@earendil-works/pi-ai 0.80.9 → 0.81.0

package/dist/utils/overflow.js.map

@@ -1 +1 @@
-{"version":3,"file":"overflow.js","sourceRoot":"","sources":["../../src/utils/overflow.ts"],"names":[],"mappings":"AAEA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAgCG;AACH,MAAM,iBAAiB,GAAG;IACzB,qBAAqB,EAAE,2BAA2B;IAClD,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,gCAAgC,EAAE,mBAAmB;IACrD,kBAAkB,EAAE,mBAAmB;IACvC,uBAAuB,EAAE,mBAAmB;IAC5C,+CAA+C,EAAE,iCAAiC;CAClF,CAAC;AAEF;;;;;;;;GAQG;AACH,MAAM,qBAAqB,GAAG;IAC7B,2CAA2C,EAAE,oFAAoF;IACjI,aAAa,EAAE,wBAAwB;IACvC,oBAAoB,EAAE,yBAAyB;CAC/C,CAAC;AAEF;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAkDG;AACH,MAAM,UAAU,iBAAiB,CAAC,OAAyB,EAAE,aAAsB,EAAW;IAC7F,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,IAAI,iBAAiB,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC,OAAO,CAAC,YAAa,CAAC,CAAC,EAAE,CAAC;YACpF,OAAO,IAAI,CAAC;QACb,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;AAAA,CACb;AAED;;GAEG;AACH,MAAM,UAAU,mBAAmB,GAAa;IAC/C,OAAO,CAAC,GAAG,iBAAiB,CAAC,CAAC;AAAA,CAC9B","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: Does NOT error, accepts overflow silently - handled 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 * - 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 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/context[_ ]length[_ ]exceeded/i, // Generic fallback\n\t/too many tokens/i, // Generic fallback\n\t/token limit exceeded/i, // Generic fallback\n\t/^4(?:00|13)\\s*(?:status code)?\\s*\\(no body\\)/i, // Cerebras: 400/413 with no body\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 two 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 *\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 *\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 && OVERFLOW_PATTERNS.some((p) => p.test(message.errorMessage!))) {\n\t\t\treturn true;\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 * 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,qBAAqB,EAAE,2BAA2B;IAClD,oBAAoB,EAAE,kDAAkD;IACxE,wCAAwC,EAAE,iBAAiB;IAC3D,6BAA6B,EAAE,uCAAuC;IACtE,4FAA4F,EAAE,sCAAsC;IACpI,yCAAyC,EAAE,kBAAkB;IAC7D,+BAA+B,EAAE,aAAa;IAC9C,oCAAoC,EAAE,OAAO;IAC7C,uCAAuC,EAAE,6BAA6B;IACtE,kEAAkE,EAAE,sBAAsB;IAC1F,iFAAiF,EAAE,cAAc;IACjG,2BAA2B,EAAE,iBAAiB;IAC9C,qCAAqC,EAAE,mBAAmB;IAC1D,kCAAkC,EAAE,YAAY;IAChD,+BAA+B,EAAE,UAAU;IAC3C,6BAA6B,EAAE,kBAAkB;IACjD,sDAAsD,EAAE,UAAU;IAClE,+EAA+E,EAAE,aAAa;IAC9F,gCAAgC,EAAE,yDAAyD;IAC3F,oDAAoD,EAAE,iCAAiC;IACvF,kCAAkC,EAAE,8BAA8B;IAClE,gCAAgC,EAAE,mBAAmB;IACrD,kBAAkB,EAAE,mBAAmB;IACvC,uBAAuB,EAAE,mBAAmB;IAC5C,+CAA+C,EAAE,iCAAiC;CAClF,CAAC;AAEF;;;;;;;;GAQG;AACH,MAAM,qBAAqB,GAAG;IAC7B,2CAA2C,EAAE,oFAAoF;IACjI,aAAa,EAAE,wBAAwB;IACvC,oBAAoB,EAAE,yBAAyB;CAC/C,CAAC;AAEF;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAmDG;AACH,MAAM,UAAU,iBAAiB,CAAC,OAAyB,EAAE,aAAsB,EAAW;IAC7F,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,IAAI,iBAAiB,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC,OAAO,CAAC,YAAa,CAAC,CAAC,EAAE,CAAC;YACpF,OAAO,IAAI,CAAC;QACb,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;AAAA,CACb;AAED;;GAEG;AACH,MAAM,UAAU,mBAAmB,GAAa;IAC/C,OAAO,CAAC,GAAG,iBAAiB,CAAC,CAAC;AAAA,CAC9B","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: Does NOT error, accepts overflow silently - handled 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 token overflow\n\t/request_too_large/i, // Anthropic request byte-size overflow (HTTP 413)\n\t/input is too long for requested model/i, // Amazon Bedrock\n\t/exceeds the context window/i, // OpenAI (Completions & Responses API)\n\t/exceeds (?:the )?(?:model'?s )?maximum context length(?: of [\\d,]+ tokens?|\\s*\\([\\d,]+\\))/i, // OpenAI-compatible proxies (LiteLLM)\n\t/input token count.*exceeds the maximum/i, // Google (Gemini)\n\t/maximum prompt length is \\d+/i, // xAI (Grok)\n\t/reduce the length of the messages/i, // Groq\n\t/maximum context length is \\d+ tokens/i, // OpenRouter (most backends)\n\t/exceeds (?:the )?maximum allowed input length of [\\d,]+ tokens?/i, // OpenRouter/Poolside\n\t/input \\(\\d+ tokens\\) is longer than the model'?s context length \\(\\d+ tokens\\)/i, // Together AI\n\t/exceeds the limit of \\d+/i, // GitHub Copilot\n\t/exceeds the available context size/i, // llama.cpp server\n\t/greater than the context length/i, // LM Studio\n\t/context window exceeds limit/i, // MiniMax\n\t/exceeded model token limit/i, // Kimi For Coding\n\t/too large for model with \\d+ maximum context length/i, // Mistral\n\t/prompt has [\\d,]+ tokens?, but the configured context size is [\\d,]+ tokens?/i, // DS4 server\n\t/model_context_window_exceeded/i, // z.ai non-standard finish_reason surfaced as error text\n\t/prompt too long; exceeded (?:max )?context length/i, // Ollama explicit overflow error\n\t/range of input length should be/i, // DashScope / Qwen Token Plan\n\t/context[_ ]length[_ ]exceeded/i, // Generic fallback\n\t/too many tokens/i, // Generic fallback\n\t/token limit exceeded/i, // Generic fallback\n\t/^4(?:00|13)\\s*(?:status code)?\\s*\\(no body\\)/i, // Cerebras: 400/413 with no body\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 two 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 *\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 *\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 && OVERFLOW_PATTERNS.some((p) => p.test(message.errorMessage!))) {\n\t\t\treturn true;\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 * Get the overflow patterns for testing purposes.\n */\nexport function getOverflowPatterns(): RegExp[] {\n\treturn [...OVERFLOW_PATTERNS];\n}\n"]}

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

@@ -1 +1 @@
-{"version":3,"file":"retry.d.ts","sourceRoot":"","sources":["../../src/utils/retry.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,gBAAgB,EAAE,MAAM,aAAa,CAAC;AAsFpD;;;;;;;;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\"out of budget\",\n\t\"quota exceeded\",\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\"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\"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 * 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;AAuFpD;;;;;;;;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\"out of budget\",\n\t\"quota exceeded\",\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\"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 * 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

@@ -58,6 +58,7 @@ const RETRYABLE_PROVIDER_ERROR_PATTERN = buildProviderErrorPattern([
     // (#4433); Bedrock/Smithy can throw an HTTP/2 no-response error (#3594).
     "ended without",
     "stream ended before message_stop",
+    "stream ended before a terminal response event",
     "http2 request did not get a response",
     // Provider-requested retry delay cap failures should flow through the outer
     // retry policy so callers can surface/abort the backoff (#1123).

package/dist/utils/retry.js.map

@@ -1 +1 @@
-{"version":3,"file":"retry.js","sourceRoot":"","sources":["../../src/utils/retry.ts"],"names":[],"mappings":"AAEA,SAAS,yBAAyB,CAAC,QAA2B,EAAU;IACvE,OAAO,IAAI,MAAM,CAAC,QAAQ,CAAC,IAAI,CAAC,GAAG,CAAC,EAAE,GAAG,CAAC,CAAC;AAAA,CAC3C;AAED,MAAM,0CAA0C,GAAG,yBAAyB,CAAC;IAC5E,8EAA8E;IAC9E,2EAA2E;IAC3E,mBAAmB;IACnB,qBAAqB;IAErB,6EAA6E;IAC7E,yDAAyD;IACzD,6BAA6B;IAC7B,mBAAmB;IAEnB,4EAA4E;IAC5E,4EAA4E;IAC5E,oBAAoB;IACpB,eAAe;IACf,gBAAgB;IAChB,SAAS;CACT,CAAC,CAAC;AAEH,MAAM,gCAAgC,GAAG,yBAAyB,CAAC;IAClE,0EAA0E;IAC1E,YAAY;IACZ,aAAa;IACb,mBAAmB;IACnB,KAAK;IACL,KAAK;IACL,KAAK;IACL,KAAK;IACL,KAAK;IACL,KAAK;IACL,sBAAsB;IACtB,eAAe;IACf,iBAAiB;IAEjB,8EAA8E;IAC9E,+CAA+C;IAC/C,2BAA2B;IAE3B,2EAA2E;IAC3E,2EAA2E;IAC3E,2EAA2E;IAC3E,gBAAgB;IAChB,mBAAmB;IACnB,qBAAqB;IACrB,kBAAkB;IAClB,mBAAmB;IACnB,cAAc;IACd,mBAAmB;IACnB,sBAAsB;IACtB,gBAAgB;IAChB,8BAA8B;IAC9B,YAAY;IACZ,SAAS;IACT,YAAY;IAEZ,+EAA+E;IAC/E,mBAAmB;IACnB,kBAAkB;IAElB,yEAAyE;IACzE,8EAA8E;IAC9E,yEAAyE;IACzE,eAAe;IACf,kCAAkC;IAClC,sCAAsC;IAEtC,4EAA4E;IAC5E,iEAAiE;IACjE,aAAa;IAEb,6EAA6E;IAC7E,6BAA6B;IAC7B,4BAA4B;IAC5B,wBAAwB;IACxB,2BAA2B;IAE3B,yCAAyC;IACzC,mBAAmB;CACnB,CAAC,CAAC;AAEH;;;;;;;;GAQG;AACH,MAAM,UAAU,yBAAyB,CAAC,OAAyB,EAAW;IAC7E,IAAI,OAAO,CAAC,UAAU,KAAK,OAAO,IAAI,CAAC,OAAO,CAAC,YAAY;QAAE,OAAO,KAAK,CAAC;IAC1E,MAAM,YAAY,GAAG,OAAO,CAAC,YAAY,CAAC;IAC1C,IAAI,0CAA0C,CAAC,IAAI,CAAC,YAAY,CAAC;QAAE,OAAO,KAAK,CAAC;IAChF,OAAO,gCAAgC,CAAC,IAAI,CAAC,YAAY,CAAC,CAAC;AAAA,CAC3D","sourcesContent":["import type { AssistantMessage } from \"../types.ts\";\n\nfunction buildProviderErrorPattern(patterns: readonly string[]): RegExp {\n\treturn new RegExp(patterns.join(\"|\"), \"i\");\n}\n\nconst NON_RETRYABLE_PROVIDER_LIMIT_ERROR_PATTERN = buildProviderErrorPattern([\n\t// OpenCode Go/free-tier limits returned as 429 JSON error types by OpenCode's\n\t// Zen API. These are subscription/account limits, not transient throttles.\n\t\"GoUsageLimitError\",\n\t\"FreeUsageLimitError\",\n\n\t// OpenCode Go subscription-limit text asks users to enable available-balance\n\t// usage after rolling/weekly/monthly limits are reached.\n\t\"Monthly usage limit reached\",\n\t\"available balance\",\n\n\t// Generic quota/budget/billing exhaustion. `insufficient_quota` is OpenAI's\n\t// quota/billing error code; the other strings cover common gateway wording.\n\t\"insufficient_quota\",\n\t\"out of budget\",\n\t\"quota exceeded\",\n\t\"billing\",\n]);\n\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\"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\"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 * Classifies whether a failed assistant message looks like a transient provider\n * or transport error, so callers can decide if the last assistant turn should be\n * restarted.\n *\n * This does not implement retry policy. Callers should first handle context\n * overflow separately, then apply their own retry budget, backoff, and reporting\n * before restarting the assistant turn.\n */\nexport function isRetryableAssistantError(message: AssistantMessage): boolean {\n\tif (message.stopReason !== \"error\" || !message.errorMessage) return false;\n\tconst errorMessage = message.errorMessage;\n\tif (NON_RETRYABLE_PROVIDER_LIMIT_ERROR_PATTERN.test(errorMessage)) return false;\n\treturn RETRYABLE_PROVIDER_ERROR_PATTERN.test(errorMessage);\n}\n"]}
+{"version":3,"file":"retry.js","sourceRoot":"","sources":["../../src/utils/retry.ts"],"names":[],"mappings":"AAEA,SAAS,yBAAyB,CAAC,QAA2B,EAAU;IACvE,OAAO,IAAI,MAAM,CAAC,QAAQ,CAAC,IAAI,CAAC,GAAG,CAAC,EAAE,GAAG,CAAC,CAAC;AAAA,CAC3C;AAED,MAAM,0CAA0C,GAAG,yBAAyB,CAAC;IAC5E,8EAA8E;IAC9E,2EAA2E;IAC3E,mBAAmB;IACnB,qBAAqB;IAErB,6EAA6E;IAC7E,yDAAyD;IACzD,6BAA6B;IAC7B,mBAAmB;IAEnB,4EAA4E;IAC5E,4EAA4E;IAC5E,oBAAoB;IACpB,eAAe;IACf,gBAAgB;IAChB,SAAS;CACT,CAAC,CAAC;AAEH,MAAM,gCAAgC,GAAG,yBAAyB,CAAC;IAClE,0EAA0E;IAC1E,YAAY;IACZ,aAAa;IACb,mBAAmB;IACnB,KAAK;IACL,KAAK;IACL,KAAK;IACL,KAAK;IACL,KAAK;IACL,KAAK;IACL,sBAAsB;IACtB,eAAe;IACf,iBAAiB;IAEjB,8EAA8E;IAC9E,+CAA+C;IAC/C,2BAA2B;IAE3B,2EAA2E;IAC3E,2EAA2E;IAC3E,2EAA2E;IAC3E,gBAAgB;IAChB,mBAAmB;IACnB,qBAAqB;IACrB,kBAAkB;IAClB,mBAAmB;IACnB,cAAc;IACd,mBAAmB;IACnB,sBAAsB;IACtB,gBAAgB;IAChB,8BAA8B;IAC9B,YAAY;IACZ,SAAS;IACT,YAAY;IAEZ,+EAA+E;IAC/E,mBAAmB;IACnB,kBAAkB;IAElB,yEAAyE;IACzE,8EAA8E;IAC9E,yEAAyE;IACzE,eAAe;IACf,kCAAkC;IAClC,+CAA+C;IAC/C,sCAAsC;IAEtC,4EAA4E;IAC5E,iEAAiE;IACjE,aAAa;IAEb,6EAA6E;IAC7E,6BAA6B;IAC7B,4BAA4B;IAC5B,wBAAwB;IACxB,2BAA2B;IAE3B,yCAAyC;IACzC,mBAAmB;CACnB,CAAC,CAAC;AAEH;;;;;;;;GAQG;AACH,MAAM,UAAU,yBAAyB,CAAC,OAAyB,EAAW;IAC7E,IAAI,OAAO,CAAC,UAAU,KAAK,OAAO,IAAI,CAAC,OAAO,CAAC,YAAY;QAAE,OAAO,KAAK,CAAC;IAC1E,MAAM,YAAY,GAAG,OAAO,CAAC,YAAY,CAAC;IAC1C,IAAI,0CAA0C,CAAC,IAAI,CAAC,YAAY,CAAC;QAAE,OAAO,KAAK,CAAC;IAChF,OAAO,gCAAgC,CAAC,IAAI,CAAC,YAAY,CAAC,CAAC;AAAA,CAC3D","sourcesContent":["import type { AssistantMessage } from \"../types.ts\";\n\nfunction buildProviderErrorPattern(patterns: readonly string[]): RegExp {\n\treturn new RegExp(patterns.join(\"|\"), \"i\");\n}\n\nconst NON_RETRYABLE_PROVIDER_LIMIT_ERROR_PATTERN = buildProviderErrorPattern([\n\t// OpenCode Go/free-tier limits returned as 429 JSON error types by OpenCode's\n\t// Zen API. These are subscription/account limits, not transient throttles.\n\t\"GoUsageLimitError\",\n\t\"FreeUsageLimitError\",\n\n\t// OpenCode Go subscription-limit text asks users to enable available-balance\n\t// usage after rolling/weekly/monthly limits are reached.\n\t\"Monthly usage limit reached\",\n\t\"available balance\",\n\n\t// Generic quota/budget/billing exhaustion. `insufficient_quota` is OpenAI's\n\t// quota/billing error code; the other strings cover common gateway wording.\n\t\"insufficient_quota\",\n\t\"out of budget\",\n\t\"quota exceeded\",\n\t\"billing\",\n]);\n\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\"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 * 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/text.d.ts

@@ -0,0 +1,6 @@
+import type { ImageContent, TextContent, ThinkingContent, ToolCall } from "../types.ts";
+type Content = TextContent | ImageContent | ThinkingContent | ToolCall;
+/** Extract and join text from message content. */
+export declare function contentText(content: string | readonly Content[], separator?: string): string;
+export {};
+//# sourceMappingURL=text.d.ts.map

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

@@ -0,0 +1 @@
+{"version":3,"file":"text.d.ts","sourceRoot":"","sources":["../../src/utils/text.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,YAAY,EAAE,WAAW,EAAE,eAAe,EAAE,QAAQ,EAAE,MAAM,aAAa,CAAC;AAExF,KAAK,OAAO,GAAG,WAAW,GAAG,YAAY,GAAG,eAAe,GAAG,QAAQ,CAAC;AAEvE,kDAAkD;AAClD,wBAAgB,WAAW,CAAC,OAAO,EAAE,MAAM,GAAG,SAAS,OAAO,EAAE,EAAE,SAAS,SAAO,GAAG,MAAM,CAM1F","sourcesContent":["import type { ImageContent, TextContent, ThinkingContent, ToolCall } from \"../types.ts\";\n\ntype Content = TextContent | ImageContent | ThinkingContent | ToolCall;\n\n/** Extract and join text from message content. */\nexport function contentText(content: string | readonly Content[], separator = \"\\n\"): string {\n\tif (typeof content === \"string\") return content;\n\treturn content\n\t\t.filter((block) => block.type === \"text\")\n\t\t.map((block) => block.text)\n\t\t.join(separator);\n}\n"]}

package/dist/utils/text.js

@@ -0,0 +1,10 @@
+/** Extract and join text from message content. */
+export function contentText(content, separator = "\n") {
+    if (typeof content === "string")
+        return content;
+    return content
+        .filter((block) => block.type === "text")
+        .map((block) => block.text)
+        .join(separator);
+}
+//# sourceMappingURL=text.js.map

package/dist/utils/text.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"text.js","sourceRoot":"","sources":["../../src/utils/text.ts"],"names":[],"mappings":"AAIA,kDAAkD;AAClD,MAAM,UAAU,WAAW,CAAC,OAAoC,EAAE,SAAS,GAAG,IAAI,EAAU;IAC3F,IAAI,OAAO,OAAO,KAAK,QAAQ;QAAE,OAAO,OAAO,CAAC;IAChD,OAAO,OAAO;SACZ,MAAM,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC,KAAK,CAAC,IAAI,KAAK,MAAM,CAAC;SACxC,GAAG,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC,KAAK,CAAC,IAAI,CAAC;SAC1B,IAAI,CAAC,SAAS,CAAC,CAAC;AAAA,CAClB","sourcesContent":["import type { ImageContent, TextContent, ThinkingContent, ToolCall } from \"../types.ts\";\n\ntype Content = TextContent | ImageContent | ThinkingContent | ToolCall;\n\n/** Extract and join text from message content. */\nexport function contentText(content: string | readonly Content[], separator = \"\\n\"): string {\n\tif (typeof content === \"string\") return content;\n\treturn content\n\t\t.filter((block) => block.type === \"text\")\n\t\t.map((block) => block.text)\n\t\t.join(separator);\n}\n"]}

package/dist/utils/uuid.d.ts

@@ -0,0 +1,3 @@
+/** Generate a time-ordered UUIDv7. */
+export declare function uuidv7(): string;
+//# sourceMappingURL=uuid.d.ts.map

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

@@ -0,0 +1 @@
+{"version":3,"file":"uuid.d.ts","sourceRoot":"","sources":["../../src/utils/uuid.ts"],"names":[],"mappings":"AAaA,sCAAsC;AACtC,wBAAgB,MAAM,IAAI,MAAM,CAiC/B","sourcesContent":["let lastTimestamp = -Infinity;\nlet sequence = 0;\n\nfunction fillRandomBytes(bytes: Uint8Array<ArrayBuffer>): void {\n\tif (globalThis.crypto?.getRandomValues) {\n\t\tglobalThis.crypto.getRandomValues(bytes);\n\t\treturn;\n\t}\n\tfor (let i = 0; i < bytes.length; i++) {\n\t\tbytes[i] = Math.floor(Math.random() * 256);\n\t}\n}\n\n/** Generate a time-ordered UUIDv7. */\nexport function uuidv7(): string {\n\tconst random = new Uint8Array(16);\n\tfillRandomBytes(random);\n\tconst timestamp = Date.now();\n\n\tif (timestamp > lastTimestamp) {\n\t\tsequence = random[6] * 0x1000000 + random[7] * 0x10000 + random[8] * 0x100 + random[9];\n\t\tlastTimestamp = timestamp;\n\t} else {\n\t\tsequence = (sequence + 1) >>> 0;\n\t\tif (sequence === 0) lastTimestamp++;\n\t}\n\n\tconst bytes = new Uint8Array(16);\n\tbytes[0] = (lastTimestamp / 0x10000000000) & 0xff;\n\tbytes[1] = (lastTimestamp / 0x100000000) & 0xff;\n\tbytes[2] = (lastTimestamp / 0x1000000) & 0xff;\n\tbytes[3] = (lastTimestamp / 0x10000) & 0xff;\n\tbytes[4] = (lastTimestamp / 0x100) & 0xff;\n\tbytes[5] = lastTimestamp & 0xff;\n\tbytes[6] = 0x70 | ((sequence >>> 28) & 0x0f);\n\tbytes[7] = (sequence >>> 20) & 0xff;\n\tbytes[8] = 0x80 | ((sequence >>> 14) & 0x3f);\n\tbytes[9] = (sequence >>> 6) & 0xff;\n\tbytes[10] = ((sequence & 0x3f) << 2) | (random[10] & 0x03);\n\tbytes[11] = random[11];\n\tbytes[12] = random[12];\n\tbytes[13] = random[13];\n\tbytes[14] = random[14];\n\tbytes[15] = random[15];\n\n\tconst hex = Array.from(bytes, (byte) => byte.toString(16).padStart(2, \"0\"));\n\treturn `${hex.slice(0, 4).join(\"\")}-${hex.slice(4, 6).join(\"\")}-${hex.slice(6, 8).join(\"\")}-${hex.slice(8, 10).join(\"\")}-${hex.slice(10, 16).join(\"\")}`;\n}\n"]}

package/dist/utils/uuid.js

@@ -0,0 +1,46 @@
+let lastTimestamp = -Infinity;
+let sequence = 0;
+function fillRandomBytes(bytes) {
+    if (globalThis.crypto?.getRandomValues) {
+        globalThis.crypto.getRandomValues(bytes);
+        return;
+    }
+    for (let i = 0; i < bytes.length; i++) {
+        bytes[i] = Math.floor(Math.random() * 256);
+    }
+}
+/** Generate a time-ordered UUIDv7. */
+export function uuidv7() {
+    const random = new Uint8Array(16);
+    fillRandomBytes(random);
+    const timestamp = Date.now();
+    if (timestamp > lastTimestamp) {
+        sequence = random[6] * 0x1000000 + random[7] * 0x10000 + random[8] * 0x100 + random[9];
+        lastTimestamp = timestamp;
+    }
+    else {
+        sequence = (sequence + 1) >>> 0;
+        if (sequence === 0)
+            lastTimestamp++;
+    }
+    const bytes = new Uint8Array(16);
+    bytes[0] = (lastTimestamp / 0x10000000000) & 0xff;
+    bytes[1] = (lastTimestamp / 0x100000000) & 0xff;
+    bytes[2] = (lastTimestamp / 0x1000000) & 0xff;
+    bytes[3] = (lastTimestamp / 0x10000) & 0xff;
+    bytes[4] = (lastTimestamp / 0x100) & 0xff;
+    bytes[5] = lastTimestamp & 0xff;
+    bytes[6] = 0x70 | ((sequence >>> 28) & 0x0f);
+    bytes[7] = (sequence >>> 20) & 0xff;
+    bytes[8] = 0x80 | ((sequence >>> 14) & 0x3f);
+    bytes[9] = (sequence >>> 6) & 0xff;
+    bytes[10] = ((sequence & 0x3f) << 2) | (random[10] & 0x03);
+    bytes[11] = random[11];
+    bytes[12] = random[12];
+    bytes[13] = random[13];
+    bytes[14] = random[14];
+    bytes[15] = random[15];
+    const hex = Array.from(bytes, (byte) => byte.toString(16).padStart(2, "0"));
+    return `${hex.slice(0, 4).join("")}-${hex.slice(4, 6).join("")}-${hex.slice(6, 8).join("")}-${hex.slice(8, 10).join("")}-${hex.slice(10, 16).join("")}`;
+}
+//# sourceMappingURL=uuid.js.map

package/dist/utils/uuid.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"uuid.js","sourceRoot":"","sources":["../../src/utils/uuid.ts"],"names":[],"mappings":"AAAA,IAAI,aAAa,GAAG,CAAC,QAAQ,CAAC;AAC9B,IAAI,QAAQ,GAAG,CAAC,CAAC;AAEjB,SAAS,eAAe,CAAC,KAA8B,EAAQ;IAC9D,IAAI,UAAU,CAAC,MAAM,EAAE,eAAe,EAAE,CAAC;QACxC,UAAU,CAAC,MAAM,CAAC,eAAe,CAAC,KAAK,CAAC,CAAC;QACzC,OAAO;IACR,CAAC;IACD,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,KAAK,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;QACvC,KAAK,CAAC,CAAC,CAAC,GAAG,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,MAAM,EAAE,GAAG,GAAG,CAAC,CAAC;IAC5C,CAAC;AAAA,CACD;AAED,sCAAsC;AACtC,MAAM,UAAU,MAAM,GAAW;IAChC,MAAM,MAAM,GAAG,IAAI,UAAU,CAAC,EAAE,CAAC,CAAC;IAClC,eAAe,CAAC,MAAM,CAAC,CAAC;IACxB,MAAM,SAAS,GAAG,IAAI,CAAC,GAAG,EAAE,CAAC;IAE7B,IAAI,SAAS,GAAG,aAAa,EAAE,CAAC;QAC/B,QAAQ,GAAG,MAAM,CAAC,CAAC,CAAC,GAAG,SAAS,GAAG,MAAM,CAAC,CAAC,CAAC,GAAG,OAAO,GAAG,MAAM,CAAC,CAAC,CAAC,GAAG,KAAK,GAAG,MAAM,CAAC,CAAC,CAAC,CAAC;QACvF,aAAa,GAAG,SAAS,CAAC;IAC3B,CAAC;SAAM,CAAC;QACP,QAAQ,GAAG,CAAC,QAAQ,GAAG,CAAC,CAAC,KAAK,CAAC,CAAC;QAChC,IAAI,QAAQ,KAAK,CAAC;YAAE,aAAa,EAAE,CAAC;IACrC,CAAC;IAED,MAAM,KAAK,GAAG,IAAI,UAAU,CAAC,EAAE,CAAC,CAAC;IACjC,KAAK,CAAC,CAAC,CAAC,GAAG,CAAC,aAAa,GAAG,aAAa,CAAC,GAAG,IAAI,CAAC;IAClD,KAAK,CAAC,CAAC,CAAC,GAAG,CAAC,aAAa,GAAG,WAAW,CAAC,GAAG,IAAI,CAAC;IAChD,KAAK,CAAC,CAAC,CAAC,GAAG,CAAC,aAAa,GAAG,SAAS,CAAC,GAAG,IAAI,CAAC;IAC9C,KAAK,CAAC,CAAC,CAAC,GAAG,CAAC,aAAa,GAAG,OAAO,CAAC,GAAG,IAAI,CAAC;IAC5C,KAAK,CAAC,CAAC,CAAC,GAAG,CAAC,aAAa,GAAG,KAAK,CAAC,GAAG,IAAI,CAAC;IAC1C,KAAK,CAAC,CAAC,CAAC,GAAG,aAAa,GAAG,IAAI,CAAC;IAChC,KAAK,CAAC,CAAC,CAAC,GAAG,IAAI,GAAG,CAAC,CAAC,QAAQ,KAAK,EAAE,CAAC,GAAG,IAAI,CAAC,CAAC;IAC7C,KAAK,CAAC,CAAC,CAAC,GAAG,CAAC,QAAQ,KAAK,EAAE,CAAC,GAAG,IAAI,CAAC;IACpC,KAAK,CAAC,CAAC,CAAC,GAAG,IAAI,GAAG,CAAC,CAAC,QAAQ,KAAK,EAAE,CAAC,GAAG,IAAI,CAAC,CAAC;IAC7C,KAAK,CAAC,CAAC,CAAC,GAAG,CAAC,QAAQ,KAAK,CAAC,CAAC,GAAG,IAAI,CAAC;IACnC,KAAK,CAAC,EAAE,CAAC,GAAG,CAAC,CAAC,QAAQ,GAAG,IAAI,CAAC,IAAI,CAAC,CAAC,GAAG,CAAC,MAAM,CAAC,EAAE,CAAC,GAAG,IAAI,CAAC,CAAC;IAC3D,KAAK,CAAC,EAAE,CAAC,GAAG,MAAM,CAAC,EAAE,CAAC,CAAC;IACvB,KAAK,CAAC,EAAE,CAAC,GAAG,MAAM,CAAC,EAAE,CAAC,CAAC;IACvB,KAAK,CAAC,EAAE,CAAC,GAAG,MAAM,CAAC,EAAE,CAAC,CAAC;IACvB,KAAK,CAAC,EAAE,CAAC,GAAG,MAAM,CAAC,EAAE,CAAC,CAAC;IACvB,KAAK,CAAC,EAAE,CAAC,GAAG,MAAM,CAAC,EAAE,CAAC,CAAC;IAEvB,MAAM,GAAG,GAAG,KAAK,CAAC,IAAI,CAAC,KAAK,EAAE,CAAC,IAAI,EAAE,EAAE,CAAC,IAAI,CAAC,QAAQ,CAAC,EAAE,CAAC,CAAC,QAAQ,CAAC,CAAC,EAAE,GAAG,CAAC,CAAC,CAAC;IAC5E,OAAO,GAAG,GAAG,CAAC,KAAK,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC,EAAE,CAAC,IAAI,GAAG,CAAC,KAAK,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC,EAAE,CAAC,IAAI,GAAG,CAAC,KAAK,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC,EAAE,CAAC,IAAI,GAAG,CAAC,KAAK,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,IAAI,CAAC,EAAE,CAAC,IAAI,GAAG,CAAC,KAAK,CAAC,EAAE,EAAE,EAAE,CAAC,CAAC,IAAI,CAAC,EAAE,CAAC,EAAE,CAAC;AAAA,CACxJ","sourcesContent":["let lastTimestamp = -Infinity;\nlet sequence = 0;\n\nfunction fillRandomBytes(bytes: Uint8Array<ArrayBuffer>): void {\n\tif (globalThis.crypto?.getRandomValues) {\n\t\tglobalThis.crypto.getRandomValues(bytes);\n\t\treturn;\n\t}\n\tfor (let i = 0; i < bytes.length; i++) {\n\t\tbytes[i] = Math.floor(Math.random() * 256);\n\t}\n}\n\n/** Generate a time-ordered UUIDv7. */\nexport function uuidv7(): string {\n\tconst random = new Uint8Array(16);\n\tfillRandomBytes(random);\n\tconst timestamp = Date.now();\n\n\tif (timestamp > lastTimestamp) {\n\t\tsequence = random[6] * 0x1000000 + random[7] * 0x10000 + random[8] * 0x100 + random[9];\n\t\tlastTimestamp = timestamp;\n\t} else {\n\t\tsequence = (sequence + 1) >>> 0;\n\t\tif (sequence === 0) lastTimestamp++;\n\t}\n\n\tconst bytes = new Uint8Array(16);\n\tbytes[0] = (lastTimestamp / 0x10000000000) & 0xff;\n\tbytes[1] = (lastTimestamp / 0x100000000) & 0xff;\n\tbytes[2] = (lastTimestamp / 0x1000000) & 0xff;\n\tbytes[3] = (lastTimestamp / 0x10000) & 0xff;\n\tbytes[4] = (lastTimestamp / 0x100) & 0xff;\n\tbytes[5] = lastTimestamp & 0xff;\n\tbytes[6] = 0x70 | ((sequence >>> 28) & 0x0f);\n\tbytes[7] = (sequence >>> 20) & 0xff;\n\tbytes[8] = 0x80 | ((sequence >>> 14) & 0x3f);\n\tbytes[9] = (sequence >>> 6) & 0xff;\n\tbytes[10] = ((sequence & 0x3f) << 2) | (random[10] & 0x03);\n\tbytes[11] = random[11];\n\tbytes[12] = random[12];\n\tbytes[13] = random[13];\n\tbytes[14] = random[14];\n\tbytes[15] = random[15];\n\n\tconst hex = Array.from(bytes, (byte) => byte.toString(16).padStart(2, \"0\"));\n\treturn `${hex.slice(0, 4).join(\"\")}-${hex.slice(4, 6).join(\"\")}-${hex.slice(6, 8).join(\"\")}-${hex.slice(8, 10).join(\"\")}-${hex.slice(10, 16).join(\"\")}`;\n}\n"]}

package/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@earendil-works/pi-ai",
-	"version": "0.80.9",
+	"version": "0.81.0",
 	"description": "Unified LLM API with automatic model discovery and provider configuration",
 	"type": "module",
 	"main": "./dist/index.js",
@@ -41,7 +41,7 @@
 		}
 	},
 	"bin": {
-		"pi-ai": "./dist/cli.js"
+		"pi-ai": "dist/cli.js"
 	},
 	"files": [
 		"dist",
@@ -49,10 +49,13 @@
 	],
 	"scripts": {
 		"clean": "shx rm -rf dist",
-		"generate-models": "node scripts/generate-models.ts",
+		"generate-models": "node scripts/generate-models.ts --strict",
+		"hydrate-model-data": "node scripts/generate-models.ts --strict --data-only",
 		"generate-model-catalog": "node scripts/generate-models.ts --strict --json-only --json-output ../../.artifacts/model-catalog",
-		"generate-image-models": "node scripts/generate-image-models.ts",
-		"build": "npm run generate-models && npm run generate-image-models && tsgo -p tsconfig.build.json",
+		"generate-image-models": "node scripts/generate-image-models.ts --strict",
+		"check:model-data": "node scripts/check-model-data.ts",
+		"build": "npm run generate-models && npm run build:offline",
+		"build:offline": "npm run check:model-data && tsgo -p tsconfig.build.json && shx rm -rf dist/providers/data && shx cp -r src/providers/data dist/providers/data",
 		"test": "vitest --run",
 		"prepublishOnly": "npm run clean && npm run build"
 	},