@asm-agent/ai 0.8.2

package/dist/utils/oauth/types.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"types.d.ts","sourceRoot":"","sources":["../../../src/utils/oauth/types.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,GAAG,EAAE,KAAK,EAAE,MAAM,gBAAgB,CAAC;AAEjD,MAAM,MAAM,gBAAgB,GAAG;IAC9B,OAAO,EAAE,MAAM,CAAC;IAChB,MAAM,EAAE,MAAM,CAAC;IACf,OAAO,EAAE,MAAM,CAAC;IAChB,CAAC,GAAG,EAAE,MAAM,GAAG,OAAO,CAAC;CACvB,CAAC;AAEF,MAAM,MAAM,eAAe,GAAG,MAAM,CAAC;AAErC,8CAA8C;AAC9C,MAAM,MAAM,aAAa,GAAG,eAAe,CAAC;AAE5C,MAAM,MAAM,WAAW,GAAG;IACzB,OAAO,EAAE,MAAM,CAAC;IAChB,WAAW,CAAC,EAAE,MAAM,CAAC;IACrB,UAAU,CAAC,EAAE,OAAO,CAAC;CACrB,CAAC;AAEF,MAAM,MAAM,aAAa,GAAG;IAC3B,GAAG,EAAE,MAAM,CAAC;IACZ,YAAY,CAAC,EAAE,MAAM,CAAC;CACtB,CAAC;AAEF,MAAM,MAAM,iBAAiB,GAAG;IAC/B,EAAE,EAAE,MAAM,CAAC;IACX,KAAK,EAAE,MAAM,CAAC;CACd,CAAC;AAEF,MAAM,MAAM,iBAAiB,GAAG;IAC/B,OAAO,EAAE,MAAM,CAAC;IAChB,OAAO,EAAE,iBAAiB,EAAE,CAAC;CAC7B,CAAC;AAEF,MAAM,WAAW,mBAAmB;IACnC,MAAM,EAAE,CAAC,IAAI,EAAE,aAAa,KAAK,IAAI,CAAC;IACtC,QAAQ,EAAE,CAAC,MAAM,EAAE,WAAW,KAAK,OAAO,CAAC,MAAM,CAAC,CAAC;IACnD,UAAU,CAAC,EAAE,CAAC,OAAO,EAAE,MAAM,KAAK,IAAI,CAAC;IACvC,iBAAiB,CAAC,EAAE,MAAM,OAAO,CAAC,MAAM,CAAC,CAAC;IAC1C,8FAA8F;IAC9F,QAAQ,CAAC,EAAE,CAAC,MAAM,EAAE,iBAAiB,KAAK,OAAO,CAAC,MAAM,GAAG,SAAS,CAAC,CAAC;IACtE,MAAM,CAAC,EAAE,WAAW,CAAC;CACrB;AAED,MAAM,WAAW,sBAAsB;IACtC,QAAQ,CAAC,EAAE,EAAE,eAAe,CAAC;IAC7B,QAAQ,CAAC,IAAI,EAAE,MAAM,CAAC;IAEtB,wDAAwD;IACxD,KAAK,CAAC,SAAS,EAAE,mBAAmB,GAAG,OAAO,CAAC,gBAAgB,CAAC,CAAC;IAEjE,iFAAiF;IACjF,kBAAkB,CAAC,EAAE,OAAO,CAAC;IAE7B,yEAAyE;IACzE,YAAY,CAAC,WAAW,EAAE,gBAAgB,GAAG,OAAO,CAAC,gBAAgB,CAAC,CAAC;IAEvE,6DAA6D;IAC7D,SAAS,CAAC,WAAW,EAAE,gBAAgB,GAAG,MAAM,CAAC;IAEjD,uEAAuE;IACvE,YAAY,CAAC,CAAC,MAAM,EAAE,KAAK,CAAC,GAAG,CAAC,EAAE,EAAE,WAAW,EAAE,gBAAgB,GAAG,KAAK,CAAC,GAAG,CAAC,EAAE,CAAC;CACjF;AAED,qDAAqD;AACrD,MAAM,WAAW,iBAAiB;IACjC,EAAE,EAAE,eAAe,CAAC;IACpB,IAAI,EAAE,MAAM,CAAC;IACb,SAAS,EAAE,OAAO,CAAC;CACnB","sourcesContent":["import type { Api, Model } from \"../../types.js\";\n\nexport type OAuthCredentials = {\n\trefresh: string;\n\taccess: string;\n\texpires: number;\n\t[key: string]: unknown;\n};\n\nexport type OAuthProviderId = string;\n\n/** @deprecated Use OAuthProviderId instead */\nexport type OAuthProvider = OAuthProviderId;\n\nexport type OAuthPrompt = {\n\tmessage: string;\n\tplaceholder?: string;\n\tallowEmpty?: boolean;\n};\n\nexport type OAuthAuthInfo = {\n\turl: string;\n\tinstructions?: string;\n};\n\nexport type OAuthSelectOption = {\n\tid: string;\n\tlabel: string;\n};\n\nexport type OAuthSelectPrompt = {\n\tmessage: string;\n\toptions: OAuthSelectOption[];\n};\n\nexport interface OAuthLoginCallbacks {\n\tonAuth: (info: OAuthAuthInfo) => void;\n\tonPrompt: (prompt: OAuthPrompt) => Promise<string>;\n\tonProgress?: (message: string) => void;\n\tonManualCodeInput?: () => Promise<string>;\n\t/** Show an interactive selector and return the selected option id, or undefined on cancel. */\n\tonSelect?: (prompt: OAuthSelectPrompt) => Promise<string | undefined>;\n\tsignal?: AbortSignal;\n}\n\nexport interface OAuthProviderInterface {\n\treadonly id: OAuthProviderId;\n\treadonly name: string;\n\n\t/** Run the login flow, return credentials to persist */\n\tlogin(callbacks: OAuthLoginCallbacks): Promise<OAuthCredentials>;\n\n\t/** Whether login uses a local callback server and supports manual code input. */\n\tusesCallbackServer?: boolean;\n\n\t/** Refresh expired credentials, return updated credentials to persist */\n\trefreshToken(credentials: OAuthCredentials): Promise<OAuthCredentials>;\n\n\t/** Convert credentials to API key string for the provider */\n\tgetApiKey(credentials: OAuthCredentials): string;\n\n\t/** Optional: modify models for this provider (e.g., update baseUrl) */\n\tmodifyModels?(models: Model<Api>[], credentials: OAuthCredentials): Model<Api>[];\n}\n\n/** @deprecated Use OAuthProviderInterface instead */\nexport interface OAuthProviderInfo {\n\tid: OAuthProviderId;\n\tname: string;\n\tavailable: boolean;\n}\n"]}

package/dist/utils/oauth/types.js

@@ -0,0 +1,2 @@
+export {};
+//# sourceMappingURL=types.js.map

package/dist/utils/oauth/types.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"types.js","sourceRoot":"","sources":["../../../src/utils/oauth/types.ts"],"names":[],"mappings":"","sourcesContent":["import type { Api, Model } from \"../../types.js\";\n\nexport type OAuthCredentials = {\n\trefresh: string;\n\taccess: string;\n\texpires: number;\n\t[key: string]: unknown;\n};\n\nexport type OAuthProviderId = string;\n\n/** @deprecated Use OAuthProviderId instead */\nexport type OAuthProvider = OAuthProviderId;\n\nexport type OAuthPrompt = {\n\tmessage: string;\n\tplaceholder?: string;\n\tallowEmpty?: boolean;\n};\n\nexport type OAuthAuthInfo = {\n\turl: string;\n\tinstructions?: string;\n};\n\nexport type OAuthSelectOption = {\n\tid: string;\n\tlabel: string;\n};\n\nexport type OAuthSelectPrompt = {\n\tmessage: string;\n\toptions: OAuthSelectOption[];\n};\n\nexport interface OAuthLoginCallbacks {\n\tonAuth: (info: OAuthAuthInfo) => void;\n\tonPrompt: (prompt: OAuthPrompt) => Promise<string>;\n\tonProgress?: (message: string) => void;\n\tonManualCodeInput?: () => Promise<string>;\n\t/** Show an interactive selector and return the selected option id, or undefined on cancel. */\n\tonSelect?: (prompt: OAuthSelectPrompt) => Promise<string | undefined>;\n\tsignal?: AbortSignal;\n}\n\nexport interface OAuthProviderInterface {\n\treadonly id: OAuthProviderId;\n\treadonly name: string;\n\n\t/** Run the login flow, return credentials to persist */\n\tlogin(callbacks: OAuthLoginCallbacks): Promise<OAuthCredentials>;\n\n\t/** Whether login uses a local callback server and supports manual code input. */\n\tusesCallbackServer?: boolean;\n\n\t/** Refresh expired credentials, return updated credentials to persist */\n\trefreshToken(credentials: OAuthCredentials): Promise<OAuthCredentials>;\n\n\t/** Convert credentials to API key string for the provider */\n\tgetApiKey(credentials: OAuthCredentials): string;\n\n\t/** Optional: modify models for this provider (e.g., update baseUrl) */\n\tmodifyModels?(models: Model<Api>[], credentials: OAuthCredentials): Model<Api>[];\n}\n\n/** @deprecated Use OAuthProviderInterface instead */\nexport interface OAuthProviderInfo {\n\tid: OAuthProviderId;\n\tname: string;\n\tavailable: boolean;\n}\n"]}

package/dist/utils/overflow.d.ts

@@ -0,0 +1,55 @@
+import type { AssistantMessage } from "../types.js";
+/**
+ * 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"
+ * - 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 (all backends): "maximum context length is X 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)"
+ *
+ * **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;AAoEpD;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA+CG;AACH,wBAAgB,iBAAiB,CAAC,OAAO,EAAE,gBAAgB,EAAE,aAAa,CAAC,EAAE,MAAM,GAAG,OAAO,CA2B5F;AAED;;GAEG;AACH,wBAAgB,mBAAmB,IAAI,MAAM,EAAE,CAE9C","sourcesContent":["import type { AssistantMessage } from \"../types.js\";\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 * - 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 * - 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 * - 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/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 (all backends)\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/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\"\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 (all backends): \"maximum context length is X 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 *\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\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\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 +1,144 @@
+/**
+ * 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"
+ * - 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"
+ * - 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)"
+ * - 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.
+ * - 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)
+    /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 (all backends)
+    /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
+    /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
+    /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"
+ * - 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 (all backends): "maximum context length is X 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)"
+ *
+ * **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) {
+    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;
+        }
+    }
+    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;;;;;;;;;;;;;;;;;;;;;;;;;;;GA2BG;AACH,MAAM,iBAAiB,GAAG;IACzB,qBAAqB,EAAE,2BAA2B;IAClD,oBAAoB,EAAE,kDAAkD;IACxE,wCAAwC,EAAE,iBAAiB;IAC3D,6BAA6B,EAAE,uCAAuC;IACtE,yCAAyC,EAAE,kBAAkB;IAC7D,+BAA+B,EAAE,aAAa;IAC9C,oCAAoC,EAAE,OAAO;IAC7C,uCAAuC,EAAE,4BAA4B;IACrE,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,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;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA+CG;AACH,MAAM,UAAU,iBAAiB,CAAC,OAAyB,EAAE,aAAsB,EAAW;IAC7F,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,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.js\";\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 * - 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 * - 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 * - 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/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 (all backends)\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/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\"\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 (all backends): \"maximum context length is X 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 *\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\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\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/sanitize-unicode.d.ts

@@ -0,0 +1,22 @@
+/**
+ * Removes unpaired Unicode surrogate characters from a string.
+ *
+ * Unpaired surrogates (high surrogates 0xD800-0xDBFF without matching low surrogates 0xDC00-0xDFFF,
+ * or vice versa) cause JSON serialization errors in many API providers.
+ *
+ * Valid emoji and other characters outside the Basic Multilingual Plane use properly paired
+ * surrogates and will NOT be affected by this function.
+ *
+ * @param text - The text to sanitize
+ * @returns The sanitized text with unpaired surrogates removed
+ *
+ * @example
+ * // Valid emoji (properly paired surrogates) are preserved
+ * sanitizeSurrogates("Hello 🙈 World") // => "Hello 🙈 World"
+ *
+ * // Unpaired high surrogate is removed
+ * const unpaired = String.fromCharCode(0xD83D); // high surrogate without low
+ * sanitizeSurrogates(`Text ${unpaired} here`) // => "Text  here"
+ */
+export declare function sanitizeSurrogates(text: string): string;
+//# sourceMappingURL=sanitize-unicode.d.ts.map

package/dist/utils/sanitize-unicode.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"sanitize-unicode.d.ts","sourceRoot":"","sources":["../../src/utils/sanitize-unicode.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;GAmBG;AACH,wBAAgB,kBAAkB,CAAC,IAAI,EAAE,MAAM,GAAG,MAAM,CAIvD","sourcesContent":["/**\n * Removes unpaired Unicode surrogate characters from a string.\n *\n * Unpaired surrogates (high surrogates 0xD800-0xDBFF without matching low surrogates 0xDC00-0xDFFF,\n * or vice versa) cause JSON serialization errors in many API providers.\n *\n * Valid emoji and other characters outside the Basic Multilingual Plane use properly paired\n * surrogates and will NOT be affected by this function.\n *\n * @param text - The text to sanitize\n * @returns The sanitized text with unpaired surrogates removed\n *\n * @example\n * // Valid emoji (properly paired surrogates) are preserved\n * sanitizeSurrogates(\"Hello 🙈 World\") // => \"Hello 🙈 World\"\n *\n * // Unpaired high surrogate is removed\n * const unpaired = String.fromCharCode(0xD83D); // high surrogate without low\n * sanitizeSurrogates(`Text ${unpaired} here`) // => \"Text  here\"\n */\nexport function sanitizeSurrogates(text: string): string {\n\t// Replace unpaired high surrogates (0xD800-0xDBFF not followed by low surrogate)\n\t// Replace unpaired low surrogates (0xDC00-0xDFFF not preceded by high surrogate)\n\treturn text.replace(/[\\uD800-\\uDBFF](?![\\uDC00-\\uDFFF])|(?<![\\uD800-\\uDBFF])[\\uDC00-\\uDFFF]/g, \"\");\n}\n"]}

package/dist/utils/sanitize-unicode.js

@@ -0,0 +1,26 @@
+/**
+ * Removes unpaired Unicode surrogate characters from a string.
+ *
+ * Unpaired surrogates (high surrogates 0xD800-0xDBFF without matching low surrogates 0xDC00-0xDFFF,
+ * or vice versa) cause JSON serialization errors in many API providers.
+ *
+ * Valid emoji and other characters outside the Basic Multilingual Plane use properly paired
+ * surrogates and will NOT be affected by this function.
+ *
+ * @param text - The text to sanitize
+ * @returns The sanitized text with unpaired surrogates removed
+ *
+ * @example
+ * // Valid emoji (properly paired surrogates) are preserved
+ * sanitizeSurrogates("Hello 🙈 World") // => "Hello 🙈 World"
+ *
+ * // Unpaired high surrogate is removed
+ * const unpaired = String.fromCharCode(0xD83D); // high surrogate without low
+ * sanitizeSurrogates(`Text ${unpaired} here`) // => "Text  here"
+ */
+export function sanitizeSurrogates(text) {
+    // Replace unpaired high surrogates (0xD800-0xDBFF not followed by low surrogate)
+    // Replace unpaired low surrogates (0xDC00-0xDFFF not preceded by high surrogate)
+    return text.replace(/[\uD800-\uDBFF](?![\uDC00-\uDFFF])|(?<![\uD800-\uDBFF])[\uDC00-\uDFFF]/g, "");
+}
+//# sourceMappingURL=sanitize-unicode.js.map

package/dist/utils/sanitize-unicode.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"sanitize-unicode.js","sourceRoot":"","sources":["../../src/utils/sanitize-unicode.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;GAmBG;AACH,MAAM,UAAU,kBAAkB,CAAC,IAAY,EAAU;IACxD,iFAAiF;IACjF,iFAAiF;IACjF,OAAO,IAAI,CAAC,OAAO,CAAC,yEAAyE,EAAE,EAAE,CAAC,CAAC;AAAA,CACnG","sourcesContent":["/**\n * Removes unpaired Unicode surrogate characters from a string.\n *\n * Unpaired surrogates (high surrogates 0xD800-0xDBFF without matching low surrogates 0xDC00-0xDFFF,\n * or vice versa) cause JSON serialization errors in many API providers.\n *\n * Valid emoji and other characters outside the Basic Multilingual Plane use properly paired\n * surrogates and will NOT be affected by this function.\n *\n * @param text - The text to sanitize\n * @returns The sanitized text with unpaired surrogates removed\n *\n * @example\n * // Valid emoji (properly paired surrogates) are preserved\n * sanitizeSurrogates(\"Hello 🙈 World\") // => \"Hello 🙈 World\"\n *\n * // Unpaired high surrogate is removed\n * const unpaired = String.fromCharCode(0xD83D); // high surrogate without low\n * sanitizeSurrogates(`Text ${unpaired} here`) // => \"Text  here\"\n */\nexport function sanitizeSurrogates(text: string): string {\n\t// Replace unpaired high surrogates (0xD800-0xDBFF not followed by low surrogate)\n\t// Replace unpaired low surrogates (0xDC00-0xDFFF not preceded by high surrogate)\n\treturn text.replace(/[\\uD800-\\uDBFF](?![\\uDC00-\\uDFFF])|(?<![\\uD800-\\uDBFF])[\\uDC00-\\uDFFF]/g, \"\");\n}\n"]}

package/dist/utils/stream-failure.d.ts

@@ -0,0 +1,55 @@
+import type { AssistantMessage } from "../types.js";
+/**
+ * Shared classification and reporting for provider stream failures, so no
+ * provider collapses a specific cause (refusal, safety filter, overload, ...)
+ * into a generic string before it is logged and persisted.
+ */
+export type StreamFailureKind = "refusal" | "safety" | "overloaded" | "rate_limit" | "server_error" | "auth" | "invalid_request" | "malformed_response" | "unknown";
+export interface StreamFailureInfo {
+    kind: StreamFailureKind;
+    /** Provider's own error/stop identifier, e.g. "overloaded_error" or "SAFETY". */
+    providerErrorType?: string;
+    status?: number;
+    requestId?: string;
+    /** Truncated raw provider payload for post-mortems. */
+    raw?: string;
+}
+export declare class StreamFailureError extends Error {
+    readonly info: StreamFailureInfo;
+    constructor(message: string, info: StreamFailureInfo);
+}
+/** Build a user-facing message like "Provider overloaded (overloaded_error, 529) [request_id: req_abc]". */
+export declare function streamFailureMessage(info: StreamFailureInfo, detail?: string): string;
+export declare function classifyStreamFailure(providerErrorType?: string, status?: number): StreamFailureKind;
+/**
+ * Failure for a stream that terminated with a provider stop/finish reason that
+ * maps to "error" (e.g. Anthropic "refusal", Gemini "SAFETY"). Providers call
+ * this instead of throwing a generic error, so the raw reason survives.
+ */
+export declare function streamFailureFromStopReason(rawStopReason: string | undefined, extra?: Pick<StreamFailureInfo, "requestId">): StreamFailureError;
+export declare function truncateRawPayload(raw: string): string;
+/**
+ * Best-effort extraction of structured failure info from any thrown value:
+ * StreamFailureError, provider SDK errors (Anthropic/OpenAI APIError, AWS SDK
+ * exceptions, Google ApiError), or plain errors.
+ */
+export declare function extractStreamFailureInfo(error: unknown): StreamFailureInfo;
+/**
+ * User-facing message for a thrown stream error: a classified one-liner with
+ * the provider's own short message, never the raw payload/trace. Unrecognized
+ * errors pass through verbatim so their text (which downstream retry matching
+ * may depend on) is preserved.
+ */
+export declare function formatStreamFailureMessage(error: unknown): string;
+/**
+ * Record a terminal stream failure on the message (structured diagnostic that
+ * persists to session JSONL) and emit one structured log line. Call from the
+ * provider's terminal catch after stopReason/errorMessage are set; no-op for
+ * user-initiated aborts.
+ */
+export declare function recordStreamFailure(model: {
+    provider: string;
+    id: string;
+    api: string;
+}, output: AssistantMessage, error: unknown): void;
+//# sourceMappingURL=stream-failure.d.ts.map

package/dist/utils/stream-failure.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"stream-failure.d.ts","sourceRoot":"","sources":["../../src/utils/stream-failure.ts"],"names":[],"mappings":"AACA,OAAO,KAAK,EAAE,gBAAgB,EAAE,MAAM,aAAa,CAAC;AAGpD;;;;GAIG;AAEH,MAAM,MAAM,iBAAiB,GAC1B,SAAS,GACT,QAAQ,GACR,YAAY,GACZ,YAAY,GACZ,cAAc,GACd,MAAM,GACN,iBAAiB,GACjB,oBAAoB,GACpB,SAAS,CAAC;AAEb,MAAM,WAAW,iBAAiB;IACjC,IAAI,EAAE,iBAAiB,CAAC;IACxB,iFAAiF;IACjF,iBAAiB,CAAC,EAAE,MAAM,CAAC;IAC3B,MAAM,CAAC,EAAE,MAAM,CAAC;IAChB,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,uDAAuD;IACvD,GAAG,CAAC,EAAE,MAAM,CAAC;CACb;AAED,qBAAa,kBAAmB,SAAQ,KAAK;IAC5C,QAAQ,CAAC,IAAI,EAAE,iBAAiB,CAAC;IAEjC,YAAY,OAAO,EAAE,MAAM,EAAE,IAAI,EAAE,iBAAiB,EAInD;CACD;AAcD,4GAA4G;AAC5G,wBAAgB,oBAAoB,CAAC,IAAI,EAAE,iBAAiB,EAAE,MAAM,CAAC,EAAE,MAAM,GAAG,MAAM,CASrF;AAED,wBAAgB,qBAAqB,CAAC,iBAAiB,CAAC,EAAE,MAAM,EAAE,MAAM,CAAC,EAAE,MAAM,GAAG,iBAAiB,CAsBpG;AAED;;;;GAIG;AACH,wBAAgB,2BAA2B,CAC1C,aAAa,EAAE,MAAM,GAAG,SAAS,EACjC,KAAK,CAAC,EAAE,IAAI,CAAC,iBAAiB,EAAE,WAAW,CAAC,GAC1C,kBAAkB,CAWpB;AAID,wBAAgB,kBAAkB,CAAC,GAAG,EAAE,MAAM,GAAG,MAAM,CAEtD;AA2DD;;;;GAIG;AACH,wBAAgB,wBAAwB,CAAC,KAAK,EAAE,OAAO,GAAG,iBAAiB,CAE1E;AAED;;;;;GAKG;AACH,wBAAgB,0BAA0B,CAAC,KAAK,EAAE,OAAO,GAAG,MAAM,CAOjE;AAID;;;;;GAKG;AACH,wBAAgB,mBAAmB,CAClC,KAAK,EAAE;IAAE,QAAQ,EAAE,MAAM,CAAC;IAAC,EAAE,EAAE,MAAM,CAAC;IAAC,GAAG,EAAE,MAAM,CAAA;CAAE,EACpD,MAAM,EAAE,gBAAgB,EACxB,KAAK,EAAE,OAAO,GACZ,IAAI,CAsBN","sourcesContent":["import { getLogger } from \"../log.js\";\nimport type { AssistantMessage } from \"../types.js\";\nimport { appendAssistantMessageDiagnostic, extractDiagnosticError } from \"./diagnostics.js\";\n\n/**\n * Shared classification and reporting for provider stream failures, so no\n * provider collapses a specific cause (refusal, safety filter, overload, ...)\n * into a generic string before it is logged and persisted.\n */\n\nexport type StreamFailureKind =\n\t| \"refusal\"\n\t| \"safety\"\n\t| \"overloaded\"\n\t| \"rate_limit\"\n\t| \"server_error\"\n\t| \"auth\"\n\t| \"invalid_request\"\n\t| \"malformed_response\"\n\t| \"unknown\";\n\nexport interface StreamFailureInfo {\n\tkind: StreamFailureKind;\n\t/** Provider's own error/stop identifier, e.g. \"overloaded_error\" or \"SAFETY\". */\n\tproviderErrorType?: string;\n\tstatus?: number;\n\trequestId?: string;\n\t/** Truncated raw provider payload for post-mortems. */\n\traw?: string;\n}\n\nexport class StreamFailureError extends Error {\n\treadonly info: StreamFailureInfo;\n\n\tconstructor(message: string, info: StreamFailureInfo) {\n\t\tsuper(message);\n\t\tthis.name = \"StreamFailureError\";\n\t\tthis.info = info;\n\t}\n}\n\nconst KIND_MESSAGES: Record<StreamFailureKind, string> = {\n\trefusal: \"Model refused to respond\",\n\tsafety: \"Response blocked by provider safety filters\",\n\toverloaded: \"Provider overloaded\",\n\trate_limit: \"Provider rate limit exceeded\",\n\tserver_error: \"Provider server error\",\n\tauth: \"Provider authentication failed\",\n\tinvalid_request: \"Provider rejected the request\",\n\tmalformed_response: \"Provider returned a malformed response\",\n\tunknown: \"Provider stream failed\",\n};\n\n/** Build a user-facing message like \"Provider overloaded (overloaded_error, 529) [request_id: req_abc]\". */\nexport function streamFailureMessage(info: StreamFailureInfo, detail?: string): string {\n\tconst qualifiers = [info.providerErrorType, info.status !== undefined ? String(info.status) : undefined]\n\t\t.filter(Boolean)\n\t\t.join(\", \");\n\tlet message = KIND_MESSAGES[info.kind];\n\tif (qualifiers) message += ` (${qualifiers})`;\n\tif (detail) message += `: ${detail}`;\n\tif (info.requestId) message += ` [request_id: ${info.requestId}]`;\n\treturn message;\n}\n\nexport function classifyStreamFailure(providerErrorType?: string, status?: number): StreamFailureKind {\n\tconst type = providerErrorType?.toLowerCase() ?? \"\";\n\tif (type === \"refusal\") return \"refusal\";\n\tif (/sensitive|safety|prohibited_content|blocklist|spii|recitation|content.?filter|guardrail|flagged/.test(type)) {\n\t\treturn \"safety\";\n\t}\n\tif (type.includes(\"overloaded\") || status === 529) return \"overloaded\";\n\tif (type.includes(\"rate_limit\") || type.includes(\"throttl\") || status === 429) return \"rate_limit\";\n\tif (/authentication|permission|unauthorized/.test(type) || status === 401 || status === 403) return \"auth\";\n\tif (type.includes(\"invalid_request\") || type.includes(\"not_found_error\") || status === 400 || status === 404) {\n\t\treturn \"invalid_request\";\n\t}\n\tif (type.includes(\"malformed\")) return \"malformed_response\";\n\tif (\n\t\ttype.includes(\"api_error\") ||\n\t\ttype.includes(\"server_error\") ||\n\t\ttype.includes(\"unavailable\") ||\n\t\t(status !== undefined && status >= 500)\n\t) {\n\t\treturn \"server_error\";\n\t}\n\treturn \"unknown\";\n}\n\n/**\n * Failure for a stream that terminated with a provider stop/finish reason that\n * maps to \"error\" (e.g. Anthropic \"refusal\", Gemini \"SAFETY\"). Providers call\n * this instead of throwing a generic error, so the raw reason survives.\n */\nexport function streamFailureFromStopReason(\n\trawStopReason: string | undefined,\n\textra?: Pick<StreamFailureInfo, \"requestId\">,\n): StreamFailureError {\n\tconst info: StreamFailureInfo = {\n\t\tkind: rawStopReason ? classifyStreamFailure(rawStopReason) : \"unknown\",\n\t\tproviderErrorType: rawStopReason,\n\t\trequestId: extra?.requestId,\n\t};\n\tif (info.kind === \"unknown\" && /malformed/i.test(rawStopReason ?? \"\")) info.kind = \"malformed_response\";\n\tconst message = rawStopReason\n\t\t? streamFailureMessage(info)\n\t\t: streamFailureMessage(info, \"stream ended with an error and no stop reason\");\n\treturn new StreamFailureError(message, info);\n}\n\nconst MAX_RAW_LENGTH = 2000;\n\nexport function truncateRawPayload(raw: string): string {\n\treturn raw.length > MAX_RAW_LENGTH ? `${raw.slice(0, MAX_RAW_LENGTH)}…` : raw;\n}\n\nfunction extractStreamFailureParts(error: unknown): { info: StreamFailureInfo; detail?: string } {\n\tif (error instanceof StreamFailureError) return { info: error.info };\n\tif (!(error instanceof Error)) return { info: { kind: \"unknown\" } };\n\n\tconst err = error as Error & {\n\t\tstatus?: unknown;\n\t\tstatusCode?: unknown;\n\t\tcode?: unknown;\n\t\trequestID?: unknown;\n\t\trequest_id?: unknown;\n\t\theaders?: unknown;\n\t\terror?: unknown;\n\t\t$metadata?: { requestId?: unknown };\n\t};\n\n\tconst status =\n\t\ttypeof err.status === \"number\" ? err.status : typeof err.statusCode === \"number\" ? err.statusCode : undefined;\n\n\t// Error bodies come nested differently per SDK: Anthropic/OpenAI expose\n\t// `error.error = {type|code, message}` (sometimes doubly nested).\n\tlet body = err.error as { type?: unknown; code?: unknown; message?: unknown; error?: unknown } | undefined;\n\tif (body && typeof body === \"object\" && body.error && typeof body.error === \"object\") {\n\t\tbody = body.error as { type?: unknown; code?: unknown; message?: unknown };\n\t}\n\tconst bodyType = body && typeof body === \"object\" ? (body.type ?? body.code) : undefined;\n\tconst bodyMessage = body && typeof body === \"object\" ? body.message : undefined;\n\tconst providerErrorType =\n\t\ttypeof bodyType === \"string\"\n\t\t\t? bodyType\n\t\t\t: typeof err.code === \"string\"\n\t\t\t\t? err.code\n\t\t\t\t: err.name !== \"Error\" && err.name !== \"StreamFailureError\"\n\t\t\t\t\t? err.name\n\t\t\t\t\t: undefined;\n\n\tconst headers = err.headers;\n\tconst headerRequestId =\n\t\theaders && typeof (headers as Headers).get === \"function\"\n\t\t\t? ((headers as Headers).get(\"request-id\") ?? (headers as Headers).get(\"x-request-id\"))\n\t\t\t: headers && typeof headers === \"object\"\n\t\t\t\t? ((headers as Record<string, unknown>)[\"request-id\"] ??\n\t\t\t\t\t(headers as Record<string, unknown>)[\"x-request-id\"])\n\t\t\t\t: undefined;\n\tconst rawRequestId = err.requestID ?? err.request_id ?? err.$metadata?.requestId ?? headerRequestId;\n\tconst requestId = typeof rawRequestId === \"string\" ? rawRequestId : undefined;\n\n\treturn {\n\t\tinfo: {\n\t\t\tkind: classifyStreamFailure(providerErrorType ?? error.message, status),\n\t\t\tproviderErrorType,\n\t\t\tstatus,\n\t\t\trequestId,\n\t\t},\n\t\tdetail: typeof bodyMessage === \"string\" ? bodyMessage : undefined,\n\t};\n}\n\n/**\n * Best-effort extraction of structured failure info from any thrown value:\n * StreamFailureError, provider SDK errors (Anthropic/OpenAI APIError, AWS SDK\n * exceptions, Google ApiError), or plain errors.\n */\nexport function extractStreamFailureInfo(error: unknown): StreamFailureInfo {\n\treturn extractStreamFailureParts(error).info;\n}\n\n/**\n * User-facing message for a thrown stream error: a classified one-liner with\n * the provider's own short message, never the raw payload/trace. Unrecognized\n * errors pass through verbatim so their text (which downstream retry matching\n * may depend on) is preserved.\n */\nexport function formatStreamFailureMessage(error: unknown): string {\n\tif (error instanceof StreamFailureError) return error.message;\n\tconst { info, detail } = extractStreamFailureParts(error);\n\tif (info.kind === \"unknown\") {\n\t\treturn error instanceof Error ? error.message : JSON.stringify(error);\n\t}\n\treturn streamFailureMessage(info, detail);\n}\n\nconst log = getLogger(\"ai.provider\");\n\n/**\n * Record a terminal stream failure on the message (structured diagnostic that\n * persists to session JSONL) and emit one structured log line. Call from the\n * provider's terminal catch after stopReason/errorMessage are set; no-op for\n * user-initiated aborts.\n */\nexport function recordStreamFailure(\n\tmodel: { provider: string; id: string; api: string },\n\toutput: AssistantMessage,\n\terror: unknown,\n): void {\n\tif (output.stopReason !== \"error\") return;\n\tconst info = extractStreamFailureInfo(error);\n\tappendAssistantMessageDiagnostic(output, {\n\t\ttype: \"provider_stream_failure\",\n\t\ttimestamp: Date.now(),\n\t\terror: extractDiagnosticError(error),\n\t\tdetails: { ...info },\n\t});\n\tconst rawMessage = error instanceof Error ? error.message : String(error);\n\tlog.error(\"provider stream failure\", {\n\t\tprovider: model.provider,\n\t\tmodel: model.id,\n\t\tapi: model.api,\n\t\tkind: info.kind,\n\t\tproviderErrorType: info.providerErrorType,\n\t\tstatus: info.status,\n\t\trequestId: info.requestId,\n\t\tmessage: output.errorMessage,\n\t\t// errorMessage is user-facing and concise; keep the raw cause for debugging.\n\t\tcause: rawMessage === output.errorMessage ? undefined : truncateRawPayload(rawMessage),\n\t});\n}\n"]}

package/dist/utils/stream-failure.js

@@ -0,0 +1,179 @@
+import { getLogger } from "../log.js";
+import { appendAssistantMessageDiagnostic, extractDiagnosticError } from "./diagnostics.js";
+export class StreamFailureError extends Error {
+    info;
+    constructor(message, info) {
+        super(message);
+        this.name = "StreamFailureError";
+        this.info = info;
+    }
+}
+const KIND_MESSAGES = {
+    refusal: "Model refused to respond",
+    safety: "Response blocked by provider safety filters",
+    overloaded: "Provider overloaded",
+    rate_limit: "Provider rate limit exceeded",
+    server_error: "Provider server error",
+    auth: "Provider authentication failed",
+    invalid_request: "Provider rejected the request",
+    malformed_response: "Provider returned a malformed response",
+    unknown: "Provider stream failed",
+};
+/** Build a user-facing message like "Provider overloaded (overloaded_error, 529) [request_id: req_abc]". */
+export function streamFailureMessage(info, detail) {
+    const qualifiers = [info.providerErrorType, info.status !== undefined ? String(info.status) : undefined]
+        .filter(Boolean)
+        .join(", ");
+    let message = KIND_MESSAGES[info.kind];
+    if (qualifiers)
+        message += ` (${qualifiers})`;
+    if (detail)
+        message += `: ${detail}`;
+    if (info.requestId)
+        message += ` [request_id: ${info.requestId}]`;
+    return message;
+}
+export function classifyStreamFailure(providerErrorType, status) {
+    const type = providerErrorType?.toLowerCase() ?? "";
+    if (type === "refusal")
+        return "refusal";
+    if (/sensitive|safety|prohibited_content|blocklist|spii|recitation|content.?filter|guardrail|flagged/.test(type)) {
+        return "safety";
+    }
+    if (type.includes("overloaded") || status === 529)
+        return "overloaded";
+    if (type.includes("rate_limit") || type.includes("throttl") || status === 429)
+        return "rate_limit";
+    if (/authentication|permission|unauthorized/.test(type) || status === 401 || status === 403)
+        return "auth";
+    if (type.includes("invalid_request") || type.includes("not_found_error") || status === 400 || status === 404) {
+        return "invalid_request";
+    }
+    if (type.includes("malformed"))
+        return "malformed_response";
+    if (type.includes("api_error") ||
+        type.includes("server_error") ||
+        type.includes("unavailable") ||
+        (status !== undefined && status >= 500)) {
+        return "server_error";
+    }
+    return "unknown";
+}
+/**
+ * Failure for a stream that terminated with a provider stop/finish reason that
+ * maps to "error" (e.g. Anthropic "refusal", Gemini "SAFETY"). Providers call
+ * this instead of throwing a generic error, so the raw reason survives.
+ */
+export function streamFailureFromStopReason(rawStopReason, extra) {
+    const info = {
+        kind: rawStopReason ? classifyStreamFailure(rawStopReason) : "unknown",
+        providerErrorType: rawStopReason,
+        requestId: extra?.requestId,
+    };
+    if (info.kind === "unknown" && /malformed/i.test(rawStopReason ?? ""))
+        info.kind = "malformed_response";
+    const message = rawStopReason
+        ? streamFailureMessage(info)
+        : streamFailureMessage(info, "stream ended with an error and no stop reason");
+    return new StreamFailureError(message, info);
+}
+const MAX_RAW_LENGTH = 2000;
+export function truncateRawPayload(raw) {
+    return raw.length > MAX_RAW_LENGTH ? `${raw.slice(0, MAX_RAW_LENGTH)}…` : raw;
+}
+function extractStreamFailureParts(error) {
+    if (error instanceof StreamFailureError)
+        return { info: error.info };
+    if (!(error instanceof Error))
+        return { info: { kind: "unknown" } };
+    const err = error;
+    const status = typeof err.status === "number" ? err.status : typeof err.statusCode === "number" ? err.statusCode : undefined;
+    // Error bodies come nested differently per SDK: Anthropic/OpenAI expose
+    // `error.error = {type|code, message}` (sometimes doubly nested).
+    let body = err.error;
+    if (body && typeof body === "object" && body.error && typeof body.error === "object") {
+        body = body.error;
+    }
+    const bodyType = body && typeof body === "object" ? (body.type ?? body.code) : undefined;
+    const bodyMessage = body && typeof body === "object" ? body.message : undefined;
+    const providerErrorType = typeof bodyType === "string"
+        ? bodyType
+        : typeof err.code === "string"
+            ? err.code
+            : err.name !== "Error" && err.name !== "StreamFailureError"
+                ? err.name
+                : undefined;
+    const headers = err.headers;
+    const headerRequestId = headers && typeof headers.get === "function"
+        ? (headers.get("request-id") ?? headers.get("x-request-id"))
+        : headers && typeof headers === "object"
+            ? (headers["request-id"] ??
+                headers["x-request-id"])
+            : undefined;
+    const rawRequestId = err.requestID ?? err.request_id ?? err.$metadata?.requestId ?? headerRequestId;
+    const requestId = typeof rawRequestId === "string" ? rawRequestId : undefined;
+    return {
+        info: {
+            kind: classifyStreamFailure(providerErrorType ?? error.message, status),
+            providerErrorType,
+            status,
+            requestId,
+        },
+        detail: typeof bodyMessage === "string" ? bodyMessage : undefined,
+    };
+}
+/**
+ * Best-effort extraction of structured failure info from any thrown value:
+ * StreamFailureError, provider SDK errors (Anthropic/OpenAI APIError, AWS SDK
+ * exceptions, Google ApiError), or plain errors.
+ */
+export function extractStreamFailureInfo(error) {
+    return extractStreamFailureParts(error).info;
+}
+/**
+ * User-facing message for a thrown stream error: a classified one-liner with
+ * the provider's own short message, never the raw payload/trace. Unrecognized
+ * errors pass through verbatim so their text (which downstream retry matching
+ * may depend on) is preserved.
+ */
+export function formatStreamFailureMessage(error) {
+    if (error instanceof StreamFailureError)
+        return error.message;
+    const { info, detail } = extractStreamFailureParts(error);
+    if (info.kind === "unknown") {
+        return error instanceof Error ? error.message : JSON.stringify(error);
+    }
+    return streamFailureMessage(info, detail);
+}
+const log = getLogger("ai.provider");
+/**
+ * Record a terminal stream failure on the message (structured diagnostic that
+ * persists to session JSONL) and emit one structured log line. Call from the
+ * provider's terminal catch after stopReason/errorMessage are set; no-op for
+ * user-initiated aborts.
+ */
+export function recordStreamFailure(model, output, error) {
+    if (output.stopReason !== "error")
+        return;
+    const info = extractStreamFailureInfo(error);
+    appendAssistantMessageDiagnostic(output, {
+        type: "provider_stream_failure",
+        timestamp: Date.now(),
+        error: extractDiagnosticError(error),
+        details: { ...info },
+    });
+    const rawMessage = error instanceof Error ? error.message : String(error);
+    log.error("provider stream failure", {
+        provider: model.provider,
+        model: model.id,
+        api: model.api,
+        kind: info.kind,
+        providerErrorType: info.providerErrorType,
+        status: info.status,
+        requestId: info.requestId,
+        message: output.errorMessage,
+        // errorMessage is user-facing and concise; keep the raw cause for debugging.
+        cause: rawMessage === output.errorMessage ? undefined : truncateRawPayload(rawMessage),
+    });
+}
+//# sourceMappingURL=stream-failure.js.map