@code-yeongyu/senpi 2026.7.23 → 2026.7.25-2

package/node_modules/@earendil-works/pi-ai/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;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,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;AACd,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: 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"]}
+{"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;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,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;AACd,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: 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"]}

package/node_modules/@earendil-works/pi-ai/dist/utils/provider-retry.d.ts

@@ -0,0 +1,16 @@
+interface ProviderRetryOptions {
+    maxRetries?: number;
+    maxRetryDelayMs?: number;
+    signal?: AbortSignal;
+}
+/**
+ * Reproduce the retry behavior used by the OpenAI and Anthropic SDKs while making
+ * their backoff sleep interruptible. Their built-in retry timers ignore the
+ * request AbortSignal, so callers must invoke the SDK with `maxRetries: 0` and
+ * wrap the request with this helper. Provider-requested delays above
+ * `maxRetryDelayMs` fail immediately (60 seconds by default); set it to zero to
+ * disable the limit.
+ */
+export declare function retryProviderRequest<T>(request: () => Promise<T>, options?: ProviderRetryOptions): Promise<T>;
+export {};
+//# sourceMappingURL=provider-retry.d.ts.map

package/node_modules/@earendil-works/pi-ai/dist/utils/provider-retry.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"provider-retry.d.ts","sourceRoot":"","sources":["../../src/utils/provider-retry.ts"],"names":[],"mappings":"AAEA,UAAU,oBAAoB;IAC7B,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,eAAe,CAAC,EAAE,MAAM,CAAC;IACzB,MAAM,CAAC,EAAE,WAAW,CAAC;CACrB;AA0FD;;;;;;;GAOG;AACH,wBAAsB,oBAAoB,CAAC,CAAC,EAC3C,OAAO,EAAE,MAAM,OAAO,CAAC,CAAC,CAAC,EACzB,OAAO,GAAE,oBAAyB,GAChC,OAAO,CAAC,CAAC,CAAC,CAiBZ"}

package/node_modules/@earendil-works/pi-ai/dist/utils/provider-retry.js

@@ -0,0 +1,94 @@
+const DEFAULT_MAX_RETRY_DELAY_MS = 60_000;
+function isProviderError(error) {
+    if (!(error instanceof Error) || !("status" in error) || !("headers" in error))
+        return false;
+    return ((error.status === undefined || typeof error.status === "number") &&
+        (error.headers === undefined || error.headers instanceof Headers));
+}
+/** Mirrors the pinned OpenAI/Anthropic SDK retry policy; review when either SDK is upgraded. */
+function isRetryableProviderError(error) {
+    const shouldRetry = error.headers?.get("x-should-retry");
+    if (shouldRetry === "true")
+        return true;
+    if (shouldRetry === "false")
+        return false;
+    if (error.status === undefined)
+        return true;
+    return (error.status === 408 ||
+        error.status === 409 ||
+        error.status === 429 ||
+        (typeof error.status === "number" && error.status >= 500));
+}
+function validateServerRetryDelayMs(delayMs, maxRetryDelayMs, providerErrorMessage) {
+    const maxDelayMs = maxRetryDelayMs ?? DEFAULT_MAX_RETRY_DELAY_MS;
+    if (maxDelayMs > 0 && delayMs > maxDelayMs) {
+        throw new Error(`Server requested ${Math.ceil(delayMs / 1000)}s retry delay (max: ${Math.ceil(maxDelayMs / 1000)}s). ${providerErrorMessage}`);
+    }
+    return delayMs;
+}
+function getRetryDelayMs(error, retryIndex, maxRetryDelayMs) {
+    const retryAfterMs = error.headers?.get("retry-after-ms");
+    if (retryAfterMs) {
+        const value = Number.parseFloat(retryAfterMs);
+        if (!Number.isNaN(value))
+            return validateServerRetryDelayMs(value, maxRetryDelayMs, error.message);
+    }
+    const retryAfter = error.headers?.get("retry-after");
+    if (retryAfter) {
+        const seconds = Number.parseFloat(retryAfter);
+        const delayMs = Number.isNaN(seconds) ? Date.parse(retryAfter) - Date.now() : seconds * 1000;
+        return validateServerRetryDelayMs(delayMs, maxRetryDelayMs, error.message);
+    }
+    const exponentialDelay = Math.min(0.5 * 2 ** retryIndex, 8) * 1000;
+    return exponentialDelay * (1 - Math.random() * 0.25);
+}
+function createAbortError() {
+    const error = new Error("Request aborted");
+    error.name = "AbortError";
+    return error;
+}
+function abortableSleep(ms, signal) {
+    return new Promise((resolve, reject) => {
+        if (signal?.aborted) {
+            reject(createAbortError());
+            return;
+        }
+        const onAbort = () => {
+            clearTimeout(timeout);
+            reject(createAbortError());
+        };
+        const timeout = setTimeout(() => {
+            signal?.removeEventListener("abort", onAbort);
+            resolve();
+        }, Math.max(0, ms));
+        signal?.addEventListener("abort", onAbort, { once: true });
+    });
+}
+/**
+ * Reproduce the retry behavior used by the OpenAI and Anthropic SDKs while making
+ * their backoff sleep interruptible. Their built-in retry timers ignore the
+ * request AbortSignal, so callers must invoke the SDK with `maxRetries: 0` and
+ * wrap the request with this helper. Provider-requested delays above
+ * `maxRetryDelayMs` fail immediately (60 seconds by default); set it to zero to
+ * disable the limit.
+ */
+export async function retryProviderRequest(request, options = {}) {
+    const maxRetries = options.maxRetries ?? 0;
+    let retriesRemaining = maxRetries;
+    for (;;) {
+        try {
+            // Each retry is a fresh SDK request, so X-Stainless-Retry-Count remains zero.
+            return await request();
+        }
+        catch (error) {
+            if (options.signal?.aborted)
+                throw createAbortError();
+            if (retriesRemaining <= 0 || !isProviderError(error) || !isRetryableProviderError(error))
+                throw error;
+            const retryIndex = maxRetries - retriesRemaining;
+            retriesRemaining--;
+            await abortableSleep(getRetryDelayMs(error, retryIndex, options.maxRetryDelayMs), options.signal);
+        }
+    }
+}
+//# sourceMappingURL=provider-retry.js.map

package/node_modules/@earendil-works/pi-ai/dist/utils/provider-retry.js.map

@@ -0,0 +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"]}

package/node_modules/@earendil-works/pi-ai/dist/utils/retry.d.ts

@@ -1,4 +1,45 @@
 import type { AssistantMessage } from "../types.ts";
+/**
+ * Retry policy: bounded attempts with exponential backoff (`baseDelayMs * 2^(attempt-1)`).
+ * Matches `settings.retry` (`enabled`, `maxRetries`, `baseDelayMs`) in coding-agent; kept
+ * here so the classifier and the policy-driven retry loop live together and stay reusable
+ * by the SDK and other callers.
+ */
+export interface RetryPolicy {
+    enabled: boolean;
+    /** Max retry attempts (0 = no retries). The initial call never counts as a retry. */
+    maxRetries: number;
+    /** Base delay in ms. Per-attempt delay is `baseDelayMs * 2^(attempt-1)` before jitter. */
+    baseDelayMs: number;
+}
+/** Optional callbacks emitted by {@link retryAssistantCall} around each retry. */
+export interface RetryCallbacks {
+    /** Emitted before the backoff sleep of each retry attempt (1-indexed). */
+    onRetryScheduled?: (attempt: number, maxAttempts: number, delayMs: number, errorMessage: string) => void | Promise<void>;
+    /** Emitted after the backoff sleep, immediately before the retried call starts. */
+    onRetryAttemptStart?: () => void | Promise<void>;
+    /** Emitted once when the loop ends: success if a later call completed normally. */
+    onRetryFinished?: (success: boolean, attempt: number, finalError?: string) => void | Promise<void>;
+}
+/**
+ * Run a single assistant-producing call with bounded retry on transient errors.
+ *
+ * Behavior:
+ * - A successful response is returned immediately. Aborts are terminal and never
+ *   retried, but reported as unsuccessful if they happen after a retry was scheduled.
+ *   Aborts during the backoff sleep are normalized to an aborted `AssistantMessage`
+ *   too, so callers do not need to care when cancellation happened.
+ * - A non-retryable error (per {@link isRetryableAssistantError}, including quota/
+ *   billing exhaustion) is returned immediately so deterministic errors fail fast.
+ * - Otherwise retries up to `maxRetries` times with exponential backoff, emitting
+ *   `onRetryScheduled` before each sleep, `onRetryAttemptStart` after each sleep before
+ *   the retried call starts, and `onRetryFinished` once at the end (whether the loop
+ *   ends in success, exhausted retries, or an aborted backoff).
+ *
+ * When `policy` is undefined or disabled, the first response is returned unchanged
+ * (equivalent to calling `produce()` directly).
+ */
+export declare function retryAssistantCall(produce: () => Promise<AssistantMessage>, policy: RetryPolicy | undefined, signal: AbortSignal | undefined, callbacks?: RetryCallbacks): Promise<AssistantMessage>;
 /**
  * Classifies whether a failed assistant message looks like a transient provider
  * or transport error, so callers can decide if the last assistant turn should be

package/node_modules/@earendil-works/pi-ai/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;AAuFpD;;;;;;;;GAQG;AACH,wBAAgB,yBAAyB,CAAC,OAAO,EAAE,gBAAgB,GAAG,OAAO,CAY5E"}
+{"version":3,"file":"retry.d.ts","sourceRoot":"","sources":["../../src/utils/retry.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,gBAAgB,EAAE,MAAM,aAAa,CAAC;AA0FpD;;;;;GAKG;AACH,MAAM,WAAW,WAAW;IAC3B,OAAO,EAAE,OAAO,CAAC;IACjB,qFAAqF;IACrF,UAAU,EAAE,MAAM,CAAC;IACnB,0FAA0F;IAC1F,WAAW,EAAE,MAAM,CAAC;CACpB;AAED,kFAAkF;AAClF,MAAM,WAAW,cAAc;IAC9B,0EAA0E;IAC1E,gBAAgB,CAAC,EAAE,CAClB,OAAO,EAAE,MAAM,EACf,WAAW,EAAE,MAAM,EACnB,OAAO,EAAE,MAAM,EACf,YAAY,EAAE,MAAM,KAChB,IAAI,GAAG,OAAO,CAAC,IAAI,CAAC,CAAC;IAC1B,mFAAmF;IACnF,mBAAmB,CAAC,EAAE,MAAM,IAAI,GAAG,OAAO,CAAC,IAAI,CAAC,CAAC;IACjD,mFAAmF;IACnF,eAAe,CAAC,EAAE,CAAC,OAAO,EAAE,OAAO,EAAE,OAAO,EAAE,MAAM,EAAE,UAAU,CAAC,EAAE,MAAM,KAAK,IAAI,GAAG,OAAO,CAAC,IAAI,CAAC,CAAC;CACnG;AA0BD;;;;;;;;;;;;;;;;;GAiBG;AACH,wBAAsB,kBAAkB,CACvC,OAAO,EAAE,MAAM,OAAO,CAAC,gBAAgB,CAAC,EACxC,MAAM,EAAE,WAAW,GAAG,SAAS,EAC/B,MAAM,EAAE,WAAW,GAAG,SAAS,EAC/B,SAAS,CAAC,EAAE,cAAc,GACxB,OAAO,CAAC,gBAAgB,CAAC,CA4C3B;AAED;;;;;;;;GAQG;AACH,wBAAgB,yBAAyB,CAAC,OAAO,EAAE,gBAAgB,GAAG,OAAO,CAY5E"}

package/node_modules/@earendil-works/pi-ai/dist/utils/retry.js

@@ -43,6 +43,9 @@ const RETRYABLE_PROVIDER_ERROR_PATTERN = buildProviderErrorPattern([
     "connection.?lost",
     "other side closed",
     "fetch failed",
+    "getaddrinfo",
+    "ENOTFOUND",
+    "EAI_AGAIN",
     "upstream.?connect",
     "reset before headers",
     "socket hang up",
@@ -71,6 +74,85 @@ const RETRYABLE_PROVIDER_ERROR_PATTERN = buildProviderErrorPattern([
     // gRPC based providers (e.g. NVIDIA NIM)
     "ResourceExhausted",
 ]);
+class RetrySleepAbortError extends Error {
+    constructor() {
+        super("Aborted");
+    }
+}
+function sleep(ms, signal) {
+    return new Promise((resolve, reject) => {
+        if (signal?.aborted) {
+            reject(new RetrySleepAbortError());
+            return;
+        }
+        const timeout = setTimeout(resolve, ms);
+        signal?.addEventListener("abort", () => {
+            clearTimeout(timeout);
+            reject(new RetrySleepAbortError());
+        }, { once: true });
+    });
+}
+/**
+ * Run a single assistant-producing call with bounded retry on transient errors.
+ *
+ * Behavior:
+ * - A successful response is returned immediately. Aborts are terminal and never
+ *   retried, but reported as unsuccessful if they happen after a retry was scheduled.
+ *   Aborts during the backoff sleep are normalized to an aborted `AssistantMessage`
+ *   too, so callers do not need to care when cancellation happened.
+ * - A non-retryable error (per {@link isRetryableAssistantError}, including quota/
+ *   billing exhaustion) is returned immediately so deterministic errors fail fast.
+ * - Otherwise retries up to `maxRetries` times with exponential backoff, emitting
+ *   `onRetryScheduled` before each sleep, `onRetryAttemptStart` after each sleep before
+ *   the retried call starts, and `onRetryFinished` once at the end (whether the loop
+ *   ends in success, exhausted retries, or an aborted backoff).
+ *
+ * When `policy` is undefined or disabled, the first response is returned unchanged
+ * (equivalent to calling `produce()` directly).
+ */
+export async function retryAssistantCall(produce, policy, signal, callbacks) {
+    const maxAttempts = policy?.enabled ? policy.maxRetries : 0;
+    let attempt = 0;
+    let lastRetry;
+    for (;;) {
+        const response = await produce();
+        // Abort: terminal but not successful. Never retry an aborted message.
+        if (response.stopReason === "aborted") {
+            if (lastRetry)
+                await callbacks?.onRetryFinished?.(false, lastRetry.attempt);
+            return response;
+        }
+        // Success: non-error, non-abort responses return as-is.
+        if (response.stopReason !== "error") {
+            if (lastRetry)
+                await callbacks?.onRetryFinished?.(true, lastRetry.attempt);
+            return response;
+        }
+        // Non-retryable, or budget exhausted: return the final error message.
+        if (attempt >= maxAttempts || !isRetryableAssistantError(response)) {
+            if (lastRetry)
+                await callbacks?.onRetryFinished?.(false, lastRetry.attempt, response.errorMessage);
+            return response;
+        }
+        attempt++;
+        lastRetry = { attempt, errorMessage: response.errorMessage || "Unknown error" };
+        const delayMs = policy.baseDelayMs * 2 ** (attempt - 1);
+        await callbacks?.onRetryScheduled?.(attempt, maxAttempts, delayMs, lastRetry.errorMessage);
+        // Normalize aborts during retry backoff to the same AssistantMessage shape as
+        // provider stream aborts, so callers do not need to care when cancellation happened.
+        try {
+            await sleep(delayMs, signal);
+        }
+        catch (error) {
+            await callbacks?.onRetryFinished?.(false, attempt, lastRetry.errorMessage);
+            if (error instanceof RetrySleepAbortError) {
+                return { ...response, stopReason: "aborted", errorMessage: undefined };
+            }
+            throw error;
+        }
+        await callbacks?.onRetryAttemptStart?.();
+    }
+}
 /**
  * Classifies whether a failed assistant message looks like a transient provider
  * or transport error, so callers can decide if the last assistant turn should be

package/node_modules/@earendil-works/pi-ai/dist/utils/retry.js.map

@@ -1 +1 @@
-{"version":3,"file":"retry.js","sourceRoot":"","sources":["../../src/utils/retry.ts"],"names":[],"mappings":"AAEA,SAAS,yBAAyB,CAAC,QAA2B;IAC7D,OAAO,IAAI,MAAM,CAAC,QAAQ,CAAC,IAAI,CAAC,GAAG,CAAC,EAAE,GAAG,CAAC,CAAC;AAC5C,CAAC;AAED,MAAM,0CAA0C,GAAG,yBAAyB,CAAC;IAC5E,8EAA8E;IAC9E,2EAA2E;IAC3E,mBAAmB;IACnB,qBAAqB;IAErB,6EAA6E;IAC7E,yDAAyD;IACzD,6BAA6B;IAC7B,mBAAmB;IAEnB,4EAA4E;IAC5E,4EAA4E;IAC5E,oBAAoB;IACpB,eAAe;IACf,gBAAgB;IAChB,SAAS;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;IAClE,IACC,OAAO,CAAC,UAAU,KAAK,OAAO;QAC9B,OAAO,CAAC,WAAW,EAAE,IAAI,KAAK,SAAS;QACvC,OAAO,CAAC,WAAW,EAAE,IAAI,KAAK,WAAW;QACzC,CAAC,OAAO,CAAC,YAAY,EACpB,CAAC;QACF,OAAO,KAAK,CAAC;IACd,CAAC;IACD,MAAM,YAAY,GAAG,OAAO,CAAC,YAAY,CAAC;IAC1C,IAAI,0CAA0C,CAAC,IAAI,CAAC,YAAY,CAAC;QAAE,OAAO,KAAK,CAAC;IAChF,OAAO,gCAAgC,CAAC,IAAI,CAAC,YAAY,CAAC,CAAC;AAC5D,CAAC","sourcesContent":["import type { AssistantMessage } from \"../types.ts\";\n\nfunction buildProviderErrorPattern(patterns: readonly string[]): RegExp {\n\treturn new RegExp(patterns.join(\"|\"), \"i\");\n}\n\nconst NON_RETRYABLE_PROVIDER_LIMIT_ERROR_PATTERN = buildProviderErrorPattern([\n\t// OpenCode Go/free-tier limits returned as 429 JSON error types by OpenCode's\n\t// Zen API. These are subscription/account limits, not transient throttles.\n\t\"GoUsageLimitError\",\n\t\"FreeUsageLimitError\",\n\n\t// OpenCode Go subscription-limit text asks users to enable available-balance\n\t// usage after rolling/weekly/monthly limits are reached.\n\t\"Monthly usage limit reached\",\n\t\"available balance\",\n\n\t// Generic quota/budget/billing exhaustion. `insufficient_quota` is OpenAI's\n\t// quota/billing error code; the other strings cover common gateway wording.\n\t\"insufficient_quota\",\n\t\"out of budget\",\n\t\"quota exceeded\",\n\t\"billing\",\n]);\n\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 (\n\t\tmessage.stopReason !== \"error\" ||\n\t\tmessage.stopDetails?.type === \"refusal\" ||\n\t\tmessage.stopDetails?.type === \"sensitive\" ||\n\t\t!message.errorMessage\n\t) {\n\t\treturn false;\n\t}\n\tconst errorMessage = message.errorMessage;\n\tif (NON_RETRYABLE_PROVIDER_LIMIT_ERROR_PATTERN.test(errorMessage)) return false;\n\treturn RETRYABLE_PROVIDER_ERROR_PATTERN.test(errorMessage);\n}\n"]}
+{"version":3,"file":"retry.js","sourceRoot":"","sources":["../../src/utils/retry.ts"],"names":[],"mappings":"AAEA,SAAS,yBAAyB,CAAC,QAA2B;IAC7D,OAAO,IAAI,MAAM,CAAC,QAAQ,CAAC,IAAI,CAAC,GAAG,CAAC,EAAE,GAAG,CAAC,CAAC;AAC5C,CAAC;AAED,MAAM,0CAA0C,GAAG,yBAAyB,CAAC;IAC5E,8EAA8E;IAC9E,2EAA2E;IAC3E,mBAAmB;IACnB,qBAAqB;IAErB,6EAA6E;IAC7E,yDAAyD;IACzD,6BAA6B;IAC7B,mBAAmB;IAEnB,4EAA4E;IAC5E,4EAA4E;IAC5E,oBAAoB;IACpB,eAAe;IACf,gBAAgB;IAChB,SAAS;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,aAAa;IACb,WAAW;IACX,WAAW;IACX,mBAAmB;IACnB,sBAAsB;IACtB,gBAAgB;IAChB,8BAA8B;IAC9B,YAAY;IACZ,SAAS;IACT,YAAY;IAEZ,+EAA+E;IAC/E,mBAAmB;IACnB,kBAAkB;IAElB,yEAAyE;IACzE,8EAA8E;IAC9E,yEAAyE;IACzE,eAAe;IACf,kCAAkC;IAClC,+CAA+C;IAC/C,sCAAsC;IAEtC,4EAA4E;IAC5E,iEAAiE;IACjE,aAAa;IAEb,6EAA6E;IAC7E,6BAA6B;IAC7B,4BAA4B;IAC5B,wBAAwB;IACxB,2BAA2B;IAE3B,yCAAyC;IACzC,mBAAmB;CACnB,CAAC,CAAC;AA+BH,MAAM,oBAAqB,SAAQ,KAAK;IACvC;QACC,KAAK,CAAC,SAAS,CAAC,CAAC;IAClB,CAAC;CACD;AAED,SAAS,KAAK,CAAC,EAAU,EAAE,MAAoB;IAC9C,OAAO,IAAI,OAAO,CAAC,CAAC,OAAO,EAAE,MAAM,EAAE,EAAE;QACtC,IAAI,MAAM,EAAE,OAAO,EAAE,CAAC;YACrB,MAAM,CAAC,IAAI,oBAAoB,EAAE,CAAC,CAAC;YACnC,OAAO;QACR,CAAC;QACD,MAAM,OAAO,GAAG,UAAU,CAAC,OAAO,EAAE,EAAE,CAAC,CAAC;QACxC,MAAM,EAAE,gBAAgB,CACvB,OAAO,EACP,GAAG,EAAE;YACJ,YAAY,CAAC,OAAO,CAAC,CAAC;YACtB,MAAM,CAAC,IAAI,oBAAoB,EAAE,CAAC,CAAC;QACpC,CAAC,EACD,EAAE,IAAI,EAAE,IAAI,EAAE,CACd,CAAC;IACH,CAAC,CAAC,CAAC;AACJ,CAAC;AAED;;;;;;;;;;;;;;;;;GAiBG;AACH,MAAM,CAAC,KAAK,UAAU,kBAAkB,CACvC,OAAwC,EACxC,MAA+B,EAC/B,MAA+B,EAC/B,SAA0B;IAE1B,MAAM,WAAW,GAAG,MAAM,EAAE,OAAO,CAAC,CAAC,CAAC,MAAM,CAAC,UAAU,CAAC,CAAC,CAAC,CAAC,CAAC;IAE5D,IAAI,OAAO,GAAG,CAAC,CAAC;IAChB,IAAI,SAAgE,CAAC;IACrE,SAAS,CAAC;QACT,MAAM,QAAQ,GAAG,MAAM,OAAO,EAAE,CAAC;QAEjC,sEAAsE;QACtE,IAAI,QAAQ,CAAC,UAAU,KAAK,SAAS,EAAE,CAAC;YACvC,IAAI,SAAS;gBAAE,MAAM,SAAS,EAAE,eAAe,EAAE,CAAC,KAAK,EAAE,SAAS,CAAC,OAAO,CAAC,CAAC;YAC5E,OAAO,QAAQ,CAAC;QACjB,CAAC;QAED,wDAAwD;QACxD,IAAI,QAAQ,CAAC,UAAU,KAAK,OAAO,EAAE,CAAC;YACrC,IAAI,SAAS;gBAAE,MAAM,SAAS,EAAE,eAAe,EAAE,CAAC,IAAI,EAAE,SAAS,CAAC,OAAO,CAAC,CAAC;YAC3E,OAAO,QAAQ,CAAC;QACjB,CAAC;QAED,sEAAsE;QACtE,IAAI,OAAO,IAAI,WAAW,IAAI,CAAC,yBAAyB,CAAC,QAAQ,CAAC,EAAE,CAAC;YACpE,IAAI,SAAS;gBAAE,MAAM,SAAS,EAAE,eAAe,EAAE,CAAC,KAAK,EAAE,SAAS,CAAC,OAAO,EAAE,QAAQ,CAAC,YAAY,CAAC,CAAC;YACnG,OAAO,QAAQ,CAAC;QACjB,CAAC;QAED,OAAO,EAAE,CAAC;QACV,SAAS,GAAG,EAAE,OAAO,EAAE,YAAY,EAAE,QAAQ,CAAC,YAAY,IAAI,eAAe,EAAE,CAAC;QAChF,MAAM,OAAO,GAAG,MAAO,CAAC,WAAW,GAAG,CAAC,IAAI,CAAC,OAAO,GAAG,CAAC,CAAC,CAAC;QACzD,MAAM,SAAS,EAAE,gBAAgB,EAAE,CAAC,OAAO,EAAE,WAAW,EAAE,OAAO,EAAE,SAAS,CAAC,YAAY,CAAC,CAAC;QAE3F,8EAA8E;QAC9E,qFAAqF;QACrF,IAAI,CAAC;YACJ,MAAM,KAAK,CAAC,OAAO,EAAE,MAAM,CAAC,CAAC;QAC9B,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YAChB,MAAM,SAAS,EAAE,eAAe,EAAE,CAAC,KAAK,EAAE,OAAO,EAAE,SAAS,CAAC,YAAY,CAAC,CAAC;YAC3E,IAAI,KAAK,YAAY,oBAAoB,EAAE,CAAC;gBAC3C,OAAO,EAAE,GAAG,QAAQ,EAAE,UAAU,EAAE,SAAS,EAAE,YAAY,EAAE,SAAS,EAAE,CAAC;YACxE,CAAC;YACD,MAAM,KAAK,CAAC;QACb,CAAC;QACD,MAAM,SAAS,EAAE,mBAAmB,EAAE,EAAE,CAAC;IAC1C,CAAC;AACF,CAAC;AAED;;;;;;;;GAQG;AACH,MAAM,UAAU,yBAAyB,CAAC,OAAyB;IAClE,IACC,OAAO,CAAC,UAAU,KAAK,OAAO;QAC9B,OAAO,CAAC,WAAW,EAAE,IAAI,KAAK,SAAS;QACvC,OAAO,CAAC,WAAW,EAAE,IAAI,KAAK,WAAW;QACzC,CAAC,OAAO,CAAC,YAAY,EACpB,CAAC;QACF,OAAO,KAAK,CAAC;IACd,CAAC;IACD,MAAM,YAAY,GAAG,OAAO,CAAC,YAAY,CAAC;IAC1C,IAAI,0CAA0C,CAAC,IAAI,CAAC,YAAY,CAAC;QAAE,OAAO,KAAK,CAAC;IAChF,OAAO,gCAAgC,CAAC,IAAI,CAAC,YAAY,CAAC,CAAC;AAC5D,CAAC","sourcesContent":["import type { AssistantMessage } from \"../types.ts\";\n\nfunction buildProviderErrorPattern(patterns: readonly string[]): RegExp {\n\treturn new RegExp(patterns.join(\"|\"), \"i\");\n}\n\nconst NON_RETRYABLE_PROVIDER_LIMIT_ERROR_PATTERN = buildProviderErrorPattern([\n\t// OpenCode Go/free-tier limits returned as 429 JSON error types by OpenCode's\n\t// Zen API. These are subscription/account limits, not transient throttles.\n\t\"GoUsageLimitError\",\n\t\"FreeUsageLimitError\",\n\n\t// OpenCode Go subscription-limit text asks users to enable available-balance\n\t// usage after rolling/weekly/monthly limits are reached.\n\t\"Monthly usage limit reached\",\n\t\"available balance\",\n\n\t// Generic quota/budget/billing exhaustion. `insufficient_quota` is OpenAI's\n\t// quota/billing error code; the other strings cover common gateway wording.\n\t\"insufficient_quota\",\n\t\"out of budget\",\n\t\"quota exceeded\",\n\t\"billing\",\n]);\n\nconst RETRYABLE_PROVIDER_ERROR_PATTERN = buildProviderErrorPattern([\n\t// Generic provider load, HTTP status, and server-side transient failures.\n\t\"overloaded\",\n\t\"rate.?limit\",\n\t\"too many requests\",\n\t\"429\",\n\t\"500\",\n\t\"502\",\n\t\"503\",\n\t\"504\",\n\t\"524\",\n\t\"service.?unavailable\",\n\t\"server.?error\",\n\t\"internal.?error\",\n\n\t// Wrapper/provider text for transient upstream failures, including OpenRouter\n\t// \"Provider returned error\" responses (#2264).\n\t\"provider.?returned.?error\",\n\n\t// Network, proxy, and fetch transport failures. This includes OpenAI Codex\n\t// raw-fetch failures such as \"upstream connect\", \"connection refused\", and\n\t// \"reset before headers\" (#733), plus OpenRouter connection drops (#3317).\n\t\"network.?error\",\n\t\"connection.?error\",\n\t\"connection.?refused\",\n\t\"connection.?lost\",\n\t\"other side closed\",\n\t\"fetch failed\",\n\t\"getaddrinfo\",\n\t\"ENOTFOUND\",\n\t\"EAI_AGAIN\",\n\t\"upstream.?connect\",\n\t\"reset before headers\",\n\t\"socket hang up\",\n\t\"socket connection was closed\",\n\t\"timed? out\",\n\t\"timeout\",\n\t\"terminated\",\n\n\t// WebSocket transports can report close/error text instead of HTTP/fetch text.\n\t\"websocket.?closed\",\n\t\"websocket.?error\",\n\n\t// Premature stream endings from SDKs and transports. Anthropic can throw\n\t// \"stream ended without ...\" and \"Anthropic stream ended before message_stop\"\n\t// (#4433); Bedrock/Smithy can throw an HTTP/2 no-response error (#3594).\n\t\"ended without\",\n\t\"stream ended before message_stop\",\n\t\"stream ended before a terminal response event\",\n\t\"http2 request did not get a response\",\n\n\t// Provider-requested retry delay cap failures should flow through the outer\n\t// retry policy so callers can surface/abort the backoff (#1123).\n\t\"retry delay\",\n\n\t// Explicit retry guidance emitted mid-stream by OpenAI Responses and Bedrock\n\t// stream exceptions (#6019).\n\t\"you can retry your request\",\n\t\"try your request again\",\n\t\"please retry your request\",\n\n\t// gRPC based providers (e.g. NVIDIA NIM)\n\t\"ResourceExhausted\",\n]);\n\n/**\n * Retry policy: bounded attempts with exponential backoff (`baseDelayMs * 2^(attempt-1)`).\n * Matches `settings.retry` (`enabled`, `maxRetries`, `baseDelayMs`) in coding-agent; kept\n * here so the classifier and the policy-driven retry loop live together and stay reusable\n * by the SDK and other callers.\n */\nexport interface RetryPolicy {\n\tenabled: boolean;\n\t/** Max retry attempts (0 = no retries). The initial call never counts as a retry. */\n\tmaxRetries: number;\n\t/** Base delay in ms. Per-attempt delay is `baseDelayMs * 2^(attempt-1)` before jitter. */\n\tbaseDelayMs: number;\n}\n\n/** Optional callbacks emitted by {@link retryAssistantCall} around each retry. */\nexport interface RetryCallbacks {\n\t/** Emitted before the backoff sleep of each retry attempt (1-indexed). */\n\tonRetryScheduled?: (\n\t\tattempt: number,\n\t\tmaxAttempts: number,\n\t\tdelayMs: number,\n\t\terrorMessage: string,\n\t) => void | Promise<void>;\n\t/** Emitted after the backoff sleep, immediately before the retried call starts. */\n\tonRetryAttemptStart?: () => void | Promise<void>;\n\t/** Emitted once when the loop ends: success if a later call completed normally. */\n\tonRetryFinished?: (success: boolean, attempt: number, finalError?: string) => void | Promise<void>;\n}\n\nclass RetrySleepAbortError extends Error {\n\tconstructor() {\n\t\tsuper(\"Aborted\");\n\t}\n}\n\nfunction sleep(ms: number, signal?: AbortSignal): Promise<void> {\n\treturn new Promise((resolve, reject) => {\n\t\tif (signal?.aborted) {\n\t\t\treject(new RetrySleepAbortError());\n\t\t\treturn;\n\t\t}\n\t\tconst timeout = setTimeout(resolve, ms);\n\t\tsignal?.addEventListener(\n\t\t\t\"abort\",\n\t\t\t() => {\n\t\t\t\tclearTimeout(timeout);\n\t\t\t\treject(new RetrySleepAbortError());\n\t\t\t},\n\t\t\t{ once: true },\n\t\t);\n\t});\n}\n\n/**\n * Run a single assistant-producing call with bounded retry on transient errors.\n *\n * Behavior:\n * - A successful response is returned immediately. Aborts are terminal and never\n *   retried, but reported as unsuccessful if they happen after a retry was scheduled.\n *   Aborts during the backoff sleep are normalized to an aborted `AssistantMessage`\n *   too, so callers do not need to care when cancellation happened.\n * - A non-retryable error (per {@link isRetryableAssistantError}, including quota/\n *   billing exhaustion) is returned immediately so deterministic errors fail fast.\n * - Otherwise retries up to `maxRetries` times with exponential backoff, emitting\n *   `onRetryScheduled` before each sleep, `onRetryAttemptStart` after each sleep before\n *   the retried call starts, and `onRetryFinished` once at the end (whether the loop\n *   ends in success, exhausted retries, or an aborted backoff).\n *\n * When `policy` is undefined or disabled, the first response is returned unchanged\n * (equivalent to calling `produce()` directly).\n */\nexport async function retryAssistantCall(\n\tproduce: () => Promise<AssistantMessage>,\n\tpolicy: RetryPolicy | undefined,\n\tsignal: AbortSignal | undefined,\n\tcallbacks?: RetryCallbacks,\n): Promise<AssistantMessage> {\n\tconst maxAttempts = policy?.enabled ? policy.maxRetries : 0;\n\n\tlet attempt = 0;\n\tlet lastRetry: { attempt: number; errorMessage: string } | undefined;\n\tfor (;;) {\n\t\tconst response = await produce();\n\n\t\t// Abort: terminal but not successful. Never retry an aborted message.\n\t\tif (response.stopReason === \"aborted\") {\n\t\t\tif (lastRetry) await callbacks?.onRetryFinished?.(false, lastRetry.attempt);\n\t\t\treturn response;\n\t\t}\n\n\t\t// Success: non-error, non-abort responses return as-is.\n\t\tif (response.stopReason !== \"error\") {\n\t\t\tif (lastRetry) await callbacks?.onRetryFinished?.(true, lastRetry.attempt);\n\t\t\treturn response;\n\t\t}\n\n\t\t// Non-retryable, or budget exhausted: return the final error message.\n\t\tif (attempt >= maxAttempts || !isRetryableAssistantError(response)) {\n\t\t\tif (lastRetry) await callbacks?.onRetryFinished?.(false, lastRetry.attempt, response.errorMessage);\n\t\t\treturn response;\n\t\t}\n\n\t\tattempt++;\n\t\tlastRetry = { attempt, errorMessage: response.errorMessage || \"Unknown error\" };\n\t\tconst delayMs = policy!.baseDelayMs * 2 ** (attempt - 1);\n\t\tawait callbacks?.onRetryScheduled?.(attempt, maxAttempts, delayMs, lastRetry.errorMessage);\n\n\t\t// Normalize aborts during retry backoff to the same AssistantMessage shape as\n\t\t// provider stream aborts, so callers do not need to care when cancellation happened.\n\t\ttry {\n\t\t\tawait sleep(delayMs, signal);\n\t\t} catch (error) {\n\t\t\tawait callbacks?.onRetryFinished?.(false, attempt, lastRetry.errorMessage);\n\t\t\tif (error instanceof RetrySleepAbortError) {\n\t\t\t\treturn { ...response, stopReason: \"aborted\", errorMessage: undefined };\n\t\t\t}\n\t\t\tthrow error;\n\t\t}\n\t\tawait callbacks?.onRetryAttemptStart?.();\n\t}\n}\n\n/**\n * Classifies whether a failed assistant message looks like a transient provider\n * or transport error, so callers can decide if the last assistant turn should be\n * restarted.\n *\n * This does not implement retry policy. Callers should first handle context\n * overflow separately, then apply their own retry budget, backoff, and reporting\n * before restarting the assistant turn.\n */\nexport function isRetryableAssistantError(message: AssistantMessage): boolean {\n\tif (\n\t\tmessage.stopReason !== \"error\" ||\n\t\tmessage.stopDetails?.type === \"refusal\" ||\n\t\tmessage.stopDetails?.type === \"sensitive\" ||\n\t\t!message.errorMessage\n\t) {\n\t\treturn false;\n\t}\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/node_modules/@earendil-works/pi-ai/dist/utils/server-fallback-receipt.d.ts

@@ -0,0 +1,30 @@
+import type { AssistantMessage } from "../types.ts";
+export declare const SERVER_FALLBACK_ABORTED_DIAGNOSTIC = "server_fallback_aborted";
+export declare const BILLING_INCOMPLETE_DIAGNOSTIC = "billing_incomplete_after_client_abort";
+export interface ServerFallbackReceipt {
+    readonly from: string;
+    readonly to: string;
+}
+/**
+ * Anthropic marks a classifier handoff with `{type:"fallback", from:{model}, to:{model}}`
+ * (`server-side-fallback-*` betas). Returns undefined for anything else, so an
+ * unrelated or malformed provider-native block falls through to normal handling.
+ */
+export declare function parseServerFallbackReceipt(block: unknown): ServerFallbackReceipt | undefined;
+/**
+ * Sticky routing serves later turns of a fallen-back conversation from the
+ * substitute model with no `fallback` block at all; a `fallback_message` entry
+ * in `usage.iterations` is the documented signal. A served-model string
+ * comparison is deliberately NOT used: gateways and Bedrock-style endpoints
+ * rewrite model ids, so a mismatch is not evidence of a fallback.
+ */
+export declare function parseStickyFallbackReceipt(usage: unknown, requestedModel: string, servedModel?: string): ServerFallbackReceipt | undefined;
+export declare function serverFallbackRefusalExplanation(receipt: ServerFallbackReceipt): string;
+/**
+ * Rewrites an aborted turn into the same shape a pre-output classifier refusal
+ * has, so `isClassifierRefusal()` routes it through the caller's fallback chain.
+ * Content is dropped: the substitute model's partial output was never requested,
+ * and replaying a half-turn containing a fallback marker breaks Anthropic replay.
+ */
+export declare function applyServerFallbackAbort(message: AssistantMessage, receipt: ServerFallbackReceipt): void;
+//# sourceMappingURL=server-fallback-receipt.d.ts.map

package/node_modules/@earendil-works/pi-ai/dist/utils/server-fallback-receipt.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"server-fallback-receipt.d.ts","sourceRoot":"","sources":["../../src/utils/server-fallback-receipt.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,gBAAgB,EAAE,MAAM,aAAa,CAAC;AAGpD,eAAO,MAAM,kCAAkC,4BAA4B,CAAC;AAC5E,eAAO,MAAM,6BAA6B,0CAA0C,CAAC;AAErF,MAAM,WAAW,qBAAqB;IACrC,QAAQ,CAAC,IAAI,EAAE,MAAM,CAAC;IACtB,QAAQ,CAAC,EAAE,EAAE,MAAM,CAAC;CACpB;AAWD;;;;GAIG;AACH,wBAAgB,0BAA0B,CAAC,KAAK,EAAE,OAAO,GAAG,qBAAqB,GAAG,SAAS,CAK5F;AAED;;;;;;GAMG;AACH,wBAAgB,0BAA0B,CACzC,KAAK,EAAE,OAAO,EACd,cAAc,EAAE,MAAM,EACtB,WAAW,CAAC,EAAE,MAAM,GAClB,qBAAqB,GAAG,SAAS,CAUnC;AAED,wBAAgB,gCAAgC,CAAC,OAAO,EAAE,qBAAqB,GAAG,MAAM,CAEvF;AAED;;;;;GAKG;AACH,wBAAgB,wBAAwB,CAAC,OAAO,EAAE,gBAAgB,EAAE,OAAO,EAAE,qBAAqB,GAAG,IAAI,CAgBxG"}

package/node_modules/@earendil-works/pi-ai/dist/utils/server-fallback-receipt.js

@@ -0,0 +1,70 @@
+import { appendAssistantMessageDiagnostic } from "./diagnostics.js";
+export const SERVER_FALLBACK_ABORTED_DIAGNOSTIC = "server_fallback_aborted";
+export const BILLING_INCOMPLETE_DIAGNOSTIC = "billing_incomplete_after_client_abort";
+function isRecord(value) {
+    return typeof value === "object" && value !== null && !Array.isArray(value);
+}
+function readModel(value) {
+    if (!isRecord(value))
+        return undefined;
+    return typeof value.model === "string" && value.model.length > 0 ? value.model : undefined;
+}
+/**
+ * Anthropic marks a classifier handoff with `{type:"fallback", from:{model}, to:{model}}`
+ * (`server-side-fallback-*` betas). Returns undefined for anything else, so an
+ * unrelated or malformed provider-native block falls through to normal handling.
+ */
+export function parseServerFallbackReceipt(block) {
+    if (!isRecord(block) || block.type !== "fallback")
+        return undefined;
+    const from = readModel(block.from);
+    const to = readModel(block.to);
+    return from !== undefined && to !== undefined ? { from, to } : undefined;
+}
+/**
+ * Sticky routing serves later turns of a fallen-back conversation from the
+ * substitute model with no `fallback` block at all; a `fallback_message` entry
+ * in `usage.iterations` is the documented signal. A served-model string
+ * comparison is deliberately NOT used: gateways and Bedrock-style endpoints
+ * rewrite model ids, so a mismatch is not evidence of a fallback.
+ */
+export function parseStickyFallbackReceipt(usage, requestedModel, servedModel) {
+    if (!isRecord(usage) || !Array.isArray(usage.iterations))
+        return undefined;
+    for (let index = usage.iterations.length - 1; index >= 0; index--) {
+        const entry = usage.iterations[index];
+        if (!isRecord(entry) || entry.type !== "fallback_message")
+            continue;
+        const entryModel = typeof entry.model === "string" && entry.model.length > 0 ? entry.model : undefined;
+        const to = entryModel ?? servedModel;
+        return to !== undefined ? { from: requestedModel, to } : undefined;
+    }
+    return undefined;
+}
+export function serverFallbackRefusalExplanation(receipt) {
+    return `Server-side fallback (${receipt.from} -> ${receipt.to}) aborted by client policy`;
+}
+/**
+ * Rewrites an aborted turn into the same shape a pre-output classifier refusal
+ * has, so `isClassifierRefusal()` routes it through the caller's fallback chain.
+ * Content is dropped: the substitute model's partial output was never requested,
+ * and replaying a half-turn containing a fallback marker breaks Anthropic replay.
+ */
+export function applyServerFallbackAbort(message, receipt) {
+    const explanation = serverFallbackRefusalExplanation(receipt);
+    message.content = [];
+    message.stopReason = "error";
+    message.stopDetails = { type: "refusal", explanation };
+    message.errorMessage = explanation;
+    appendAssistantMessageDiagnostic(message, {
+        type: SERVER_FALLBACK_ABORTED_DIAGNOSTIC,
+        timestamp: Date.now(),
+        details: { from: receipt.from, to: receipt.to },
+    });
+    appendAssistantMessageDiagnostic(message, {
+        type: BILLING_INCOMPLETE_DIAGNOSTIC,
+        timestamp: Date.now(),
+        details: { reason: "per-attempt usage does not arrive after a client abort" },
+    });
+}
+//# sourceMappingURL=server-fallback-receipt.js.map

package/node_modules/@earendil-works/pi-ai/dist/utils/server-fallback-receipt.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"server-fallback-receipt.js","sourceRoot":"","sources":["../../src/utils/server-fallback-receipt.ts"],"names":[],"mappings":"AACA,OAAO,EAAE,gCAAgC,EAAE,MAAM,kBAAkB,CAAC;AAEpE,MAAM,CAAC,MAAM,kCAAkC,GAAG,yBAAyB,CAAC;AAC5E,MAAM,CAAC,MAAM,6BAA6B,GAAG,uCAAuC,CAAC;AAOrF,SAAS,QAAQ,CAAC,KAAc;IAC/B,OAAO,OAAO,KAAK,KAAK,QAAQ,IAAI,KAAK,KAAK,IAAI,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC,CAAC;AAC7E,CAAC;AAED,SAAS,SAAS,CAAC,KAAc;IAChC,IAAI,CAAC,QAAQ,CAAC,KAAK,CAAC;QAAE,OAAO,SAAS,CAAC;IACvC,OAAO,OAAO,KAAK,CAAC,KAAK,KAAK,QAAQ,IAAI,KAAK,CAAC,KAAK,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC,KAAK,CAAC,CAAC,CAAC,SAAS,CAAC;AAC5F,CAAC;AAED;;;;GAIG;AACH,MAAM,UAAU,0BAA0B,CAAC,KAAc;IACxD,IAAI,CAAC,QAAQ,CAAC,KAAK,CAAC,IAAI,KAAK,CAAC,IAAI,KAAK,UAAU;QAAE,OAAO,SAAS,CAAC;IACpE,MAAM,IAAI,GAAG,SAAS,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC;IACnC,MAAM,EAAE,GAAG,SAAS,CAAC,KAAK,CAAC,EAAE,CAAC,CAAC;IAC/B,OAAO,IAAI,KAAK,SAAS,IAAI,EAAE,KAAK,SAAS,CAAC,CAAC,CAAC,EAAE,IAAI,EAAE,EAAE,EAAE,CAAC,CAAC,CAAC,SAAS,CAAC;AAC1E,CAAC;AAED;;;;;;GAMG;AACH,MAAM,UAAU,0BAA0B,CACzC,KAAc,EACd,cAAsB,EACtB,WAAoB;IAEpB,IAAI,CAAC,QAAQ,CAAC,KAAK,CAAC,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC,UAAU,CAAC;QAAE,OAAO,SAAS,CAAC;IAC3E,KAAK,IAAI,KAAK,GAAG,KAAK,CAAC,UAAU,CAAC,MAAM,GAAG,CAAC,EAAE,KAAK,IAAI,CAAC,EAAE,KAAK,EAAE,EAAE,CAAC;QACnE,MAAM,KAAK,GAAY,KAAK,CAAC,UAAU,CAAC,KAAK,CAAC,CAAC;QAC/C,IAAI,CAAC,QAAQ,CAAC,KAAK,CAAC,IAAI,KAAK,CAAC,IAAI,KAAK,kBAAkB;YAAE,SAAS;QACpE,MAAM,UAAU,GAAG,OAAO,KAAK,CAAC,KAAK,KAAK,QAAQ,IAAI,KAAK,CAAC,KAAK,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC,KAAK,CAAC,CAAC,CAAC,SAAS,CAAC;QACvG,MAAM,EAAE,GAAG,UAAU,IAAI,WAAW,CAAC;QACrC,OAAO,EAAE,KAAK,SAAS,CAAC,CAAC,CAAC,EAAE,IAAI,EAAE,cAAc,EAAE,EAAE,EAAE,CAAC,CAAC,CAAC,SAAS,CAAC;IACpE,CAAC;IACD,OAAO,SAAS,CAAC;AAClB,CAAC;AAED,MAAM,UAAU,gCAAgC,CAAC,OAA8B;IAC9E,OAAO,yBAAyB,OAAO,CAAC,IAAI,OAAO,OAAO,CAAC,EAAE,4BAA4B,CAAC;AAC3F,CAAC;AAED;;;;;GAKG;AACH,MAAM,UAAU,wBAAwB,CAAC,OAAyB,EAAE,OAA8B;IACjG,MAAM,WAAW,GAAG,gCAAgC,CAAC,OAAO,CAAC,CAAC;IAC9D,OAAO,CAAC,OAAO,GAAG,EAAE,CAAC;IACrB,OAAO,CAAC,UAAU,GAAG,OAAO,CAAC;IAC7B,OAAO,CAAC,WAAW,GAAG,EAAE,IAAI,EAAE,SAAS,EAAE,WAAW,EAAE,CAAC;IACvD,OAAO,CAAC,YAAY,GAAG,WAAW,CAAC;IACnC,gCAAgC,CAAC,OAAO,EAAE;QACzC,IAAI,EAAE,kCAAkC;QACxC,SAAS,EAAE,IAAI,CAAC,GAAG,EAAE;QACrB,OAAO,EAAE,EAAE,IAAI,EAAE,OAAO,CAAC,IAAI,EAAE,EAAE,EAAE,OAAO,CAAC,EAAE,EAAE;KAC/C,CAAC,CAAC;IACH,gCAAgC,CAAC,OAAO,EAAE;QACzC,IAAI,EAAE,6BAA6B;QACnC,SAAS,EAAE,IAAI,CAAC,GAAG,EAAE;QACrB,OAAO,EAAE,EAAE,MAAM,EAAE,wDAAwD,EAAE;KAC7E,CAAC,CAAC;AACJ,CAAC","sourcesContent":["import type { AssistantMessage } from \"../types.ts\";\nimport { appendAssistantMessageDiagnostic } from \"./diagnostics.ts\";\n\nexport const SERVER_FALLBACK_ABORTED_DIAGNOSTIC = \"server_fallback_aborted\";\nexport const BILLING_INCOMPLETE_DIAGNOSTIC = \"billing_incomplete_after_client_abort\";\n\nexport interface ServerFallbackReceipt {\n\treadonly from: string;\n\treadonly to: string;\n}\n\nfunction isRecord(value: unknown): value is Record<string, unknown> {\n\treturn typeof value === \"object\" && value !== null && !Array.isArray(value);\n}\n\nfunction readModel(value: unknown): string | undefined {\n\tif (!isRecord(value)) return undefined;\n\treturn typeof value.model === \"string\" && value.model.length > 0 ? value.model : undefined;\n}\n\n/**\n * Anthropic marks a classifier handoff with `{type:\"fallback\", from:{model}, to:{model}}`\n * (`server-side-fallback-*` betas). Returns undefined for anything else, so an\n * unrelated or malformed provider-native block falls through to normal handling.\n */\nexport function parseServerFallbackReceipt(block: unknown): ServerFallbackReceipt | undefined {\n\tif (!isRecord(block) || block.type !== \"fallback\") return undefined;\n\tconst from = readModel(block.from);\n\tconst to = readModel(block.to);\n\treturn from !== undefined && to !== undefined ? { from, to } : undefined;\n}\n\n/**\n * Sticky routing serves later turns of a fallen-back conversation from the\n * substitute model with no `fallback` block at all; a `fallback_message` entry\n * in `usage.iterations` is the documented signal. A served-model string\n * comparison is deliberately NOT used: gateways and Bedrock-style endpoints\n * rewrite model ids, so a mismatch is not evidence of a fallback.\n */\nexport function parseStickyFallbackReceipt(\n\tusage: unknown,\n\trequestedModel: string,\n\tservedModel?: string,\n): ServerFallbackReceipt | undefined {\n\tif (!isRecord(usage) || !Array.isArray(usage.iterations)) return undefined;\n\tfor (let index = usage.iterations.length - 1; index >= 0; index--) {\n\t\tconst entry: unknown = usage.iterations[index];\n\t\tif (!isRecord(entry) || entry.type !== \"fallback_message\") continue;\n\t\tconst entryModel = typeof entry.model === \"string\" && entry.model.length > 0 ? entry.model : undefined;\n\t\tconst to = entryModel ?? servedModel;\n\t\treturn to !== undefined ? { from: requestedModel, to } : undefined;\n\t}\n\treturn undefined;\n}\n\nexport function serverFallbackRefusalExplanation(receipt: ServerFallbackReceipt): string {\n\treturn `Server-side fallback (${receipt.from} -> ${receipt.to}) aborted by client policy`;\n}\n\n/**\n * Rewrites an aborted turn into the same shape a pre-output classifier refusal\n * has, so `isClassifierRefusal()` routes it through the caller's fallback chain.\n * Content is dropped: the substitute model's partial output was never requested,\n * and replaying a half-turn containing a fallback marker breaks Anthropic replay.\n */\nexport function applyServerFallbackAbort(message: AssistantMessage, receipt: ServerFallbackReceipt): void {\n\tconst explanation = serverFallbackRefusalExplanation(receipt);\n\tmessage.content = [];\n\tmessage.stopReason = \"error\";\n\tmessage.stopDetails = { type: \"refusal\", explanation };\n\tmessage.errorMessage = explanation;\n\tappendAssistantMessageDiagnostic(message, {\n\t\ttype: SERVER_FALLBACK_ABORTED_DIAGNOSTIC,\n\t\ttimestamp: Date.now(),\n\t\tdetails: { from: receipt.from, to: receipt.to },\n\t});\n\tappendAssistantMessageDiagnostic(message, {\n\t\ttype: BILLING_INCOMPLETE_DIAGNOSTIC,\n\t\ttimestamp: Date.now(),\n\t\tdetails: { reason: \"per-attempt usage does not arrive after a client abort\" },\n\t});\n}\n"]}

package/node_modules/@earendil-works/pi-ai/dist/utils/text.d.ts

@@ -0,0 +1,6 @@
+import type { ImageContent, ProviderNativeContent, TextContent, ThinkingContent, ToolCall } from "../types.ts";
+type Content = TextContent | ImageContent | ThinkingContent | ToolCall | ProviderNativeContent;
+/** 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/node_modules/@earendil-works/pi-ai/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,qBAAqB,EAAE,WAAW,EAAE,eAAe,EAAE,QAAQ,EAAE,MAAM,aAAa,CAAC;AAE/G,KAAK,OAAO,GAAG,WAAW,GAAG,YAAY,GAAG,eAAe,GAAG,QAAQ,GAAG,qBAAqB,CAAC;AAE/F,kDAAkD;AAClD,wBAAgB,WAAW,CAAC,OAAO,EAAE,MAAM,GAAG,SAAS,OAAO,EAAE,EAAE,SAAS,SAAO,GAAG,MAAM,CAM1F"}

package/node_modules/@earendil-works/pi-ai/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/node_modules/@earendil-works/pi-ai/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;IACjF,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;AACnB,CAAC","sourcesContent":["import type { ImageContent, ProviderNativeContent, TextContent, ThinkingContent, ToolCall } from \"../types.ts\";\n\ntype Content = TextContent | ImageContent | ThinkingContent | ToolCall | ProviderNativeContent;\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/node_modules/@earendil-works/pi-ai/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/node_modules/@earendil-works/pi-ai/dist/utils/uuid.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"uuid.d.ts","sourceRoot":"","sources":["../../src/utils/uuid.ts"],"names":[],"mappings":"AAcA,sCAAsC;AACtC,wBAAgB,MAAM,IAAI,MAAM,CAgC/B"}

package/node_modules/@earendil-works/{pi-agent-core/dist/harness/session → pi-ai/dist/utils}/uuid.js

@@ -10,6 +10,7 @@ function fillRandomBytes(bytes) {
         bytes[i] = Math.floor(Math.random() * 256);
     }
 }
+/** Generate a time-ordered UUIDv7. */
 export function uuidv7() {
     const random = new Uint8Array(16);
     fillRandomBytes(random);
@@ -20,9 +21,8 @@ export function uuidv7() {
     }
     else {
         sequence = (sequence + 1) >>> 0;
-        if (sequence === 0) {
+        if (sequence === 0)
             lastTimestamp++;
-        }
     }
     const bytes = new Uint8Array(16);
     bytes[0] = (lastTimestamp / 0x10000000000) & 0xff;

package/node_modules/@earendil-works/pi-ai/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;IACtD,MAAM,MAAM,GAAG,UAAU,CAAC,MAAM,CAAC;IACjC,IAAI,MAAM,EAAE,eAAe,EAAE,CAAC;QAC7B,MAAM,CAAC,eAAe,CAAC,KAAK,CAAC,CAAC;QAC9B,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;AACF,CAAC;AAED,sCAAsC;AACtC,MAAM,UAAU,MAAM;IACrB,MAAM,MAAM,GAA4B,IAAI,UAAU,CAAC,EAAE,CAAC,CAAC;IAC3D,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,GAA4B,IAAI,UAAU,CAAC,EAAE,CAAC,CAAC;IAC1D,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,OAAO,UAAU,CAAC,KAAK,CAAC,CAAC;AAC1B,CAAC;AAED,SAAS,UAAU,CAAC,KAA8B;IACjD,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;AACzJ,CAAC","sourcesContent":["let lastTimestamp = -Infinity;\nlet sequence = 0;\n\nfunction fillRandomBytes(bytes: Uint8Array<ArrayBuffer>): void {\n\tconst crypto = globalThis.crypto;\n\tif (crypto?.getRandomValues) {\n\t\tcrypto.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: Uint8Array<ArrayBuffer> = 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: Uint8Array<ArrayBuffer> = 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\treturn formatUuid(bytes);\n}\n\nfunction formatUuid(bytes: Uint8Array<ArrayBuffer>): string {\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"]}