,EAAE,OAAO,EAAE,MAAM,GAAG,MAAM,CAIlH;AAED,kFAAkF;AAClF,MAAM,WAAW,cAAc;IAC9B,0EAA0E;IAC1E,gBAAgB,CAAC,EAAE,CAClB,OAAO,EAAE,MAAM,EACf,WAAW,EAAE,MAAM,EACnB,OAAO,EAAE,MAAM,EACf,YAAY,EAAE,MAAM,KAChB,IAAI,GAAG,OAAO,CAAC,IAAI,CAAC,CAAC;IAC1B,mFAAmF;IACnF,mBAAmB,CAAC,EAAE,MAAM,IAAI,GAAG,OAAO,CAAC,IAAI,CAAC,CAAC;IACjD,mFAAmF;IACnF,eAAe,CAAC,EAAE,CAAC,OAAO,EAAE,OAAO,EAAE,OAAO,EAAE,MAAM,EAAE,UAAU,CAAC,EAAE,MAAM,KAAK,IAAI,GAAG,OAAO,CAAC,IAAI,CAAC,CAAC;CACnG;AA0BD;;;;;;;;;;;;;;;;;GAiBG;AACH,wBAAsB,kBAAkB,CACvC,OAAO,EAAE,MAAM,OAAO,CAAC,gBAAgB,CAAC,EACxC,MAAM,EAAE,WAAW,GAAG,SAAS,EAC/B,MAAM,EAAE,WAAW,GAAG,SAAS,EAC/B,SAAS,CAAC,EAAE,cAAc,GACxB,OAAO,CAAC,gBAAgB,CAAC,CA6C3B;AAED;;;;;;;;GAQG;AACH,wBAAgB,yBAAyB,CAAC,OAAO,EAAE,gBAAgB,GAAG,OAAO,CAK5E"}

package/dist/utils/retry.js

@@ -19,6 +19,9 @@ const NON_RETRYABLE_PROVIDER_LIMIT_ERROR_PATTERN = buildProviderErrorPattern([
     // Sign in with ChatGPT: the subscription's shared usage limit, which resets
     // after hours rather than seconds.
     "subscription_sharing_usage_limit_exceeded",
+    // Sign in with ChatGPT on a ChatGPT Business plan: a usage limit set by the
+    // workspace admin. Only the admin can raise it.
+    "subscription_sharing_workplace_individual_usage_limit_exceeded",
 ]);
 const RETRYABLE_PROVIDER_ERROR_PATTERN = buildProviderErrorPattern([
     // Generic provider load, HTTP status, and server-side transient failures.
@@ -29,13 +32,8 @@ const RETRYABLE_PROVIDER_ERROR_PATTERN = buildProviderErrorPattern([
     "model is at capacity",
     "rate.?limit",
     "too many requests",
-    "429",
-    "500",
-    "502",
-    "503",
-    "504",
-    "520",
-    "524",
+    // Status codes must not be part of a longer word or number, e.g. a request id or byte count (#11057).
+    "(?<![a-z0-9])(?:429|500|502|503|504|520|524)(?![a-z0-9])",
     // Also matches "The service is temporarily unavailable" (#10965).
     "service.?(?:is.?)?(?:temporarily.?)?unavailable",
     "server.?error",
@@ -48,6 +46,11 @@ const RETRYABLE_PROVIDER_ERROR_PATTERN = buildProviderErrorPattern([
     // Gateway failure relaying the upstream response stream, sent as an OpenAI
     // Responses `response.failed` event without an error code (#10920).
     "unable to forward upstream",
+    // Other gateway upstream failures relayed as `response.failed` error codes (#10920).
+    "upstream.?request.?failed",
+    "upstream_stream_read_error",
+    "stream was interrupted",
+    "http/?2 stream failed",
     // Network, proxy, and fetch transport failures. This includes OpenAI Codex
     // raw-fetch failures such as "upstream connect", "connection refused", and
     // "reset before headers" (#733), plus OpenRouter connection drops (#3317).
@@ -57,6 +60,9 @@ const RETRYABLE_PROVIDER_ERROR_PATTERN = buildProviderErrorPattern([
     "connection.?lost",
     "other side closed",
     "fetch failed",
+    // node-fetch system errors, e.g. the Google ADC OAuth token refresh used by
+    // google-vertex. The reason is often empty (#10996).
+    "request to \\S+ failed, reason:",
     "getaddrinfo",
     "ENOTFOUND",
     "EAI_AGAIN",
@@ -77,6 +83,9 @@ const RETRYABLE_PROVIDER_ERROR_PATTERN = buildProviderErrorPattern([
     "stream ended before message_stop",
     "stream ended before a terminal response event",
     "http2 request did not get a response",
+    // Codex CLI mid-stream disconnect, relayed by proxies that wrap the Codex CLI,
+    // e.g. "stream disconnected before completion: stream closed before response.completed" (#11057).
+    "stream disconnected before completion",
     // Node ERR_HTTP2_STREAM_CANCEL: the HTTP/2 session died before the request was
     // sent, e.g. after the Bedrock SDK's 5-minute session timeout (#10379).
     "pending stream has been canceled",

package/dist/utils/retry.js.map

@@ -1 +1 @@
-{"version":3,"file":"retry.js","sourceRoot":"","sources":["../../src/utils/retry.ts"],"names":[],"mappings":"AAEA,SAAS,yBAAyB,CAAC,QAA2B;IAC7D,OAAO,IAAI,MAAM,CAAC,QAAQ,CAAC,IAAI,CAAC,GAAG,CAAC,EAAE,GAAG,CAAC,CAAC;AAC5C,CAAC;AAED,MAAM,0CAA0C,GAAG,yBAAyB,CAAC;IAC5E,8EAA8E;IAC9E,2EAA2E;IAC3E,mBAAmB;IACnB,qBAAqB;IAErB,6EAA6E;IAC7E,yDAAyD;IACzD,6BAA6B;IAC7B,mBAAmB;IAEnB,4EAA4E;IAC5E,4EAA4E;IAC5E,oBAAoB;IACpB,eAAe;IACf,gBAAgB;IAChB,SAAS;IAET,4EAA4E;IAC5E,mCAAmC;IACnC,2CAA2C;CAC3C,CAAC,CAAC;AAEH,MAAM,gCAAgC,GAAG,yBAAyB,CAAC;IAClE,0EAA0E;IAC1E,YAAY;IACZ,aAAa;IACb,4BAA4B;IAC5B,oCAAoC;IACpC,sBAAsB;IACtB,aAAa;IACb,mBAAmB;IACnB,KAAK;IACL,KAAK;IACL,KAAK;IACL,KAAK;IACL,KAAK;IACL,KAAK;IACL,KAAK;IACL,kEAAkE;IAClE,iDAAiD;IACjD,eAAe;IACf,iBAAiB;IAEjB,8EAA8E;IAC9E,0EAA0E;IAC1E,sDAAsD;IACtD,oCAAoC;IACpC,uDAAuD;IACvD,2EAA2E;IAC3E,oEAAoE;IACpE,4BAA4B;IAE5B,2EAA2E;IAC3E,2EAA2E;IAC3E,2EAA2E;IAC3E,gBAAgB;IAChB,mBAAmB;IACnB,qBAAqB;IACrB,kBAAkB;IAClB,mBAAmB;IACnB,cAAc;IACd,aAAa;IACb,WAAW;IACX,WAAW;IACX,mBAAmB;IACnB,sBAAsB;IACtB,gBAAgB;IAChB,8BAA8B;IAC9B,YAAY;IACZ,SAAS;IACT,YAAY;IAEZ,+EAA+E;IAC/E,mBAAmB;IACnB,kBAAkB;IAElB,yEAAyE;IACzE,8EAA8E;IAC9E,yEAAyE;IACzE,eAAe;IACf,kCAAkC;IAClC,+CAA+C;IAC/C,sCAAsC;IACtC,+EAA+E;IAC/E,wEAAwE;IACxE,kCAAkC;IAElC,4EAA4E;IAC5E,iEAAiE;IACjE,aAAa;IAEb,6EAA6E;IAC7E,6BAA6B;IAC7B,4BAA4B;IAC5B,wBAAwB;IACxB,2BAA2B;IAE3B,yCAAyC;IACzC,mBAAmB;IAEnB,0EAA0E;IAC1E,qEAAqE;IACrE,wCAAwC;IACxC,uCAAuC;CACvC,CAAC,CAAC;AAmBH,MAAM,CAAC,MAAM,gCAAgC,GAAG,MAAM,CAAC;AAEvD,MAAM,UAAU,YAAY,CAAC,MAA4D,EAAE,OAAe;IACzG,MAAM,KAAK,GAAG,MAAM,CAAC,WAAW,GAAG,CAAC,IAAI,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,OAAO,GAAG,CAAC,CAAC,CAAC;IACjE,MAAM,SAAS,GAAG,MAAM,CAAC,aAAa,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,MAAM,CAAC,gBAAgB,CAAC;IAChF,OAAO,IAAI,CAAC,GAAG,CAAC,SAAS,EAAE,MAAM,CAAC,eAAe,IAAI,gCAAgC,CAAC,CAAC;AACxF,CAAC;AAiBD,MAAM,oBAAqB,SAAQ,KAAK;IACvC;QACC,KAAK,CAAC,SAAS,CAAC,CAAC;IAClB,CAAC;CACD;AAED,SAAS,KAAK,CAAC,EAAU,EAAE,MAAoB;IAC9C,OAAO,IAAI,OAAO,CAAC,CAAC,OAAO,EAAE,MAAM,EAAE,EAAE;QACtC,IAAI,MAAM,EAAE,OAAO,EAAE,CAAC;YACrB,MAAM,CAAC,IAAI,oBAAoB,EAAE,CAAC,CAAC;YACnC,OAAO;QACR,CAAC;QACD,MAAM,OAAO,GAAG,UAAU,CAAC,OAAO,EAAE,EAAE,CAAC,CAAC;QACxC,MAAM,EAAE,gBAAgB,CACvB,OAAO,EACP,GAAG,EAAE;YACJ,YAAY,CAAC,OAAO,CAAC,CAAC;YACtB,MAAM,CAAC,IAAI,oBAAoB,EAAE,CAAC,CAAC;QACpC,CAAC,EACD,EAAE,IAAI,EAAE,IAAI,EAAE,CACd,CAAC;IACH,CAAC,CAAC,CAAC;AACJ,CAAC;AAED;;;;;;;;;;;;;;;;;GAiBG;AACH,MAAM,CAAC,KAAK,UAAU,kBAAkB,CACvC,OAAwC,EACxC,MAA+B,EAC/B,MAA+B,EAC/B,SAA0B;IAE1B,MAAM,WAAW,GAAG,MAAM,EAAE,OAAO,CAAC,CAAC,CAAC,MAAM,CAAC,UAAU,CAAC,CAAC,CAAC,CAAC,CAAC;IAE5D,IAAI,OAAO,GAAG,CAAC,CAAC;IAChB,IAAI,SAAgE,CAAC;IACrE,SAAS,CAAC;QACT,MAAM,QAAQ,GAAG,MAAM,OAAO,EAAE,CAAC;QAEjC,sEAAsE;QACtE,IAAI,QAAQ,CAAC,UAAU,KAAK,SAAS,EAAE,CAAC;YACvC,IAAI,SAAS;gBAAE,MAAM,SAAS,EAAE,eAAe,EAAE,CAAC,KAAK,EAAE,SAAS,CAAC,OAAO,CAAC,CAAC;YAC5E,OAAO,QAAQ,CAAC;QACjB,CAAC;QAED,wDAAwD;QACxD,IAAI,QAAQ,CAAC,UAAU,KAAK,OAAO,EAAE,CAAC;YACrC,IAAI,SAAS;gBAAE,MAAM,SAAS,EAAE,eAAe,EAAE,CAAC,IAAI,EAAE,SAAS,CAAC,OAAO,CAAC,CAAC;YAC3E,OAAO,QAAQ,CAAC;QACjB,CAAC;QAED,sEAAsE;QACtE,IAAI,OAAO,IAAI,WAAW,IAAI,CAAC,yBAAyB,CAAC,QAAQ,CAAC,EAAE,CAAC;YACpE,IAAI,SAAS;gBAAE,MAAM,SAAS,EAAE,eAAe,EAAE,CAAC,KAAK,EAAE,SAAS,CAAC,OAAO,EAAE,QAAQ,CAAC,YAAY,CAAC,CAAC;YACnG,OAAO,QAAQ,CAAC;QACjB,CAAC;QAED,OAAO,EAAE,CAAC;QACV,SAAS,GAAG,EAAE,OAAO,EAAE,YAAY,EAAE,QAAQ,CAAC,YAAY,IAAI,eAAe,EAAE,CAAC;QAChF,MAAM,OAAO,GAAG,YAAY,CAAC,MAAO,EAAE,OAAO,CAAC,CAAC;QAC/C,MAAM,SAAS,EAAE,gBAAgB,EAAE,CAAC,OAAO,EAAE,WAAW,EAAE,OAAO,EAAE,SAAS,CAAC,YAAY,CAAC,CAAC;QAE3F,8EAA8E;QAC9E,qFAAqF;QACrF,IAAI,CAAC;YACJ,MAAM,KAAK,CAAC,OAAO,EAAE,MAAM,CAAC,CAAC;QAC9B,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YAChB,MAAM,SAAS,EAAE,eAAe,EAAE,CAAC,KAAK,EAAE,OAAO,EAAE,SAAS,CAAC,YAAY,CAAC,CAAC;YAC3E,IAAI,KAAK,YAAY,oBAAoB,EAAE,CAAC;gBAC3C,MAAM,EAAE,YAAY,EAAE,aAAa,EAAE,GAAG,IAAI,EAAE,GAAG,QAAQ,CAAC;gBAC1D,OAAO,EAAE,GAAG,IAAI,EAAE,UAAU,EAAE,SAAS,EAAE,CAAC;YAC3C,CAAC;YACD,MAAM,KAAK,CAAC;QACb,CAAC;QACD,MAAM,SAAS,EAAE,mBAAmB,EAAE,EAAE,CAAC;IAC1C,CAAC;AACF,CAAC;AAED;;;;;;;;GAQG;AACH,MAAM,UAAU,yBAAyB,CAAC,OAAyB;IAClE,IAAI,OAAO,CAAC,UAAU,KAAK,OAAO,IAAI,CAAC,OAAO,CAAC,YAAY;QAAE,OAAO,KAAK,CAAC;IAC1E,MAAM,YAAY,GAAG,OAAO,CAAC,YAAY,CAAC;IAC1C,IAAI,0CAA0C,CAAC,IAAI,CAAC,YAAY,CAAC;QAAE,OAAO,KAAK,CAAC;IAChF,OAAO,gCAAgC,CAAC,IAAI,CAAC,YAAY,CAAC,CAAC;AAC5D,CAAC","sourcesContent":["import type { AssistantMessage } from \"../types.ts\";\n\nfunction buildProviderErrorPattern(patterns: readonly string[]): RegExp {\n\treturn new RegExp(patterns.join(\"|\"), \"i\");\n}\n\nconst NON_RETRYABLE_PROVIDER_LIMIT_ERROR_PATTERN = buildProviderErrorPattern([\n\t// OpenCode Go/free-tier limits returned as 429 JSON error types by OpenCode's\n\t// Zen API. These are subscription/account limits, not transient throttles.\n\t\"GoUsageLimitError\",\n\t\"FreeUsageLimitError\",\n\n\t// OpenCode Go subscription-limit text asks users to enable available-balance\n\t// usage after rolling/weekly/monthly limits are reached.\n\t\"Monthly usage limit reached\",\n\t\"available balance\",\n\n\t// Generic quota/budget/billing exhaustion. `insufficient_quota` is OpenAI's\n\t// quota/billing error code; the other strings cover common gateway wording.\n\t\"insufficient_quota\",\n\t\"out of budget\",\n\t\"quota exceeded\",\n\t\"billing\",\n\n\t// Sign in with ChatGPT: the subscription's shared usage limit, which resets\n\t// after hours rather than seconds.\n\t\"subscription_sharing_usage_limit_exceeded\",\n]);\n\nconst RETRYABLE_PROVIDER_ERROR_PATTERN = buildProviderErrorPattern([\n\t// Generic provider load, HTTP status, and server-side transient failures.\n\t\"overloaded\",\n\t\"server_busy\",\n\t\"servers are currently busy\",\n\t\"currently experiencing high demand\",\n\t\"model is at capacity\",\n\t\"rate.?limit\",\n\t\"too many requests\",\n\t\"429\",\n\t\"500\",\n\t\"502\",\n\t\"503\",\n\t\"504\",\n\t\"520\",\n\t\"524\",\n\t// Also matches \"The service is temporarily unavailable\" (#10965).\n\t\"service.?(?:is.?)?(?:temporarily.?)?unavailable\",\n\t\"server.?error\",\n\t\"internal.?error\",\n\n\t// Wrapper/provider text for transient upstream failures, including OpenRouter\n\t// \"Provider returned error\" responses (#2264) and gateway wording such as\n\t// \"The upstream provider returned an error\" (#10842).\n\t\"provider.?returned.?(?:an.?)?error\",\n\t\"exceeded request buffer limit while retrying upstream\",\n\t// Gateway failure relaying the upstream response stream, sent as an OpenAI\n\t// Responses `response.failed` event without an error code (#10920).\n\t\"unable to forward upstream\",\n\n\t// Network, proxy, and fetch transport failures. This includes OpenAI Codex\n\t// raw-fetch failures such as \"upstream connect\", \"connection refused\", and\n\t// \"reset before headers\" (#733), plus OpenRouter connection drops (#3317).\n\t\"network.?error\",\n\t\"connection.?error\",\n\t\"connection.?refused\",\n\t\"connection.?lost\",\n\t\"other side closed\",\n\t\"fetch failed\",\n\t\"getaddrinfo\",\n\t\"ENOTFOUND\",\n\t\"EAI_AGAIN\",\n\t\"upstream.?connect\",\n\t\"reset before headers\",\n\t\"socket hang up\",\n\t\"socket connection was closed\",\n\t\"timed? out\",\n\t\"timeout\",\n\t\"terminated\",\n\n\t// WebSocket transports can report close/error text instead of HTTP/fetch text.\n\t\"websocket.?closed\",\n\t\"websocket.?error\",\n\n\t// Premature stream endings from SDKs and transports. Anthropic can throw\n\t// \"stream ended without ...\" and \"Anthropic stream ended before message_stop\"\n\t// (#4433); Bedrock/Smithy can throw an HTTP/2 no-response error (#3594).\n\t\"ended without\",\n\t\"stream ended before message_stop\",\n\t\"stream ended before a terminal response event\",\n\t\"http2 request did not get a response\",\n\t// Node ERR_HTTP2_STREAM_CANCEL: the HTTP/2 session died before the request was\n\t// sent, e.g. after the Bedrock SDK's 5-minute session timeout (#10379).\n\t\"pending stream has been canceled\",\n\n\t// Provider-requested retry delay cap failures should flow through the outer\n\t// retry policy so callers can surface/abort the backoff (#1123).\n\t\"retry delay\",\n\n\t// Explicit retry guidance emitted mid-stream by OpenAI Responses and Bedrock\n\t// stream exceptions (#6019).\n\t\"you can retry your request\",\n\t\"try your request again\",\n\t\"please retry your request\",\n\n\t// gRPC based providers (e.g. NVIDIA NIM)\n\t\"ResourceExhausted\",\n\n\t// Sign in with ChatGPT: usage or user data temporarily unavailable. Usage\n\t// failures can arrive mid-stream without an HTTP 503 in the message.\n\t\"subscription_sharing_usage_unavailable\",\n\t\"subscription_sharing_user_unavailable\",\n]);\n\n/**\n * Retry policy: bounded attempts with exponential backoff (`baseDelayMs * 2^(attempt-1)`).\n * `maxAgentDelayMs` caps each computed delay and defaults to 60 seconds.\n * Matches `settings.retry` (`enabled`, `maxRetries`, `baseDelayMs`, `maxAgentDelayMs`) in coding-agent; kept\n * here so the classifier and the policy-driven retry loop live together and stay reusable\n * by the SDK and other callers.\n */\nexport interface RetryPolicy {\n\tenabled: boolean;\n\t/** Max retry attempts (0 = no retries). The initial call never counts as a retry. */\n\tmaxRetries: number;\n\t/** Base delay in ms. Per-attempt delay is `baseDelayMs * 2^(attempt-1)` before jitter. */\n\tbaseDelayMs: number;\n\t/** Optional cap for agent-level retry delays in ms. Defaults to 60 seconds. */\n\tmaxAgentDelayMs?: number;\n}\n\nexport const DEFAULT_MAX_AGENT_RETRY_DELAY_MS = 60_000;\n\nexport function retryDelayMs(policy: Pick<RetryPolicy, \"baseDelayMs\" | \"maxAgentDelayMs\">, attempt: number): number {\n\tconst delay = policy.baseDelayMs * 2 ** Math.max(0, attempt - 1);\n\tconst safeDelay = Number.isSafeInteger(delay) ? delay : Number.MAX_SAFE_INTEGER;\n\treturn Math.min(safeDelay, policy.maxAgentDelayMs ?? DEFAULT_MAX_AGENT_RETRY_DELAY_MS);\n}\n\n/** Optional callbacks emitted by {@link retryAssistantCall} around each retry. */\nexport interface RetryCallbacks {\n\t/** Emitted before the backoff sleep of each retry attempt (1-indexed). */\n\tonRetryScheduled?: (\n\t\tattempt: number,\n\t\tmaxAttempts: number,\n\t\tdelayMs: number,\n\t\terrorMessage: string,\n\t) => void | Promise<void>;\n\t/** Emitted after the backoff sleep, immediately before the retried call starts. */\n\tonRetryAttemptStart?: () => void | Promise<void>;\n\t/** Emitted once when the loop ends: success if a later call completed normally. */\n\tonRetryFinished?: (success: boolean, attempt: number, finalError?: string) => void | Promise<void>;\n}\n\nclass RetrySleepAbortError extends Error {\n\tconstructor() {\n\t\tsuper(\"Aborted\");\n\t}\n}\n\nfunction sleep(ms: number, signal?: AbortSignal): Promise<void> {\n\treturn new Promise((resolve, reject) => {\n\t\tif (signal?.aborted) {\n\t\t\treject(new RetrySleepAbortError());\n\t\t\treturn;\n\t\t}\n\t\tconst timeout = setTimeout(resolve, ms);\n\t\tsignal?.addEventListener(\n\t\t\t\"abort\",\n\t\t\t() => {\n\t\t\t\tclearTimeout(timeout);\n\t\t\t\treject(new RetrySleepAbortError());\n\t\t\t},\n\t\t\t{ once: true },\n\t\t);\n\t});\n}\n\n/**\n * Run a single assistant-producing call with bounded retry on transient errors.\n *\n * Behavior:\n * - A successful response is returned immediately. Aborts are terminal and never\n *   retried, but reported as unsuccessful if they happen after a retry was scheduled.\n *   Aborts during the backoff sleep are normalized to an aborted `AssistantMessage`\n *   too, so callers do not need to care when cancellation happened.\n * - A non-retryable error (per {@link isRetryableAssistantError}, including quota/\n *   billing exhaustion) is returned immediately so deterministic errors fail fast.\n * - Otherwise retries up to `maxRetries` times with exponential backoff, emitting\n *   `onRetryScheduled` before each sleep, `onRetryAttemptStart` after each sleep before\n *   the retried call starts, and `onRetryFinished` once at the end (whether the loop\n *   ends in success, exhausted retries, or an aborted backoff).\n *\n * When `policy` is undefined or disabled, the first response is returned unchanged\n * (equivalent to calling `produce()` directly).\n */\nexport async function retryAssistantCall(\n\tproduce: () => Promise<AssistantMessage>,\n\tpolicy: RetryPolicy | undefined,\n\tsignal: AbortSignal | undefined,\n\tcallbacks?: RetryCallbacks,\n): Promise<AssistantMessage> {\n\tconst maxAttempts = policy?.enabled ? policy.maxRetries : 0;\n\n\tlet attempt = 0;\n\tlet lastRetry: { attempt: number; errorMessage: string } | undefined;\n\tfor (;;) {\n\t\tconst response = await produce();\n\n\t\t// Abort: terminal but not successful. Never retry an aborted message.\n\t\tif (response.stopReason === \"aborted\") {\n\t\t\tif (lastRetry) await callbacks?.onRetryFinished?.(false, lastRetry.attempt);\n\t\t\treturn response;\n\t\t}\n\n\t\t// Success: non-error, non-abort responses return as-is.\n\t\tif (response.stopReason !== \"error\") {\n\t\t\tif (lastRetry) await callbacks?.onRetryFinished?.(true, lastRetry.attempt);\n\t\t\treturn response;\n\t\t}\n\n\t\t// Non-retryable, or budget exhausted: return the final error message.\n\t\tif (attempt >= maxAttempts || !isRetryableAssistantError(response)) {\n\t\t\tif (lastRetry) await callbacks?.onRetryFinished?.(false, lastRetry.attempt, response.errorMessage);\n\t\t\treturn response;\n\t\t}\n\n\t\tattempt++;\n\t\tlastRetry = { attempt, errorMessage: response.errorMessage || \"Unknown error\" };\n\t\tconst delayMs = retryDelayMs(policy!, attempt);\n\t\tawait callbacks?.onRetryScheduled?.(attempt, maxAttempts, delayMs, lastRetry.errorMessage);\n\n\t\t// Normalize aborts during retry backoff to the same AssistantMessage shape as\n\t\t// provider stream aborts, so callers do not need to care when cancellation happened.\n\t\ttry {\n\t\t\tawait sleep(delayMs, signal);\n\t\t} catch (error) {\n\t\t\tawait callbacks?.onRetryFinished?.(false, attempt, lastRetry.errorMessage);\n\t\t\tif (error instanceof RetrySleepAbortError) {\n\t\t\t\tconst { errorMessage: _errorMessage, ...rest } = response;\n\t\t\t\treturn { ...rest, stopReason: \"aborted\" };\n\t\t\t}\n\t\t\tthrow error;\n\t\t}\n\t\tawait callbacks?.onRetryAttemptStart?.();\n\t}\n}\n\n/**\n * Classifies whether a failed assistant message looks like a transient provider\n * or transport error, so callers can decide if the last assistant turn should be\n * restarted.\n *\n * This does not implement retry policy. Callers should first handle context\n * overflow separately, then apply their own retry budget, backoff, and reporting\n * before restarting the assistant turn.\n */\nexport function isRetryableAssistantError(message: AssistantMessage): boolean {\n\tif (message.stopReason !== \"error\" || !message.errorMessage) return false;\n\tconst errorMessage = message.errorMessage;\n\tif (NON_RETRYABLE_PROVIDER_LIMIT_ERROR_PATTERN.test(errorMessage)) return false;\n\treturn RETRYABLE_PROVIDER_ERROR_PATTERN.test(errorMessage);\n}\n"]}
+{"version":3,"file":"retry.js","sourceRoot":"","sources":["../../src/utils/retry.ts"],"names":[],"mappings":"AAEA,SAAS,yBAAyB,CAAC,QAA2B;IAC7D,OAAO,IAAI,MAAM,CAAC,QAAQ,CAAC,IAAI,CAAC,GAAG,CAAC,EAAE,GAAG,CAAC,CAAC;AAC5C,CAAC;AAED,MAAM,0CAA0C,GAAG,yBAAyB,CAAC;IAC5E,8EAA8E;IAC9E,2EAA2E;IAC3E,mBAAmB;IACnB,qBAAqB;IAErB,6EAA6E;IAC7E,yDAAyD;IACzD,6BAA6B;IAC7B,mBAAmB;IAEnB,4EAA4E;IAC5E,4EAA4E;IAC5E,oBAAoB;IACpB,eAAe;IACf,gBAAgB;IAChB,SAAS;IAET,4EAA4E;IAC5E,mCAAmC;IACnC,2CAA2C;IAC3C,4EAA4E;IAC5E,gDAAgD;IAChD,gEAAgE;CAChE,CAAC,CAAC;AAEH,MAAM,gCAAgC,GAAG,yBAAyB,CAAC;IAClE,0EAA0E;IAC1E,YAAY;IACZ,aAAa;IACb,4BAA4B;IAC5B,oCAAoC;IACpC,sBAAsB;IACtB,aAAa;IACb,mBAAmB;IACnB,sGAAsG;IACtG,0DAA0D;IAC1D,kEAAkE;IAClE,iDAAiD;IACjD,eAAe;IACf,iBAAiB;IAEjB,8EAA8E;IAC9E,0EAA0E;IAC1E,sDAAsD;IACtD,oCAAoC;IACpC,uDAAuD;IACvD,2EAA2E;IAC3E,oEAAoE;IACpE,4BAA4B;IAC5B,qFAAqF;IACrF,2BAA2B;IAC3B,4BAA4B;IAC5B,wBAAwB;IACxB,uBAAuB;IAEvB,2EAA2E;IAC3E,2EAA2E;IAC3E,2EAA2E;IAC3E,gBAAgB;IAChB,mBAAmB;IACnB,qBAAqB;IACrB,kBAAkB;IAClB,mBAAmB;IACnB,cAAc;IACd,4EAA4E;IAC5E,qDAAqD;IACrD,iCAAiC;IACjC,aAAa;IACb,WAAW;IACX,WAAW;IACX,mBAAmB;IACnB,sBAAsB;IACtB,gBAAgB;IAChB,8BAA8B;IAC9B,YAAY;IACZ,SAAS;IACT,YAAY;IAEZ,+EAA+E;IAC/E,mBAAmB;IACnB,kBAAkB;IAElB,yEAAyE;IACzE,8EAA8E;IAC9E,yEAAyE;IACzE,eAAe;IACf,kCAAkC;IAClC,+CAA+C;IAC/C,sCAAsC;IACtC,+EAA+E;IAC/E,kGAAkG;IAClG,uCAAuC;IACvC,+EAA+E;IAC/E,wEAAwE;IACxE,kCAAkC;IAElC,4EAA4E;IAC5E,iEAAiE;IACjE,aAAa;IAEb,6EAA6E;IAC7E,6BAA6B;IAC7B,4BAA4B;IAC5B,wBAAwB;IACxB,2BAA2B;IAE3B,yCAAyC;IACzC,mBAAmB;IAEnB,0EAA0E;IAC1E,qEAAqE;IACrE,wCAAwC;IACxC,uCAAuC;CACvC,CAAC,CAAC;AAmBH,MAAM,CAAC,MAAM,gCAAgC,GAAG,MAAM,CAAC;AAEvD,MAAM,UAAU,YAAY,CAAC,MAA4D,EAAE,OAAe;IACzG,MAAM,KAAK,GAAG,MAAM,CAAC,WAAW,GAAG,CAAC,IAAI,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,OAAO,GAAG,CAAC,CAAC,CAAC;IACjE,MAAM,SAAS,GAAG,MAAM,CAAC,aAAa,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,MAAM,CAAC,gBAAgB,CAAC;IAChF,OAAO,IAAI,CAAC,GAAG,CAAC,SAAS,EAAE,MAAM,CAAC,eAAe,IAAI,gCAAgC,CAAC,CAAC;AACxF,CAAC;AAiBD,MAAM,oBAAqB,SAAQ,KAAK;IACvC;QACC,KAAK,CAAC,SAAS,CAAC,CAAC;IAClB,CAAC;CACD;AAED,SAAS,KAAK,CAAC,EAAU,EAAE,MAAoB;IAC9C,OAAO,IAAI,OAAO,CAAC,CAAC,OAAO,EAAE,MAAM,EAAE,EAAE;QACtC,IAAI,MAAM,EAAE,OAAO,EAAE,CAAC;YACrB,MAAM,CAAC,IAAI,oBAAoB,EAAE,CAAC,CAAC;YACnC,OAAO;QACR,CAAC;QACD,MAAM,OAAO,GAAG,UAAU,CAAC,OAAO,EAAE,EAAE,CAAC,CAAC;QACxC,MAAM,EAAE,gBAAgB,CACvB,OAAO,EACP,GAAG,EAAE;YACJ,YAAY,CAAC,OAAO,CAAC,CAAC;YACtB,MAAM,CAAC,IAAI,oBAAoB,EAAE,CAAC,CAAC;QACpC,CAAC,EACD,EAAE,IAAI,EAAE,IAAI,EAAE,CACd,CAAC;IACH,CAAC,CAAC,CAAC;AACJ,CAAC;AAED;;;;;;;;;;;;;;;;;GAiBG;AACH,MAAM,CAAC,KAAK,UAAU,kBAAkB,CACvC,OAAwC,EACxC,MAA+B,EAC/B,MAA+B,EAC/B,SAA0B;IAE1B,MAAM,WAAW,GAAG,MAAM,EAAE,OAAO,CAAC,CAAC,CAAC,MAAM,CAAC,UAAU,CAAC,CAAC,CAAC,CAAC,CAAC;IAE5D,IAAI,OAAO,GAAG,CAAC,CAAC;IAChB,IAAI,SAAgE,CAAC;IACrE,SAAS,CAAC;QACT,MAAM,QAAQ,GAAG,MAAM,OAAO,EAAE,CAAC;QAEjC,sEAAsE;QACtE,IAAI,QAAQ,CAAC,UAAU,KAAK,SAAS,EAAE,CAAC;YACvC,IAAI,SAAS;gBAAE,MAAM,SAAS,EAAE,eAAe,EAAE,CAAC,KAAK,EAAE,SAAS,CAAC,OAAO,CAAC,CAAC;YAC5E,OAAO,QAAQ,CAAC;QACjB,CAAC;QAED,wDAAwD;QACxD,IAAI,QAAQ,CAAC,UAAU,KAAK,OAAO,EAAE,CAAC;YACrC,IAAI,SAAS;gBAAE,MAAM,SAAS,EAAE,eAAe,EAAE,CAAC,IAAI,EAAE,SAAS,CAAC,OAAO,CAAC,CAAC;YAC3E,OAAO,QAAQ,CAAC;QACjB,CAAC;QAED,sEAAsE;QACtE,IAAI,OAAO,IAAI,WAAW,IAAI,CAAC,yBAAyB,CAAC,QAAQ,CAAC,EAAE,CAAC;YACpE,IAAI,SAAS;gBAAE,MAAM,SAAS,EAAE,eAAe,EAAE,CAAC,KAAK,EAAE,SAAS,CAAC,OAAO,EAAE,QAAQ,CAAC,YAAY,CAAC,CAAC;YACnG,OAAO,QAAQ,CAAC;QACjB,CAAC;QAED,OAAO,EAAE,CAAC;QACV,SAAS,GAAG,EAAE,OAAO,EAAE,YAAY,EAAE,QAAQ,CAAC,YAAY,IAAI,eAAe,EAAE,CAAC;QAChF,MAAM,OAAO,GAAG,YAAY,CAAC,MAAO,EAAE,OAAO,CAAC,CAAC;QAC/C,MAAM,SAAS,EAAE,gBAAgB,EAAE,CAAC,OAAO,EAAE,WAAW,EAAE,OAAO,EAAE,SAAS,CAAC,YAAY,CAAC,CAAC;QAE3F,8EAA8E;QAC9E,qFAAqF;QACrF,IAAI,CAAC;YACJ,MAAM,KAAK,CAAC,OAAO,EAAE,MAAM,CAAC,CAAC;QAC9B,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YAChB,MAAM,SAAS,EAAE,eAAe,EAAE,CAAC,KAAK,EAAE,OAAO,EAAE,SAAS,CAAC,YAAY,CAAC,CAAC;YAC3E,IAAI,KAAK,YAAY,oBAAoB,EAAE,CAAC;gBAC3C,MAAM,EAAE,YAAY,EAAE,aAAa,EAAE,GAAG,IAAI,EAAE,GAAG,QAAQ,CAAC;gBAC1D,OAAO,EAAE,GAAG,IAAI,EAAE,UAAU,EAAE,SAAS,EAAE,CAAC;YAC3C,CAAC;YACD,MAAM,KAAK,CAAC;QACb,CAAC;QACD,MAAM,SAAS,EAAE,mBAAmB,EAAE,EAAE,CAAC;IAC1C,CAAC;AACF,CAAC;AAED;;;;;;;;GAQG;AACH,MAAM,UAAU,yBAAyB,CAAC,OAAyB;IAClE,IAAI,OAAO,CAAC,UAAU,KAAK,OAAO,IAAI,CAAC,OAAO,CAAC,YAAY;QAAE,OAAO,KAAK,CAAC;IAC1E,MAAM,YAAY,GAAG,OAAO,CAAC,YAAY,CAAC;IAC1C,IAAI,0CAA0C,CAAC,IAAI,CAAC,YAAY,CAAC;QAAE,OAAO,KAAK,CAAC;IAChF,OAAO,gCAAgC,CAAC,IAAI,CAAC,YAAY,CAAC,CAAC;AAC5D,CAAC","sourcesContent":["import type { AssistantMessage } from \"../types.ts\";\n\nfunction buildProviderErrorPattern(patterns: readonly string[]): RegExp {\n\treturn new RegExp(patterns.join(\"|\"), \"i\");\n}\n\nconst NON_RETRYABLE_PROVIDER_LIMIT_ERROR_PATTERN = buildProviderErrorPattern([\n\t// OpenCode Go/free-tier limits returned as 429 JSON error types by OpenCode's\n\t// Zen API. These are subscription/account limits, not transient throttles.\n\t\"GoUsageLimitError\",\n\t\"FreeUsageLimitError\",\n\n\t// OpenCode Go subscription-limit text asks users to enable available-balance\n\t// usage after rolling/weekly/monthly limits are reached.\n\t\"Monthly usage limit reached\",\n\t\"available balance\",\n\n\t// Generic quota/budget/billing exhaustion. `insufficient_quota` is OpenAI's\n\t// quota/billing error code; the other strings cover common gateway wording.\n\t\"insufficient_quota\",\n\t\"out of budget\",\n\t\"quota exceeded\",\n\t\"billing\",\n\n\t// Sign in with ChatGPT: the subscription's shared usage limit, which resets\n\t// after hours rather than seconds.\n\t\"subscription_sharing_usage_limit_exceeded\",\n\t// Sign in with ChatGPT on a ChatGPT Business plan: a usage limit set by the\n\t// workspace admin. Only the admin can raise it.\n\t\"subscription_sharing_workplace_individual_usage_limit_exceeded\",\n]);\n\nconst RETRYABLE_PROVIDER_ERROR_PATTERN = buildProviderErrorPattern([\n\t// Generic provider load, HTTP status, and server-side transient failures.\n\t\"overloaded\",\n\t\"server_busy\",\n\t\"servers are currently busy\",\n\t\"currently experiencing high demand\",\n\t\"model is at capacity\",\n\t\"rate.?limit\",\n\t\"too many requests\",\n\t// Status codes must not be part of a longer word or number, e.g. a request id or byte count (#11057).\n\t\"(?<![a-z0-9])(?:429|500|502|503|504|520|524)(?![a-z0-9])\",\n\t// Also matches \"The service is temporarily unavailable\" (#10965).\n\t\"service.?(?:is.?)?(?:temporarily.?)?unavailable\",\n\t\"server.?error\",\n\t\"internal.?error\",\n\n\t// Wrapper/provider text for transient upstream failures, including OpenRouter\n\t// \"Provider returned error\" responses (#2264) and gateway wording such as\n\t// \"The upstream provider returned an error\" (#10842).\n\t\"provider.?returned.?(?:an.?)?error\",\n\t\"exceeded request buffer limit while retrying upstream\",\n\t// Gateway failure relaying the upstream response stream, sent as an OpenAI\n\t// Responses `response.failed` event without an error code (#10920).\n\t\"unable to forward upstream\",\n\t// Other gateway upstream failures relayed as `response.failed` error codes (#10920).\n\t\"upstream.?request.?failed\",\n\t\"upstream_stream_read_error\",\n\t\"stream was interrupted\",\n\t\"http/?2 stream failed\",\n\n\t// Network, proxy, and fetch transport failures. This includes OpenAI Codex\n\t// raw-fetch failures such as \"upstream connect\", \"connection refused\", and\n\t// \"reset before headers\" (#733), plus OpenRouter connection drops (#3317).\n\t\"network.?error\",\n\t\"connection.?error\",\n\t\"connection.?refused\",\n\t\"connection.?lost\",\n\t\"other side closed\",\n\t\"fetch failed\",\n\t// node-fetch system errors, e.g. the Google ADC OAuth token refresh used by\n\t// google-vertex. The reason is often empty (#10996).\n\t\"request to \\\\S+ failed, reason:\",\n\t\"getaddrinfo\",\n\t\"ENOTFOUND\",\n\t\"EAI_AGAIN\",\n\t\"upstream.?connect\",\n\t\"reset before headers\",\n\t\"socket hang up\",\n\t\"socket connection was closed\",\n\t\"timed? out\",\n\t\"timeout\",\n\t\"terminated\",\n\n\t// WebSocket transports can report close/error text instead of HTTP/fetch text.\n\t\"websocket.?closed\",\n\t\"websocket.?error\",\n\n\t// Premature stream endings from SDKs and transports. Anthropic can throw\n\t// \"stream ended without ...\" and \"Anthropic stream ended before message_stop\"\n\t// (#4433); Bedrock/Smithy can throw an HTTP/2 no-response error (#3594).\n\t\"ended without\",\n\t\"stream ended before message_stop\",\n\t\"stream ended before a terminal response event\",\n\t\"http2 request did not get a response\",\n\t// Codex CLI mid-stream disconnect, relayed by proxies that wrap the Codex CLI,\n\t// e.g. \"stream disconnected before completion: stream closed before response.completed\" (#11057).\n\t\"stream disconnected before completion\",\n\t// Node ERR_HTTP2_STREAM_CANCEL: the HTTP/2 session died before the request was\n\t// sent, e.g. after the Bedrock SDK's 5-minute session timeout (#10379).\n\t\"pending stream has been canceled\",\n\n\t// Provider-requested retry delay cap failures should flow through the outer\n\t// retry policy so callers can surface/abort the backoff (#1123).\n\t\"retry delay\",\n\n\t// Explicit retry guidance emitted mid-stream by OpenAI Responses and Bedrock\n\t// stream exceptions (#6019).\n\t\"you can retry your request\",\n\t\"try your request again\",\n\t\"please retry your request\",\n\n\t// gRPC based providers (e.g. NVIDIA NIM)\n\t\"ResourceExhausted\",\n\n\t// Sign in with ChatGPT: usage or user data temporarily unavailable. Usage\n\t// failures can arrive mid-stream without an HTTP 503 in the message.\n\t\"subscription_sharing_usage_unavailable\",\n\t\"subscription_sharing_user_unavailable\",\n]);\n\n/**\n * Retry policy: bounded attempts with exponential backoff (`baseDelayMs * 2^(attempt-1)`).\n * `maxAgentDelayMs` caps each computed delay and defaults to 60 seconds.\n * Matches `settings.retry` (`enabled`, `maxRetries`, `baseDelayMs`, `maxAgentDelayMs`) in coding-agent; kept\n * here so the classifier and the policy-driven retry loop live together and stay reusable\n * by the SDK and other callers.\n */\nexport interface RetryPolicy {\n\tenabled: boolean;\n\t/** Max retry attempts (0 = no retries). The initial call never counts as a retry. */\n\tmaxRetries: number;\n\t/** Base delay in ms. Per-attempt delay is `baseDelayMs * 2^(attempt-1)` before jitter. */\n\tbaseDelayMs: number;\n\t/** Optional cap for agent-level retry delays in ms. Defaults to 60 seconds. */\n\tmaxAgentDelayMs?: number;\n}\n\nexport const DEFAULT_MAX_AGENT_RETRY_DELAY_MS = 60_000;\n\nexport function retryDelayMs(policy: Pick<RetryPolicy, \"baseDelayMs\" | \"maxAgentDelayMs\">, attempt: number): number {\n\tconst delay = policy.baseDelayMs * 2 ** Math.max(0, attempt - 1);\n\tconst safeDelay = Number.isSafeInteger(delay) ? delay : Number.MAX_SAFE_INTEGER;\n\treturn Math.min(safeDelay, policy.maxAgentDelayMs ?? DEFAULT_MAX_AGENT_RETRY_DELAY_MS);\n}\n\n/** Optional callbacks emitted by {@link retryAssistantCall} around each retry. */\nexport interface RetryCallbacks {\n\t/** Emitted before the backoff sleep of each retry attempt (1-indexed). */\n\tonRetryScheduled?: (\n\t\tattempt: number,\n\t\tmaxAttempts: number,\n\t\tdelayMs: number,\n\t\terrorMessage: string,\n\t) => void | Promise<void>;\n\t/** Emitted after the backoff sleep, immediately before the retried call starts. */\n\tonRetryAttemptStart?: () => void | Promise<void>;\n\t/** Emitted once when the loop ends: success if a later call completed normally. */\n\tonRetryFinished?: (success: boolean, attempt: number, finalError?: string) => void | Promise<void>;\n}\n\nclass RetrySleepAbortError extends Error {\n\tconstructor() {\n\t\tsuper(\"Aborted\");\n\t}\n}\n\nfunction sleep(ms: number, signal?: AbortSignal): Promise<void> {\n\treturn new Promise((resolve, reject) => {\n\t\tif (signal?.aborted) {\n\t\t\treject(new RetrySleepAbortError());\n\t\t\treturn;\n\t\t}\n\t\tconst timeout = setTimeout(resolve, ms);\n\t\tsignal?.addEventListener(\n\t\t\t\"abort\",\n\t\t\t() => {\n\t\t\t\tclearTimeout(timeout);\n\t\t\t\treject(new RetrySleepAbortError());\n\t\t\t},\n\t\t\t{ once: true },\n\t\t);\n\t});\n}\n\n/**\n * Run a single assistant-producing call with bounded retry on transient errors.\n *\n * Behavior:\n * - A successful response is returned immediately. Aborts are terminal and never\n *   retried, but reported as unsuccessful if they happen after a retry was scheduled.\n *   Aborts during the backoff sleep are normalized to an aborted `AssistantMessage`\n *   too, so callers do not need to care when cancellation happened.\n * - A non-retryable error (per {@link isRetryableAssistantError}, including quota/\n *   billing exhaustion) is returned immediately so deterministic errors fail fast.\n * - Otherwise retries up to `maxRetries` times with exponential backoff, emitting\n *   `onRetryScheduled` before each sleep, `onRetryAttemptStart` after each sleep before\n *   the retried call starts, and `onRetryFinished` once at the end (whether the loop\n *   ends in success, exhausted retries, or an aborted backoff).\n *\n * When `policy` is undefined or disabled, the first response is returned unchanged\n * (equivalent to calling `produce()` directly).\n */\nexport async function retryAssistantCall(\n\tproduce: () => Promise<AssistantMessage>,\n\tpolicy: RetryPolicy | undefined,\n\tsignal: AbortSignal | undefined,\n\tcallbacks?: RetryCallbacks,\n): Promise<AssistantMessage> {\n\tconst maxAttempts = policy?.enabled ? policy.maxRetries : 0;\n\n\tlet attempt = 0;\n\tlet lastRetry: { attempt: number; errorMessage: string } | undefined;\n\tfor (;;) {\n\t\tconst response = await produce();\n\n\t\t// Abort: terminal but not successful. Never retry an aborted message.\n\t\tif (response.stopReason === \"aborted\") {\n\t\t\tif (lastRetry) await callbacks?.onRetryFinished?.(false, lastRetry.attempt);\n\t\t\treturn response;\n\t\t}\n\n\t\t// Success: non-error, non-abort responses return as-is.\n\t\tif (response.stopReason !== \"error\") {\n\t\t\tif (lastRetry) await callbacks?.onRetryFinished?.(true, lastRetry.attempt);\n\t\t\treturn response;\n\t\t}\n\n\t\t// Non-retryable, or budget exhausted: return the final error message.\n\t\tif (attempt >= maxAttempts || !isRetryableAssistantError(response)) {\n\t\t\tif (lastRetry) await callbacks?.onRetryFinished?.(false, lastRetry.attempt, response.errorMessage);\n\t\t\treturn response;\n\t\t}\n\n\t\tattempt++;\n\t\tlastRetry = { attempt, errorMessage: response.errorMessage || \"Unknown error\" };\n\t\tconst delayMs = retryDelayMs(policy!, attempt);\n\t\tawait callbacks?.onRetryScheduled?.(attempt, maxAttempts, delayMs, lastRetry.errorMessage);\n\n\t\t// Normalize aborts during retry backoff to the same AssistantMessage shape as\n\t\t// provider stream aborts, so callers do not need to care when cancellation happened.\n\t\ttry {\n\t\t\tawait sleep(delayMs, signal);\n\t\t} catch (error) {\n\t\t\tawait callbacks?.onRetryFinished?.(false, attempt, lastRetry.errorMessage);\n\t\t\tif (error instanceof RetrySleepAbortError) {\n\t\t\t\tconst { errorMessage: _errorMessage, ...rest } = response;\n\t\t\t\treturn { ...rest, stopReason: \"aborted\" };\n\t\t\t}\n\t\t\tthrow error;\n\t\t}\n\t\tawait callbacks?.onRetryAttemptStart?.();\n\t}\n}\n\n/**\n * Classifies whether a failed assistant message looks like a transient provider\n * or transport error, so callers can decide if the last assistant turn should be\n * restarted.\n *\n * This does not implement retry policy. Callers should first handle context\n * overflow separately, then apply their own retry budget, backoff, and reporting\n * before restarting the assistant turn.\n */\nexport function isRetryableAssistantError(message: AssistantMessage): boolean {\n\tif (message.stopReason !== \"error\" || !message.errorMessage) return false;\n\tconst errorMessage = message.errorMessage;\n\tif (NON_RETRYABLE_PROVIDER_LIMIT_ERROR_PATTERN.test(errorMessage)) return false;\n\treturn RETRYABLE_PROVIDER_ERROR_PATTERN.test(errorMessage);\n}\n"]}

package/dist/utils/transcript.d.ts

@@ -68,20 +68,34 @@ export declare function getDeclaredTools(messages: TranscriptMessages): Tool[];
 export declare function hasToolRedefinitions(messages: TranscriptMessages): boolean;
 /** Whether tool history contains a removal or same-name redeclaration that an addition-only transport cannot replay. */
 export declare function hasNonAdditiveToolChanges(messages: TranscriptMessages): boolean;
+/**
+ * How a transport can express tool changes after the first request.
+ *
+ * - `"none"`: only the top-level tool list exists; it always holds the current tools.
+ * - `"additions"`: tools can be added at a later message, but not removed or redefined.
+ * - `"changes"`: every tool, including the initial ones, is declared in place, and a later
+ *   declaration under the same name replaces the earlier one. Removals are not expressible on
+ *   the wire, so the transport announces them as text and the removed tools stay declared.
+ */
+export type TranscriptToolSupport = "none" | "additions" | "changes";
 export interface TranscriptTools {
     /** Tools sent in the top-level request field. */
     requestTools: Tool[];
     /**
-     * Whether later system messages carry their own `toolsAdded` as in-place additions.
-     * When false, `requestTools` already holds the complete current tool set.
+     * Number of leading system messages whose tool changes `requestTools` already holds. Every
+     * later system message carries its own tool changes in place.
      */
-    anchorsAdditions: boolean;
+    foldedSystemMessages: number;
 }
 /**
- * Split tool declarations between the top-level request field and in-place additions.
- * Transports that can anchor additions at a system message keep the initial tools at the
- * top and load later ones where they appear; that only works when no tool was removed or
- * redeclared, so everything else sends the current tool list.
+ * Split tool declarations between the top-level request field and in-place changes.
+ *
+ * With `"changes"`, the top-level list stays empty and every system message, including the
+ * leading one, declares its tools in place: an in-place declaration can replace an earlier
+ * in-place one, but not a top-level one. With `"additions"`, a removal or redefinition cannot be
+ * expressed in place, so the top-level list becomes the tool state after the last such change;
+ * later additions stay anchored, and the cached prefix is invalidated once per removal instead
+ * of on every later addition. With `"none"`, the top-level list is the current tool set.
  */
-export declare function resolveTranscriptTools(messages: TranscriptMessages, supportsToolAdditions: boolean): TranscriptTools;
+export declare function resolveTranscriptTools(messages: TranscriptMessages, support: TranscriptToolSupport): TranscriptTools;
 //# sourceMappingURL=transcript.d.ts.map

package/dist/utils/transcript.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"transcript.d.ts","sourceRoot":"","sources":["../../src/utils/transcript.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,OAAO,EAAE,OAAO,EAAE,aAAa,EAAE,IAAI,EAAE,aAAa,EAAE,iBAAiB,EAAE,MAAM,aAAa,CAAC;AAG3G,YAAY,EAAE,iBAAiB,EAAE,MAAM,aAAa,CAAC;AAErD;;;GAGG;AACH,wBAAgB,0BAA0B,CACzC,YAAY,EAAE,MAAM,GAAG,SAAS,EAChC,KAAK,EAAE,IAAI,EAAE,GAAG,SAAS,GACvB,aAAa,GAAG,SAAS,CAU3B;AAED;;;;GAIG;AACH,wBAAgB,gBAAgB,CAAC,OAAO,EAAE,OAAO,GAAG,iBAAiB,CAIpE;AAED;;;GAGG;AACH,MAAM,MAAM,kBAAkB,GAAG,SAAS;IAAE,IAAI,EAAE,MAAM,CAAA;CAAE,EAAE,CAAC;AAM7D,4EAA4E;AAC5E,wBAAgB,uBAAuB,CAAC,QAAQ,EAAE,kBAAkB,GAAG,aAAa,GAAG,SAAS,CAG/F;AAED,+FAA+F;AAC/F,wBAAgB,2BAA2B,CAAC,QAAQ,EAAE,OAAO,EAAE,GAAG,OAAO,EAAE,CAE1E;AAED,kFAAkF;AAClF,wBAAgB,eAAe,CAAC,QAAQ,EAAE,kBAAkB,GAAG,IAAI,EAAE,CAQpE;AAED;;;;GAIG;AACH,wBAAgB,uBAAuB,CAAC,QAAQ,EAAE,kBAAkB,GAAG,aAAa,GAAG,SAAS,CAuB/F;AAED,kFAAkF;AAClF,wBAAgB,sBAAsB,CAAC,QAAQ,EAAE,kBAAkB,GAAG,MAAM,CAG3E;AAED;;;GAGG;AACH,wBAAgB,sBAAsB,CAAC,OAAO,EAAE,iBAAiB,GAAG,iBAAiB,CAIpF;AAED,gGAAgG;AAChG,wBAAgB,iBAAiB,CAChC,OAAO,EAAE,iBAAiB,EAC1B,8BAA8B,EAAE,OAAO,GAAG,SAAS,GACjD,iBAAiB,CAEnB;AAED,wGAAwG;AACxG,wBAAgB,iBAAiB,CAAC,IAAI,EAAE,IAAI,GAAG,IAAI,CAOlD;AAED;;;;;;;GAOG;AACH,wBAAgB,iBAAiB,CAAC,IAAI,EAAE,IAAI,EAAE,KAAK,EAAE,IAAI,GAAG,OAAO,CAElE;AAED,MAAM,WAAW,gBAAgB;IAChC,UAAU,EAAE,IAAI,EAAE,CAAC;IACnB,YAAY,EAAE,aAAa,EAAE,CAAC;CAC9B;AAED,mGAAmG;AACnG,wBAAgB,mBAAmB,CAAC,QAAQ,EAAE,SAAS,IAAI,EAAE,EAAE,OAAO,EAAE,SAAS,IAAI,EAAE,GAAG,gBAAgB,CAiBzG;AAED,wFAAwF;AACxF,wBAAgB,gBAAgB,CAAC,QAAQ,EAAE,kBAAkB,GAAG,IAAI,EAAE,CAOrE;AAED;;;;;;GAMG;AACH,wBAAgB,oBAAoB,CAAC,QAAQ,EAAE,kBAAkB,GAAG,OAAO,CAW1E;AAED,wHAAwH;AACxH,wBAAgB,yBAAyB,CAAC,QAAQ,EAAE,kBAAkB,GAAG,OAAO,CAW/E;AAED,MAAM,WAAW,eAAe;IAC/B,iDAAiD;IACjD,YAAY,EAAE,IAAI,EAAE,CAAC;IACrB;;;OAGG;IACH,gBAAgB,EAAE,OAAO,CAAC;CAC1B;AAED;;;;;GAKG;AACH,wBAAgB,sBAAsB,CAAC,QAAQ,EAAE,kBAAkB,EAAE,qBAAqB,EAAE,OAAO,GAAG,eAAe,CAQpH"}
+{"version":3,"file":"transcript.d.ts","sourceRoot":"","sources":["../../src/utils/transcript.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,OAAO,EAAE,OAAO,EAAE,aAAa,EAAE,IAAI,EAAE,aAAa,EAAE,iBAAiB,EAAE,MAAM,aAAa,CAAC;AAG3G,YAAY,EAAE,iBAAiB,EAAE,MAAM,aAAa,CAAC;AAErD;;;GAGG;AACH,wBAAgB,0BAA0B,CACzC,YAAY,EAAE,MAAM,GAAG,SAAS,EAChC,KAAK,EAAE,IAAI,EAAE,GAAG,SAAS,GACvB,aAAa,GAAG,SAAS,CAU3B;AAED;;;;GAIG;AACH,wBAAgB,gBAAgB,CAAC,OAAO,EAAE,OAAO,GAAG,iBAAiB,CAIpE;AAED;;;GAGG;AACH,MAAM,MAAM,kBAAkB,GAAG,SAAS;IAAE,IAAI,EAAE,MAAM,CAAA;CAAE,EAAE,CAAC;AAM7D,4EAA4E;AAC5E,wBAAgB,uBAAuB,CAAC,QAAQ,EAAE,kBAAkB,GAAG,aAAa,GAAG,SAAS,CAG/F;AAED,+FAA+F;AAC/F,wBAAgB,2BAA2B,CAAC,QAAQ,EAAE,OAAO,EAAE,GAAG,OAAO,EAAE,CAE1E;AAED,kFAAkF;AAClF,wBAAgB,eAAe,CAAC,QAAQ,EAAE,kBAAkB,GAAG,IAAI,EAAE,CAQpE;AAED;;;;GAIG;AACH,wBAAgB,uBAAuB,CAAC,QAAQ,EAAE,kBAAkB,GAAG,aAAa,GAAG,SAAS,CAuB/F;AAED,kFAAkF;AAClF,wBAAgB,sBAAsB,CAAC,QAAQ,EAAE,kBAAkB,GAAG,MAAM,CAG3E;AAED;;;GAGG;AACH,wBAAgB,sBAAsB,CAAC,OAAO,EAAE,iBAAiB,GAAG,iBAAiB,CAIpF;AAED,gGAAgG;AAChG,wBAAgB,iBAAiB,CAChC,OAAO,EAAE,iBAAiB,EAC1B,8BAA8B,EAAE,OAAO,GAAG,SAAS,GACjD,iBAAiB,CAEnB;AAED,wGAAwG;AACxG,wBAAgB,iBAAiB,CAAC,IAAI,EAAE,IAAI,GAAG,IAAI,CAOlD;AAED;;;;;;;GAOG;AACH,wBAAgB,iBAAiB,CAAC,IAAI,EAAE,IAAI,EAAE,KAAK,EAAE,IAAI,GAAG,OAAO,CAElE;AAED,MAAM,WAAW,gBAAgB;IAChC,UAAU,EAAE,IAAI,EAAE,CAAC;IACnB,YAAY,EAAE,aAAa,EAAE,CAAC;CAC9B;AAED,mGAAmG;AACnG,wBAAgB,mBAAmB,CAAC,QAAQ,EAAE,SAAS,IAAI,EAAE,EAAE,OAAO,EAAE,SAAS,IAAI,EAAE,GAAG,gBAAgB,CAiBzG;AAED,wFAAwF;AACxF,wBAAgB,gBAAgB,CAAC,QAAQ,EAAE,kBAAkB,GAAG,IAAI,EAAE,CAOrE;AAED;;;;;;GAMG;AACH,wBAAgB,oBAAoB,CAAC,QAAQ,EAAE,kBAAkB,GAAG,OAAO,CAW1E;AAED,wHAAwH;AACxH,wBAAgB,yBAAyB,CAAC,QAAQ,EAAE,kBAAkB,GAAG,OAAO,CAE/E;AAmBD;;;;;;;;GAQG;AACH,MAAM,MAAM,qBAAqB,GAAG,MAAM,GAAG,WAAW,GAAG,SAAS,CAAC;AAErE,MAAM,WAAW,eAAe;IAC/B,iDAAiD;IACjD,YAAY,EAAE,IAAI,EAAE,CAAC;IACrB;;;OAGG;IACH,oBAAoB,EAAE,MAAM,CAAC;CAC7B;AAED;;;;;;;;;GASG;AACH,wBAAgB,sBAAsB,CAAC,QAAQ,EAAE,kBAAkB,EAAE,OAAO,EAAE,qBAAqB,GAAG,eAAe,CAWpH"}

package/dist/utils/transcript.js

@@ -175,33 +175,44 @@ export function hasToolRedefinitions(messages) {
 }
 /** Whether tool history contains a removal or same-name redeclaration that an addition-only transport cannot replay. */
 export function hasNonAdditiveToolChanges(messages) {
+    return findLastNonAdditiveToolChange(messages.filter(isSystemMessage)) >= 0;
+}
+/** Index of the last system message with a removal or same-name redeclaration, or -1. */
+function findLastNonAdditiveToolChange(systemMessages) {
+    // A name declared before counts as a redeclaration even after its removal: Kimi K3 ignores a
+    // removed tool that a later tool message declares again (verified October 2026).
     const declared = new Set();
-    for (const message of messages) {
-        if (!isSystemMessage(message))
-            continue;
+    let last = -1;
+    for (let index = 0; index < systemMessages.length; index++) {
+        const message = systemMessages[index];
         if ((message.toolsRemoved?.length ?? 0) > 0)
-            return true;
+            last = index;
         for (const tool of message.toolsAdded ?? []) {
             if (declared.has(tool.name))
-                return true;
+                last = index;
             declared.add(tool.name);
         }
     }
-    return false;
+    return last;
 }
 /**
- * Split tool declarations between the top-level request field and in-place additions.
- * Transports that can anchor additions at a system message keep the initial tools at the
- * top and load later ones where they appear; that only works when no tool was removed or
- * redeclared, so everything else sends the current tool list.
+ * Split tool declarations between the top-level request field and in-place changes.
+ *
+ * With `"changes"`, the top-level list stays empty and every system message, including the
+ * leading one, declares its tools in place: an in-place declaration can replace an earlier
+ * in-place one, but not a top-level one. With `"additions"`, a removal or redefinition cannot be
+ * expressed in place, so the top-level list becomes the tool state after the last such change;
+ * later additions stay anchored, and the cached prefix is invalidated once per removal instead
+ * of on every later addition. With `"none"`, the top-level list is the current tool set.
  */
-export function resolveTranscriptTools(messages, supportsToolAdditions) {
-    const anchorsAdditions = supportsToolAdditions && !hasNonAdditiveToolChanges(messages);
-    return {
-        requestTools: anchorsAdditions
-            ? (getInitialSystemMessage(messages)?.toolsAdded ?? [])
-            : getCurrentTools(messages),
-        anchorsAdditions,
-    };
+export function resolveTranscriptTools(messages, support) {
+    const systemMessages = messages.filter(isSystemMessage);
+    if (support === "none") {
+        return { requestTools: getCurrentTools(systemMessages), foldedSystemMessages: systemMessages.length };
+    }
+    if (support === "changes")
+        return { requestTools: [], foldedSystemMessages: 0 };
+    const folded = Math.max(getInitialSystemMessage(messages) ? 1 : 0, findLastNonAdditiveToolChange(systemMessages) + 1);
+    return { requestTools: getCurrentTools(systemMessages.slice(0, folded)), foldedSystemMessages: folded };
 }
 //# sourceMappingURL=transcript.js.map