@handy_wote/pi-ai 0.81.1

package/dist/utils/overflow.d.ts

@@ -0,0 +1,59 @@
+import type { AssistantMessage } from "../types.ts";
+/**
+ * Check if an assistant message represents a context overflow error.
+ *
+ * This handles two cases:
+ * 1. Error-based overflow: Most providers return stopReason "error" with a
+ *    specific error message pattern.
+ * 2. Silent overflow: Some providers accept overflow requests and return
+ *    successfully. For these, we check if usage.input exceeds the context window.
+ *
+ * ## Reliability by Provider
+ *
+ * **Reliable detection (returns error with detectable message):**
+ * - Anthropic: "prompt is too long: X tokens > Y maximum" or "request_too_large"
+ * - 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)"
+ * - Google Gemini: "input token count exceeds the maximum"
+ * - xAI (Grok): "maximum prompt length is X but request contains Y"
+ * - Groq: "reduce the length of the messages"
+ * - Cerebras: 400/413 status code (no body)
+ * - Mistral: "Prompt contains X tokens ... too large for model with Y maximum context length"
+ * - OpenRouter (most backends): "maximum context length is X tokens"
+ * - OpenRouter/Poolside: "Input length X exceeds the maximum allowed input length of Y tokens."
+ * - Together AI: "The input (X tokens) is longer than the model's context length (Y tokens)."
+ * - llama.cpp: "exceeds the available context size"
+ * - LM Studio: "greater than the context length"
+ * - Kimi For Coding: "exceeded model token limit: X (requested: Y)"
+ * - DS4: "Prompt has X tokens, but the configured context size is Y tokens"
+ * - DashScope/Qwen: "Range of input length should be [1, X]"
+ *
+ * **Unreliable detection:**
+ * - z.ai: Sometimes accepts overflow silently (detectable via usage.input > contextWindow),
+ *   sometimes returns rate limit errors. Pass contextWindow param to detect silent overflow.
+ * - Xiaomi MiMo: Truncates input to fit contextWindow then returns stopReason "length" with
+ *   output=0. Pass contextWindow param to detect via the "filled context + zero output" signal.
+ * - Ollama: May truncate input silently for some setups, but may also return explicit
+ *   overflow errors that match the patterns above. Silent truncation still cannot be
+ *   detected here because we do not know the expected token count.
+ *
+ * ## Custom Providers
+ *
+ * If you've added custom models via settings.json, this function may not detect
+ * overflow errors from those providers. To add support:
+ *
+ * 1. Send a request that exceeds the model's context window
+ * 2. Check the errorMessage in the response
+ * 3. Create a regex pattern that matches the error
+ * 4. The pattern should be added to OVERFLOW_PATTERNS in this file, or
+ *    check the errorMessage yourself before calling this function
+ *
+ * @param message - The assistant message to check
+ * @param contextWindow - Optional context window size for detecting silent overflow (z.ai)
+ * @returns true if the message indicates a context overflow
+ */
+export declare function isContextOverflow(message: AssistantMessage, contextWindow?: number): boolean;
+/**
+ * Get the overflow patterns for testing purposes.
+ */
+export declare function getOverflowPatterns(): RegExp[];
+//# sourceMappingURL=overflow.d.ts.map

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

@@ -0,0 +1 @@
+{"version":3,"file":"overflow.d.ts","sourceRoot":"","sources":["../../src/utils/overflow.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,gBAAgB,EAAE,MAAM,aAAa,CAAC;AA+EpD;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAmDG;AACH,wBAAgB,iBAAiB,CAAC,OAAO,EAAE,gBAAgB,EAAE,aAAa,CAAC,EAAE,MAAM,GAAG,OAAO,CA6B5F;AAED;;GAEG;AACH,wBAAgB,mBAAmB,IAAI,MAAM,EAAE,CAE9C","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/overflow.js

@@ -0,0 +1,161 @@
+/**
+ * Regex patterns to detect context overflow errors from different providers.
+ *
+ * These patterns match error messages returned when the input exceeds
+ * the model's context window.
+ *
+ * Provider-specific patterns (with example error messages):
+ *
+ * - Anthropic: "prompt is too long: 213462 tokens > 200000 maximum"
+ * - Anthropic: "413 {\"error\":{\"type\":\"request_too_large\",\"message\":\"Request exceeds the maximum size\"}}"
+ * - OpenAI: "Your input exceeds the context window of this model"
+ * - OpenAI/LiteLLM: "Requested token count exceeds the model's maximum context length of 131072 tokens"
+ * - OpenAI-compatible: "Input length (265330) exceeds model's maximum context length (262144)."
+ * - Google: "The input token count (1196265) exceeds the maximum number of tokens allowed (1048575)"
+ * - xAI: "This model's maximum prompt length is 131072 but the request contains 537812 tokens"
+ * - Groq: "Please reduce the length of the messages or completion"
+ * - OpenRouter: "This endpoint's maximum context length is X tokens. However, you requested about Y tokens"
+ * - OpenRouter/Poolside: "Input length X exceeds the maximum allowed input length of Y tokens."
+ * - Together AI: "The input (X tokens) is longer than the model's context length (Y tokens)."
+ * - llama.cpp: "the request exceeds the available context size, try increasing it"
+ * - LM Studio: "tokens to keep from the initial prompt is greater than the context length"
+ * - GitHub Copilot: "prompt token count of X exceeds the limit of Y"
+ * - MiniMax: "invalid params, context window exceeds limit"
+ * - Kimi For Coding: "Your request exceeded model token limit: X (requested: Y)"
+ * - DS4: "Prompt has X tokens, but the configured context size is Y tokens"
+ * - Cerebras: "400/413 status code (no body)"
+ * - Mistral: "Prompt contains X tokens ... too large for model with Y maximum context length"
+ * - z.ai: Does NOT error, accepts overflow silently - handled via usage.input > contextWindow
+ * - Xiaomi MiMo: Truncates input to fill contextWindow exactly, then returns finish_reason "length"
+ *   with output=0 (no room left to generate). Detected via stopReason "length" + zero output +
+ *   input filling the context window.
+ * - DashScope/Qwen: "Range of input length should be [1, X]" (HTTP 400 invalid_parameter_error)
+ * - Ollama: Some deployments truncate silently, others return errors like "prompt too long; exceeded max context length by X tokens"
+ */
+const OVERFLOW_PATTERNS = [
+    /prompt is too long/i, // Anthropic token overflow
+    /request_too_large/i, // Anthropic request byte-size overflow (HTTP 413)
+    /input is too long for requested model/i, // Amazon Bedrock
+    /exceeds the context window/i, // OpenAI (Completions & Responses API)
+    /exceeds (?:the )?(?:model'?s )?maximum context length(?: of [\d,]+ tokens?|\s*\([\d,]+\))/i, // OpenAI-compatible proxies (LiteLLM)
+    /input token count.*exceeds the maximum/i, // Google (Gemini)
+    /maximum prompt length is \d+/i, // xAI (Grok)
+    /reduce the length of the messages/i, // Groq
+    /maximum context length is \d+ tokens/i, // OpenRouter (most backends)
+    /exceeds (?:the )?maximum allowed input length of [\d,]+ tokens?/i, // OpenRouter/Poolside
+    /input \(\d+ tokens\) is longer than the model'?s context length \(\d+ tokens\)/i, // Together AI
+    /exceeds the limit of \d+/i, // GitHub Copilot
+    /exceeds the available context size/i, // llama.cpp server
+    /greater than the context length/i, // LM Studio
+    /context window exceeds limit/i, // MiniMax
+    /exceeded model token limit/i, // Kimi For Coding
+    /too large for model with \d+ maximum context length/i, // Mistral
+    /prompt has [\d,]+ tokens?, but the configured context size is [\d,]+ tokens?/i, // DS4 server
+    /model_context_window_exceeded/i, // z.ai non-standard finish_reason surfaced as error text
+    /prompt too long; exceeded (?:max )?context length/i, // Ollama explicit overflow error
+    /range of input length should be/i, // DashScope / Qwen Token Plan
+    /context[_ ]length[_ ]exceeded/i, // Generic fallback
+    /too many tokens/i, // Generic fallback
+    /token limit exceeded/i, // Generic fallback
+    /^4(?:00|13)\s*(?:status code)?\s*\(no body\)/i, // Cerebras: 400/413 with no body
+];
+/**
+ * Patterns that indicate non-overflow errors (e.g. rate limiting, server errors).
+ * Error messages matching any of these are excluded from overflow detection
+ * even if they also match an OVERFLOW_PATTERN.
+ *
+ * Example: Bedrock formats throttling errors as "ThrottlingException: Too many tokens,
+ * please wait before trying again." which would match the /too many tokens/i overflow
+ * pattern without this exclusion.
+ */
+const NON_OVERFLOW_PATTERNS = [
+    /^(Throttling error|Service unavailable):/i, // AWS Bedrock non-overflow errors (human-readable prefixes from formatBedrockError)
+    /rate limit/i, // Generic rate limiting
+    /too many requests/i, // Generic HTTP 429 style
+];
+/**
+ * Check if an assistant message represents a context overflow error.
+ *
+ * This handles two cases:
+ * 1. Error-based overflow: Most providers return stopReason "error" with a
+ *    specific error message pattern.
+ * 2. Silent overflow: Some providers accept overflow requests and return
+ *    successfully. For these, we check if usage.input exceeds the context window.
+ *
+ * ## Reliability by Provider
+ *
+ * **Reliable detection (returns error with detectable message):**
+ * - Anthropic: "prompt is too long: X tokens > Y maximum" or "request_too_large"
+ * - 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)"
+ * - Google Gemini: "input token count exceeds the maximum"
+ * - xAI (Grok): "maximum prompt length is X but request contains Y"
+ * - Groq: "reduce the length of the messages"
+ * - Cerebras: 400/413 status code (no body)
+ * - Mistral: "Prompt contains X tokens ... too large for model with Y maximum context length"
+ * - OpenRouter (most backends): "maximum context length is X tokens"
+ * - OpenRouter/Poolside: "Input length X exceeds the maximum allowed input length of Y tokens."
+ * - Together AI: "The input (X tokens) is longer than the model's context length (Y tokens)."
+ * - llama.cpp: "exceeds the available context size"
+ * - LM Studio: "greater than the context length"
+ * - Kimi For Coding: "exceeded model token limit: X (requested: Y)"
+ * - DS4: "Prompt has X tokens, but the configured context size is Y tokens"
+ * - DashScope/Qwen: "Range of input length should be [1, X]"
+ *
+ * **Unreliable detection:**
+ * - z.ai: Sometimes accepts overflow silently (detectable via usage.input > contextWindow),
+ *   sometimes returns rate limit errors. Pass contextWindow param to detect silent overflow.
+ * - Xiaomi MiMo: Truncates input to fit contextWindow then returns stopReason "length" with
+ *   output=0. Pass contextWindow param to detect via the "filled context + zero output" signal.
+ * - Ollama: May truncate input silently for some setups, but may also return explicit
+ *   overflow errors that match the patterns above. Silent truncation still cannot be
+ *   detected here because we do not know the expected token count.
+ *
+ * ## Custom Providers
+ *
+ * If you've added custom models via settings.json, this function may not detect
+ * overflow errors from those providers. To add support:
+ *
+ * 1. Send a request that exceeds the model's context window
+ * 2. Check the errorMessage in the response
+ * 3. Create a regex pattern that matches the error
+ * 4. The pattern should be added to OVERFLOW_PATTERNS in this file, or
+ *    check the errorMessage yourself before calling this function
+ *
+ * @param message - The assistant message to check
+ * @param contextWindow - Optional context window size for detecting silent overflow (z.ai)
+ * @returns true if the message indicates a context overflow
+ */
+export function isContextOverflow(message, contextWindow) {
+    // Case 1: Check error message patterns
+    if (message.stopReason === "error" && message.errorMessage) {
+        // Skip messages matching known non-overflow patterns (e.g. throttling / rate-limit)
+        const isNonOverflow = NON_OVERFLOW_PATTERNS.some((p) => p.test(message.errorMessage));
+        if (!isNonOverflow && OVERFLOW_PATTERNS.some((p) => p.test(message.errorMessage))) {
+            return true;
+        }
+    }
+    // Case 2: Silent overflow (z.ai style) - successful but usage exceeds context
+    if (contextWindow && message.stopReason === "stop") {
+        const inputTokens = message.usage.input + message.usage.cacheRead;
+        if (inputTokens > contextWindow) {
+            return true;
+        }
+    }
+    // Case 3: Length-stop overflow (Xiaomi MiMo style) - server truncates oversized input
+    // to fit the context window, leaving no room for output. Returns stopReason "length"
+    // with output=0 and input+cacheRead filling the context window.
+    if (contextWindow && message.stopReason === "length" && message.usage.output === 0) {
+        const inputTokens = message.usage.input + message.usage.cacheRead;
+        if (inputTokens >= contextWindow * 0.99) {
+            return true;
+        }
+    }
+    return false;
+}
+/**
+ * Get the overflow patterns for testing purposes.
+ */
+export function getOverflowPatterns() {
+    return [...OVERFLOW_PATTERNS];
+}
+//# sourceMappingURL=overflow.js.map

package/dist/utils/overflow.js.map

@@ -0,0 +1 @@
+{"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/provider-env.d.ts

@@ -0,0 +1,7 @@
+import type { ProviderEnv } from "../types.ts";
+/**
+ * Resolve a provider env value from scoped overrides, normal process.env, then
+ * the duplicated Bun sandbox fallback for direct pi-ai consumers.
+ */
+export declare function getProviderEnvValue(name: string, env?: ProviderEnv): string | undefined;
+//# sourceMappingURL=provider-env.d.ts.map

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

@@ -0,0 +1 @@
+{"version":3,"file":"provider-env.d.ts","sourceRoot":"","sources":["../../src/utils/provider-env.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,WAAW,EAAE,MAAM,aAAa,CAAC;AAwC/C;;;GAGG;AACH,wBAAgB,mBAAmB,CAAC,IAAI,EAAE,MAAM,EAAE,GAAG,CAAC,EAAE,WAAW,GAAG,MAAM,GAAG,SAAS,CAOvF","sourcesContent":["import type { ProviderEnv } from \"../types.ts\";\n\nlet procEnvCache: Map<string, string> | null = null;\n\n/**\n * Fallback for https://github.com/oven-sh/bun/issues/27802.\n * Bun compiled binaries can expose an empty process.env inside Linux sandboxes\n * even though /proc/self/environ contains the environment.\n *\n * This intentionally duplicates restoreSandboxEnv() in\n * packages/coding-agent/src/bun/restore-sandbox-env.ts. The ai package can be\n * used directly, without going through that entrypoint, so provider env lookup\n * must not depend on process.env having been patched.\n */\nfunction getBunSandboxEnvValue(name: string): string | undefined {\n\tif (typeof process === \"undefined\" || !process.versions?.bun || Object.keys(process.env).length > 0) {\n\t\treturn undefined;\n\t}\n\n\tif (procEnvCache === null) {\n\t\tprocEnvCache = new Map();\n\t\ttry {\n\t\t\tconst { readFileSync } = require(\"node:fs\") as {\n\t\t\t\treadFileSync(path: string, encoding: BufferEncoding): string;\n\t\t\t};\n\t\t\tconst data = readFileSync(\"/proc/self/environ\", \"utf-8\");\n\t\t\tfor (const entry of data.split(\"\\0\")) {\n\t\t\t\tconst idx = entry.indexOf(\"=\");\n\t\t\t\tif (idx > 0) {\n\t\t\t\t\tprocEnvCache.set(entry.slice(0, idx), entry.slice(idx + 1));\n\t\t\t\t}\n\t\t\t}\n\t\t} catch {\n\t\t\t// /proc/self/environ may not exist or may not be readable.\n\t\t}\n\t}\n\n\treturn procEnvCache.get(name);\n}\n\n/**\n * Resolve a provider env value from scoped overrides, normal process.env, then\n * the duplicated Bun sandbox fallback for direct pi-ai consumers.\n */\nexport function getProviderEnvValue(name: string, env?: ProviderEnv): string | undefined {\n\treturn (\n\t\tenv?.[name] ||\n\t\t(typeof process !== \"undefined\" ? process.env[name] : undefined) ||\n\t\tgetBunSandboxEnvValue(name) ||\n\t\tundefined\n\t);\n}\n"]}

package/dist/utils/provider-env.js

@@ -0,0 +1,44 @@
+let procEnvCache = null;
+/**
+ * Fallback for https://github.com/oven-sh/bun/issues/27802.
+ * Bun compiled binaries can expose an empty process.env inside Linux sandboxes
+ * even though /proc/self/environ contains the environment.
+ *
+ * This intentionally duplicates restoreSandboxEnv() in
+ * packages/coding-agent/src/bun/restore-sandbox-env.ts. The ai package can be
+ * used directly, without going through that entrypoint, so provider env lookup
+ * must not depend on process.env having been patched.
+ */
+function getBunSandboxEnvValue(name) {
+    if (typeof process === "undefined" || !process.versions?.bun || Object.keys(process.env).length > 0) {
+        return undefined;
+    }
+    if (procEnvCache === null) {
+        procEnvCache = new Map();
+        try {
+            const { readFileSync } = require("node:fs");
+            const data = readFileSync("/proc/self/environ", "utf-8");
+            for (const entry of data.split("\0")) {
+                const idx = entry.indexOf("=");
+                if (idx > 0) {
+                    procEnvCache.set(entry.slice(0, idx), entry.slice(idx + 1));
+                }
+            }
+        }
+        catch {
+            // /proc/self/environ may not exist or may not be readable.
+        }
+    }
+    return procEnvCache.get(name);
+}
+/**
+ * Resolve a provider env value from scoped overrides, normal process.env, then
+ * the duplicated Bun sandbox fallback for direct pi-ai consumers.
+ */
+export function getProviderEnvValue(name, env) {
+    return (env?.[name] ||
+        (typeof process !== "undefined" ? process.env[name] : undefined) ||
+        getBunSandboxEnvValue(name) ||
+        undefined);
+}
+//# sourceMappingURL=provider-env.js.map

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

@@ -0,0 +1 @@
+{"version":3,"file":"provider-env.js","sourceRoot":"","sources":["../../src/utils/provider-env.ts"],"names":[],"mappings":"AAEA,IAAI,YAAY,GAA+B,IAAI,CAAC;AAEpD;;;;;;;;;GASG;AACH,SAAS,qBAAqB,CAAC,IAAY,EAAsB;IAChE,IAAI,OAAO,OAAO,KAAK,WAAW,IAAI,CAAC,OAAO,CAAC,QAAQ,EAAE,GAAG,IAAI,MAAM,CAAC,IAAI,CAAC,OAAO,CAAC,GAAG,CAAC,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;QACrG,OAAO,SAAS,CAAC;IAClB,CAAC;IAED,IAAI,YAAY,KAAK,IAAI,EAAE,CAAC;QAC3B,YAAY,GAAG,IAAI,GAAG,EAAE,CAAC;QACzB,IAAI,CAAC;YACJ,MAAM,EAAE,YAAY,EAAE,GAAG,OAAO,CAAC,SAAS,CAEzC,CAAC;YACF,MAAM,IAAI,GAAG,YAAY,CAAC,oBAAoB,EAAE,OAAO,CAAC,CAAC;YACzD,KAAK,MAAM,KAAK,IAAI,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,EAAE,CAAC;gBACtC,MAAM,GAAG,GAAG,KAAK,CAAC,OAAO,CAAC,GAAG,CAAC,CAAC;gBAC/B,IAAI,GAAG,GAAG,CAAC,EAAE,CAAC;oBACb,YAAY,CAAC,GAAG,CAAC,KAAK,CAAC,KAAK,CAAC,CAAC,EAAE,GAAG,CAAC,EAAE,KAAK,CAAC,KAAK,CAAC,GAAG,GAAG,CAAC,CAAC,CAAC,CAAC;gBAC7D,CAAC;YACF,CAAC;QACF,CAAC;QAAC,MAAM,CAAC;YACR,2DAA2D;QAC5D,CAAC;IACF,CAAC;IAED,OAAO,YAAY,CAAC,GAAG,CAAC,IAAI,CAAC,CAAC;AAAA,CAC9B;AAED;;;GAGG;AACH,MAAM,UAAU,mBAAmB,CAAC,IAAY,EAAE,GAAiB,EAAsB;IACxF,OAAO,CACN,GAAG,EAAE,CAAC,IAAI,CAAC;QACX,CAAC,OAAO,OAAO,KAAK,WAAW,CAAC,CAAC,CAAC,OAAO,CAAC,GAAG,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,SAAS,CAAC;QAChE,qBAAqB,CAAC,IAAI,CAAC;QAC3B,SAAS,CACT,CAAC;AAAA,CACF","sourcesContent":["import type { ProviderEnv } from \"../types.ts\";\n\nlet procEnvCache: Map<string, string> | null = null;\n\n/**\n * Fallback for https://github.com/oven-sh/bun/issues/27802.\n * Bun compiled binaries can expose an empty process.env inside Linux sandboxes\n * even though /proc/self/environ contains the environment.\n *\n * This intentionally duplicates restoreSandboxEnv() in\n * packages/coding-agent/src/bun/restore-sandbox-env.ts. The ai package can be\n * used directly, without going through that entrypoint, so provider env lookup\n * must not depend on process.env having been patched.\n */\nfunction getBunSandboxEnvValue(name: string): string | undefined {\n\tif (typeof process === \"undefined\" || !process.versions?.bun || Object.keys(process.env).length > 0) {\n\t\treturn undefined;\n\t}\n\n\tif (procEnvCache === null) {\n\t\tprocEnvCache = new Map();\n\t\ttry {\n\t\t\tconst { readFileSync } = require(\"node:fs\") as {\n\t\t\t\treadFileSync(path: string, encoding: BufferEncoding): string;\n\t\t\t};\n\t\t\tconst data = readFileSync(\"/proc/self/environ\", \"utf-8\");\n\t\t\tfor (const entry of data.split(\"\\0\")) {\n\t\t\t\tconst idx = entry.indexOf(\"=\");\n\t\t\t\tif (idx > 0) {\n\t\t\t\t\tprocEnvCache.set(entry.slice(0, idx), entry.slice(idx + 1));\n\t\t\t\t}\n\t\t\t}\n\t\t} catch {\n\t\t\t// /proc/self/environ may not exist or may not be readable.\n\t\t}\n\t}\n\n\treturn procEnvCache.get(name);\n}\n\n/**\n * Resolve a provider env value from scoped overrides, normal process.env, then\n * the duplicated Bun sandbox fallback for direct pi-ai consumers.\n */\nexport function getProviderEnvValue(name: string, env?: ProviderEnv): string | undefined {\n\treturn (\n\t\tenv?.[name] ||\n\t\t(typeof process !== \"undefined\" ? process.env[name] : undefined) ||\n\t\tgetBunSandboxEnvValue(name) ||\n\t\tundefined\n\t);\n}\n"]}

package/dist/utils/retry.d.ts

@@ -0,0 +1,53 @@
+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
+ * restarted.
+ *
+ * This does not implement retry policy. Callers should first handle context
+ * overflow separately, then apply their own retry budget, backoff, and reporting
+ * before restarting the assistant turn.
+ */
+export declare function isRetryableAssistantError(message: AssistantMessage): boolean;
+//# sourceMappingURL=retry.d.ts.map

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

@@ -0,0 +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;;;;;GAKG;AACH,MAAM,WAAW,WAAW;IAC3B,OAAO,EAAE,OAAO,CAAC;IACjB,qFAAqF;IACrF,UAAU,EAAE,MAAM,CAAC;IACnB,0FAA0F;IAC1F,WAAW,EAAE,MAAM,CAAC;CACpB;AAED,kFAAkF;AAClF,MAAM,WAAW,cAAc;IAC9B,0EAA0E;IAC1E,gBAAgB,CAAC,EAAE,CAClB,OAAO,EAAE,MAAM,EACf,WAAW,EAAE,MAAM,EACnB,OAAO,EAAE,MAAM,EACf,YAAY,EAAE,MAAM,KAChB,IAAI,GAAG,OAAO,CAAC,IAAI,CAAC,CAAC;IAC1B,mFAAmF;IACnF,mBAAmB,CAAC,EAAE,MAAM,IAAI,GAAG,OAAO,CAAC,IAAI,CAAC,CAAC;IACjD,mFAAmF;IACnF,eAAe,CAAC,EAAE,CAAC,OAAO,EAAE,OAAO,EAAE,OAAO,EAAE,MAAM,EAAE,UAAU,CAAC,EAAE,MAAM,KAAK,IAAI,GAAG,OAAO,CAAC,IAAI,CAAC,CAAC;CACnG;AA0BD;;;;;;;;;;;;;;;;;GAiBG;AACH,wBAAsB,kBAAkB,CACvC,OAAO,EAAE,MAAM,OAAO,CAAC,gBAAgB,CAAC,EACxC,MAAM,EAAE,WAAW,GAAG,SAAS,EAC/B,MAAM,EAAE,WAAW,GAAG,SAAS,EAC/B,SAAS,CAAC,EAAE,cAAc,GACxB,OAAO,CAAC,gBAAgB,CAAC,CA4C3B;AAED;;;;;;;;GAQG;AACH,wBAAgB,yBAAyB,CAAC,OAAO,EAAE,gBAAgB,GAAG,OAAO,CAK5E","sourcesContent":["import type { AssistantMessage } from \"../types.ts\";\n\nfunction buildProviderErrorPattern(patterns: readonly string[]): RegExp {\n\treturn new RegExp(patterns.join(\"|\"), \"i\");\n}\n\nconst NON_RETRYABLE_PROVIDER_LIMIT_ERROR_PATTERN = buildProviderErrorPattern([\n\t// OpenCode Go/free-tier limits returned as 429 JSON error types by OpenCode's\n\t// Zen API. These are subscription/account limits, not transient throttles.\n\t\"GoUsageLimitError\",\n\t\"FreeUsageLimitError\",\n\n\t// OpenCode Go subscription-limit text asks users to enable available-balance\n\t// usage after rolling/weekly/monthly limits are reached.\n\t\"Monthly usage limit reached\",\n\t\"available balance\",\n\n\t// Generic quota/budget/billing exhaustion. `insufficient_quota` is OpenAI's\n\t// quota/billing error code; the other strings cover common gateway wording.\n\t\"insufficient_quota\",\n\t\"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 * Retry policy: bounded attempts with exponential backoff (`baseDelayMs * 2^(attempt-1)`).\n * Matches `settings.retry` (`enabled`, `maxRetries`, `baseDelayMs`) in coding-agent; kept\n * here so the classifier and the policy-driven retry loop live together and stay reusable\n * by the SDK and other callers.\n */\nexport interface RetryPolicy {\n\tenabled: boolean;\n\t/** Max retry attempts (0 = no retries). The initial call never counts as a retry. */\n\tmaxRetries: number;\n\t/** Base delay in ms. Per-attempt delay is `baseDelayMs * 2^(attempt-1)` before jitter. */\n\tbaseDelayMs: number;\n}\n\n/** Optional callbacks emitted by {@link retryAssistantCall} around each retry. */\nexport interface RetryCallbacks {\n\t/** Emitted before the backoff sleep of each retry attempt (1-indexed). */\n\tonRetryScheduled?: (\n\t\tattempt: number,\n\t\tmaxAttempts: number,\n\t\tdelayMs: number,\n\t\terrorMessage: string,\n\t) => void | Promise<void>;\n\t/** Emitted after the backoff sleep, immediately before the retried call starts. */\n\tonRetryAttemptStart?: () => void | Promise<void>;\n\t/** Emitted once when the loop ends: success if a later call completed normally. */\n\tonRetryFinished?: (success: boolean, attempt: number, finalError?: string) => void | Promise<void>;\n}\n\nclass RetrySleepAbortError extends Error {\n\tconstructor() {\n\t\tsuper(\"Aborted\");\n\t}\n}\n\nfunction sleep(ms: number, signal?: AbortSignal): Promise<void> {\n\treturn new Promise((resolve, reject) => {\n\t\tif (signal?.aborted) {\n\t\t\treject(new RetrySleepAbortError());\n\t\t\treturn;\n\t\t}\n\t\tconst timeout = setTimeout(resolve, ms);\n\t\tsignal?.addEventListener(\n\t\t\t\"abort\",\n\t\t\t() => {\n\t\t\t\tclearTimeout(timeout);\n\t\t\t\treject(new RetrySleepAbortError());\n\t\t\t},\n\t\t\t{ once: true },\n\t\t);\n\t});\n}\n\n/**\n * Run a single assistant-producing call with bounded retry on transient errors.\n *\n * Behavior:\n * - A successful response is returned immediately. Aborts are terminal and never\n *   retried, but reported as unsuccessful if they happen after a retry was scheduled.\n *   Aborts during the backoff sleep are normalized to an aborted `AssistantMessage`\n *   too, so callers do not need to care when cancellation happened.\n * - A non-retryable error (per {@link isRetryableAssistantError}, including quota/\n *   billing exhaustion) is returned immediately so deterministic errors fail fast.\n * - Otherwise retries up to `maxRetries` times with exponential backoff, emitting\n *   `onRetryScheduled` before each sleep, `onRetryAttemptStart` after each sleep before\n *   the retried call starts, and `onRetryFinished` once at the end (whether the loop\n *   ends in success, exhausted retries, or an aborted backoff).\n *\n * When `policy` is undefined or disabled, the first response is returned unchanged\n * (equivalent to calling `produce()` directly).\n */\nexport async function retryAssistantCall(\n\tproduce: () => Promise<AssistantMessage>,\n\tpolicy: RetryPolicy | undefined,\n\tsignal: AbortSignal | undefined,\n\tcallbacks?: RetryCallbacks,\n): Promise<AssistantMessage> {\n\tconst maxAttempts = policy?.enabled ? policy.maxRetries : 0;\n\n\tlet attempt = 0;\n\tlet lastRetry: { attempt: number; errorMessage: string } | undefined;\n\tfor (;;) {\n\t\tconst response = await produce();\n\n\t\t// Abort: terminal but not successful. Never retry an aborted message.\n\t\tif (response.stopReason === \"aborted\") {\n\t\t\tif (lastRetry) await callbacks?.onRetryFinished?.(false, lastRetry.attempt);\n\t\t\treturn response;\n\t\t}\n\n\t\t// Success: non-error, non-abort responses return as-is.\n\t\tif (response.stopReason !== \"error\") {\n\t\t\tif (lastRetry) await callbacks?.onRetryFinished?.(true, lastRetry.attempt);\n\t\t\treturn response;\n\t\t}\n\n\t\t// Non-retryable, or budget exhausted: return the final error message.\n\t\tif (attempt >= maxAttempts || !isRetryableAssistantError(response)) {\n\t\t\tif (lastRetry) await callbacks?.onRetryFinished?.(false, lastRetry.attempt, response.errorMessage);\n\t\t\treturn response;\n\t\t}\n\n\t\tattempt++;\n\t\tlastRetry = { attempt, errorMessage: response.errorMessage || \"Unknown error\" };\n\t\tconst delayMs = policy!.baseDelayMs * 2 ** (attempt - 1);\n\t\tawait callbacks?.onRetryScheduled?.(attempt, maxAttempts, delayMs, lastRetry.errorMessage);\n\n\t\t// Normalize aborts during retry backoff to the same AssistantMessage shape as\n\t\t// provider stream aborts, so callers do not need to care when cancellation happened.\n\t\ttry {\n\t\t\tawait sleep(delayMs, signal);\n\t\t} catch (error) {\n\t\t\tawait callbacks?.onRetryFinished?.(false, attempt, lastRetry.errorMessage);\n\t\t\tif (error instanceof RetrySleepAbortError) {\n\t\t\t\treturn { ...response, stopReason: \"aborted\", errorMessage: undefined };\n\t\t\t}\n\t\t\tthrow error;\n\t\t}\n\t\tawait callbacks?.onRetryAttemptStart?.();\n\t}\n}\n\n/**\n * Classifies whether a failed assistant message looks like a transient provider\n * or transport error, so callers can decide if the last assistant turn should be\n * restarted.\n *\n * This does not implement retry policy. Callers should first handle context\n * overflow separately, then apply their own retry budget, backoff, and reporting\n * before restarting the assistant turn.\n */\nexport function isRetryableAssistantError(message: AssistantMessage): boolean {\n\tif (message.stopReason !== \"error\" || !message.errorMessage) return false;\n\tconst errorMessage = message.errorMessage;\n\tif (NON_RETRYABLE_PROVIDER_LIMIT_ERROR_PATTERN.test(errorMessage)) return false;\n\treturn RETRYABLE_PROVIDER_ERROR_PATTERN.test(errorMessage);\n}\n"]}

package/dist/utils/retry.js

@@ -0,0 +1,170 @@
+function buildProviderErrorPattern(patterns) {
+    return new RegExp(patterns.join("|"), "i");
+}
+const NON_RETRYABLE_PROVIDER_LIMIT_ERROR_PATTERN = buildProviderErrorPattern([
+    // OpenCode Go/free-tier limits returned as 429 JSON error types by OpenCode's
+    // Zen API. These are subscription/account limits, not transient throttles.
+    "GoUsageLimitError",
+    "FreeUsageLimitError",
+    // OpenCode Go subscription-limit text asks users to enable available-balance
+    // usage after rolling/weekly/monthly limits are reached.
+    "Monthly usage limit reached",
+    "available balance",
+    // Generic quota/budget/billing exhaustion. `insufficient_quota` is OpenAI's
+    // quota/billing error code; the other strings cover common gateway wording.
+    "insufficient_quota",
+    "out of budget",
+    "quota exceeded",
+    "billing",
+]);
+const RETRYABLE_PROVIDER_ERROR_PATTERN = buildProviderErrorPattern([
+    // Generic provider load, HTTP status, and server-side transient failures.
+    "overloaded",
+    "rate.?limit",
+    "too many requests",
+    "429",
+    "500",
+    "502",
+    "503",
+    "504",
+    "524",
+    "service.?unavailable",
+    "server.?error",
+    "internal.?error",
+    // Wrapper/provider text for transient upstream failures, including OpenRouter
+    // "Provider returned error" responses (#2264).
+    "provider.?returned.?error",
+    // Network, proxy, and fetch transport failures. This includes OpenAI Codex
+    // raw-fetch failures such as "upstream connect", "connection refused", and
+    // "reset before headers" (#733), plus OpenRouter connection drops (#3317).
+    "network.?error",
+    "connection.?error",
+    "connection.?refused",
+    "connection.?lost",
+    "other side closed",
+    "fetch failed",
+    "upstream.?connect",
+    "reset before headers",
+    "socket hang up",
+    "socket connection was closed",
+    "timed? out",
+    "timeout",
+    "terminated",
+    // WebSocket transports can report close/error text instead of HTTP/fetch text.
+    "websocket.?closed",
+    "websocket.?error",
+    // Premature stream endings from SDKs and transports. Anthropic can throw
+    // "stream ended without ..." and "Anthropic stream ended before message_stop"
+    // (#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).
+    "retry delay",
+    // Explicit retry guidance emitted mid-stream by OpenAI Responses and Bedrock
+    // stream exceptions (#6019).
+    "you can retry your request",
+    "try your request again",
+    "please retry your request",
+    // 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
+ * restarted.
+ *
+ * This does not implement retry policy. Callers should first handle context
+ * overflow separately, then apply their own retry budget, backoff, and reporting
+ * before restarting the assistant turn.
+ */
+export function isRetryableAssistantError(message) {
+    if (message.stopReason !== "error" || !message.errorMessage)
+        return false;
+    const errorMessage = message.errorMessage;
+    if (NON_RETRYABLE_PROVIDER_LIMIT_ERROR_PATTERN.test(errorMessage))
+        return false;
+    return RETRYABLE_PROVIDER_ERROR_PATTERN.test(errorMessage);
+}
+//# sourceMappingURL=retry.js.map

package/dist/utils/retry.js.map

@@ -0,0 +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,+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,cAAc;QACb,KAAK,CAAC,SAAS,CAAC,CAAC;IAAA,CACjB;CACD;AAED,SAAS,KAAK,CAAC,EAAU,EAAE,MAAoB,EAAiB;IAC/D,OAAO,IAAI,OAAO,CAAC,CAAC,OAAO,EAAE,MAAM,EAAE,EAAE,CAAC;QACvC,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,CAAC;YACL,YAAY,CAAC,OAAO,CAAC,CAAC;YACtB,MAAM,CAAC,IAAI,oBAAoB,EAAE,CAAC,CAAC;QAAA,CACnC,EACD,EAAE,IAAI,EAAE,IAAI,EAAE,CACd,CAAC;IAAA,CACF,CAAC,CAAC;AAAA,CACH;AAED;;;;;;;;;;;;;;;;;GAiBG;AACH,MAAM,CAAC,KAAK,UAAU,kBAAkB,CACvC,OAAwC,EACxC,MAA+B,EAC/B,MAA+B,EAC/B,SAA0B,EACE;IAC5B,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;AAAA,CACD;AAED;;;;;;;;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 * Retry policy: bounded attempts with exponential backoff (`baseDelayMs * 2^(attempt-1)`).\n * Matches `settings.retry` (`enabled`, `maxRetries`, `baseDelayMs`) in coding-agent; kept\n * here so the classifier and the policy-driven retry loop live together and stay reusable\n * by the SDK and other callers.\n */\nexport interface RetryPolicy {\n\tenabled: boolean;\n\t/** Max retry attempts (0 = no retries). The initial call never counts as a retry. */\n\tmaxRetries: number;\n\t/** Base delay in ms. Per-attempt delay is `baseDelayMs * 2^(attempt-1)` before jitter. */\n\tbaseDelayMs: number;\n}\n\n/** Optional callbacks emitted by {@link retryAssistantCall} around each retry. */\nexport interface RetryCallbacks {\n\t/** Emitted before the backoff sleep of each retry attempt (1-indexed). */\n\tonRetryScheduled?: (\n\t\tattempt: number,\n\t\tmaxAttempts: number,\n\t\tdelayMs: number,\n\t\terrorMessage: string,\n\t) => void | Promise<void>;\n\t/** Emitted after the backoff sleep, immediately before the retried call starts. */\n\tonRetryAttemptStart?: () => void | Promise<void>;\n\t/** Emitted once when the loop ends: success if a later call completed normally. */\n\tonRetryFinished?: (success: boolean, attempt: number, finalError?: string) => void | Promise<void>;\n}\n\nclass RetrySleepAbortError extends Error {\n\tconstructor() {\n\t\tsuper(\"Aborted\");\n\t}\n}\n\nfunction sleep(ms: number, signal?: AbortSignal): Promise<void> {\n\treturn new Promise((resolve, reject) => {\n\t\tif (signal?.aborted) {\n\t\t\treject(new RetrySleepAbortError());\n\t\t\treturn;\n\t\t}\n\t\tconst timeout = setTimeout(resolve, ms);\n\t\tsignal?.addEventListener(\n\t\t\t\"abort\",\n\t\t\t() => {\n\t\t\t\tclearTimeout(timeout);\n\t\t\t\treject(new RetrySleepAbortError());\n\t\t\t},\n\t\t\t{ once: true },\n\t\t);\n\t});\n}\n\n/**\n * Run a single assistant-producing call with bounded retry on transient errors.\n *\n * Behavior:\n * - A successful response is returned immediately. Aborts are terminal and never\n *   retried, but reported as unsuccessful if they happen after a retry was scheduled.\n *   Aborts during the backoff sleep are normalized to an aborted `AssistantMessage`\n *   too, so callers do not need to care when cancellation happened.\n * - A non-retryable error (per {@link isRetryableAssistantError}, including quota/\n *   billing exhaustion) is returned immediately so deterministic errors fail fast.\n * - Otherwise retries up to `maxRetries` times with exponential backoff, emitting\n *   `onRetryScheduled` before each sleep, `onRetryAttemptStart` after each sleep before\n *   the retried call starts, and `onRetryFinished` once at the end (whether the loop\n *   ends in success, exhausted retries, or an aborted backoff).\n *\n * When `policy` is undefined or disabled, the first response is returned unchanged\n * (equivalent to calling `produce()` directly).\n */\nexport async function retryAssistantCall(\n\tproduce: () => Promise<AssistantMessage>,\n\tpolicy: RetryPolicy | undefined,\n\tsignal: AbortSignal | undefined,\n\tcallbacks?: RetryCallbacks,\n): Promise<AssistantMessage> {\n\tconst maxAttempts = policy?.enabled ? policy.maxRetries : 0;\n\n\tlet attempt = 0;\n\tlet lastRetry: { attempt: number; errorMessage: string } | undefined;\n\tfor (;;) {\n\t\tconst response = await produce();\n\n\t\t// Abort: terminal but not successful. Never retry an aborted message.\n\t\tif (response.stopReason === \"aborted\") {\n\t\t\tif (lastRetry) await callbacks?.onRetryFinished?.(false, lastRetry.attempt);\n\t\t\treturn response;\n\t\t}\n\n\t\t// Success: non-error, non-abort responses return as-is.\n\t\tif (response.stopReason !== \"error\") {\n\t\t\tif (lastRetry) await callbacks?.onRetryFinished?.(true, lastRetry.attempt);\n\t\t\treturn response;\n\t\t}\n\n\t\t// Non-retryable, or budget exhausted: return the final error message.\n\t\tif (attempt >= maxAttempts || !isRetryableAssistantError(response)) {\n\t\t\tif (lastRetry) await callbacks?.onRetryFinished?.(false, lastRetry.attempt, response.errorMessage);\n\t\t\treturn response;\n\t\t}\n\n\t\tattempt++;\n\t\tlastRetry = { attempt, errorMessage: response.errorMessage || \"Unknown error\" };\n\t\tconst delayMs = policy!.baseDelayMs * 2 ** (attempt - 1);\n\t\tawait callbacks?.onRetryScheduled?.(attempt, maxAttempts, delayMs, lastRetry.errorMessage);\n\n\t\t// Normalize aborts during retry backoff to the same AssistantMessage shape as\n\t\t// provider stream aborts, so callers do not need to care when cancellation happened.\n\t\ttry {\n\t\t\tawait sleep(delayMs, signal);\n\t\t} catch (error) {\n\t\t\tawait callbacks?.onRetryFinished?.(false, attempt, lastRetry.errorMessage);\n\t\t\tif (error instanceof RetrySleepAbortError) {\n\t\t\t\treturn { ...response, stopReason: \"aborted\", errorMessage: undefined };\n\t\t\t}\n\t\t\tthrow error;\n\t\t}\n\t\tawait callbacks?.onRetryAttemptStart?.();\n\t}\n}\n\n/**\n * Classifies whether a failed assistant message looks like a transient provider\n * or transport error, so callers can decide if the last assistant turn should be\n * restarted.\n *\n * This does not implement retry policy. Callers should first handle context\n * overflow separately, then apply their own retry budget, backoff, and reporting\n * before restarting the assistant turn.\n */\nexport function isRetryableAssistantError(message: AssistantMessage): boolean {\n\tif (message.stopReason !== \"error\" || !message.errorMessage) return false;\n\tconst errorMessage = message.errorMessage;\n\tif (NON_RETRYABLE_PROVIDER_LIMIT_ERROR_PATTERN.test(errorMessage)) return false;\n\treturn RETRYABLE_PROVIDER_ERROR_PATTERN.test(errorMessage);\n}\n"]}