omk-ai 1.2.4 → 1.3.0

package/dist/stream.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"stream.d.ts","sourceRoot":"","sources":["../src/stream.ts"],"names":[],"mappings":"AAAA,OAAO,kCAAkC,CAAC;AAI1C,OAAO,KAAK,EACX,GAAG,EACH,gBAAgB,EAChB,2BAA2B,EAC3B,OAAO,EACP,KAAK,EACL,qBAAqB,EACrB,mBAAmB,EAEnB,MAAM,YAAY,CAAC;AAEpB,OAAO,EAAE,YAAY,EAAE,MAAM,mBAAmB,CAAC;AAwBjD,wBAAgB,MAAM,CAAC,IAAI,SAAS,GAAG,EACtC,KAAK,EAAE,KAAK,CAAC,IAAI,CAAC,EAClB,OAAO,EAAE,OAAO,EAChB,OAAO,CAAC,EAAE,qBAAqB,GAC7B,2BAA2B,CAG7B;AAED,wBAAsB,QAAQ,CAAC,IAAI,SAAS,GAAG,EAC9C,KAAK,EAAE,KAAK,CAAC,IAAI,CAAC,EAClB,OAAO,EAAE,OAAO,EAChB,OAAO,CAAC,EAAE,qBAAqB,GAC7B,OAAO,CAAC,gBAAgB,CAAC,CAG3B;AAoBD,wBAAgB,YAAY,CAAC,IAAI,SAAS,GAAG,EAC5C,KAAK,EAAE,KAAK,CAAC,IAAI,CAAC,EAClB,OAAO,EAAE,OAAO,EAChB,OAAO,CAAC,EAAE,mBAAmB,GAC3B,2BAA2B,CAG7B;AAID,mGAAmG;AACnG,wBAAgB,iBAAiB,CAAC,EAAE,EAAE,OAAO,GAAG,OAAO,CAGtD;AAED,wBAAsB,cAAc,CAAC,IAAI,SAAS,GAAG,EACpD,KAAK,EAAE,KAAK,CAAC,IAAI,CAAC,EAClB,OAAO,EAAE,OAAO,EAChB,OAAO,CAAC,EAAE,mBAAmB,GAC3B,OAAO,CAAC,gBAAgB,CAAC,CAG3B","sourcesContent":["import \"./providers/register-builtins.ts\";\n\nimport { getApiProvider } from \"./api-registry.ts\";\nimport { getEnvApiKey } from \"./env-api-keys.ts\";\nimport type {\n\tApi,\n\tAssistantMessage,\n\tAssistantMessageEventStream,\n\tContext,\n\tModel,\n\tProviderStreamOptions,\n\tSimpleStreamOptions,\n\tStreamOptions,\n} from \"./types.ts\";\n\nexport { getEnvApiKey } from \"./env-api-keys.ts\";\n\nfunction hasExplicitApiKey(apiKey: string | undefined): apiKey is string {\n\treturn typeof apiKey === \"string\" && apiKey.trim().length > 0;\n}\n\nfunction withEnvApiKey<TOptions extends StreamOptions>(\n\tmodel: Model<Api>,\n\toptions: TOptions | undefined,\n): TOptions | undefined {\n\tif (hasExplicitApiKey(options?.apiKey)) return options;\n\tconst apiKey = getEnvApiKey(model.provider);\n\tif (!apiKey) return options;\n\treturn { ...options, apiKey } as TOptions;\n}\n\nfunction resolveApiProvider(api: Api) {\n\tconst provider = getApiProvider(api);\n\tif (!provider) {\n\t\tthrow new Error(`No API provider registered for api: ${api}`);\n\t}\n\treturn provider;\n}\n\nexport function stream<TApi extends Api>(\n\tmodel: Model<TApi>,\n\tcontext: Context,\n\toptions?: ProviderStreamOptions,\n): AssistantMessageEventStream {\n\tconst provider = resolveApiProvider(model.api);\n\treturn provider.stream(model, context, withEnvApiKey(model, options) as StreamOptions);\n}\n\nexport async function complete<TApi extends Api>(\n\tmodel: Model<TApi>,\n\tcontext: Context,\n\toptions?: ProviderStreamOptions,\n): Promise<AssistantMessage> {\n\tconst s = stream(model, context, options);\n\treturn s.result();\n}\n\n/**\n * Brand identifying the built-in stream function.\n *\n * Callers decide whether provider credentials are mandatory by asking \"is this\n * still the built-in stream function?\", and a bare `fn === streamSimple`\n * answers that with reference identity. Reference identity is not dependable\n * here: this package can legitimately load more than once in one process (a\n * workspace symlink beside an installed copy, or two dependents resolving\n * different versions), producing two distinct `streamSimple` functions that\n * behave identically. The comparison then reports \"custom stream function\" for\n * what is really the built-in one, and a credential check silently relaxes.\n *\n * `Symbol.for` keys into the per-realm global registry, so every copy of this\n * module brands its own `streamSimple` with the same symbol and the check\n * survives duplication.\n */\nconst BUILTIN_STREAM_FN = Symbol.for(\"omk-ai.streamSimple\");\n\nexport function streamSimple<TApi extends Api>(\n\tmodel: Model<TApi>,\n\tcontext: Context,\n\toptions?: SimpleStreamOptions,\n): AssistantMessageEventStream {\n\tconst provider = resolveApiProvider(model.api);\n\treturn provider.streamSimple(model, context, withEnvApiKey(model, options));\n}\n\nObject.defineProperty(streamSimple, BUILTIN_STREAM_FN, { value: true });\n\n/** True when `fn` is this package's built-in stream function, including a duplicate copy of it. */\nexport function isBuiltinStreamFn(fn: unknown): boolean {\n\tif (typeof fn !== \"function\") return false;\n\treturn Reflect.get(fn, BUILTIN_STREAM_FN) === true;\n}\n\nexport async function completeSimple<TApi extends Api>(\n\tmodel: Model<TApi>,\n\tcontext: Context,\n\toptions?: SimpleStreamOptions,\n): Promise<AssistantMessage> {\n\tconst s = streamSimple(model, context, options);\n\treturn s.result();\n}\n"]}
+{"version":3,"file":"stream.d.ts","sourceRoot":"","sources":["../src/stream.ts"],"names":[],"mappings":"AAAA,OAAO,kCAAkC,CAAC;AAI1C,OAAO,KAAK,EACX,GAAG,EACH,gBAAgB,EAChB,2BAA2B,EAC3B,OAAO,EACP,KAAK,EACL,qBAAqB,EACrB,mBAAmB,EAEnB,MAAM,YAAY,CAAC;AAEpB,OAAO,EAAE,YAAY,EAAE,MAAM,mBAAmB,CAAC;AAwBjD,wBAAgB,MAAM,CAAC,IAAI,SAAS,GAAG,EACtC,KAAK,EAAE,KAAK,CAAC,IAAI,CAAC,EAClB,OAAO,EAAE,OAAO,EAChB,OAAO,CAAC,EAAE,qBAAqB,GAC7B,2BAA2B,CAG7B;AAED,wBAAsB,QAAQ,CAAC,IAAI,SAAS,GAAG,EAC9C,KAAK,EAAE,KAAK,CAAC,IAAI,CAAC,EAClB,OAAO,EAAE,OAAO,EAChB,OAAO,CAAC,EAAE,qBAAqB,GAC7B,OAAO,CAAC,gBAAgB,CAAC,CAM3B;AAoBD,wBAAgB,YAAY,CAAC,IAAI,SAAS,GAAG,EAC5C,KAAK,EAAE,KAAK,CAAC,IAAI,CAAC,EAClB,OAAO,EAAE,OAAO,EAChB,OAAO,CAAC,EAAE,mBAAmB,GAC3B,2BAA2B,CAG7B;AAID,mGAAmG;AACnG,wBAAgB,iBAAiB,CAAC,EAAE,EAAE,OAAO,GAAG,OAAO,CAGtD;AAED,wBAAsB,cAAc,CAAC,IAAI,SAAS,GAAG,EACpD,KAAK,EAAE,KAAK,CAAC,IAAI,CAAC,EAClB,OAAO,EAAE,OAAO,EAChB,OAAO,CAAC,EAAE,mBAAmB,GAC3B,OAAO,CAAC,gBAAgB,CAAC,CAM3B","sourcesContent":["import \"./providers/register-builtins.ts\";\n\nimport { getApiProvider } from \"./api-registry.ts\";\nimport { getEnvApiKey } from \"./env-api-keys.ts\";\nimport type {\n\tApi,\n\tAssistantMessage,\n\tAssistantMessageEventStream,\n\tContext,\n\tModel,\n\tProviderStreamOptions,\n\tSimpleStreamOptions,\n\tStreamOptions,\n} from \"./types.ts\";\n\nexport { getEnvApiKey } from \"./env-api-keys.ts\";\n\nfunction hasExplicitApiKey(apiKey: string | undefined): apiKey is string {\n\treturn typeof apiKey === \"string\" && apiKey.trim().length > 0;\n}\n\nfunction withEnvApiKey<TOptions extends StreamOptions>(\n\tmodel: Model<Api>,\n\toptions: TOptions | undefined,\n): TOptions | undefined {\n\tif (hasExplicitApiKey(options?.apiKey)) return options;\n\tconst apiKey = getEnvApiKey(model.provider);\n\tif (!apiKey) return options;\n\treturn { ...options, apiKey } as TOptions;\n}\n\nfunction resolveApiProvider(api: Api) {\n\tconst provider = getApiProvider(api);\n\tif (!provider) {\n\t\tthrow new Error(`No API provider registered for api: ${api}`);\n\t}\n\treturn provider;\n}\n\nexport function stream<TApi extends Api>(\n\tmodel: Model<TApi>,\n\tcontext: Context,\n\toptions?: ProviderStreamOptions,\n): AssistantMessageEventStream {\n\tconst provider = resolveApiProvider(model.api);\n\treturn provider.stream(model, context, withEnvApiKey(model, options) as StreamOptions);\n}\n\nexport async function complete<TApi extends Api>(\n\tmodel: Model<TApi>,\n\tcontext: Context,\n\toptions?: ProviderStreamOptions,\n): Promise<AssistantMessage> {\n\tconst s = stream(model, context, options);\n\t// This API exposes only the final message; consume deltas as they arrive.\n\tfor await (const _event of s) {\n\t}\n\treturn s.result();\n}\n\n/**\n * Brand identifying the built-in stream function.\n *\n * Callers decide whether provider credentials are mandatory by asking \"is this\n * still the built-in stream function?\", and a bare `fn === streamSimple`\n * answers that with reference identity. Reference identity is not dependable\n * here: this package can legitimately load more than once in one process (a\n * workspace symlink beside an installed copy, or two dependents resolving\n * different versions), producing two distinct `streamSimple` functions that\n * behave identically. The comparison then reports \"custom stream function\" for\n * what is really the built-in one, and a credential check silently relaxes.\n *\n * `Symbol.for` keys into the per-realm global registry, so every copy of this\n * module brands its own `streamSimple` with the same symbol and the check\n * survives duplication.\n */\nconst BUILTIN_STREAM_FN = Symbol.for(\"omk-ai.streamSimple\");\n\nexport function streamSimple<TApi extends Api>(\n\tmodel: Model<TApi>,\n\tcontext: Context,\n\toptions?: SimpleStreamOptions,\n): AssistantMessageEventStream {\n\tconst provider = resolveApiProvider(model.api);\n\treturn provider.streamSimple(model, context, withEnvApiKey(model, options));\n}\n\nObject.defineProperty(streamSimple, BUILTIN_STREAM_FN, { value: true });\n\n/** True when `fn` is this package's built-in stream function, including a duplicate copy of it. */\nexport function isBuiltinStreamFn(fn: unknown): boolean {\n\tif (typeof fn !== \"function\") return false;\n\treturn Reflect.get(fn, BUILTIN_STREAM_FN) === true;\n}\n\nexport async function completeSimple<TApi extends Api>(\n\tmodel: Model<TApi>,\n\tcontext: Context,\n\toptions?: SimpleStreamOptions,\n): Promise<AssistantMessage> {\n\tconst s = streamSimple(model, context, options);\n\t// Keep completion-only callers from retaining the full event backlog.\n\tfor await (const _event of s) {\n\t}\n\treturn s.result();\n}\n"]}

package/dist/stream.js

@@ -26,6 +26,9 @@ export function stream(model, context, options) {
 }
 export async function complete(model, context, options) {
     const s = stream(model, context, options);
+    // This API exposes only the final message; consume deltas as they arrive.
+    for await (const _event of s) {
+    }
     return s.result();
 }
 /**
@@ -58,6 +61,9 @@ export function isBuiltinStreamFn(fn) {
 }
 export async function completeSimple(model, context, options) {
     const s = streamSimple(model, context, options);
+    // Keep completion-only callers from retaining the full event backlog.
+    for await (const _event of s) {
+    }
     return s.result();
 }
 //# sourceMappingURL=stream.js.map

package/dist/stream.js.map

@@ -1 +1 @@
-{"version":3,"file":"stream.js","sourceRoot":"","sources":["../src/stream.ts"],"names":[],"mappings":"AAAA,OAAO,kCAAkC,CAAC;AAE1C,OAAO,EAAE,cAAc,EAAE,MAAM,mBAAmB,CAAC;AACnD,OAAO,EAAE,YAAY,EAAE,MAAM,mBAAmB,CAAC;AAYjD,OAAO,EAAE,YAAY,EAAE,MAAM,mBAAmB,CAAC;AAEjD,SAAS,iBAAiB,CAAC,MAA0B,EAAoB;IACxE,OAAO,OAAO,MAAM,KAAK,QAAQ,IAAI,MAAM,CAAC,IAAI,EAAE,CAAC,MAAM,GAAG,CAAC,CAAC;AAAA,CAC9D;AAED,SAAS,aAAa,CACrB,KAAiB,EACjB,OAA6B,EACN;IACvB,IAAI,iBAAiB,CAAC,OAAO,EAAE,MAAM,CAAC;QAAE,OAAO,OAAO,CAAC;IACvD,MAAM,MAAM,GAAG,YAAY,CAAC,KAAK,CAAC,QAAQ,CAAC,CAAC;IAC5C,IAAI,CAAC,MAAM;QAAE,OAAO,OAAO,CAAC;IAC5B,OAAO,EAAE,GAAG,OAAO,EAAE,MAAM,EAAc,CAAC;AAAA,CAC1C;AAED,SAAS,kBAAkB,CAAC,GAAQ,EAAE;IACrC,MAAM,QAAQ,GAAG,cAAc,CAAC,GAAG,CAAC,CAAC;IACrC,IAAI,CAAC,QAAQ,EAAE,CAAC;QACf,MAAM,IAAI,KAAK,CAAC,uCAAuC,GAAG,EAAE,CAAC,CAAC;IAC/D,CAAC;IACD,OAAO,QAAQ,CAAC;AAAA,CAChB;AAED,MAAM,UAAU,MAAM,CACrB,KAAkB,EAClB,OAAgB,EAChB,OAA+B,EACD;IAC9B,MAAM,QAAQ,GAAG,kBAAkB,CAAC,KAAK,CAAC,GAAG,CAAC,CAAC;IAC/C,OAAO,QAAQ,CAAC,MAAM,CAAC,KAAK,EAAE,OAAO,EAAE,aAAa,CAAC,KAAK,EAAE,OAAO,CAAkB,CAAC,CAAC;AAAA,CACvF;AAED,MAAM,CAAC,KAAK,UAAU,QAAQ,CAC7B,KAAkB,EAClB,OAAgB,EAChB,OAA+B,EACH;IAC5B,MAAM,CAAC,GAAG,MAAM,CAAC,KAAK,EAAE,OAAO,EAAE,OAAO,CAAC,CAAC;IAC1C,OAAO,CAAC,CAAC,MAAM,EAAE,CAAC;AAAA,CAClB;AAED;;;;;;;;;;;;;;;GAeG;AACH,MAAM,iBAAiB,GAAG,MAAM,CAAC,GAAG,CAAC,qBAAqB,CAAC,CAAC;AAE5D,MAAM,UAAU,YAAY,CAC3B,KAAkB,EAClB,OAAgB,EAChB,OAA6B,EACC;IAC9B,MAAM,QAAQ,GAAG,kBAAkB,CAAC,KAAK,CAAC,GAAG,CAAC,CAAC;IAC/C,OAAO,QAAQ,CAAC,YAAY,CAAC,KAAK,EAAE,OAAO,EAAE,aAAa,CAAC,KAAK,EAAE,OAAO,CAAC,CAAC,CAAC;AAAA,CAC5E;AAED,MAAM,CAAC,cAAc,CAAC,YAAY,EAAE,iBAAiB,EAAE,EAAE,KAAK,EAAE,IAAI,EAAE,CAAC,CAAC;AAExE,mGAAmG;AACnG,MAAM,UAAU,iBAAiB,CAAC,EAAW,EAAW;IACvD,IAAI,OAAO,EAAE,KAAK,UAAU;QAAE,OAAO,KAAK,CAAC;IAC3C,OAAO,OAAO,CAAC,GAAG,CAAC,EAAE,EAAE,iBAAiB,CAAC,KAAK,IAAI,CAAC;AAAA,CACnD;AAED,MAAM,CAAC,KAAK,UAAU,cAAc,CACnC,KAAkB,EAClB,OAAgB,EAChB,OAA6B,EACD;IAC5B,MAAM,CAAC,GAAG,YAAY,CAAC,KAAK,EAAE,OAAO,EAAE,OAAO,CAAC,CAAC;IAChD,OAAO,CAAC,CAAC,MAAM,EAAE,CAAC;AAAA,CAClB","sourcesContent":["import \"./providers/register-builtins.ts\";\n\nimport { getApiProvider } from \"./api-registry.ts\";\nimport { getEnvApiKey } from \"./env-api-keys.ts\";\nimport type {\n\tApi,\n\tAssistantMessage,\n\tAssistantMessageEventStream,\n\tContext,\n\tModel,\n\tProviderStreamOptions,\n\tSimpleStreamOptions,\n\tStreamOptions,\n} from \"./types.ts\";\n\nexport { getEnvApiKey } from \"./env-api-keys.ts\";\n\nfunction hasExplicitApiKey(apiKey: string | undefined): apiKey is string {\n\treturn typeof apiKey === \"string\" && apiKey.trim().length > 0;\n}\n\nfunction withEnvApiKey<TOptions extends StreamOptions>(\n\tmodel: Model<Api>,\n\toptions: TOptions | undefined,\n): TOptions | undefined {\n\tif (hasExplicitApiKey(options?.apiKey)) return options;\n\tconst apiKey = getEnvApiKey(model.provider);\n\tif (!apiKey) return options;\n\treturn { ...options, apiKey } as TOptions;\n}\n\nfunction resolveApiProvider(api: Api) {\n\tconst provider = getApiProvider(api);\n\tif (!provider) {\n\t\tthrow new Error(`No API provider registered for api: ${api}`);\n\t}\n\treturn provider;\n}\n\nexport function stream<TApi extends Api>(\n\tmodel: Model<TApi>,\n\tcontext: Context,\n\toptions?: ProviderStreamOptions,\n): AssistantMessageEventStream {\n\tconst provider = resolveApiProvider(model.api);\n\treturn provider.stream(model, context, withEnvApiKey(model, options) as StreamOptions);\n}\n\nexport async function complete<TApi extends Api>(\n\tmodel: Model<TApi>,\n\tcontext: Context,\n\toptions?: ProviderStreamOptions,\n): Promise<AssistantMessage> {\n\tconst s = stream(model, context, options);\n\treturn s.result();\n}\n\n/**\n * Brand identifying the built-in stream function.\n *\n * Callers decide whether provider credentials are mandatory by asking \"is this\n * still the built-in stream function?\", and a bare `fn === streamSimple`\n * answers that with reference identity. Reference identity is not dependable\n * here: this package can legitimately load more than once in one process (a\n * workspace symlink beside an installed copy, or two dependents resolving\n * different versions), producing two distinct `streamSimple` functions that\n * behave identically. The comparison then reports \"custom stream function\" for\n * what is really the built-in one, and a credential check silently relaxes.\n *\n * `Symbol.for` keys into the per-realm global registry, so every copy of this\n * module brands its own `streamSimple` with the same symbol and the check\n * survives duplication.\n */\nconst BUILTIN_STREAM_FN = Symbol.for(\"omk-ai.streamSimple\");\n\nexport function streamSimple<TApi extends Api>(\n\tmodel: Model<TApi>,\n\tcontext: Context,\n\toptions?: SimpleStreamOptions,\n): AssistantMessageEventStream {\n\tconst provider = resolveApiProvider(model.api);\n\treturn provider.streamSimple(model, context, withEnvApiKey(model, options));\n}\n\nObject.defineProperty(streamSimple, BUILTIN_STREAM_FN, { value: true });\n\n/** True when `fn` is this package's built-in stream function, including a duplicate copy of it. */\nexport function isBuiltinStreamFn(fn: unknown): boolean {\n\tif (typeof fn !== \"function\") return false;\n\treturn Reflect.get(fn, BUILTIN_STREAM_FN) === true;\n}\n\nexport async function completeSimple<TApi extends Api>(\n\tmodel: Model<TApi>,\n\tcontext: Context,\n\toptions?: SimpleStreamOptions,\n): Promise<AssistantMessage> {\n\tconst s = streamSimple(model, context, options);\n\treturn s.result();\n}\n"]}
+{"version":3,"file":"stream.js","sourceRoot":"","sources":["../src/stream.ts"],"names":[],"mappings":"AAAA,OAAO,kCAAkC,CAAC;AAE1C,OAAO,EAAE,cAAc,EAAE,MAAM,mBAAmB,CAAC;AACnD,OAAO,EAAE,YAAY,EAAE,MAAM,mBAAmB,CAAC;AAYjD,OAAO,EAAE,YAAY,EAAE,MAAM,mBAAmB,CAAC;AAEjD,SAAS,iBAAiB,CAAC,MAA0B,EAAoB;IACxE,OAAO,OAAO,MAAM,KAAK,QAAQ,IAAI,MAAM,CAAC,IAAI,EAAE,CAAC,MAAM,GAAG,CAAC,CAAC;AAAA,CAC9D;AAED,SAAS,aAAa,CACrB,KAAiB,EACjB,OAA6B,EACN;IACvB,IAAI,iBAAiB,CAAC,OAAO,EAAE,MAAM,CAAC;QAAE,OAAO,OAAO,CAAC;IACvD,MAAM,MAAM,GAAG,YAAY,CAAC,KAAK,CAAC,QAAQ,CAAC,CAAC;IAC5C,IAAI,CAAC,MAAM;QAAE,OAAO,OAAO,CAAC;IAC5B,OAAO,EAAE,GAAG,OAAO,EAAE,MAAM,EAAc,CAAC;AAAA,CAC1C;AAED,SAAS,kBAAkB,CAAC,GAAQ,EAAE;IACrC,MAAM,QAAQ,GAAG,cAAc,CAAC,GAAG,CAAC,CAAC;IACrC,IAAI,CAAC,QAAQ,EAAE,CAAC;QACf,MAAM,IAAI,KAAK,CAAC,uCAAuC,GAAG,EAAE,CAAC,CAAC;IAC/D,CAAC;IACD,OAAO,QAAQ,CAAC;AAAA,CAChB;AAED,MAAM,UAAU,MAAM,CACrB,KAAkB,EAClB,OAAgB,EAChB,OAA+B,EACD;IAC9B,MAAM,QAAQ,GAAG,kBAAkB,CAAC,KAAK,CAAC,GAAG,CAAC,CAAC;IAC/C,OAAO,QAAQ,CAAC,MAAM,CAAC,KAAK,EAAE,OAAO,EAAE,aAAa,CAAC,KAAK,EAAE,OAAO,CAAkB,CAAC,CAAC;AAAA,CACvF;AAED,MAAM,CAAC,KAAK,UAAU,QAAQ,CAC7B,KAAkB,EAClB,OAAgB,EAChB,OAA+B,EACH;IAC5B,MAAM,CAAC,GAAG,MAAM,CAAC,KAAK,EAAE,OAAO,EAAE,OAAO,CAAC,CAAC;IAC1C,0EAA0E;IAC1E,IAAI,KAAK,EAAE,MAAM,MAAM,IAAI,CAAC,EAAE,CAAC;IAC/B,CAAC;IACD,OAAO,CAAC,CAAC,MAAM,EAAE,CAAC;AAAA,CAClB;AAED;;;;;;;;;;;;;;;GAeG;AACH,MAAM,iBAAiB,GAAG,MAAM,CAAC,GAAG,CAAC,qBAAqB,CAAC,CAAC;AAE5D,MAAM,UAAU,YAAY,CAC3B,KAAkB,EAClB,OAAgB,EAChB,OAA6B,EACC;IAC9B,MAAM,QAAQ,GAAG,kBAAkB,CAAC,KAAK,CAAC,GAAG,CAAC,CAAC;IAC/C,OAAO,QAAQ,CAAC,YAAY,CAAC,KAAK,EAAE,OAAO,EAAE,aAAa,CAAC,KAAK,EAAE,OAAO,CAAC,CAAC,CAAC;AAAA,CAC5E;AAED,MAAM,CAAC,cAAc,CAAC,YAAY,EAAE,iBAAiB,EAAE,EAAE,KAAK,EAAE,IAAI,EAAE,CAAC,CAAC;AAExE,mGAAmG;AACnG,MAAM,UAAU,iBAAiB,CAAC,EAAW,EAAW;IACvD,IAAI,OAAO,EAAE,KAAK,UAAU;QAAE,OAAO,KAAK,CAAC;IAC3C,OAAO,OAAO,CAAC,GAAG,CAAC,EAAE,EAAE,iBAAiB,CAAC,KAAK,IAAI,CAAC;AAAA,CACnD;AAED,MAAM,CAAC,KAAK,UAAU,cAAc,CACnC,KAAkB,EAClB,OAAgB,EAChB,OAA6B,EACD;IAC5B,MAAM,CAAC,GAAG,YAAY,CAAC,KAAK,EAAE,OAAO,EAAE,OAAO,CAAC,CAAC;IAChD,sEAAsE;IACtE,IAAI,KAAK,EAAE,MAAM,MAAM,IAAI,CAAC,EAAE,CAAC;IAC/B,CAAC;IACD,OAAO,CAAC,CAAC,MAAM,EAAE,CAAC;AAAA,CAClB","sourcesContent":["import \"./providers/register-builtins.ts\";\n\nimport { getApiProvider } from \"./api-registry.ts\";\nimport { getEnvApiKey } from \"./env-api-keys.ts\";\nimport type {\n\tApi,\n\tAssistantMessage,\n\tAssistantMessageEventStream,\n\tContext,\n\tModel,\n\tProviderStreamOptions,\n\tSimpleStreamOptions,\n\tStreamOptions,\n} from \"./types.ts\";\n\nexport { getEnvApiKey } from \"./env-api-keys.ts\";\n\nfunction hasExplicitApiKey(apiKey: string | undefined): apiKey is string {\n\treturn typeof apiKey === \"string\" && apiKey.trim().length > 0;\n}\n\nfunction withEnvApiKey<TOptions extends StreamOptions>(\n\tmodel: Model<Api>,\n\toptions: TOptions | undefined,\n): TOptions | undefined {\n\tif (hasExplicitApiKey(options?.apiKey)) return options;\n\tconst apiKey = getEnvApiKey(model.provider);\n\tif (!apiKey) return options;\n\treturn { ...options, apiKey } as TOptions;\n}\n\nfunction resolveApiProvider(api: Api) {\n\tconst provider = getApiProvider(api);\n\tif (!provider) {\n\t\tthrow new Error(`No API provider registered for api: ${api}`);\n\t}\n\treturn provider;\n}\n\nexport function stream<TApi extends Api>(\n\tmodel: Model<TApi>,\n\tcontext: Context,\n\toptions?: ProviderStreamOptions,\n): AssistantMessageEventStream {\n\tconst provider = resolveApiProvider(model.api);\n\treturn provider.stream(model, context, withEnvApiKey(model, options) as StreamOptions);\n}\n\nexport async function complete<TApi extends Api>(\n\tmodel: Model<TApi>,\n\tcontext: Context,\n\toptions?: ProviderStreamOptions,\n): Promise<AssistantMessage> {\n\tconst s = stream(model, context, options);\n\t// This API exposes only the final message; consume deltas as they arrive.\n\tfor await (const _event of s) {\n\t}\n\treturn s.result();\n}\n\n/**\n * Brand identifying the built-in stream function.\n *\n * Callers decide whether provider credentials are mandatory by asking \"is this\n * still the built-in stream function?\", and a bare `fn === streamSimple`\n * answers that with reference identity. Reference identity is not dependable\n * here: this package can legitimately load more than once in one process (a\n * workspace symlink beside an installed copy, or two dependents resolving\n * different versions), producing two distinct `streamSimple` functions that\n * behave identically. The comparison then reports \"custom stream function\" for\n * what is really the built-in one, and a credential check silently relaxes.\n *\n * `Symbol.for` keys into the per-realm global registry, so every copy of this\n * module brands its own `streamSimple` with the same symbol and the check\n * survives duplication.\n */\nconst BUILTIN_STREAM_FN = Symbol.for(\"omk-ai.streamSimple\");\n\nexport function streamSimple<TApi extends Api>(\n\tmodel: Model<TApi>,\n\tcontext: Context,\n\toptions?: SimpleStreamOptions,\n): AssistantMessageEventStream {\n\tconst provider = resolveApiProvider(model.api);\n\treturn provider.streamSimple(model, context, withEnvApiKey(model, options));\n}\n\nObject.defineProperty(streamSimple, BUILTIN_STREAM_FN, { value: true });\n\n/** True when `fn` is this package's built-in stream function, including a duplicate copy of it. */\nexport function isBuiltinStreamFn(fn: unknown): boolean {\n\tif (typeof fn !== \"function\") return false;\n\treturn Reflect.get(fn, BUILTIN_STREAM_FN) === true;\n}\n\nexport async function completeSimple<TApi extends Api>(\n\tmodel: Model<TApi>,\n\tcontext: Context,\n\toptions?: SimpleStreamOptions,\n): Promise<AssistantMessage> {\n\tconst s = streamSimple(model, context, options);\n\t// Keep completion-only callers from retaining the full event backlog.\n\tfor await (const _event of s) {\n\t}\n\treturn s.result();\n}\n"]}

package/dist/utils/provider-retry-delay.d.ts

@@ -0,0 +1,4 @@
+export declare function validateProviderRetryOptions(maxRetries: number, maxRetryDelayMs: number | undefined): void;
+/** Invalid server metadata falls back to bounded exponential jitter, not a 1ms timer. */
+export declare function computeProviderRetryDelay(headers: Headers | undefined, retryIndex: number, maxRetryDelayMs: number | undefined, providerErrorMessage: string, nowMs: number, randomValue: number): number;
+//# sourceMappingURL=provider-retry-delay.d.ts.map

package/dist/utils/provider-retry-delay.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"provider-retry-delay.d.ts","sourceRoot":"","sources":["../../src/utils/provider-retry-delay.ts"],"names":[],"mappings":"AAGA,wBAAgB,4BAA4B,CAAC,UAAU,EAAE,MAAM,EAAE,eAAe,EAAE,MAAM,GAAG,SAAS,GAAG,IAAI,CAO1G;AAQD,yFAAyF;AACzF,wBAAgB,yBAAyB,CACxC,OAAO,EAAE,OAAO,GAAG,SAAS,EAC5B,UAAU,EAAE,MAAM,EAClB,eAAe,EAAE,MAAM,GAAG,SAAS,EACnC,oBAAoB,EAAE,MAAM,EAC5B,KAAK,EAAE,MAAM,EACb,WAAW,EAAE,MAAM,GACjB,MAAM,CA+BR","sourcesContent":["const DEFAULT_MAX_RETRY_DELAY_MS = 60_000;\nconst DECIMAL = /^(?:\\d+(?:\\.\\d*)?|\\.\\d+)$/;\n\nexport function validateProviderRetryOptions(maxRetries: number, maxRetryDelayMs: number | undefined): void {\n\tif (!Number.isSafeInteger(maxRetries) || maxRetries < 0) {\n\t\tthrow new RangeError(\"maxRetries must be a non-negative safe integer\");\n\t}\n\tif (maxRetryDelayMs !== undefined && (!Number.isFinite(maxRetryDelayMs) || maxRetryDelayMs < 0)) {\n\t\tthrow new RangeError(\"maxRetryDelayMs must be finite and non-negative; zero disables the server-delay cap\");\n\t}\n}\n\nfunction decimal(raw: string | null): number | undefined {\n\tif (raw === null || !DECIMAL.test(raw.trim())) return undefined;\n\tconst value = Number(raw.trim());\n\treturn Number.isFinite(value) && value >= 0 ? value : undefined;\n}\n\n/** Invalid server metadata falls back to bounded exponential jitter, not a 1ms timer. */\nexport function computeProviderRetryDelay(\n\theaders: Headers | undefined,\n\tretryIndex: number,\n\tmaxRetryDelayMs: number | undefined,\n\tproviderErrorMessage: string,\n\tnowMs: number,\n\trandomValue: number,\n): number {\n\tif (!Number.isSafeInteger(retryIndex) || retryIndex < 0) throw new RangeError(\"Invalid retryIndex\");\n\tif (!Number.isFinite(nowMs)) throw new RangeError(\"Invalid wall clock\");\n\tif (!Number.isFinite(randomValue) || randomValue < 0 || randomValue >= 1)\n\t\tthrow new RangeError(\"Invalid jitter draw\");\n\tvalidateProviderRetryOptions(0, maxRetryDelayMs);\n\tlet serverDelay = decimal(headers?.get(\"retry-after-ms\") ?? null);\n\tif (serverDelay === undefined) {\n\t\tconst raw = headers?.get(\"retry-after\")?.trim() ?? null;\n\t\tconst seconds = decimal(raw);\n\t\tif (seconds !== undefined) {\n\t\t\tconst milliseconds = seconds * 1000;\n\t\t\tif (Number.isFinite(milliseconds)) serverDelay = milliseconds;\n\t\t} else if (raw !== null && /^[A-Za-z]{3,9},?\\s/.test(raw)) {\n\t\t\t// Accept HTTP date forms; do not reinterpret a numeric suffix or negative\n\t\t\t// duration as a permissively parsed calendar date.\n\t\t\tconst date = Date.parse(raw);\n\t\t\tconst difference = date - nowMs;\n\t\t\tif (Number.isFinite(date) && Number.isFinite(difference)) serverDelay = Math.max(0, difference);\n\t\t}\n\t}\n\tif (serverDelay !== undefined) {\n\t\tconst cap = maxRetryDelayMs ?? DEFAULT_MAX_RETRY_DELAY_MS;\n\t\tif (cap > 0 && serverDelay > cap) {\n\t\t\tthrow new Error(\n\t\t\t\t`Server requested ${Math.ceil(serverDelay / 1000)}s retry delay (max: ${Math.ceil(cap / 1000)}s). ${providerErrorMessage}`,\n\t\t\t);\n\t\t}\n\t\treturn serverDelay;\n\t}\n\treturn Math.min(500 * 2 ** Math.min(retryIndex, 4), 8000) * (1 - randomValue * 0.25);\n}\n"]}

package/dist/utils/provider-retry-delay.js

@@ -0,0 +1,53 @@
+const DEFAULT_MAX_RETRY_DELAY_MS = 60_000;
+const DECIMAL = /^(?:\d+(?:\.\d*)?|\.\d+)$/;
+export function validateProviderRetryOptions(maxRetries, maxRetryDelayMs) {
+    if (!Number.isSafeInteger(maxRetries) || maxRetries < 0) {
+        throw new RangeError("maxRetries must be a non-negative safe integer");
+    }
+    if (maxRetryDelayMs !== undefined && (!Number.isFinite(maxRetryDelayMs) || maxRetryDelayMs < 0)) {
+        throw new RangeError("maxRetryDelayMs must be finite and non-negative; zero disables the server-delay cap");
+    }
+}
+function decimal(raw) {
+    if (raw === null || !DECIMAL.test(raw.trim()))
+        return undefined;
+    const value = Number(raw.trim());
+    return Number.isFinite(value) && value >= 0 ? value : undefined;
+}
+/** Invalid server metadata falls back to bounded exponential jitter, not a 1ms timer. */
+export function computeProviderRetryDelay(headers, retryIndex, maxRetryDelayMs, providerErrorMessage, nowMs, randomValue) {
+    if (!Number.isSafeInteger(retryIndex) || retryIndex < 0)
+        throw new RangeError("Invalid retryIndex");
+    if (!Number.isFinite(nowMs))
+        throw new RangeError("Invalid wall clock");
+    if (!Number.isFinite(randomValue) || randomValue < 0 || randomValue >= 1)
+        throw new RangeError("Invalid jitter draw");
+    validateProviderRetryOptions(0, maxRetryDelayMs);
+    let serverDelay = decimal(headers?.get("retry-after-ms") ?? null);
+    if (serverDelay === undefined) {
+        const raw = headers?.get("retry-after")?.trim() ?? null;
+        const seconds = decimal(raw);
+        if (seconds !== undefined) {
+            const milliseconds = seconds * 1000;
+            if (Number.isFinite(milliseconds))
+                serverDelay = milliseconds;
+        }
+        else if (raw !== null && /^[A-Za-z]{3,9},?\s/.test(raw)) {
+            // Accept HTTP date forms; do not reinterpret a numeric suffix or negative
+            // duration as a permissively parsed calendar date.
+            const date = Date.parse(raw);
+            const difference = date - nowMs;
+            if (Number.isFinite(date) && Number.isFinite(difference))
+                serverDelay = Math.max(0, difference);
+        }
+    }
+    if (serverDelay !== undefined) {
+        const cap = maxRetryDelayMs ?? DEFAULT_MAX_RETRY_DELAY_MS;
+        if (cap > 0 && serverDelay > cap) {
+            throw new Error(`Server requested ${Math.ceil(serverDelay / 1000)}s retry delay (max: ${Math.ceil(cap / 1000)}s). ${providerErrorMessage}`);
+        }
+        return serverDelay;
+    }
+    return Math.min(500 * 2 ** Math.min(retryIndex, 4), 8000) * (1 - randomValue * 0.25);
+}
+//# sourceMappingURL=provider-retry-delay.js.map

package/dist/utils/provider-retry-delay.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"provider-retry-delay.js","sourceRoot":"","sources":["../../src/utils/provider-retry-delay.ts"],"names":[],"mappings":"AAAA,MAAM,0BAA0B,GAAG,MAAM,CAAC;AAC1C,MAAM,OAAO,GAAG,2BAA2B,CAAC;AAE5C,MAAM,UAAU,4BAA4B,CAAC,UAAkB,EAAE,eAAmC,EAAQ;IAC3G,IAAI,CAAC,MAAM,CAAC,aAAa,CAAC,UAAU,CAAC,IAAI,UAAU,GAAG,CAAC,EAAE,CAAC;QACzD,MAAM,IAAI,UAAU,CAAC,gDAAgD,CAAC,CAAC;IACxE,CAAC;IACD,IAAI,eAAe,KAAK,SAAS,IAAI,CAAC,CAAC,MAAM,CAAC,QAAQ,CAAC,eAAe,CAAC,IAAI,eAAe,GAAG,CAAC,CAAC,EAAE,CAAC;QACjG,MAAM,IAAI,UAAU,CAAC,qFAAqF,CAAC,CAAC;IAC7G,CAAC;AAAA,CACD;AAED,SAAS,OAAO,CAAC,GAAkB,EAAsB;IACxD,IAAI,GAAG,KAAK,IAAI,IAAI,CAAC,OAAO,CAAC,IAAI,CAAC,GAAG,CAAC,IAAI,EAAE,CAAC;QAAE,OAAO,SAAS,CAAC;IAChE,MAAM,KAAK,GAAG,MAAM,CAAC,GAAG,CAAC,IAAI,EAAE,CAAC,CAAC;IACjC,OAAO,MAAM,CAAC,QAAQ,CAAC,KAAK,CAAC,IAAI,KAAK,IAAI,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,SAAS,CAAC;AAAA,CAChE;AAED,yFAAyF;AACzF,MAAM,UAAU,yBAAyB,CACxC,OAA4B,EAC5B,UAAkB,EAClB,eAAmC,EACnC,oBAA4B,EAC5B,KAAa,EACb,WAAmB,EACV;IACT,IAAI,CAAC,MAAM,CAAC,aAAa,CAAC,UAAU,CAAC,IAAI,UAAU,GAAG,CAAC;QAAE,MAAM,IAAI,UAAU,CAAC,oBAAoB,CAAC,CAAC;IACpG,IAAI,CAAC,MAAM,CAAC,QAAQ,CAAC,KAAK,CAAC;QAAE,MAAM,IAAI,UAAU,CAAC,oBAAoB,CAAC,CAAC;IACxE,IAAI,CAAC,MAAM,CAAC,QAAQ,CAAC,WAAW,CAAC,IAAI,WAAW,GAAG,CAAC,IAAI,WAAW,IAAI,CAAC;QACvE,MAAM,IAAI,UAAU,CAAC,qBAAqB,CAAC,CAAC;IAC7C,4BAA4B,CAAC,CAAC,EAAE,eAAe,CAAC,CAAC;IACjD,IAAI,WAAW,GAAG,OAAO,CAAC,OAAO,EAAE,GAAG,CAAC,gBAAgB,CAAC,IAAI,IAAI,CAAC,CAAC;IAClE,IAAI,WAAW,KAAK,SAAS,EAAE,CAAC;QAC/B,MAAM,GAAG,GAAG,OAAO,EAAE,GAAG,CAAC,aAAa,CAAC,EAAE,IAAI,EAAE,IAAI,IAAI,CAAC;QACxD,MAAM,OAAO,GAAG,OAAO,CAAC,GAAG,CAAC,CAAC;QAC7B,IAAI,OAAO,KAAK,SAAS,EAAE,CAAC;YAC3B,MAAM,YAAY,GAAG,OAAO,GAAG,IAAI,CAAC;YACpC,IAAI,MAAM,CAAC,QAAQ,CAAC,YAAY,CAAC;gBAAE,WAAW,GAAG,YAAY,CAAC;QAC/D,CAAC;aAAM,IAAI,GAAG,KAAK,IAAI,IAAI,oBAAoB,CAAC,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC;YAC3D,0EAA0E;YAC1E,mDAAmD;YACnD,MAAM,IAAI,GAAG,IAAI,CAAC,KAAK,CAAC,GAAG,CAAC,CAAC;YAC7B,MAAM,UAAU,GAAG,IAAI,GAAG,KAAK,CAAC;YAChC,IAAI,MAAM,CAAC,QAAQ,CAAC,IAAI,CAAC,IAAI,MAAM,CAAC,QAAQ,CAAC,UAAU,CAAC;gBAAE,WAAW,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,UAAU,CAAC,CAAC;QACjG,CAAC;IACF,CAAC;IACD,IAAI,WAAW,KAAK,SAAS,EAAE,CAAC;QAC/B,MAAM,GAAG,GAAG,eAAe,IAAI,0BAA0B,CAAC;QAC1D,IAAI,GAAG,GAAG,CAAC,IAAI,WAAW,GAAG,GAAG,EAAE,CAAC;YAClC,MAAM,IAAI,KAAK,CACd,oBAAoB,IAAI,CAAC,IAAI,CAAC,WAAW,GAAG,IAAI,CAAC,uBAAuB,IAAI,CAAC,IAAI,CAAC,GAAG,GAAG,IAAI,CAAC,OAAO,oBAAoB,EAAE,CAC1H,CAAC;QACH,CAAC;QACD,OAAO,WAAW,CAAC;IACpB,CAAC;IACD,OAAO,IAAI,CAAC,GAAG,CAAC,GAAG,GAAG,CAAC,IAAI,IAAI,CAAC,GAAG,CAAC,UAAU,EAAE,CAAC,CAAC,EAAE,IAAI,CAAC,GAAG,CAAC,CAAC,GAAG,WAAW,GAAG,IAAI,CAAC,CAAC;AAAA,CACrF","sourcesContent":["const DEFAULT_MAX_RETRY_DELAY_MS = 60_000;\nconst DECIMAL = /^(?:\\d+(?:\\.\\d*)?|\\.\\d+)$/;\n\nexport function validateProviderRetryOptions(maxRetries: number, maxRetryDelayMs: number | undefined): void {\n\tif (!Number.isSafeInteger(maxRetries) || maxRetries < 0) {\n\t\tthrow new RangeError(\"maxRetries must be a non-negative safe integer\");\n\t}\n\tif (maxRetryDelayMs !== undefined && (!Number.isFinite(maxRetryDelayMs) || maxRetryDelayMs < 0)) {\n\t\tthrow new RangeError(\"maxRetryDelayMs must be finite and non-negative; zero disables the server-delay cap\");\n\t}\n}\n\nfunction decimal(raw: string | null): number | undefined {\n\tif (raw === null || !DECIMAL.test(raw.trim())) return undefined;\n\tconst value = Number(raw.trim());\n\treturn Number.isFinite(value) && value >= 0 ? value : undefined;\n}\n\n/** Invalid server metadata falls back to bounded exponential jitter, not a 1ms timer. */\nexport function computeProviderRetryDelay(\n\theaders: Headers | undefined,\n\tretryIndex: number,\n\tmaxRetryDelayMs: number | undefined,\n\tproviderErrorMessage: string,\n\tnowMs: number,\n\trandomValue: number,\n): number {\n\tif (!Number.isSafeInteger(retryIndex) || retryIndex < 0) throw new RangeError(\"Invalid retryIndex\");\n\tif (!Number.isFinite(nowMs)) throw new RangeError(\"Invalid wall clock\");\n\tif (!Number.isFinite(randomValue) || randomValue < 0 || randomValue >= 1)\n\t\tthrow new RangeError(\"Invalid jitter draw\");\n\tvalidateProviderRetryOptions(0, maxRetryDelayMs);\n\tlet serverDelay = decimal(headers?.get(\"retry-after-ms\") ?? null);\n\tif (serverDelay === undefined) {\n\t\tconst raw = headers?.get(\"retry-after\")?.trim() ?? null;\n\t\tconst seconds = decimal(raw);\n\t\tif (seconds !== undefined) {\n\t\t\tconst milliseconds = seconds * 1000;\n\t\t\tif (Number.isFinite(milliseconds)) serverDelay = milliseconds;\n\t\t} else if (raw !== null && /^[A-Za-z]{3,9},?\\s/.test(raw)) {\n\t\t\t// Accept HTTP date forms; do not reinterpret a numeric suffix or negative\n\t\t\t// duration as a permissively parsed calendar date.\n\t\t\tconst date = Date.parse(raw);\n\t\t\tconst difference = date - nowMs;\n\t\t\tif (Number.isFinite(date) && Number.isFinite(difference)) serverDelay = Math.max(0, difference);\n\t\t}\n\t}\n\tif (serverDelay !== undefined) {\n\t\tconst cap = maxRetryDelayMs ?? DEFAULT_MAX_RETRY_DELAY_MS;\n\t\tif (cap > 0 && serverDelay > cap) {\n\t\t\tthrow new Error(\n\t\t\t\t`Server requested ${Math.ceil(serverDelay / 1000)}s retry delay (max: ${Math.ceil(cap / 1000)}s). ${providerErrorMessage}`,\n\t\t\t);\n\t\t}\n\t\treturn serverDelay;\n\t}\n\treturn Math.min(500 * 2 ** Math.min(retryIndex, 4), 8000) * (1 - randomValue * 0.25);\n}\n"]}

package/dist/utils/provider-retry-sleep.d.ts

@@ -0,0 +1,6 @@
+/** Largest portable Node timer delay; larger delays are re-armed in chunks. */
+export declare const MAX_RETRY_TIMER_MS = 2147483647;
+export declare function createProviderAbortError(): Error;
+/** A long server delay must never overflow setTimeout into an immediate retry. */
+export declare function sleepProviderRetry(ms: number, signal?: AbortSignal): Promise<void>;
+//# sourceMappingURL=provider-retry-sleep.d.ts.map

package/dist/utils/provider-retry-sleep.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"provider-retry-sleep.d.ts","sourceRoot":"","sources":["../../src/utils/provider-retry-sleep.ts"],"names":[],"mappings":"AAAA,+EAA+E;AAC/E,eAAO,MAAM,kBAAkB,aAAgB,CAAC;AAEhD,wBAAgB,wBAAwB,IAAI,KAAK,CAIhD;AAED,kFAAkF;AAClF,wBAAgB,kBAAkB,CAAC,EAAE,EAAE,MAAM,EAAE,MAAM,CAAC,EAAE,WAAW,GAAG,OAAO,CAAC,IAAI,CAAC,CAoClF","sourcesContent":["/** Largest portable Node timer delay; larger delays are re-armed in chunks. */\nexport const MAX_RETRY_TIMER_MS = 2_147_483_647;\n\nexport function createProviderAbortError(): Error {\n\tconst error = new Error(\"Request aborted\");\n\terror.name = \"AbortError\";\n\treturn error;\n}\n\n/** A long server delay must never overflow setTimeout into an immediate retry. */\nexport function sleepProviderRetry(ms: number, signal?: AbortSignal): Promise<void> {\n\tif (!Number.isFinite(ms) || ms < 0)\n\t\treturn Promise.reject(new RangeError(\"Retry sleep must be finite and non-negative\"));\n\treturn new Promise((resolve, reject) => {\n\t\tif (signal?.aborted) {\n\t\t\treject(createProviderAbortError());\n\t\t\treturn;\n\t\t}\n\t\tconst startedAt = performance.now();\n\t\tlet timer: ReturnType<typeof setTimeout> | undefined;\n\t\tlet finished = false;\n\t\tconst cleanup = (): void => {\n\t\t\tif (timer !== undefined) clearTimeout(timer);\n\t\t\tsignal?.removeEventListener(\"abort\", onAbort);\n\t\t};\n\t\tconst onAbort = (): void => {\n\t\t\tif (finished) return;\n\t\t\tfinished = true;\n\t\t\tcleanup();\n\t\t\treject(createProviderAbortError());\n\t\t};\n\t\tconst tick = (): void => {\n\t\t\tif (finished) return;\n\t\t\tconst remaining = ms - (performance.now() - startedAt);\n\t\t\tif (remaining <= 0) {\n\t\t\t\tfinished = true;\n\t\t\t\tcleanup();\n\t\t\t\tresolve();\n\t\t\t\treturn;\n\t\t\t}\n\t\t\ttimer = setTimeout(tick, Math.min(MAX_RETRY_TIMER_MS, Math.max(1, Math.ceil(remaining))));\n\t\t};\n\t\tsignal?.addEventListener(\"abort\", onAbort, { once: true });\n\t\tif (signal?.aborted) onAbort();\n\t\telse timer = setTimeout(tick, Math.min(MAX_RETRY_TIMER_MS, Math.max(0, Math.ceil(ms))));\n\t});\n}\n"]}

package/dist/utils/provider-retry-sleep.js

@@ -0,0 +1,51 @@
+/** Largest portable Node timer delay; larger delays are re-armed in chunks. */
+export const MAX_RETRY_TIMER_MS = 2_147_483_647;
+export function createProviderAbortError() {
+    const error = new Error("Request aborted");
+    error.name = "AbortError";
+    return error;
+}
+/** A long server delay must never overflow setTimeout into an immediate retry. */
+export function sleepProviderRetry(ms, signal) {
+    if (!Number.isFinite(ms) || ms < 0)
+        return Promise.reject(new RangeError("Retry sleep must be finite and non-negative"));
+    return new Promise((resolve, reject) => {
+        if (signal?.aborted) {
+            reject(createProviderAbortError());
+            return;
+        }
+        const startedAt = performance.now();
+        let timer;
+        let finished = false;
+        const cleanup = () => {
+            if (timer !== undefined)
+                clearTimeout(timer);
+            signal?.removeEventListener("abort", onAbort);
+        };
+        const onAbort = () => {
+            if (finished)
+                return;
+            finished = true;
+            cleanup();
+            reject(createProviderAbortError());
+        };
+        const tick = () => {
+            if (finished)
+                return;
+            const remaining = ms - (performance.now() - startedAt);
+            if (remaining <= 0) {
+                finished = true;
+                cleanup();
+                resolve();
+                return;
+            }
+            timer = setTimeout(tick, Math.min(MAX_RETRY_TIMER_MS, Math.max(1, Math.ceil(remaining))));
+        };
+        signal?.addEventListener("abort", onAbort, { once: true });
+        if (signal?.aborted)
+            onAbort();
+        else
+            timer = setTimeout(tick, Math.min(MAX_RETRY_TIMER_MS, Math.max(0, Math.ceil(ms))));
+    });
+}
+//# sourceMappingURL=provider-retry-sleep.js.map

package/dist/utils/provider-retry-sleep.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"provider-retry-sleep.js","sourceRoot":"","sources":["../../src/utils/provider-retry-sleep.ts"],"names":[],"mappings":"AAAA,+EAA+E;AAC/E,MAAM,CAAC,MAAM,kBAAkB,GAAG,aAAa,CAAC;AAEhD,MAAM,UAAU,wBAAwB,GAAU;IACjD,MAAM,KAAK,GAAG,IAAI,KAAK,CAAC,iBAAiB,CAAC,CAAC;IAC3C,KAAK,CAAC,IAAI,GAAG,YAAY,CAAC;IAC1B,OAAO,KAAK,CAAC;AAAA,CACb;AAED,kFAAkF;AAClF,MAAM,UAAU,kBAAkB,CAAC,EAAU,EAAE,MAAoB,EAAiB;IACnF,IAAI,CAAC,MAAM,CAAC,QAAQ,CAAC,EAAE,CAAC,IAAI,EAAE,GAAG,CAAC;QACjC,OAAO,OAAO,CAAC,MAAM,CAAC,IAAI,UAAU,CAAC,6CAA6C,CAAC,CAAC,CAAC;IACtF,OAAO,IAAI,OAAO,CAAC,CAAC,OAAO,EAAE,MAAM,EAAE,EAAE,CAAC;QACvC,IAAI,MAAM,EAAE,OAAO,EAAE,CAAC;YACrB,MAAM,CAAC,wBAAwB,EAAE,CAAC,CAAC;YACnC,OAAO;QACR,CAAC;QACD,MAAM,SAAS,GAAG,WAAW,CAAC,GAAG,EAAE,CAAC;QACpC,IAAI,KAAgD,CAAC;QACrD,IAAI,QAAQ,GAAG,KAAK,CAAC;QACrB,MAAM,OAAO,GAAG,GAAS,EAAE,CAAC;YAC3B,IAAI,KAAK,KAAK,SAAS;gBAAE,YAAY,CAAC,KAAK,CAAC,CAAC;YAC7C,MAAM,EAAE,mBAAmB,CAAC,OAAO,EAAE,OAAO,CAAC,CAAC;QAAA,CAC9C,CAAC;QACF,MAAM,OAAO,GAAG,GAAS,EAAE,CAAC;YAC3B,IAAI,QAAQ;gBAAE,OAAO;YACrB,QAAQ,GAAG,IAAI,CAAC;YAChB,OAAO,EAAE,CAAC;YACV,MAAM,CAAC,wBAAwB,EAAE,CAAC,CAAC;QAAA,CACnC,CAAC;QACF,MAAM,IAAI,GAAG,GAAS,EAAE,CAAC;YACxB,IAAI,QAAQ;gBAAE,OAAO;YACrB,MAAM,SAAS,GAAG,EAAE,GAAG,CAAC,WAAW,CAAC,GAAG,EAAE,GAAG,SAAS,CAAC,CAAC;YACvD,IAAI,SAAS,IAAI,CAAC,EAAE,CAAC;gBACpB,QAAQ,GAAG,IAAI,CAAC;gBAChB,OAAO,EAAE,CAAC;gBACV,OAAO,EAAE,CAAC;gBACV,OAAO;YACR,CAAC;YACD,KAAK,GAAG,UAAU,CAAC,IAAI,EAAE,IAAI,CAAC,GAAG,CAAC,kBAAkB,EAAE,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,IAAI,CAAC,IAAI,CAAC,SAAS,CAAC,CAAC,CAAC,CAAC,CAAC;QAAA,CAC1F,CAAC;QACF,MAAM,EAAE,gBAAgB,CAAC,OAAO,EAAE,OAAO,EAAE,EAAE,IAAI,EAAE,IAAI,EAAE,CAAC,CAAC;QAC3D,IAAI,MAAM,EAAE,OAAO;YAAE,OAAO,EAAE,CAAC;;YAC1B,KAAK,GAAG,UAAU,CAAC,IAAI,EAAE,IAAI,CAAC,GAAG,CAAC,kBAAkB,EAAE,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,IAAI,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC,CAAC;IAAA,CACxF,CAAC,CAAC;AAAA,CACH","sourcesContent":["/** Largest portable Node timer delay; larger delays are re-armed in chunks. */\nexport const MAX_RETRY_TIMER_MS = 2_147_483_647;\n\nexport function createProviderAbortError(): Error {\n\tconst error = new Error(\"Request aborted\");\n\terror.name = \"AbortError\";\n\treturn error;\n}\n\n/** A long server delay must never overflow setTimeout into an immediate retry. */\nexport function sleepProviderRetry(ms: number, signal?: AbortSignal): Promise<void> {\n\tif (!Number.isFinite(ms) || ms < 0)\n\t\treturn Promise.reject(new RangeError(\"Retry sleep must be finite and non-negative\"));\n\treturn new Promise((resolve, reject) => {\n\t\tif (signal?.aborted) {\n\t\t\treject(createProviderAbortError());\n\t\t\treturn;\n\t\t}\n\t\tconst startedAt = performance.now();\n\t\tlet timer: ReturnType<typeof setTimeout> | undefined;\n\t\tlet finished = false;\n\t\tconst cleanup = (): void => {\n\t\t\tif (timer !== undefined) clearTimeout(timer);\n\t\t\tsignal?.removeEventListener(\"abort\", onAbort);\n\t\t};\n\t\tconst onAbort = (): void => {\n\t\t\tif (finished) return;\n\t\t\tfinished = true;\n\t\t\tcleanup();\n\t\t\treject(createProviderAbortError());\n\t\t};\n\t\tconst tick = (): void => {\n\t\t\tif (finished) return;\n\t\t\tconst remaining = ms - (performance.now() - startedAt);\n\t\t\tif (remaining <= 0) {\n\t\t\t\tfinished = true;\n\t\t\t\tcleanup();\n\t\t\t\tresolve();\n\t\t\t\treturn;\n\t\t\t}\n\t\t\ttimer = setTimeout(tick, Math.min(MAX_RETRY_TIMER_MS, Math.max(1, Math.ceil(remaining))));\n\t\t};\n\t\tsignal?.addEventListener(\"abort\", onAbort, { once: true });\n\t\tif (signal?.aborted) onAbort();\n\t\telse timer = setTimeout(tick, Math.min(MAX_RETRY_TIMER_MS, Math.max(0, Math.ceil(ms))));\n\t});\n}\n"]}

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

@@ -1 +1 @@
-{"version":3,"file":"provider-retry.d.ts","sourceRoot":"","sources":["../../src/utils/provider-retry.ts"],"names":[],"mappings":"AAEA,UAAU,oBAAoB;IAC7B,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,eAAe,CAAC,EAAE,MAAM,CAAC;IACzB,MAAM,CAAC,EAAE,WAAW,CAAC;CACrB;AA0FD;;;;;;;GAOG;AACH,wBAAsB,oBAAoB,CAAC,CAAC,EAC3C,OAAO,EAAE,MAAM,OAAO,CAAC,CAAC,CAAC,EACzB,OAAO,GAAE,oBAAyB,GAChC,OAAO,CAAC,CAAC,CAAC,CAiBZ","sourcesContent":["const DEFAULT_MAX_RETRY_DELAY_MS = 60_000;\n\ninterface ProviderRetryOptions {\n\tmaxRetries?: number;\n\tmaxRetryDelayMs?: number;\n\tsignal?: AbortSignal;\n}\n\ninterface ProviderError extends Error {\n\tstatus: number | undefined;\n\theaders: Headers | undefined;\n}\n\nfunction isProviderError(error: unknown): error is ProviderError {\n\tif (!(error instanceof Error) || !(\"status\" in error) || !(\"headers\" in error)) return false;\n\treturn (\n\t\t(error.status === undefined || typeof error.status === \"number\") &&\n\t\t(error.headers === undefined || error.headers instanceof Headers)\n\t);\n}\n\n/** Mirrors the pinned OpenAI/Anthropic SDK retry policy; review when either SDK is upgraded. */\nfunction isRetryableProviderError(error: ProviderError): boolean {\n\tconst shouldRetry = error.headers?.get(\"x-should-retry\");\n\tif (shouldRetry === \"true\") return true;\n\tif (shouldRetry === \"false\") return false;\n\n\tif (error.status === undefined) return true;\n\treturn (\n\t\terror.status === 408 ||\n\t\terror.status === 409 ||\n\t\terror.status === 429 ||\n\t\t(typeof error.status === \"number\" && error.status >= 500)\n\t);\n}\n\nfunction validateServerRetryDelayMs(\n\tdelayMs: number,\n\tmaxRetryDelayMs: number | undefined,\n\tproviderErrorMessage: string,\n): number {\n\tconst maxDelayMs = maxRetryDelayMs ?? DEFAULT_MAX_RETRY_DELAY_MS;\n\tif (maxDelayMs > 0 && delayMs > maxDelayMs) {\n\t\tthrow new Error(\n\t\t\t`Server requested ${Math.ceil(delayMs / 1000)}s retry delay (max: ${Math.ceil(maxDelayMs / 1000)}s). ${providerErrorMessage}`,\n\t\t);\n\t}\n\treturn delayMs;\n}\n\nfunction getRetryDelayMs(error: ProviderError, retryIndex: number, maxRetryDelayMs: number | undefined): number {\n\tconst retryAfterMs = error.headers?.get(\"retry-after-ms\");\n\tif (retryAfterMs) {\n\t\tconst value = Number.parseFloat(retryAfterMs);\n\t\tif (!Number.isNaN(value)) return validateServerRetryDelayMs(value, maxRetryDelayMs, error.message);\n\t}\n\n\tconst retryAfter = error.headers?.get(\"retry-after\");\n\tif (retryAfter) {\n\t\tconst seconds = Number.parseFloat(retryAfter);\n\t\tconst delayMs = Number.isNaN(seconds) ? Date.parse(retryAfter) - Date.now() : seconds * 1000;\n\t\treturn validateServerRetryDelayMs(delayMs, maxRetryDelayMs, error.message);\n\t}\n\n\tconst exponentialDelay = Math.min(0.5 * 2 ** retryIndex, 8) * 1000;\n\treturn exponentialDelay * (1 - Math.random() * 0.25);\n}\n\nfunction createAbortError(): Error {\n\tconst error = new Error(\"Request aborted\");\n\terror.name = \"AbortError\";\n\treturn error;\n}\n\nfunction abortableSleep(ms: number, signal?: AbortSignal): Promise<void> {\n\treturn new Promise((resolve, reject) => {\n\t\tif (signal?.aborted) {\n\t\t\treject(createAbortError());\n\t\t\treturn;\n\t\t}\n\n\t\tconst onAbort = () => {\n\t\t\tclearTimeout(timeout);\n\t\t\treject(createAbortError());\n\t\t};\n\t\tconst timeout = setTimeout(\n\t\t\t() => {\n\t\t\t\tsignal?.removeEventListener(\"abort\", onAbort);\n\t\t\t\tresolve();\n\t\t\t},\n\t\t\tMath.max(0, ms),\n\t\t);\n\t\tsignal?.addEventListener(\"abort\", onAbort, { once: true });\n\t});\n}\n\n/**\n * Reproduce the retry behavior used by the OpenAI and Anthropic SDKs while making\n * their backoff sleep interruptible. Their built-in retry timers ignore the\n * request AbortSignal, so callers must invoke the SDK with `maxRetries: 0` and\n * wrap the request with this helper. Provider-requested delays above\n * `maxRetryDelayMs` fail immediately (60 seconds by default); set it to zero to\n * disable the limit.\n */\nexport async function retryProviderRequest<T>(\n\trequest: () => Promise<T>,\n\toptions: ProviderRetryOptions = {},\n): Promise<T> {\n\tconst maxRetries = options.maxRetries ?? 0;\n\tlet retriesRemaining = maxRetries;\n\n\tfor (;;) {\n\t\ttry {\n\t\t\t// Each retry is a fresh SDK request, so X-Stainless-Retry-Count remains zero.\n\t\t\treturn await request();\n\t\t} catch (error) {\n\t\t\tif (options.signal?.aborted) throw createAbortError();\n\t\t\tif (retriesRemaining <= 0 || !isProviderError(error) || !isRetryableProviderError(error)) throw error;\n\n\t\t\tconst retryIndex = maxRetries - retriesRemaining;\n\t\t\tretriesRemaining--;\n\t\t\tawait abortableSleep(getRetryDelayMs(error, retryIndex, options.maxRetryDelayMs), options.signal);\n\t\t}\n\t}\n}\n"]}
+{"version":3,"file":"provider-retry.d.ts","sourceRoot":"","sources":["../../src/utils/provider-retry.ts"],"names":[],"mappings":"AAMA,UAAU,oBAAoB;IAC7B,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,eAAe,CAAC,EAAE,MAAM,CAAC;IACzB,MAAM,CAAC,EAAE,WAAW,CAAC;CACrB;AAyCD;;;;;;;GAOG;AACH,wBAAsB,oBAAoB,CAAC,CAAC,EAC3C,OAAO,EAAE,MAAM,OAAO,CAAC,CAAC,CAAC,EACzB,OAAO,GAAE,oBAAyB,GAChC,OAAO,CAAC,CAAC,CAAC,CAmBZ","sourcesContent":["import { computeProviderRetryDelay, validateProviderRetryOptions } from \"./provider-retry-delay.ts\";\nimport {\n\tsleepProviderRetry as abortableSleep,\n\tcreateProviderAbortError as createAbortError,\n} from \"./provider-retry-sleep.ts\";\n\ninterface ProviderRetryOptions {\n\tmaxRetries?: number;\n\tmaxRetryDelayMs?: number;\n\tsignal?: AbortSignal;\n}\n\ninterface ProviderError extends Error {\n\tstatus: number | undefined;\n\theaders: Headers | undefined;\n}\n\nfunction isProviderError(error: unknown): error is ProviderError {\n\tif (!(error instanceof Error) || !(\"status\" in error) || !(\"headers\" in error)) return false;\n\treturn (\n\t\t(error.status === undefined || typeof error.status === \"number\") &&\n\t\t(error.headers === undefined || error.headers instanceof Headers)\n\t);\n}\n\n/** Mirrors the pinned OpenAI/Anthropic SDK retry policy; review when either SDK is upgraded. */\nfunction isRetryableProviderError(error: ProviderError): boolean {\n\tconst shouldRetry = error.headers?.get(\"x-should-retry\");\n\tif (shouldRetry === \"true\") return true;\n\tif (shouldRetry === \"false\") return false;\n\n\tif (error.status === undefined) return true;\n\treturn (\n\t\terror.status === 408 ||\n\t\terror.status === 409 ||\n\t\terror.status === 429 ||\n\t\t(typeof error.status === \"number\" && error.status >= 500)\n\t);\n}\n\nfunction getRetryDelayMs(error: ProviderError, retryIndex: number, maxRetryDelayMs: number | undefined): number {\n\treturn computeProviderRetryDelay(\n\t\terror.headers,\n\t\tretryIndex,\n\t\tmaxRetryDelayMs,\n\t\terror.message,\n\t\tDate.now(),\n\t\tMath.random(),\n\t);\n}\n\n/**\n * Reproduce the retry behavior used by the OpenAI and Anthropic SDKs while making\n * their backoff sleep interruptible. Their built-in retry timers ignore the\n * request AbortSignal, so callers must invoke the SDK with `maxRetries: 0` and\n * wrap the request with this helper. Provider-requested delays above\n * `maxRetryDelayMs` fail immediately (60 seconds by default); set it to zero to\n * disable the limit.\n */\nexport async function retryProviderRequest<T>(\n\trequest: () => Promise<T>,\n\toptions: ProviderRetryOptions = {},\n): Promise<T> {\n\tconst maxRetries = options.maxRetries ?? 0;\n\tvalidateProviderRetryOptions(maxRetries, options.maxRetryDelayMs);\n\tlet retriesRemaining = maxRetries;\n\n\tfor (;;) {\n\t\tif (options.signal?.aborted) throw createAbortError();\n\t\ttry {\n\t\t\t// Each retry is a fresh SDK request, so X-Stainless-Retry-Count remains zero.\n\t\t\treturn await request();\n\t\t} catch (error) {\n\t\t\tif (options.signal?.aborted) throw createAbortError();\n\t\t\tif (retriesRemaining <= 0 || !isProviderError(error) || !isRetryableProviderError(error)) throw error;\n\n\t\t\tconst retryIndex = maxRetries - retriesRemaining;\n\t\t\tretriesRemaining--;\n\t\t\tawait abortableSleep(getRetryDelayMs(error, retryIndex, options.maxRetryDelayMs), options.signal);\n\t\t}\n\t}\n}\n"]}

package/dist/utils/provider-retry.js

@@ -1,4 +1,5 @@
-const DEFAULT_MAX_RETRY_DELAY_MS = 60_000;
+import { computeProviderRetryDelay, validateProviderRetryOptions } from "./provider-retry-delay.js";
+import { sleepProviderRetry as abortableSleep, createProviderAbortError as createAbortError, } from "./provider-retry-sleep.js";
 function isProviderError(error) {
     if (!(error instanceof Error) || !("status" in error) || !("headers" in error))
         return false;
@@ -19,50 +20,8 @@ function isRetryableProviderError(error) {
         error.status === 429 ||
         (typeof error.status === "number" && error.status >= 500));
 }
-function validateServerRetryDelayMs(delayMs, maxRetryDelayMs, providerErrorMessage) {
-    const maxDelayMs = maxRetryDelayMs ?? DEFAULT_MAX_RETRY_DELAY_MS;
-    if (maxDelayMs > 0 && delayMs > maxDelayMs) {
-        throw new Error(`Server requested ${Math.ceil(delayMs / 1000)}s retry delay (max: ${Math.ceil(maxDelayMs / 1000)}s). ${providerErrorMessage}`);
-    }
-    return delayMs;
-}
 function getRetryDelayMs(error, retryIndex, maxRetryDelayMs) {
-    const retryAfterMs = error.headers?.get("retry-after-ms");
-    if (retryAfterMs) {
-        const value = Number.parseFloat(retryAfterMs);
-        if (!Number.isNaN(value))
-            return validateServerRetryDelayMs(value, maxRetryDelayMs, error.message);
-    }
-    const retryAfter = error.headers?.get("retry-after");
-    if (retryAfter) {
-        const seconds = Number.parseFloat(retryAfter);
-        const delayMs = Number.isNaN(seconds) ? Date.parse(retryAfter) - Date.now() : seconds * 1000;
-        return validateServerRetryDelayMs(delayMs, maxRetryDelayMs, error.message);
-    }
-    const exponentialDelay = Math.min(0.5 * 2 ** retryIndex, 8) * 1000;
-    return exponentialDelay * (1 - Math.random() * 0.25);
-}
-function createAbortError() {
-    const error = new Error("Request aborted");
-    error.name = "AbortError";
-    return error;
-}
-function abortableSleep(ms, signal) {
-    return new Promise((resolve, reject) => {
-        if (signal?.aborted) {
-            reject(createAbortError());
-            return;
-        }
-        const onAbort = () => {
-            clearTimeout(timeout);
-            reject(createAbortError());
-        };
-        const timeout = setTimeout(() => {
-            signal?.removeEventListener("abort", onAbort);
-            resolve();
-        }, Math.max(0, ms));
-        signal?.addEventListener("abort", onAbort, { once: true });
-    });
+    return computeProviderRetryDelay(error.headers, retryIndex, maxRetryDelayMs, error.message, Date.now(), Math.random());
 }
 /**
  * Reproduce the retry behavior used by the OpenAI and Anthropic SDKs while making
@@ -74,8 +33,11 @@ function abortableSleep(ms, signal) {
  */
 export async function retryProviderRequest(request, options = {}) {
     const maxRetries = options.maxRetries ?? 0;
+    validateProviderRetryOptions(maxRetries, options.maxRetryDelayMs);
     let retriesRemaining = maxRetries;
     for (;;) {
+        if (options.signal?.aborted)
+            throw createAbortError();
         try {
             // Each retry is a fresh SDK request, so X-Stainless-Retry-Count remains zero.
             return await request();

package/dist/utils/provider-retry.js.map

@@ -1 +1 @@
-{"version":3,"file":"provider-retry.js","sourceRoot":"","sources":["../../src/utils/provider-retry.ts"],"names":[],"mappings":"AAAA,MAAM,0BAA0B,GAAG,MAAM,CAAC;AAa1C,SAAS,eAAe,CAAC,KAAc,EAA0B;IAChE,IAAI,CAAC,CAAC,KAAK,YAAY,KAAK,CAAC,IAAI,CAAC,CAAC,QAAQ,IAAI,KAAK,CAAC,IAAI,CAAC,CAAC,SAAS,IAAI,KAAK,CAAC;QAAE,OAAO,KAAK,CAAC;IAC7F,OAAO,CACN,CAAC,KAAK,CAAC,MAAM,KAAK,SAAS,IAAI,OAAO,KAAK,CAAC,MAAM,KAAK,QAAQ,CAAC;QAChE,CAAC,KAAK,CAAC,OAAO,KAAK,SAAS,IAAI,KAAK,CAAC,OAAO,YAAY,OAAO,CAAC,CACjE,CAAC;AAAA,CACF;AAED,gGAAgG;AAChG,SAAS,wBAAwB,CAAC,KAAoB,EAAW;IAChE,MAAM,WAAW,GAAG,KAAK,CAAC,OAAO,EAAE,GAAG,CAAC,gBAAgB,CAAC,CAAC;IACzD,IAAI,WAAW,KAAK,MAAM;QAAE,OAAO,IAAI,CAAC;IACxC,IAAI,WAAW,KAAK,OAAO;QAAE,OAAO,KAAK,CAAC;IAE1C,IAAI,KAAK,CAAC,MAAM,KAAK,SAAS;QAAE,OAAO,IAAI,CAAC;IAC5C,OAAO,CACN,KAAK,CAAC,MAAM,KAAK,GAAG;QACpB,KAAK,CAAC,MAAM,KAAK,GAAG;QACpB,KAAK,CAAC,MAAM,KAAK,GAAG;QACpB,CAAC,OAAO,KAAK,CAAC,MAAM,KAAK,QAAQ,IAAI,KAAK,CAAC,MAAM,IAAI,GAAG,CAAC,CACzD,CAAC;AAAA,CACF;AAED,SAAS,0BAA0B,CAClC,OAAe,EACf,eAAmC,EACnC,oBAA4B,EACnB;IACT,MAAM,UAAU,GAAG,eAAe,IAAI,0BAA0B,CAAC;IACjE,IAAI,UAAU,GAAG,CAAC,IAAI,OAAO,GAAG,UAAU,EAAE,CAAC;QAC5C,MAAM,IAAI,KAAK,CACd,oBAAoB,IAAI,CAAC,IAAI,CAAC,OAAO,GAAG,IAAI,CAAC,uBAAuB,IAAI,CAAC,IAAI,CAAC,UAAU,GAAG,IAAI,CAAC,OAAO,oBAAoB,EAAE,CAC7H,CAAC;IACH,CAAC;IACD,OAAO,OAAO,CAAC;AAAA,CACf;AAED,SAAS,eAAe,CAAC,KAAoB,EAAE,UAAkB,EAAE,eAAmC,EAAU;IAC/G,MAAM,YAAY,GAAG,KAAK,CAAC,OAAO,EAAE,GAAG,CAAC,gBAAgB,CAAC,CAAC;IAC1D,IAAI,YAAY,EAAE,CAAC;QAClB,MAAM,KAAK,GAAG,MAAM,CAAC,UAAU,CAAC,YAAY,CAAC,CAAC;QAC9C,IAAI,CAAC,MAAM,CAAC,KAAK,CAAC,KAAK,CAAC;YAAE,OAAO,0BAA0B,CAAC,KAAK,EAAE,eAAe,EAAE,KAAK,CAAC,OAAO,CAAC,CAAC;IACpG,CAAC;IAED,MAAM,UAAU,GAAG,KAAK,CAAC,OAAO,EAAE,GAAG,CAAC,aAAa,CAAC,CAAC;IACrD,IAAI,UAAU,EAAE,CAAC;QAChB,MAAM,OAAO,GAAG,MAAM,CAAC,UAAU,CAAC,UAAU,CAAC,CAAC;QAC9C,MAAM,OAAO,GAAG,MAAM,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,KAAK,CAAC,UAAU,CAAC,GAAG,IAAI,CAAC,GAAG,EAAE,CAAC,CAAC,CAAC,OAAO,GAAG,IAAI,CAAC;QAC7F,OAAO,0BAA0B,CAAC,OAAO,EAAE,eAAe,EAAE,KAAK,CAAC,OAAO,CAAC,CAAC;IAC5E,CAAC;IAED,MAAM,gBAAgB,GAAG,IAAI,CAAC,GAAG,CAAC,GAAG,GAAG,CAAC,IAAI,UAAU,EAAE,CAAC,CAAC,GAAG,IAAI,CAAC;IACnE,OAAO,gBAAgB,GAAG,CAAC,CAAC,GAAG,IAAI,CAAC,MAAM,EAAE,GAAG,IAAI,CAAC,CAAC;AAAA,CACrD;AAED,SAAS,gBAAgB,GAAU;IAClC,MAAM,KAAK,GAAG,IAAI,KAAK,CAAC,iBAAiB,CAAC,CAAC;IAC3C,KAAK,CAAC,IAAI,GAAG,YAAY,CAAC;IAC1B,OAAO,KAAK,CAAC;AAAA,CACb;AAED,SAAS,cAAc,CAAC,EAAU,EAAE,MAAoB,EAAiB;IACxE,OAAO,IAAI,OAAO,CAAC,CAAC,OAAO,EAAE,MAAM,EAAE,EAAE,CAAC;QACvC,IAAI,MAAM,EAAE,OAAO,EAAE,CAAC;YACrB,MAAM,CAAC,gBAAgB,EAAE,CAAC,CAAC;YAC3B,OAAO;QACR,CAAC;QAED,MAAM,OAAO,GAAG,GAAG,EAAE,CAAC;YACrB,YAAY,CAAC,OAAO,CAAC,CAAC;YACtB,MAAM,CAAC,gBAAgB,EAAE,CAAC,CAAC;QAAA,CAC3B,CAAC;QACF,MAAM,OAAO,GAAG,UAAU,CACzB,GAAG,EAAE,CAAC;YACL,MAAM,EAAE,mBAAmB,CAAC,OAAO,EAAE,OAAO,CAAC,CAAC;YAC9C,OAAO,EAAE,CAAC;QAAA,CACV,EACD,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,EAAE,CAAC,CACf,CAAC;QACF,MAAM,EAAE,gBAAgB,CAAC,OAAO,EAAE,OAAO,EAAE,EAAE,IAAI,EAAE,IAAI,EAAE,CAAC,CAAC;IAAA,CAC3D,CAAC,CAAC;AAAA,CACH;AAED;;;;;;;GAOG;AACH,MAAM,CAAC,KAAK,UAAU,oBAAoB,CACzC,OAAyB,EACzB,OAAO,GAAyB,EAAE,EACrB;IACb,MAAM,UAAU,GAAG,OAAO,CAAC,UAAU,IAAI,CAAC,CAAC;IAC3C,IAAI,gBAAgB,GAAG,UAAU,CAAC;IAElC,SAAS,CAAC;QACT,IAAI,CAAC;YACJ,8EAA8E;YAC9E,OAAO,MAAM,OAAO,EAAE,CAAC;QACxB,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YAChB,IAAI,OAAO,CAAC,MAAM,EAAE,OAAO;gBAAE,MAAM,gBAAgB,EAAE,CAAC;YACtD,IAAI,gBAAgB,IAAI,CAAC,IAAI,CAAC,eAAe,CAAC,KAAK,CAAC,IAAI,CAAC,wBAAwB,CAAC,KAAK,CAAC;gBAAE,MAAM,KAAK,CAAC;YAEtG,MAAM,UAAU,GAAG,UAAU,GAAG,gBAAgB,CAAC;YACjD,gBAAgB,EAAE,CAAC;YACnB,MAAM,cAAc,CAAC,eAAe,CAAC,KAAK,EAAE,UAAU,EAAE,OAAO,CAAC,eAAe,CAAC,EAAE,OAAO,CAAC,MAAM,CAAC,CAAC;QACnG,CAAC;IACF,CAAC;AAAA,CACD","sourcesContent":["const DEFAULT_MAX_RETRY_DELAY_MS = 60_000;\n\ninterface ProviderRetryOptions {\n\tmaxRetries?: number;\n\tmaxRetryDelayMs?: number;\n\tsignal?: AbortSignal;\n}\n\ninterface ProviderError extends Error {\n\tstatus: number | undefined;\n\theaders: Headers | undefined;\n}\n\nfunction isProviderError(error: unknown): error is ProviderError {\n\tif (!(error instanceof Error) || !(\"status\" in error) || !(\"headers\" in error)) return false;\n\treturn (\n\t\t(error.status === undefined || typeof error.status === \"number\") &&\n\t\t(error.headers === undefined || error.headers instanceof Headers)\n\t);\n}\n\n/** Mirrors the pinned OpenAI/Anthropic SDK retry policy; review when either SDK is upgraded. */\nfunction isRetryableProviderError(error: ProviderError): boolean {\n\tconst shouldRetry = error.headers?.get(\"x-should-retry\");\n\tif (shouldRetry === \"true\") return true;\n\tif (shouldRetry === \"false\") return false;\n\n\tif (error.status === undefined) return true;\n\treturn (\n\t\terror.status === 408 ||\n\t\terror.status === 409 ||\n\t\terror.status === 429 ||\n\t\t(typeof error.status === \"number\" && error.status >= 500)\n\t);\n}\n\nfunction validateServerRetryDelayMs(\n\tdelayMs: number,\n\tmaxRetryDelayMs: number | undefined,\n\tproviderErrorMessage: string,\n): number {\n\tconst maxDelayMs = maxRetryDelayMs ?? DEFAULT_MAX_RETRY_DELAY_MS;\n\tif (maxDelayMs > 0 && delayMs > maxDelayMs) {\n\t\tthrow new Error(\n\t\t\t`Server requested ${Math.ceil(delayMs / 1000)}s retry delay (max: ${Math.ceil(maxDelayMs / 1000)}s). ${providerErrorMessage}`,\n\t\t);\n\t}\n\treturn delayMs;\n}\n\nfunction getRetryDelayMs(error: ProviderError, retryIndex: number, maxRetryDelayMs: number | undefined): number {\n\tconst retryAfterMs = error.headers?.get(\"retry-after-ms\");\n\tif (retryAfterMs) {\n\t\tconst value = Number.parseFloat(retryAfterMs);\n\t\tif (!Number.isNaN(value)) return validateServerRetryDelayMs(value, maxRetryDelayMs, error.message);\n\t}\n\n\tconst retryAfter = error.headers?.get(\"retry-after\");\n\tif (retryAfter) {\n\t\tconst seconds = Number.parseFloat(retryAfter);\n\t\tconst delayMs = Number.isNaN(seconds) ? Date.parse(retryAfter) - Date.now() : seconds * 1000;\n\t\treturn validateServerRetryDelayMs(delayMs, maxRetryDelayMs, error.message);\n\t}\n\n\tconst exponentialDelay = Math.min(0.5 * 2 ** retryIndex, 8) * 1000;\n\treturn exponentialDelay * (1 - Math.random() * 0.25);\n}\n\nfunction createAbortError(): Error {\n\tconst error = new Error(\"Request aborted\");\n\terror.name = \"AbortError\";\n\treturn error;\n}\n\nfunction abortableSleep(ms: number, signal?: AbortSignal): Promise<void> {\n\treturn new Promise((resolve, reject) => {\n\t\tif (signal?.aborted) {\n\t\t\treject(createAbortError());\n\t\t\treturn;\n\t\t}\n\n\t\tconst onAbort = () => {\n\t\t\tclearTimeout(timeout);\n\t\t\treject(createAbortError());\n\t\t};\n\t\tconst timeout = setTimeout(\n\t\t\t() => {\n\t\t\t\tsignal?.removeEventListener(\"abort\", onAbort);\n\t\t\t\tresolve();\n\t\t\t},\n\t\t\tMath.max(0, ms),\n\t\t);\n\t\tsignal?.addEventListener(\"abort\", onAbort, { once: true });\n\t});\n}\n\n/**\n * Reproduce the retry behavior used by the OpenAI and Anthropic SDKs while making\n * their backoff sleep interruptible. Their built-in retry timers ignore the\n * request AbortSignal, so callers must invoke the SDK with `maxRetries: 0` and\n * wrap the request with this helper. Provider-requested delays above\n * `maxRetryDelayMs` fail immediately (60 seconds by default); set it to zero to\n * disable the limit.\n */\nexport async function retryProviderRequest<T>(\n\trequest: () => Promise<T>,\n\toptions: ProviderRetryOptions = {},\n): Promise<T> {\n\tconst maxRetries = options.maxRetries ?? 0;\n\tlet retriesRemaining = maxRetries;\n\n\tfor (;;) {\n\t\ttry {\n\t\t\t// Each retry is a fresh SDK request, so X-Stainless-Retry-Count remains zero.\n\t\t\treturn await request();\n\t\t} catch (error) {\n\t\t\tif (options.signal?.aborted) throw createAbortError();\n\t\t\tif (retriesRemaining <= 0 || !isProviderError(error) || !isRetryableProviderError(error)) throw error;\n\n\t\t\tconst retryIndex = maxRetries - retriesRemaining;\n\t\t\tretriesRemaining--;\n\t\t\tawait abortableSleep(getRetryDelayMs(error, retryIndex, options.maxRetryDelayMs), options.signal);\n\t\t}\n\t}\n}\n"]}
+{"version":3,"file":"provider-retry.js","sourceRoot":"","sources":["../../src/utils/provider-retry.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,yBAAyB,EAAE,4BAA4B,EAAE,MAAM,2BAA2B,CAAC;AACpG,OAAO,EACN,kBAAkB,IAAI,cAAc,EACpC,wBAAwB,IAAI,gBAAgB,GAC5C,MAAM,2BAA2B,CAAC;AAanC,SAAS,eAAe,CAAC,KAAc,EAA0B;IAChE,IAAI,CAAC,CAAC,KAAK,YAAY,KAAK,CAAC,IAAI,CAAC,CAAC,QAAQ,IAAI,KAAK,CAAC,IAAI,CAAC,CAAC,SAAS,IAAI,KAAK,CAAC;QAAE,OAAO,KAAK,CAAC;IAC7F,OAAO,CACN,CAAC,KAAK,CAAC,MAAM,KAAK,SAAS,IAAI,OAAO,KAAK,CAAC,MAAM,KAAK,QAAQ,CAAC;QAChE,CAAC,KAAK,CAAC,OAAO,KAAK,SAAS,IAAI,KAAK,CAAC,OAAO,YAAY,OAAO,CAAC,CACjE,CAAC;AAAA,CACF;AAED,gGAAgG;AAChG,SAAS,wBAAwB,CAAC,KAAoB,EAAW;IAChE,MAAM,WAAW,GAAG,KAAK,CAAC,OAAO,EAAE,GAAG,CAAC,gBAAgB,CAAC,CAAC;IACzD,IAAI,WAAW,KAAK,MAAM;QAAE,OAAO,IAAI,CAAC;IACxC,IAAI,WAAW,KAAK,OAAO;QAAE,OAAO,KAAK,CAAC;IAE1C,IAAI,KAAK,CAAC,MAAM,KAAK,SAAS;QAAE,OAAO,IAAI,CAAC;IAC5C,OAAO,CACN,KAAK,CAAC,MAAM,KAAK,GAAG;QACpB,KAAK,CAAC,MAAM,KAAK,GAAG;QACpB,KAAK,CAAC,MAAM,KAAK,GAAG;QACpB,CAAC,OAAO,KAAK,CAAC,MAAM,KAAK,QAAQ,IAAI,KAAK,CAAC,MAAM,IAAI,GAAG,CAAC,CACzD,CAAC;AAAA,CACF;AAED,SAAS,eAAe,CAAC,KAAoB,EAAE,UAAkB,EAAE,eAAmC,EAAU;IAC/G,OAAO,yBAAyB,CAC/B,KAAK,CAAC,OAAO,EACb,UAAU,EACV,eAAe,EACf,KAAK,CAAC,OAAO,EACb,IAAI,CAAC,GAAG,EAAE,EACV,IAAI,CAAC,MAAM,EAAE,CACb,CAAC;AAAA,CACF;AAED;;;;;;;GAOG;AACH,MAAM,CAAC,KAAK,UAAU,oBAAoB,CACzC,OAAyB,EACzB,OAAO,GAAyB,EAAE,EACrB;IACb,MAAM,UAAU,GAAG,OAAO,CAAC,UAAU,IAAI,CAAC,CAAC;IAC3C,4BAA4B,CAAC,UAAU,EAAE,OAAO,CAAC,eAAe,CAAC,CAAC;IAClE,IAAI,gBAAgB,GAAG,UAAU,CAAC;IAElC,SAAS,CAAC;QACT,IAAI,OAAO,CAAC,MAAM,EAAE,OAAO;YAAE,MAAM,gBAAgB,EAAE,CAAC;QACtD,IAAI,CAAC;YACJ,8EAA8E;YAC9E,OAAO,MAAM,OAAO,EAAE,CAAC;QACxB,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YAChB,IAAI,OAAO,CAAC,MAAM,EAAE,OAAO;gBAAE,MAAM,gBAAgB,EAAE,CAAC;YACtD,IAAI,gBAAgB,IAAI,CAAC,IAAI,CAAC,eAAe,CAAC,KAAK,CAAC,IAAI,CAAC,wBAAwB,CAAC,KAAK,CAAC;gBAAE,MAAM,KAAK,CAAC;YAEtG,MAAM,UAAU,GAAG,UAAU,GAAG,gBAAgB,CAAC;YACjD,gBAAgB,EAAE,CAAC;YACnB,MAAM,cAAc,CAAC,eAAe,CAAC,KAAK,EAAE,UAAU,EAAE,OAAO,CAAC,eAAe,CAAC,EAAE,OAAO,CAAC,MAAM,CAAC,CAAC;QACnG,CAAC;IACF,CAAC;AAAA,CACD","sourcesContent":["import { computeProviderRetryDelay, validateProviderRetryOptions } from \"./provider-retry-delay.ts\";\nimport {\n\tsleepProviderRetry as abortableSleep,\n\tcreateProviderAbortError as createAbortError,\n} from \"./provider-retry-sleep.ts\";\n\ninterface ProviderRetryOptions {\n\tmaxRetries?: number;\n\tmaxRetryDelayMs?: number;\n\tsignal?: AbortSignal;\n}\n\ninterface ProviderError extends Error {\n\tstatus: number | undefined;\n\theaders: Headers | undefined;\n}\n\nfunction isProviderError(error: unknown): error is ProviderError {\n\tif (!(error instanceof Error) || !(\"status\" in error) || !(\"headers\" in error)) return false;\n\treturn (\n\t\t(error.status === undefined || typeof error.status === \"number\") &&\n\t\t(error.headers === undefined || error.headers instanceof Headers)\n\t);\n}\n\n/** Mirrors the pinned OpenAI/Anthropic SDK retry policy; review when either SDK is upgraded. */\nfunction isRetryableProviderError(error: ProviderError): boolean {\n\tconst shouldRetry = error.headers?.get(\"x-should-retry\");\n\tif (shouldRetry === \"true\") return true;\n\tif (shouldRetry === \"false\") return false;\n\n\tif (error.status === undefined) return true;\n\treturn (\n\t\terror.status === 408 ||\n\t\terror.status === 409 ||\n\t\terror.status === 429 ||\n\t\t(typeof error.status === \"number\" && error.status >= 500)\n\t);\n}\n\nfunction getRetryDelayMs(error: ProviderError, retryIndex: number, maxRetryDelayMs: number | undefined): number {\n\treturn computeProviderRetryDelay(\n\t\terror.headers,\n\t\tretryIndex,\n\t\tmaxRetryDelayMs,\n\t\terror.message,\n\t\tDate.now(),\n\t\tMath.random(),\n\t);\n}\n\n/**\n * Reproduce the retry behavior used by the OpenAI and Anthropic SDKs while making\n * their backoff sleep interruptible. Their built-in retry timers ignore the\n * request AbortSignal, so callers must invoke the SDK with `maxRetries: 0` and\n * wrap the request with this helper. Provider-requested delays above\n * `maxRetryDelayMs` fail immediately (60 seconds by default); set it to zero to\n * disable the limit.\n */\nexport async function retryProviderRequest<T>(\n\trequest: () => Promise<T>,\n\toptions: ProviderRetryOptions = {},\n): Promise<T> {\n\tconst maxRetries = options.maxRetries ?? 0;\n\tvalidateProviderRetryOptions(maxRetries, options.maxRetryDelayMs);\n\tlet retriesRemaining = maxRetries;\n\n\tfor (;;) {\n\t\tif (options.signal?.aborted) throw createAbortError();\n\t\ttry {\n\t\t\t// Each retry is a fresh SDK request, so X-Stainless-Retry-Count remains zero.\n\t\t\treturn await request();\n\t\t} catch (error) {\n\t\t\tif (options.signal?.aborted) throw createAbortError();\n\t\t\tif (retriesRemaining <= 0 || !isProviderError(error) || !isRetryableProviderError(error)) throw error;\n\n\t\t\tconst retryIndex = maxRetries - retriesRemaining;\n\t\t\tretriesRemaining--;\n\t\t\tawait abortableSleep(getRetryDelayMs(error, retryIndex, options.maxRetryDelayMs), options.signal);\n\t\t}\n\t}\n}\n"]}

package/dist/utils/retry-backoff.d.ts

@@ -0,0 +1,8 @@
+/**
+ * Backoff before 1-indexed retry `attempt`: `baseDelayMs * 2^(attempt-1)`, capped at the longest
+ * delay a Node timer holds, because a longer timer fires after 1 ms. The base converts as the old
+ * arithmetic did; `+Infinity` takes the cap, and a base that converts to NaN or a negative number
+ * uses the 2 s default. A NaN attempt yields the cap rather than an immediate retry.
+ */
+export declare function retryBackoffDelayMs(baseDelayMs: number, attempt: number): number;
+//# sourceMappingURL=retry-backoff.d.ts.map

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

@@ -0,0 +1 @@
+{"version":3,"file":"retry-backoff.d.ts","sourceRoot":"","sources":["../../src/utils/retry-backoff.ts"],"names":[],"mappings":"AAKA;;;;;GAKG;AACH,wBAAgB,mBAAmB,CAAC,WAAW,EAAE,MAAM,EAAE,OAAO,EAAE,MAAM,GAAG,MAAM,CAOhF","sourcesContent":["import { MAX_RETRY_TIMER_MS } from \"./provider-retry-sleep.ts\";\n\n/** The coding-agent `retry.baseDelayMs` default, which {@link RetryPolicy} mirrors. */\nconst DEFAULT_RETRY_BASE_DELAY_MS = 2000;\n\n/**\n * Backoff before 1-indexed retry `attempt`: `baseDelayMs * 2^(attempt-1)`, capped at the longest\n * delay a Node timer holds, because a longer timer fires after 1 ms. The base converts as the old\n * arithmetic did; `+Infinity` takes the cap, and a base that converts to NaN or a negative number\n * uses the 2 s default. A NaN attempt yields the cap rather than an immediate retry.\n */\nexport function retryBackoffDelayMs(baseDelayMs: number, attempt: number): number {\n\tconst requested = Number(baseDelayMs);\n\t// `>= 0` is false for NaN; +Infinity passes and takes the cap.\n\tconst base = requested >= 0 ? Math.min(requested, MAX_RETRY_TIMER_MS) : DEFAULT_RETRY_BASE_DELAY_MS;\n\tif (base === 0) return 0;\n\tconst delay = base * 2 ** (attempt - 1);\n\treturn delay <= MAX_RETRY_TIMER_MS ? delay : MAX_RETRY_TIMER_MS;\n}\n"]}

package/dist/utils/retry-backoff.js

@@ -0,0 +1,19 @@
+import { MAX_RETRY_TIMER_MS } from "./provider-retry-sleep.js";
+/** The coding-agent `retry.baseDelayMs` default, which {@link RetryPolicy} mirrors. */
+const DEFAULT_RETRY_BASE_DELAY_MS = 2000;
+/**
+ * Backoff before 1-indexed retry `attempt`: `baseDelayMs * 2^(attempt-1)`, capped at the longest
+ * delay a Node timer holds, because a longer timer fires after 1 ms. The base converts as the old
+ * arithmetic did; `+Infinity` takes the cap, and a base that converts to NaN or a negative number
+ * uses the 2 s default. A NaN attempt yields the cap rather than an immediate retry.
+ */
+export function retryBackoffDelayMs(baseDelayMs, attempt) {
+    const requested = Number(baseDelayMs);
+    // `>= 0` is false for NaN; +Infinity passes and takes the cap.
+    const base = requested >= 0 ? Math.min(requested, MAX_RETRY_TIMER_MS) : DEFAULT_RETRY_BASE_DELAY_MS;
+    if (base === 0)
+        return 0;
+    const delay = base * 2 ** (attempt - 1);
+    return delay <= MAX_RETRY_TIMER_MS ? delay : MAX_RETRY_TIMER_MS;
+}
+//# sourceMappingURL=retry-backoff.js.map

package/dist/utils/retry-backoff.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"retry-backoff.js","sourceRoot":"","sources":["../../src/utils/retry-backoff.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,kBAAkB,EAAE,MAAM,2BAA2B,CAAC;AAE/D,uFAAuF;AACvF,MAAM,2BAA2B,GAAG,IAAI,CAAC;AAEzC;;;;;GAKG;AACH,MAAM,UAAU,mBAAmB,CAAC,WAAmB,EAAE,OAAe,EAAU;IACjF,MAAM,SAAS,GAAG,MAAM,CAAC,WAAW,CAAC,CAAC;IACtC,+DAA+D;IAC/D,MAAM,IAAI,GAAG,SAAS,IAAI,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,GAAG,CAAC,SAAS,EAAE,kBAAkB,CAAC,CAAC,CAAC,CAAC,2BAA2B,CAAC;IACpG,IAAI,IAAI,KAAK,CAAC;QAAE,OAAO,CAAC,CAAC;IACzB,MAAM,KAAK,GAAG,IAAI,GAAG,CAAC,IAAI,CAAC,OAAO,GAAG,CAAC,CAAC,CAAC;IACxC,OAAO,KAAK,IAAI,kBAAkB,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,kBAAkB,CAAC;AAAA,CAChE","sourcesContent":["import { MAX_RETRY_TIMER_MS } from \"./provider-retry-sleep.ts\";\n\n/** The coding-agent `retry.baseDelayMs` default, which {@link RetryPolicy} mirrors. */\nconst DEFAULT_RETRY_BASE_DELAY_MS = 2000;\n\n/**\n * Backoff before 1-indexed retry `attempt`: `baseDelayMs * 2^(attempt-1)`, capped at the longest\n * delay a Node timer holds, because a longer timer fires after 1 ms. The base converts as the old\n * arithmetic did; `+Infinity` takes the cap, and a base that converts to NaN or a negative number\n * uses the 2 s default. A NaN attempt yields the cap rather than an immediate retry.\n */\nexport function retryBackoffDelayMs(baseDelayMs: number, attempt: number): number {\n\tconst requested = Number(baseDelayMs);\n\t// `>= 0` is false for NaN; +Infinity passes and takes the cap.\n\tconst base = requested >= 0 ? Math.min(requested, MAX_RETRY_TIMER_MS) : DEFAULT_RETRY_BASE_DELAY_MS;\n\tif (base === 0) return 0;\n\tconst delay = base * 2 ** (attempt - 1);\n\treturn delay <= MAX_RETRY_TIMER_MS ? delay : MAX_RETRY_TIMER_MS;\n}\n"]}

package/dist/utils/retry.d.ts

@@ -1,15 +1,23 @@
 import type { AssistantMessage } from "../types.ts";
 /**
- * Retry policy: bounded attempts with exponential backoff (`baseDelayMs * 2^(attempt-1)`).
+ * Retry policy: bounded attempts with exponential backoff (`baseDelayMs * 2^(attempt-1)`, at most
+ * 2,147,483,647 ms).
  * Matches `settings.retry` (`enabled`, `maxRetries`, `baseDelayMs`) in coding-agent; kept
  * here so the classifier and the policy-driven retry loop live together and stay reusable
  * by the SDK and other callers.
  */
 export interface RetryPolicy {
     enabled: boolean;
-    /** Max retry attempts (0 = no retries). The initial call never counts as a retry. */
+    /**
+     * Max retry attempts (0 = no retries). The initial call never counts as a retry. A count that
+     * converts to NaN means no retries.
+     */
     maxRetries: number;
-    /** Base delay in ms. Per-attempt delay is `baseDelayMs * 2^(attempt-1)` before jitter. */
+    /**
+     * Base delay in ms. Per-attempt delay is `baseDelayMs * 2^(attempt-1)`, capped at 2,147,483,647 ms;
+     * no jitter. A base that converts to NaN or a negative number, an omitted one included, uses 2 s;
+     * `+Infinity` takes the cap.
+     */
     baseDelayMs: number;
 }
 /** Optional callbacks emitted by {@link retryAssistantCall} around each retry. */

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;AA6FpD;;;;;GAKG;AACH,MAAM,WAAW,WAAW;IAC3B,OAAO,EAAE,OAAO,CAAC;IACjB,qFAAqF;IACrF,UAAU,EAAE,MAAM,CAAC;IACnB,0FAA0F;IAC1F,WAAW,EAAE,MAAM,CAAC;CACpB;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,CA4C3B;AAED;;;;;;;;GAQG;AACH,wBAAgB,yBAAyB,CAAC,OAAO,EAAE,gBAAgB,GAAG,OAAO,CAK5E","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\"insufficient balance\",\n\t\"out of budget\",\n\t\"quota exceeded\",\n\t// Codex token-plan wording: \"Your token-plan 1-week quota has been exhausted\".\n\t\"quota\\\\s+(?:has been\\\\s+|was\\\\s+|is\\\\s+)?exhaust\",\n\t\"billing\",\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\"rate.?limit\",\n\t\"too many requests\",\n\t\"429\",\n\t\"500\",\n\t\"502\",\n\t\"503\",\n\t\"504\",\n\t\"524\",\n\t\"service.?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).\n\t\"provider.?returned.?error\",\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\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\n/**\n * Retry policy: bounded attempts with exponential backoff (`baseDelayMs * 2^(attempt-1)`).\n * Matches `settings.retry` (`enabled`, `maxRetries`, `baseDelayMs`) 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}\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 = policy!.baseDelayMs * 2 ** (attempt - 1);\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\treturn { ...response, stopReason: \"aborted\", errorMessage: undefined };\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.d.ts","sourceRoot":"","sources":["../../src/utils/retry.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,gBAAgB,EAAE,MAAM,aAAa,CAAC;AA8FpD;;;;;;GAMG;AACH,MAAM,WAAW,WAAW;IAC3B,OAAO,EAAE,OAAO,CAAC;IACjB;;;OAGG;IACH,UAAU,EAAE,MAAM,CAAC;IACnB;;;;OAIG;IACH,WAAW,EAAE,MAAM,CAAC;CACpB;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;AA4BD;;;;;;;;;;;;;;;;;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","sourcesContent":["import type { AssistantMessage } from \"../types.ts\";\nimport { retryBackoffDelayMs } from \"./retry-backoff.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\"insufficient balance\",\n\t\"out of budget\",\n\t\"quota exceeded\",\n\t// Codex token-plan wording: \"Your token-plan 1-week quota has been exhausted\".\n\t\"quota\\\\s+(?:has been\\\\s+|was\\\\s+|is\\\\s+)?exhaust\",\n\t\"billing\",\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\"rate.?limit\",\n\t\"too many requests\",\n\t\"429\",\n\t\"500\",\n\t\"502\",\n\t\"503\",\n\t\"504\",\n\t\"524\",\n\t\"service.?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).\n\t\"provider.?returned.?error\",\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\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\n/**\n * Retry policy: bounded attempts with exponential backoff (`baseDelayMs * 2^(attempt-1)`, at most\n * 2,147,483,647 ms).\n * Matches `settings.retry` (`enabled`, `maxRetries`, `baseDelayMs`) 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/**\n\t * Max retry attempts (0 = no retries). The initial call never counts as a retry. A count that\n\t * converts to NaN means no retries.\n\t */\n\tmaxRetries: number;\n\t/**\n\t * Base delay in ms. Per-attempt delay is `baseDelayMs * 2^(attempt-1)`, capped at 2,147,483,647 ms;\n\t * no jitter. A base that converts to NaN or a negative number, an omitted one included, uses 2 s;\n\t * `+Infinity` takes the cap.\n\t */\n\tbaseDelayMs: number;\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\n/** One timer suffices: {@link retryBackoffDelayMs} never exceeds the longest delay a timer holds. */\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\tlet timeout: ReturnType<typeof setTimeout> | undefined;\n\t\tconst onAbort = (): void => {\n\t\t\tclearTimeout(timeout);\n\t\t\treject(new RetrySleepAbortError());\n\t\t};\n\t\ttimeout = setTimeout(() => {\n\t\t\tsignal?.removeEventListener(\"abort\", onAbort);\n\t\t\tresolve();\n\t\t}, ms);\n\t\tsignal?.addEventListener(\"abort\", onAbort, { once: true });\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\t// `> 0` also rejects a count that converts to NaN, which `attempt >= maxAttempts` never exhausts.\n\tconst maxAttempts = policy?.enabled && policy.maxRetries > 0 ? 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 = retryBackoffDelayMs(policy!.baseDelayMs, 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\treturn { ...response, stopReason: \"aborted\", errorMessage: undefined };\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/retry.js

@@ -1,3 +1,4 @@
+import { retryBackoffDelayMs } from "./retry-backoff.js";
 function buildProviderErrorPattern(patterns) {
     return new RegExp(patterns.join("|"), "i");
 }
@@ -82,17 +83,23 @@ class RetrySleepAbortError extends Error {
         super("Aborted");
     }
 }
+/** One timer suffices: {@link retryBackoffDelayMs} never exceeds the longest delay a timer holds. */
 function sleep(ms, signal) {
     return new Promise((resolve, reject) => {
         if (signal?.aborted) {
             reject(new RetrySleepAbortError());
             return;
         }
-        const timeout = setTimeout(resolve, ms);
-        signal?.addEventListener("abort", () => {
+        let timeout;
+        const onAbort = () => {
             clearTimeout(timeout);
             reject(new RetrySleepAbortError());
-        }, { once: true });
+        };
+        timeout = setTimeout(() => {
+            signal?.removeEventListener("abort", onAbort);
+            resolve();
+        }, ms);
+        signal?.addEventListener("abort", onAbort, { once: true });
     });
 }
 /**
@@ -114,7 +121,8 @@ function sleep(ms, signal) {
  * (equivalent to calling `produce()` directly).
  */
 export async function retryAssistantCall(produce, policy, signal, callbacks) {
-    const maxAttempts = policy?.enabled ? policy.maxRetries : 0;
+    // `> 0` also rejects a count that converts to NaN, which `attempt >= maxAttempts` never exhausts.
+    const maxAttempts = policy?.enabled && policy.maxRetries > 0 ? policy.maxRetries : 0;
     let attempt = 0;
     let lastRetry;
     for (;;) {
@@ -139,7 +147,7 @@ export async function retryAssistantCall(produce, policy, signal, callbacks) {
         }
         attempt++;
         lastRetry = { attempt, errorMessage: response.errorMessage || "Unknown error" };
-        const delayMs = policy.baseDelayMs * 2 ** (attempt - 1);
+        const delayMs = retryBackoffDelayMs(policy.baseDelayMs, attempt);
         await callbacks?.onRetryScheduled?.(attempt, maxAttempts, delayMs, lastRetry.errorMessage);
         // Normalize aborts during retry backoff to the same AssistantMessage shape as
         // provider stream aborts, so callers do not need to care when cancellation happened.