@bastani/pi-ai 0.9.25-alpha.2 → 0.9.25-alpha.4

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

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

package/dist/utils/overflow.js

@@ -25,7 +25,7 @@
  * - DS4: "Prompt has X tokens, but the configured context size is Y tokens"
  * - Cerebras: "400/413 status code (no body)"
  * - Mistral: "Prompt contains X tokens ... too large for model with Y maximum context length"
- * - z.ai: `{"code":"1261","message":"Prompt too long"}` or silent overflow via usage.input > contextWindow
+ * - z.ai: `{"code":"1261","message":"Prompt too long"}`, `{"code":"1261","message":"Prompt exceeds max length"}` (CN endpoint), or silent overflow via usage.input > contextWindow
  * - Xiaomi MiMo: Truncates input to fill contextWindow exactly, then returns finish_reason "length"
  *   with output=0 (no room left to generate). Detected via stopReason "length" + zero output +
  *   input filling the context window.
@@ -34,6 +34,7 @@
  */
 const OVERFLOW_PATTERNS = [
     /prompt (?:is )?too long/i, // Anthropic and z.ai token overflow
+    /prompt exceeds max length/i, // z.ai CN endpoint token overflow
     /request_too_large/i, // Anthropic request byte-size overflow (HTTP 413)
     /input is too long for requested model/i, // Amazon Bedrock
     /exceeds the context window/i, // OpenAI (Completions & Responses API)

package/dist/utils/overflow.js.map

@@ -1 +1 @@
-{"version":3,"file":"overflow.js","sourceRoot":"","sources":["../../src/utils/overflow.ts"],"names":[],"mappings":"AAEA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAiCG;AACH,MAAM,iBAAiB,GAAG;IACzB,0BAA0B,EAAE,oCAAoC;IAChE,oBAAoB,EAAE,kDAAkD;IACxE,wCAAwC,EAAE,iBAAiB;IAC3D,6BAA6B,EAAE,uCAAuC;IACtE,4FAA4F,EAAE,sCAAsC;IACpI,yCAAyC,EAAE,kBAAkB;IAC7D,+BAA+B,EAAE,aAAa;IAC9C,oCAAoC,EAAE,OAAO;IAC7C,uCAAuC,EAAE,6BAA6B;IACtE,kEAAkE,EAAE,sBAAsB;IAC1F,iFAAiF,EAAE,cAAc;IACjG,2BAA2B,EAAE,iBAAiB;IAC9C,qCAAqC,EAAE,mBAAmB;IAC1D,kCAAkC,EAAE,YAAY;IAChD,+BAA+B,EAAE,UAAU;IAC3C,6BAA6B,EAAE,kBAAkB;IACjD,sDAAsD,EAAE,UAAU;IAClE,+EAA+E,EAAE,aAAa;IAC9F,gCAAgC,EAAE,yDAAyD;IAC3F,oDAAoD,EAAE,iCAAiC;IACvF,kCAAkC,EAAE,8BAA8B;IAClE,gCAAgC,EAAE,mBAAmB;IACrD,kBAAkB,EAAE,mBAAmB;IACvC,uBAAuB,EAAE,mBAAmB;CAC5C,CAAC;AAEF,SAAS,0BAA0B,CAAC,YAAoB;IACvD,MAAM,UAAU,GAAG,YAAY,CAAC,IAAI,EAAE,CAAC,WAAW,EAAE,CAAC;IACrD,OAAO,CACN,UAAU,KAAK,2BAA2B;QAC1C,UAAU,KAAK,2BAA2B;QAC1C,UAAU,KAAK,eAAe;QAC9B,UAAU,KAAK,eAAe,CAC9B,CAAC;AACH,CAAC;AAED;;;;;;;;GAQG;AACH,MAAM,qBAAqB,GAAG;IAC7B,2CAA2C,EAAE,oFAAoF;IACjI,aAAa,EAAE,wBAAwB;IACvC,oBAAoB,EAAE,yBAAyB;CAC/C,CAAC;AAEF;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAsDG;AACH,MAAM,UAAU,iBAAiB,CAAC,OAAyB,EAAE,aAAsB;IAClF,uCAAuC;IACvC,IAAI,OAAO,CAAC,UAAU,KAAK,OAAO,IAAI,OAAO,CAAC,YAAY,EAAE,CAAC;QAC5D,oFAAoF;QACpF,MAAM,aAAa,GAAG,qBAAqB,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC,OAAO,CAAC,YAAa,CAAC,CAAC,CAAC;QACvF,IAAI,CAAC,aAAa,EAAE,CAAC;YACpB,IAAI,iBAAiB,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC,OAAO,CAAC,YAAa,CAAC,CAAC,EAAE,CAAC;gBAClE,OAAO,IAAI,CAAC;YACb,CAAC;YACD,IAAI,OAAO,CAAC,QAAQ,KAAK,UAAU,IAAI,0BAA0B,CAAC,OAAO,CAAC,YAAY,CAAC,EAAE,CAAC;gBACzF,OAAO,IAAI,CAAC;YACb,CAAC;QACF,CAAC;IACF,CAAC;IAED,8EAA8E;IAC9E,IAAI,aAAa,IAAI,OAAO,CAAC,UAAU,KAAK,MAAM,EAAE,CAAC;QACpD,MAAM,WAAW,GAAG,OAAO,CAAC,KAAK,CAAC,KAAK,GAAG,OAAO,CAAC,KAAK,CAAC,SAAS,CAAC;QAClE,IAAI,WAAW,GAAG,aAAa,EAAE,CAAC;YACjC,OAAO,IAAI,CAAC;QACb,CAAC;IACF,CAAC;IAED,sFAAsF;IACtF,qFAAqF;IACrF,gEAAgE;IAChE,IAAI,aAAa,IAAI,OAAO,CAAC,UAAU,KAAK,QAAQ,IAAI,OAAO,CAAC,KAAK,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;QACpF,MAAM,WAAW,GAAG,OAAO,CAAC,KAAK,CAAC,KAAK,GAAG,OAAO,CAAC,KAAK,CAAC,SAAS,CAAC;QAClE,IAAI,WAAW,IAAI,aAAa,GAAG,IAAI,EAAE,CAAC;YACzC,OAAO,IAAI,CAAC;QACb,CAAC;IACF,CAAC;IAED,OAAO,KAAK,CAAC;AACd,CAAC;AAED;;;;;GAKG;AACH,MAAM,UAAU,mBAAmB,CAAC,OAAyB,EAAE,gBAAwB;IACtF,OAAO,OAAO,CAAC,UAAU,KAAK,QAAQ,IAAI,gBAAgB,GAAG,CAAC,IAAI,OAAO,CAAC,KAAK,CAAC,MAAM,GAAG,gBAAgB,CAAC;AAC3G,CAAC;AAED;;GAEG;AACH,MAAM,UAAU,mBAAmB;IAClC,OAAO,CAAC,GAAG,iBAAiB,CAAC,CAAC;AAC/B,CAAC","sourcesContent":["import type { AssistantMessage } from \"../types.ts\";\n\n/**\n * Regex patterns to detect context overflow errors from different providers.\n *\n * These patterns match error messages returned when the input exceeds\n * the model's context window.\n *\n * Provider-specific patterns (with example error messages):\n *\n * - Anthropic: \"prompt is too long: 213462 tokens > 200000 maximum\"\n * - Anthropic: \"413 {\\\"error\\\":{\\\"type\\\":\\\"request_too_large\\\",\\\"message\\\":\\\"Request exceeds the maximum size\\\"}}\"\n * - OpenAI: \"Your input exceeds the context window of this model\"\n * - OpenAI/LiteLLM: \"Requested token count exceeds the model's maximum context length of 131072 tokens\"\n * - OpenAI-compatible: \"Input length (265330) exceeds model's maximum context length (262144).\"\n * - Google: \"The input token count (1196265) exceeds the maximum number of tokens allowed (1048575)\"\n * - xAI: \"This model's maximum prompt length is 131072 but the request contains 537812 tokens\"\n * - Groq: \"Please reduce the length of the messages or completion\"\n * - OpenRouter: \"This endpoint's maximum context length is X tokens. However, you requested about Y tokens\"\n * - OpenRouter/Poolside: \"Input length X exceeds the maximum allowed input length of Y tokens.\"\n * - Together AI: \"The input (X tokens) is longer than the model's context length (Y tokens).\"\n * - llama.cpp: \"the request exceeds the available context size, try increasing it\"\n * - LM Studio: \"tokens to keep from the initial prompt is greater than the context length\"\n * - GitHub Copilot: \"prompt token count of X exceeds the limit of Y\"\n * - MiniMax: \"invalid params, context window exceeds limit\"\n * - Kimi For Coding: \"Your request exceeded model token limit: X (requested: Y)\"\n * - DS4: \"Prompt has X tokens, but the configured context size is Y tokens\"\n * - Cerebras: \"400/413 status code (no body)\"\n * - Mistral: \"Prompt contains X tokens ... too large for model with Y maximum context length\"\n * - z.ai: `{\"code\":\"1261\",\"message\":\"Prompt too long\"}` or silent overflow via usage.input > contextWindow\n * - Xiaomi MiMo: Truncates input to fill contextWindow exactly, then returns finish_reason \"length\"\n *   with output=0 (no room left to generate). Detected via stopReason \"length\" + zero output +\n *   input filling the context window.\n * - DashScope/Qwen: \"Range of input length should be [1, X]\" (HTTP 400 invalid_parameter_error)\n * - Ollama: Some deployments truncate silently, others return errors like \"prompt too long; exceeded max context length by X tokens\"\n */\nconst OVERFLOW_PATTERNS = [\n\t/prompt (?:is )?too long/i, // Anthropic and z.ai token overflow\n\t/request_too_large/i, // Anthropic request byte-size overflow (HTTP 413)\n\t/input is too long for requested model/i, // Amazon Bedrock\n\t/exceeds the context window/i, // OpenAI (Completions & Responses API)\n\t/exceeds (?:the )?(?:model'?s )?maximum context length(?: of [\\d,]+ tokens?|\\s*\\([\\d,]+\\))/i, // OpenAI-compatible proxies (LiteLLM)\n\t/input token count.*exceeds the maximum/i, // Google (Gemini)\n\t/maximum prompt length is \\d+/i, // xAI (Grok)\n\t/reduce the length of the messages/i, // Groq\n\t/maximum context length is \\d+ tokens/i, // OpenRouter (most backends)\n\t/exceeds (?:the )?maximum allowed input length of [\\d,]+ tokens?/i, // OpenRouter/Poolside\n\t/input \\(\\d+ tokens\\) is longer than the model'?s context length \\(\\d+ tokens\\)/i, // Together AI\n\t/exceeds the limit of \\d+/i, // GitHub Copilot\n\t/exceeds the available context size/i, // llama.cpp server\n\t/greater than the context length/i, // LM Studio\n\t/context window exceeds limit/i, // MiniMax\n\t/exceeded model token limit/i, // Kimi For Coding\n\t/too large for model with \\d+ maximum context length/i, // Mistral\n\t/prompt has [\\d,]+ tokens?, but the configured context size is [\\d,]+ tokens?/i, // DS4 server\n\t/model_context_window_exceeded/i, // z.ai non-standard finish_reason surfaced as error text\n\t/prompt too long; exceeded (?:max )?context length/i, // Ollama explicit overflow error\n\t/range of input length should be/i, // DashScope / Qwen Token Plan\n\t/context[_ ]length[_ ]exceeded/i, // Generic fallback\n\t/too many tokens/i, // Generic fallback\n\t/token limit exceeded/i, // Generic fallback\n];\n\nfunction isCerebrasBodylessOverflow(errorMessage: string): boolean {\n\tconst normalized = errorMessage.trim().toLowerCase();\n\treturn (\n\t\tnormalized === \"400 status code (no body)\" ||\n\t\tnormalized === \"413 status code (no body)\" ||\n\t\tnormalized === \"400 (no body)\" ||\n\t\tnormalized === \"413 (no body)\"\n\t);\n}\n\n/**\n * Patterns that indicate non-overflow errors (e.g. rate limiting, server errors).\n * Error messages matching any of these are excluded from overflow detection\n * even if they also match an OVERFLOW_PATTERN.\n *\n * Example: Bedrock formats throttling errors as \"ThrottlingException: Too many tokens,\n * please wait before trying again.\" which would match the /too many tokens/i overflow\n * pattern without this exclusion.\n */\nconst NON_OVERFLOW_PATTERNS = [\n\t/^(Throttling error|Service unavailable):/i, // AWS Bedrock non-overflow errors (human-readable prefixes from formatBedrockError)\n\t/rate limit/i, // Generic rate limiting\n\t/too many requests/i, // Generic HTTP 429 style\n];\n\n/**\n * Check if an assistant message represents a context overflow error.\n *\n * This handles three cases:\n * 1. Error-based overflow: Most providers return stopReason \"error\" with a\n *    specific error message pattern.\n * 2. Silent overflow: Some providers accept overflow requests and return\n *    successfully. For these, we check if usage.input exceeds the context window.\n * 3. Length-stop overflow: Xiaomi MiMo can return \"length\" with zero output when\n *    the input fills the context window.\n *\n * ## Reliability by Provider\n *\n * **Reliable detection (returns error with detectable message):**\n * - Anthropic: \"prompt is too long: X tokens > Y maximum\" or \"request_too_large\"\n * - OpenAI (Completions & Responses): \"exceeds the context window\", \"exceeds the model's maximum context length of X tokens\", or \"exceeds model's maximum context length (X)\"\n * - Google Gemini: \"input token count exceeds the maximum\"\n * - xAI (Grok): \"maximum prompt length is X but request contains Y\"\n * - Groq: \"reduce the length of the messages\"\n * - Cerebras: 400/413 status code (no body)\n * - Mistral: \"Prompt contains X tokens ... too large for model with Y maximum context length\"\n * - OpenRouter (most backends): \"maximum context length is X tokens\"\n * - OpenRouter/Poolside: \"Input length X exceeds the maximum allowed input length of Y tokens.\"\n * - Together AI: \"The input (X tokens) is longer than the model's context length (Y tokens).\"\n * - llama.cpp: \"exceeds the available context size\"\n * - LM Studio: \"greater than the context length\"\n * - Kimi For Coding: \"exceeded model token limit: X (requested: Y)\"\n * - DS4: \"Prompt has X tokens, but the configured context size is Y tokens\"\n * - DashScope/Qwen: \"Range of input length should be [1, X]\"\n * - z.ai: \"Prompt too long\"\n *\n * **Unreliable detection:**\n * - z.ai: Sometimes accepts overflow silently (detectable via usage.input > contextWindow),\n *   sometimes returns rate limit errors. Pass contextWindow param to detect silent overflow.\n * - Xiaomi MiMo: Truncates input to fit contextWindow then returns stopReason \"length\" with\n *   output=0. Pass contextWindow param to detect via the \"filled context + zero output\" signal.\n * - Ollama: May truncate input silently for some setups, but may also return explicit\n *   overflow errors that match the patterns above. Silent truncation still cannot be\n *   detected here because we do not know the expected token count.\n *\n * ## Custom Providers\n *\n * If you've added custom models via settings.json, this function may not detect\n * overflow errors from those providers. To add support:\n *\n * 1. Send a request that exceeds the model's context window\n * 2. Check the errorMessage in the response\n * 3. Create a regex pattern that matches the error\n * 4. The pattern should be added to OVERFLOW_PATTERNS in this file, or\n *    check the errorMessage yourself before calling this function\n *\n * @param message - The assistant message to check\n * @param contextWindow - Optional context window size for detecting silent overflow (z.ai)\n * @returns true if the message indicates a context overflow\n */\nexport function isContextOverflow(message: AssistantMessage, contextWindow?: number): boolean {\n\t// Case 1: Check error message patterns\n\tif (message.stopReason === \"error\" && message.errorMessage) {\n\t\t// Skip messages matching known non-overflow patterns (e.g. throttling / rate-limit)\n\t\tconst isNonOverflow = NON_OVERFLOW_PATTERNS.some((p) => p.test(message.errorMessage!));\n\t\tif (!isNonOverflow) {\n\t\t\tif (OVERFLOW_PATTERNS.some((p) => p.test(message.errorMessage!))) {\n\t\t\t\treturn true;\n\t\t\t}\n\t\t\tif (message.provider === \"cerebras\" && isCerebrasBodylessOverflow(message.errorMessage)) {\n\t\t\t\treturn true;\n\t\t\t}\n\t\t}\n\t}\n\n\t// Case 2: Silent overflow (z.ai style) - successful but usage exceeds context\n\tif (contextWindow && message.stopReason === \"stop\") {\n\t\tconst inputTokens = message.usage.input + message.usage.cacheRead;\n\t\tif (inputTokens > contextWindow) {\n\t\t\treturn true;\n\t\t}\n\t}\n\n\t// Case 3: Length-stop overflow (Xiaomi MiMo style) - server truncates oversized input\n\t// to fit the context window, leaving no room for output. Returns stopReason \"length\"\n\t// with output=0 and input+cacheRead filling the context window.\n\tif (contextWindow && message.stopReason === \"length\" && message.usage.output === 0) {\n\t\tconst inputTokens = message.usage.input + message.usage.cacheRead;\n\t\tif (inputTokens >= contextWindow * 0.99) {\n\t\t\treturn true;\n\t\t}\n\t}\n\n\treturn false;\n}\n\n/**\n * Check whether a length stop ended below the caller or model's intended output limit.\n * Such responses may be caused by context pressure or provider-side truncation, so callers\n * can make one bounded compact-and-retry attempt. `desiredMaxOutput` must be the original\n * limit before any context-based clamping.\n */\nexport function isRecoverableLength(message: AssistantMessage, desiredMaxOutput: number): boolean {\n\treturn message.stopReason === \"length\" && desiredMaxOutput > 0 && message.usage.output < desiredMaxOutput;\n}\n\n/**\n * Get the overflow patterns for testing purposes.\n */\nexport function getOverflowPatterns(): RegExp[] {\n\treturn [...OVERFLOW_PATTERNS];\n}\n"]}
+{"version":3,"file":"overflow.js","sourceRoot":"","sources":["../../src/utils/overflow.ts"],"names":[],"mappings":"AAEA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAiCG;AACH,MAAM,iBAAiB,GAAG;IACzB,0BAA0B,EAAE,oCAAoC;IAChE,4BAA4B,EAAE,kCAAkC;IAChE,oBAAoB,EAAE,kDAAkD;IACxE,wCAAwC,EAAE,iBAAiB;IAC3D,6BAA6B,EAAE,uCAAuC;IACtE,4FAA4F,EAAE,sCAAsC;IACpI,yCAAyC,EAAE,kBAAkB;IAC7D,+BAA+B,EAAE,aAAa;IAC9C,oCAAoC,EAAE,OAAO;IAC7C,uCAAuC,EAAE,6BAA6B;IACtE,kEAAkE,EAAE,sBAAsB;IAC1F,iFAAiF,EAAE,cAAc;IACjG,2BAA2B,EAAE,iBAAiB;IAC9C,qCAAqC,EAAE,mBAAmB;IAC1D,kCAAkC,EAAE,YAAY;IAChD,+BAA+B,EAAE,UAAU;IAC3C,6BAA6B,EAAE,kBAAkB;IACjD,sDAAsD,EAAE,UAAU;IAClE,+EAA+E,EAAE,aAAa;IAC9F,gCAAgC,EAAE,yDAAyD;IAC3F,oDAAoD,EAAE,iCAAiC;IACvF,kCAAkC,EAAE,8BAA8B;IAClE,gCAAgC,EAAE,mBAAmB;IACrD,kBAAkB,EAAE,mBAAmB;IACvC,uBAAuB,EAAE,mBAAmB;CAC5C,CAAC;AAEF,SAAS,0BAA0B,CAAC,YAAoB;IACvD,MAAM,UAAU,GAAG,YAAY,CAAC,IAAI,EAAE,CAAC,WAAW,EAAE,CAAC;IACrD,OAAO,CACN,UAAU,KAAK,2BAA2B;QAC1C,UAAU,KAAK,2BAA2B;QAC1C,UAAU,KAAK,eAAe;QAC9B,UAAU,KAAK,eAAe,CAC9B,CAAC;AACH,CAAC;AAED;;;;;;;;GAQG;AACH,MAAM,qBAAqB,GAAG;IAC7B,2CAA2C,EAAE,oFAAoF;IACjI,aAAa,EAAE,wBAAwB;IACvC,oBAAoB,EAAE,yBAAyB;CAC/C,CAAC;AAEF;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAsDG;AACH,MAAM,UAAU,iBAAiB,CAAC,OAAyB,EAAE,aAAsB;IAClF,uCAAuC;IACvC,IAAI,OAAO,CAAC,UAAU,KAAK,OAAO,IAAI,OAAO,CAAC,YAAY,EAAE,CAAC;QAC5D,oFAAoF;QACpF,MAAM,aAAa,GAAG,qBAAqB,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC,OAAO,CAAC,YAAa,CAAC,CAAC,CAAC;QACvF,IAAI,CAAC,aAAa,EAAE,CAAC;YACpB,IAAI,iBAAiB,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC,OAAO,CAAC,YAAa,CAAC,CAAC,EAAE,CAAC;gBAClE,OAAO,IAAI,CAAC;YACb,CAAC;YACD,IAAI,OAAO,CAAC,QAAQ,KAAK,UAAU,IAAI,0BAA0B,CAAC,OAAO,CAAC,YAAY,CAAC,EAAE,CAAC;gBACzF,OAAO,IAAI,CAAC;YACb,CAAC;QACF,CAAC;IACF,CAAC;IAED,8EAA8E;IAC9E,IAAI,aAAa,IAAI,OAAO,CAAC,UAAU,KAAK,MAAM,EAAE,CAAC;QACpD,MAAM,WAAW,GAAG,OAAO,CAAC,KAAK,CAAC,KAAK,GAAG,OAAO,CAAC,KAAK,CAAC,SAAS,CAAC;QAClE,IAAI,WAAW,GAAG,aAAa,EAAE,CAAC;YACjC,OAAO,IAAI,CAAC;QACb,CAAC;IACF,CAAC;IAED,sFAAsF;IACtF,qFAAqF;IACrF,gEAAgE;IAChE,IAAI,aAAa,IAAI,OAAO,CAAC,UAAU,KAAK,QAAQ,IAAI,OAAO,CAAC,KAAK,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;QACpF,MAAM,WAAW,GAAG,OAAO,CAAC,KAAK,CAAC,KAAK,GAAG,OAAO,CAAC,KAAK,CAAC,SAAS,CAAC;QAClE,IAAI,WAAW,IAAI,aAAa,GAAG,IAAI,EAAE,CAAC;YACzC,OAAO,IAAI,CAAC;QACb,CAAC;IACF,CAAC;IAED,OAAO,KAAK,CAAC;AACd,CAAC;AAED;;;;;GAKG;AACH,MAAM,UAAU,mBAAmB,CAAC,OAAyB,EAAE,gBAAwB;IACtF,OAAO,OAAO,CAAC,UAAU,KAAK,QAAQ,IAAI,gBAAgB,GAAG,CAAC,IAAI,OAAO,CAAC,KAAK,CAAC,MAAM,GAAG,gBAAgB,CAAC;AAC3G,CAAC;AAED;;GAEG;AACH,MAAM,UAAU,mBAAmB;IAClC,OAAO,CAAC,GAAG,iBAAiB,CAAC,CAAC;AAC/B,CAAC","sourcesContent":["import type { AssistantMessage } from \"../types.ts\";\n\n/**\n * Regex patterns to detect context overflow errors from different providers.\n *\n * These patterns match error messages returned when the input exceeds\n * the model's context window.\n *\n * Provider-specific patterns (with example error messages):\n *\n * - Anthropic: \"prompt is too long: 213462 tokens > 200000 maximum\"\n * - Anthropic: \"413 {\\\"error\\\":{\\\"type\\\":\\\"request_too_large\\\",\\\"message\\\":\\\"Request exceeds the maximum size\\\"}}\"\n * - OpenAI: \"Your input exceeds the context window of this model\"\n * - OpenAI/LiteLLM: \"Requested token count exceeds the model's maximum context length of 131072 tokens\"\n * - OpenAI-compatible: \"Input length (265330) exceeds model's maximum context length (262144).\"\n * - Google: \"The input token count (1196265) exceeds the maximum number of tokens allowed (1048575)\"\n * - xAI: \"This model's maximum prompt length is 131072 but the request contains 537812 tokens\"\n * - Groq: \"Please reduce the length of the messages or completion\"\n * - OpenRouter: \"This endpoint's maximum context length is X tokens. However, you requested about Y tokens\"\n * - OpenRouter/Poolside: \"Input length X exceeds the maximum allowed input length of Y tokens.\"\n * - Together AI: \"The input (X tokens) is longer than the model's context length (Y tokens).\"\n * - llama.cpp: \"the request exceeds the available context size, try increasing it\"\n * - LM Studio: \"tokens to keep from the initial prompt is greater than the context length\"\n * - GitHub Copilot: \"prompt token count of X exceeds the limit of Y\"\n * - MiniMax: \"invalid params, context window exceeds limit\"\n * - Kimi For Coding: \"Your request exceeded model token limit: X (requested: Y)\"\n * - DS4: \"Prompt has X tokens, but the configured context size is Y tokens\"\n * - Cerebras: \"400/413 status code (no body)\"\n * - Mistral: \"Prompt contains X tokens ... too large for model with Y maximum context length\"\n * - z.ai: `{\"code\":\"1261\",\"message\":\"Prompt too long\"}`, `{\"code\":\"1261\",\"message\":\"Prompt exceeds max length\"}` (CN endpoint), or silent overflow via usage.input > contextWindow\n * - Xiaomi MiMo: Truncates input to fill contextWindow exactly, then returns finish_reason \"length\"\n *   with output=0 (no room left to generate). Detected via stopReason \"length\" + zero output +\n *   input filling the context window.\n * - DashScope/Qwen: \"Range of input length should be [1, X]\" (HTTP 400 invalid_parameter_error)\n * - Ollama: Some deployments truncate silently, others return errors like \"prompt too long; exceeded max context length by X tokens\"\n */\nconst OVERFLOW_PATTERNS = [\n\t/prompt (?:is )?too long/i, // Anthropic and z.ai token overflow\n\t/prompt exceeds max length/i, // z.ai CN endpoint token overflow\n\t/request_too_large/i, // Anthropic request byte-size overflow (HTTP 413)\n\t/input is too long for requested model/i, // Amazon Bedrock\n\t/exceeds the context window/i, // OpenAI (Completions & Responses API)\n\t/exceeds (?:the )?(?:model'?s )?maximum context length(?: of [\\d,]+ tokens?|\\s*\\([\\d,]+\\))/i, // OpenAI-compatible proxies (LiteLLM)\n\t/input token count.*exceeds the maximum/i, // Google (Gemini)\n\t/maximum prompt length is \\d+/i, // xAI (Grok)\n\t/reduce the length of the messages/i, // Groq\n\t/maximum context length is \\d+ tokens/i, // OpenRouter (most backends)\n\t/exceeds (?:the )?maximum allowed input length of [\\d,]+ tokens?/i, // OpenRouter/Poolside\n\t/input \\(\\d+ tokens\\) is longer than the model'?s context length \\(\\d+ tokens\\)/i, // Together AI\n\t/exceeds the limit of \\d+/i, // GitHub Copilot\n\t/exceeds the available context size/i, // llama.cpp server\n\t/greater than the context length/i, // LM Studio\n\t/context window exceeds limit/i, // MiniMax\n\t/exceeded model token limit/i, // Kimi For Coding\n\t/too large for model with \\d+ maximum context length/i, // Mistral\n\t/prompt has [\\d,]+ tokens?, but the configured context size is [\\d,]+ tokens?/i, // DS4 server\n\t/model_context_window_exceeded/i, // z.ai non-standard finish_reason surfaced as error text\n\t/prompt too long; exceeded (?:max )?context length/i, // Ollama explicit overflow error\n\t/range of input length should be/i, // DashScope / Qwen Token Plan\n\t/context[_ ]length[_ ]exceeded/i, // Generic fallback\n\t/too many tokens/i, // Generic fallback\n\t/token limit exceeded/i, // Generic fallback\n];\n\nfunction isCerebrasBodylessOverflow(errorMessage: string): boolean {\n\tconst normalized = errorMessage.trim().toLowerCase();\n\treturn (\n\t\tnormalized === \"400 status code (no body)\" ||\n\t\tnormalized === \"413 status code (no body)\" ||\n\t\tnormalized === \"400 (no body)\" ||\n\t\tnormalized === \"413 (no body)\"\n\t);\n}\n\n/**\n * Patterns that indicate non-overflow errors (e.g. rate limiting, server errors).\n * Error messages matching any of these are excluded from overflow detection\n * even if they also match an OVERFLOW_PATTERN.\n *\n * Example: Bedrock formats throttling errors as \"ThrottlingException: Too many tokens,\n * please wait before trying again.\" which would match the /too many tokens/i overflow\n * pattern without this exclusion.\n */\nconst NON_OVERFLOW_PATTERNS = [\n\t/^(Throttling error|Service unavailable):/i, // AWS Bedrock non-overflow errors (human-readable prefixes from formatBedrockError)\n\t/rate limit/i, // Generic rate limiting\n\t/too many requests/i, // Generic HTTP 429 style\n];\n\n/**\n * Check if an assistant message represents a context overflow error.\n *\n * This handles three cases:\n * 1. Error-based overflow: Most providers return stopReason \"error\" with a\n *    specific error message pattern.\n * 2. Silent overflow: Some providers accept overflow requests and return\n *    successfully. For these, we check if usage.input exceeds the context window.\n * 3. Length-stop overflow: Xiaomi MiMo can return \"length\" with zero output when\n *    the input fills the context window.\n *\n * ## Reliability by Provider\n *\n * **Reliable detection (returns error with detectable message):**\n * - Anthropic: \"prompt is too long: X tokens > Y maximum\" or \"request_too_large\"\n * - OpenAI (Completions & Responses): \"exceeds the context window\", \"exceeds the model's maximum context length of X tokens\", or \"exceeds model's maximum context length (X)\"\n * - Google Gemini: \"input token count exceeds the maximum\"\n * - xAI (Grok): \"maximum prompt length is X but request contains Y\"\n * - Groq: \"reduce the length of the messages\"\n * - Cerebras: 400/413 status code (no body)\n * - Mistral: \"Prompt contains X tokens ... too large for model with Y maximum context length\"\n * - OpenRouter (most backends): \"maximum context length is X tokens\"\n * - OpenRouter/Poolside: \"Input length X exceeds the maximum allowed input length of Y tokens.\"\n * - Together AI: \"The input (X tokens) is longer than the model's context length (Y tokens).\"\n * - llama.cpp: \"exceeds the available context size\"\n * - LM Studio: \"greater than the context length\"\n * - Kimi For Coding: \"exceeded model token limit: X (requested: Y)\"\n * - DS4: \"Prompt has X tokens, but the configured context size is Y tokens\"\n * - DashScope/Qwen: \"Range of input length should be [1, X]\"\n * - z.ai: \"Prompt too long\"\n *\n * **Unreliable detection:**\n * - z.ai: Sometimes accepts overflow silently (detectable via usage.input > contextWindow),\n *   sometimes returns rate limit errors. Pass contextWindow param to detect silent overflow.\n * - Xiaomi MiMo: Truncates input to fit contextWindow then returns stopReason \"length\" with\n *   output=0. Pass contextWindow param to detect via the \"filled context + zero output\" signal.\n * - Ollama: May truncate input silently for some setups, but may also return explicit\n *   overflow errors that match the patterns above. Silent truncation still cannot be\n *   detected here because we do not know the expected token count.\n *\n * ## Custom Providers\n *\n * If you've added custom models via settings.json, this function may not detect\n * overflow errors from those providers. To add support:\n *\n * 1. Send a request that exceeds the model's context window\n * 2. Check the errorMessage in the response\n * 3. Create a regex pattern that matches the error\n * 4. The pattern should be added to OVERFLOW_PATTERNS in this file, or\n *    check the errorMessage yourself before calling this function\n *\n * @param message - The assistant message to check\n * @param contextWindow - Optional context window size for detecting silent overflow (z.ai)\n * @returns true if the message indicates a context overflow\n */\nexport function isContextOverflow(message: AssistantMessage, contextWindow?: number): boolean {\n\t// Case 1: Check error message patterns\n\tif (message.stopReason === \"error\" && message.errorMessage) {\n\t\t// Skip messages matching known non-overflow patterns (e.g. throttling / rate-limit)\n\t\tconst isNonOverflow = NON_OVERFLOW_PATTERNS.some((p) => p.test(message.errorMessage!));\n\t\tif (!isNonOverflow) {\n\t\t\tif (OVERFLOW_PATTERNS.some((p) => p.test(message.errorMessage!))) {\n\t\t\t\treturn true;\n\t\t\t}\n\t\t\tif (message.provider === \"cerebras\" && isCerebrasBodylessOverflow(message.errorMessage)) {\n\t\t\t\treturn true;\n\t\t\t}\n\t\t}\n\t}\n\n\t// Case 2: Silent overflow (z.ai style) - successful but usage exceeds context\n\tif (contextWindow && message.stopReason === \"stop\") {\n\t\tconst inputTokens = message.usage.input + message.usage.cacheRead;\n\t\tif (inputTokens > contextWindow) {\n\t\t\treturn true;\n\t\t}\n\t}\n\n\t// Case 3: Length-stop overflow (Xiaomi MiMo style) - server truncates oversized input\n\t// to fit the context window, leaving no room for output. Returns stopReason \"length\"\n\t// with output=0 and input+cacheRead filling the context window.\n\tif (contextWindow && message.stopReason === \"length\" && message.usage.output === 0) {\n\t\tconst inputTokens = message.usage.input + message.usage.cacheRead;\n\t\tif (inputTokens >= contextWindow * 0.99) {\n\t\t\treturn true;\n\t\t}\n\t}\n\n\treturn false;\n}\n\n/**\n * Check whether a length stop ended below the caller or model's intended output limit.\n * Such responses may be caused by context pressure or provider-side truncation, so callers\n * can make one bounded compact-and-retry attempt. `desiredMaxOutput` must be the original\n * limit before any context-based clamping.\n */\nexport function isRecoverableLength(message: AssistantMessage, desiredMaxOutput: number): boolean {\n\treturn message.stopReason === \"length\" && desiredMaxOutput > 0 && message.usage.output < desiredMaxOutput;\n}\n\n/**\n * Get the overflow patterns for testing purposes.\n */\nexport function getOverflowPatterns(): RegExp[] {\n\treturn [...OVERFLOW_PATTERNS];\n}\n"]}

package/dist/utils/provider-retry.js

@@ -30,14 +30,15 @@ function getRetryDelayMs(error, retryIndex, maxRetryDelayMs) {
     const retryAfterMs = error.headers?.get("retry-after-ms");
     if (retryAfterMs) {
         const value = Number.parseFloat(retryAfterMs);
-        if (!Number.isNaN(value))
+        if (Number.isFinite(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);
+        if (Number.isFinite(delayMs))
+            return validateServerRetryDelayMs(delayMs, maxRetryDelayMs, error.message);
     }
     const exponentialDelay = Math.min(0.5 * 2 ** retryIndex, 8) * 1000;
     return exponentialDelay * (1 - Math.random() * 0.25);

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;IACtC,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;AACH,CAAC;AAED,gGAAgG;AAChG,SAAS,wBAAwB,CAAC,KAAoB;IACrD,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;AACH,CAAC;AAED,SAAS,0BAA0B,CAClC,OAAe,EACf,eAAmC,EACnC,oBAA4B;IAE5B,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;AAChB,CAAC;AAED,SAAS,eAAe,CAAC,KAAoB,EAAE,UAAkB,EAAE,eAAmC;IACrG,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;AACtD,CAAC;AAED,SAAS,gBAAgB;IACxB,MAAM,KAAK,GAAG,IAAI,KAAK,CAAC,iBAAiB,CAAC,CAAC;IAC3C,KAAK,CAAC,IAAI,GAAG,YAAY,CAAC;IAC1B,OAAO,KAAK,CAAC;AACd,CAAC;AAED,SAAS,cAAc,CAAC,EAAU,EAAE,MAAoB;IACvD,OAAO,IAAI,OAAO,CAAC,CAAC,OAAO,EAAE,MAAM,EAAE,EAAE;QACtC,IAAI,MAAM,EAAE,OAAO,EAAE,CAAC;YACrB,MAAM,CAAC,gBAAgB,EAAE,CAAC,CAAC;YAC3B,OAAO;QACR,CAAC;QAED,MAAM,OAAO,GAAG,GAAG,EAAE;YACpB,YAAY,CAAC,OAAO,CAAC,CAAC;YACtB,MAAM,CAAC,gBAAgB,EAAE,CAAC,CAAC;QAC5B,CAAC,CAAC;QACF,MAAM,OAAO,GAAG,UAAU,CACzB,GAAG,EAAE;YACJ,MAAM,EAAE,mBAAmB,CAAC,OAAO,EAAE,OAAO,CAAC,CAAC;YAC9C,OAAO,EAAE,CAAC;QACX,CAAC,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;IAC5D,CAAC,CAAC,CAAC;AACJ,CAAC;AAED;;;;;;;GAOG;AACH,MAAM,CAAC,KAAK,UAAU,oBAAoB,CACzC,OAAyB,EACzB,OAAO,GAAyB,EAAE;IAElC,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;AACF,CAAC","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,MAAM,0BAA0B,GAAG,MAAM,CAAC;AAa1C,SAAS,eAAe,CAAC,KAAc;IACtC,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;AACH,CAAC;AAED,gGAAgG;AAChG,SAAS,wBAAwB,CAAC,KAAoB;IACrD,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;AACH,CAAC;AAED,SAAS,0BAA0B,CAClC,OAAe,EACf,eAAmC,EACnC,oBAA4B;IAE5B,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;AAChB,CAAC;AAED,SAAS,eAAe,CAAC,KAAoB,EAAE,UAAkB,EAAE,eAAmC;IACrG,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,MAAM,CAAC,QAAQ,CAAC,KAAK,CAAC;YAAE,OAAO,0BAA0B,CAAC,KAAK,EAAE,eAAe,EAAE,KAAK,CAAC,OAAO,CAAC,CAAC;IACtG,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,IAAI,MAAM,CAAC,QAAQ,CAAC,OAAO,CAAC;YAAE,OAAO,0BAA0B,CAAC,OAAO,EAAE,eAAe,EAAE,KAAK,CAAC,OAAO,CAAC,CAAC;IAC1G,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;AACtD,CAAC;AAED,SAAS,gBAAgB;IACxB,MAAM,KAAK,GAAG,IAAI,KAAK,CAAC,iBAAiB,CAAC,CAAC;IAC3C,KAAK,CAAC,IAAI,GAAG,YAAY,CAAC;IAC1B,OAAO,KAAK,CAAC;AACd,CAAC;AAED,SAAS,cAAc,CAAC,EAAU,EAAE,MAAoB;IACvD,OAAO,IAAI,OAAO,CAAC,CAAC,OAAO,EAAE,MAAM,EAAE,EAAE;QACtC,IAAI,MAAM,EAAE,OAAO,EAAE,CAAC;YACrB,MAAM,CAAC,gBAAgB,EAAE,CAAC,CAAC;YAC3B,OAAO;QACR,CAAC;QAED,MAAM,OAAO,GAAG,GAAG,EAAE;YACpB,YAAY,CAAC,OAAO,CAAC,CAAC;YACtB,MAAM,CAAC,gBAAgB,EAAE,CAAC,CAAC;QAC5B,CAAC,CAAC;QACF,MAAM,OAAO,GAAG,UAAU,CACzB,GAAG,EAAE;YACJ,MAAM,EAAE,mBAAmB,CAAC,OAAO,EAAE,OAAO,CAAC,CAAC;YAC9C,OAAO,EAAE,CAAC;QACX,CAAC,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;IAC5D,CAAC,CAAC,CAAC;AACJ,CAAC;AAED;;;;;;;GAOG;AACH,MAAM,CAAC,KAAK,UAAU,oBAAoB,CACzC,OAAyB,EACzB,OAAO,GAAyB,EAAE;IAElC,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;AACF,CAAC","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.isFinite(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\tif (Number.isFinite(delayMs)) return 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"]}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@bastani/pi-ai",
-  "version": "0.9.25-alpha.2",
+  "version": "0.9.25-alpha.4",
   "description": "Bastani-branded unified LLM API with automatic model discovery and provider configuration",
   "type": "module",
   "main": "./dist/index.js",
@@ -77,7 +77,7 @@
   "dependencies": {
     "@anthropic-ai/sdk": "0.124.0",
     "@aws-sdk/client-bedrock-runtime": "3.1127.0",
-    "@earendil-works/pi-telemetry": "0.99.1",
+    "@earendil-works/pi-telemetry": "0.99.2",
     "@google/genai": "2.21.0",
     "@smithy/node-http-handler": "4.12.1",
     "http-proxy-agent": "9.1.0",