@jecruz/merc-ai 0.1.4 → 0.2.0

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

@@ -1 +1 @@
-{"version":3,"file":"overflow.d.ts","sourceRoot":"","sources":["../../src/utils/overflow.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,gBAAgB,EAAE,MAAM,aAAa,CAAC;AAyEpD;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAgDG;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.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 * - OpenAI/LiteLLM: \"Requested token count exceeds the model's maximum context length of 131072 tokens\"\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 * - DS4: \"Prompt has X tokens, but the configured context size is Y tokens\"\n * - Cerebras: \"400/413 status code (no body)\"\n * - Mistral: \"Prompt contains X tokens ... too large for model with Y maximum context length\"\n * - z.ai: Does NOT error, accepts overflow silently - handled via usage.input > contextWindow\n * - Xiaomi MiMo: Truncates input to fill contextWindow exactly, then returns finish_reason \"length\"\n *   with output=0 (no room left to generate). Detected via stopReason \"length\" + zero output +\n *   input filling the context window.\n * - Ollama: Some deployments truncate silently, others return errors like \"prompt too long; exceeded max context length by X tokens\"\n */\nconst OVERFLOW_PATTERNS = [\n\t/prompt is too long/i, // Anthropic token overflow\n\t/request_too_large/i, // Anthropic request byte-size overflow (HTTP 413)\n\t/input is too long for requested model/i, // Amazon Bedrock\n\t/exceeds the context window/i, // OpenAI (Completions & Responses API)\n\t/exceeds (?:the )?(?:model'?s )?maximum context length(?: of [\\d,]+ tokens?|\\s*\\([\\d,]+\\))/i, // OpenAI-compatible proxies (LiteLLM)\n\t/input token count.*exceeds the maximum/i, // Google (Gemini)\n\t/maximum prompt length is \\d+/i, // xAI (Grok)\n\t/reduce the length of the messages/i, // Groq\n\t/maximum context length is \\d+ tokens/i, // OpenRouter (all backends)\n\t/exceeds (?:the )?maximum allowed input length of [\\d,]+ tokens?/i, // OpenRouter/Poolside\n\t/exceeds the limit of \\d+/i, // GitHub Copilot\n\t/exceeds the available context size/i, // llama.cpp server\n\t/greater than the context length/i, // LM Studio\n\t/context window exceeds limit/i, // MiniMax\n\t/exceeded model token limit/i, // Kimi For Coding\n\t/too large for model with \\d+ maximum context length/i, // Mistral\n\t/prompt has [\\d,]+ tokens?, but the configured context size is [\\d,]+ tokens?/i, // DS4 server\n\t/model_context_window_exceeded/i, // z.ai non-standard finish_reason surfaced as error text\n\t/prompt too long; exceeded (?:max )?context length/i, // Ollama explicit overflow error\n\t/context[_ ]length[_ ]exceeded/i, // Generic fallback\n\t/too many tokens/i, // Generic fallback\n\t/token limit exceeded/i, // Generic fallback\n\t/^4(?:00|13)\\s*(?:status code)?\\s*\\(no body\\)/i, // Cerebras: 400/413 with no body\n];\n\n/**\n * Patterns that indicate non-overflow errors (e.g. rate limiting, server errors).\n * Error messages matching any of these are excluded from overflow detection\n * even if they also match an OVERFLOW_PATTERN.\n *\n * Example: Bedrock formats throttling errors as \"ThrottlingException: Too many tokens,\n * please wait before trying again.\" which would match the /too many tokens/i overflow\n * pattern without this exclusion.\n */\nconst NON_OVERFLOW_PATTERNS = [\n\t/^(Throttling error|Service unavailable):/i, // AWS Bedrock non-overflow errors (human-readable prefixes from formatBedrockError)\n\t/rate limit/i, // Generic rate limiting\n\t/too many requests/i, // Generic HTTP 429 style\n];\n\n/**\n * Check if an assistant message represents a context overflow error.\n *\n * This handles two cases:\n * 1. Error-based overflow: Most providers return stopReason \"error\" with a\n *    specific error message pattern.\n * 2. Silent overflow: Some providers accept overflow requests and return\n *    successfully. For these, we check if usage.input exceeds the context window.\n *\n * ## Reliability by Provider\n *\n * **Reliable detection (returns error with detectable message):**\n * - Anthropic: \"prompt is too long: X tokens > Y maximum\" or \"request_too_large\"\n * - OpenAI (Completions & Responses): \"exceeds the context window\" or \"exceeds the model's maximum context length of X tokens\"\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 * - DS4: \"Prompt has X tokens, but the configured context size is Y tokens\"\n *\n * **Unreliable detection:**\n * - z.ai: Sometimes accepts overflow silently (detectable via usage.input > contextWindow),\n *   sometimes returns rate limit errors. Pass contextWindow param to detect silent overflow.\n * - Xiaomi MiMo: Truncates input to fit contextWindow then returns stopReason \"length\" with\n *   output=0. Pass contextWindow param to detect via the \"filled context + zero output\" signal.\n * - Ollama: May truncate input silently for some setups, but may also return explicit\n *   overflow errors that match the patterns above. Silent truncation still cannot be\n *   detected here because we do not know the expected token count.\n *\n * ## Custom Providers\n *\n * If you've added custom models via settings.json, this function may not detect\n * overflow errors from those providers. To add support:\n *\n * 1. Send a request that exceeds the model's context window\n * 2. Check the errorMessage in the response\n * 3. Create a regex pattern that matches the error\n * 4. The pattern should be added to OVERFLOW_PATTERNS in this file, or\n *    check the errorMessage yourself before calling this function\n *\n * @param message - The assistant message to check\n * @param contextWindow - Optional context window size for detecting silent overflow (z.ai)\n * @returns true if the message indicates a context overflow\n */\nexport function isContextOverflow(message: AssistantMessage, contextWindow?: number): boolean {\n\t// Case 1: Check error message patterns\n\tif (message.stopReason === \"error\" && message.errorMessage) {\n\t\t// Skip messages matching known non-overflow patterns (e.g. throttling / rate-limit)\n\t\tconst isNonOverflow = NON_OVERFLOW_PATTERNS.some((p) => p.test(message.errorMessage!));\n\t\tif (!isNonOverflow && OVERFLOW_PATTERNS.some((p) => p.test(message.errorMessage!))) {\n\t\t\treturn true;\n\t\t}\n\t}\n\n\t// Case 2: Silent overflow (z.ai style) - successful but usage exceeds context\n\tif (contextWindow && message.stopReason === \"stop\") {\n\t\tconst inputTokens = message.usage.input + message.usage.cacheRead;\n\t\tif (inputTokens > contextWindow) {\n\t\t\treturn true;\n\t\t}\n\t}\n\n\t// Case 3: Length-stop overflow (Xiaomi MiMo style) - server truncates oversized input\n\t// to fit the context window, leaving no room for output. Returns stopReason \"length\"\n\t// with output=0 and input+cacheRead filling the context window.\n\tif (contextWindow && message.stopReason === \"length\" && message.usage.output === 0) {\n\t\tconst inputTokens = message.usage.input + message.usage.cacheRead;\n\t\tif (inputTokens >= contextWindow * 0.99) {\n\t\t\treturn true;\n\t\t}\n\t}\n\n\treturn false;\n}\n\n/**\n * Get the overflow patterns for testing purposes.\n */\nexport function getOverflowPatterns(): RegExp[] {\n\treturn [...OVERFLOW_PATTERNS];\n}\n"]}
+{"version":3,"file":"overflow.d.ts","sourceRoot":"","sources":["../../src/utils/overflow.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,gBAAgB,EAAE,MAAM,aAAa,CAAC;AAgGpD;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAkDG;AACH,wBAAgB,iBAAiB,CAAC,OAAO,EAAE,gBAAgB,EAAE,aAAa,CAAC,EAAE,MAAM,GAAG,OAAO,CAqC5F;AAED;;;;;GAKG;AACH,wBAAgB,mBAAmB,CAAC,OAAO,EAAE,gBAAgB,EAAE,gBAAgB,EAAE,MAAM,GAAG,OAAO,CAEhG;AAED;;GAEG;AACH,wBAAgB,mBAAmB,IAAI,MAAM,EAAE,CAE9C","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 or the provider's request byte-size limit.\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 * - 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 * - GitHub Copilot: 'github-copilot API error (413): {\"message\":\"failed to parse request\",\"code\":\"\"}'\n * - MiniMax: \"invalid params, context window exceeds limit\"\n * - Kimi For Coding: \"Your request exceeded model token limit: X (requested: Y)\"\n * - DS4: \"Prompt has X tokens, but the configured context size is Y tokens\"\n * - Cerebras: \"400/413 status code (no body)\"\n * - Mistral: \"Prompt contains X tokens ... too large for model with Y maximum context length\"\n * - z.ai: `{\"code\":\"1261\",\"message\":\"Prompt too long\"}`, `{\"code\":\"1261\",\"message\":\"Prompt exceeds max length\"}` (CN endpoint), or silent overflow via usage.input > contextWindow\n * - Xiaomi MiMo: Truncates input to fill contextWindow exactly, then returns finish_reason \"length\"\n *   with output=0 (no room left to generate). Detected via stopReason \"length\" + zero output +\n *   input filling the context window.\n * - Ollama: Some deployments truncate silently, others return errors like \"prompt too long; exceeded max context length by X tokens\"\n */\nconst OVERFLOW_PATTERNS = [\n\t/prompt (?:is )?too long/i, // Anthropic and z.ai token overflow\n\t/prompt exceeds max length/i, // z.ai CN endpoint token overflow\n\t/request_too_large/i, // Anthropic request byte-size overflow (HTTP 413)\n\t/input is too long for requested model/i, // Amazon Bedrock\n\t/exceeds the context window/i, // OpenAI (Completions & Responses API)\n\t/exceeds (?:the )?(?:model'?s )?maximum context length(?: of [\\d,]+ tokens?|\\s*\\([\\d,]+\\))/i, // OpenAI-compatible proxies (LiteLLM)\n\t/input token count.*exceeds the maximum/i, // Google (Gemini)\n\t/maximum prompt length is \\d+/i, // xAI (Grok)\n\t/reduce the length of the messages/i, // Groq\n\t/maximum context length is \\d+ tokens/i, // OpenRouter (all backends)\n\t/exceeds (?:the )?maximum allowed input length of [\\d,]+ tokens?/i, // OpenRouter/Poolside\n\t/exceeds the limit of \\d+/i, // GitHub Copilot\n\t/exceeds the available context size/i, // llama.cpp server\n\t/greater than the context length/i, // LM Studio\n\t/context window exceeds limit/i, // MiniMax\n\t/exceeded model token limit/i, // Kimi For Coding\n\t/too large for model with \\d+ maximum context length/i, // Mistral\n\t/prompt has [\\d,]+ tokens?, but the configured context size is [\\d,]+ tokens?/i, // DS4 server\n\t/model_context_window_exceeded/i, // z.ai non-standard finish_reason surfaced as error text\n\t/prompt too long; exceeded (?:max )?context length/i, // Ollama explicit overflow error\n\t/context[_ ]length[_ ]exceeded/i, // Generic fallback\n\t/too many tokens/i, // Generic fallback\n\t/token limit exceeded/i, // Generic fallback\n\t// HTTP 413 means request-size overflow regardless of provider or body text.\n\t// Bare SDK status; formatted provider headers are parsed separately below.\n\t/^413(?:\\s|:|$)/i,\n];\n\n/** Match the first single-line API-error header with linear work and unrestricted provider names. */\nfunction hasRequestSizeOverflowHeader(message: string): boolean {\n\tconst header = \" api error (\";\n\tfor (let index = 0; index < message.length; index++) {\n\t\tconst code = message.charCodeAt(index);\n\t\tif (code === 10 || code === 13) return false;\n\t\tif (code !== 32 || index === 0) continue;\n\t\t// The fixed-size comparison cannot backtrack over the provider prefix.\n\t\tif (message.slice(index, index + header.length).toLowerCase() === header) {\n\t\t\t// Stop at the first header so a different status's body cannot supply a 413 header.\n\t\t\treturn message.startsWith(\"413):\", index + header.length);\n\t\t}\n\t}\n\treturn false;\n}\n\nconst CEREBRAS_BODYLESS_OVERFLOW_PATTERN = /^4(?:00|13)\\s*(?:status code)?\\s*\\(no body\\)/i;\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 * - GitHub Copilot: HTTP 413 \"failed to parse request\" (request byte-size overflow)\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 * - DS4: \"Prompt has X tokens, but the configured context size is Y tokens\"\n * - z.ai: \"Prompt too long\"\n *\n * **Unreliable detection:**\n * - z.ai: Sometimes accepts overflow silently (detectable via usage.input > contextWindow),\n *   sometimes returns rate limit errors. Pass contextWindow param to detect silent overflow.\n * - Xiaomi MiMo: Truncates input to fit contextWindow then returns stopReason \"length\" with\n *   output=0. Pass contextWindow param to detect via the \"filled context + zero output\" signal.\n * - Ollama: May truncate input silently for some setups, but may also return explicit\n *   overflow errors that match the patterns above. Silent truncation still cannot be\n *   detected here because we do not know the expected token count.\n *\n * ## Custom Providers\n *\n * If you've added custom models via settings.json, this function may not detect\n * overflow errors from those providers. To add support:\n *\n * 1. Send a request that exceeds the model's context window\n * 2. Check the errorMessage in the response\n * 3. Create a regex pattern that matches the error\n * 4. The pattern should be added to OVERFLOW_PATTERNS in this file, or\n *    check the errorMessage yourself before calling this function\n *\n * @param message - The assistant message to check\n * @param contextWindow - Optional context window size for detecting silent overflow (z.ai)\n * @returns true if the message indicates a context overflow\n */\nexport function isContextOverflow(message: AssistantMessage, contextWindow?: number): boolean {\n\t// Case 1: Check error message patterns\n\tif (message.stopReason === \"error\" && message.errorMessage) {\n\t\t// Skip messages matching known non-overflow patterns (e.g. throttling / rate-limit)\n\t\tconst isNonOverflow = NON_OVERFLOW_PATTERNS.some((p) => p.test(message.errorMessage!));\n\t\tif (!isNonOverflow) {\n\t\t\tif (\n\t\t\t\thasRequestSizeOverflowHeader(message.errorMessage) ||\n\t\t\t\tOVERFLOW_PATTERNS.some((p) => p.test(message.errorMessage!))\n\t\t\t) {\n\t\t\t\treturn true;\n\t\t\t}\n\t\t\tif (message.provider === \"cerebras\" && CEREBRAS_BODYLESS_OVERFLOW_PATTERN.test(message.errorMessage)) {\n\t\t\t\treturn true;\n\t\t\t}\n\t\t}\n\t}\n\n\t// Case 2: Silent overflow (z.ai style) - successful but usage exceeds context\n\tif (contextWindow && message.stopReason === \"stop\") {\n\t\tconst inputTokens = message.usage.input + message.usage.cacheRead;\n\t\tif (inputTokens > contextWindow) {\n\t\t\treturn true;\n\t\t}\n\t}\n\n\t// Case 3: Length-stop overflow (Xiaomi MiMo style) - server truncates oversized input\n\t// to fit the context window, leaving no room for output. Returns stopReason \"length\"\n\t// with output=0 and input+cacheRead filling the context window.\n\tif (contextWindow && message.stopReason === \"length\" && message.usage.output === 0) {\n\t\tconst inputTokens = message.usage.input + message.usage.cacheRead;\n\t\tif (inputTokens >= contextWindow * 0.99) {\n\t\t\treturn true;\n\t\t}\n\t}\n\n\treturn false;\n}\n\n/**\n * Check whether a length stop ended below the caller or model's intended output limit.\n * Such responses may be caused by context pressure or provider-side truncation, so callers\n * can make one bounded compact-and-retry attempt. `desiredMaxOutput` must be the original\n * limit before any context-based clamping.\n */\nexport function isRecoverableLength(message: AssistantMessage, desiredMaxOutput: number): boolean {\n\treturn message.stopReason === \"length\" && desiredMaxOutput > 0 && message.usage.output < desiredMaxOutput;\n}\n\n/**\n * Get the overflow patterns for testing purposes.\n */\nexport function getOverflowPatterns(): RegExp[] {\n\treturn [...OVERFLOW_PATTERNS];\n}\n"]}

package/dist/utils/overflow.js

@@ -2,7 +2,7 @@
  * 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.
+ * the model's context window or the provider's request byte-size limit.
  *
  * Provider-specific patterns (with example error messages):
  *
@@ -10,6 +10,7 @@
  * - 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"
@@ -17,19 +18,21 @@
  * - 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"
+ * - GitHub Copilot: 'github-copilot API error (413): {"message":"failed to parse request","code":""}'
  * - 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
+ * - z.ai: `{"code":"1261","message":"Prompt too long"}`, `{"code":"1261","message":"Prompt exceeds max length"}` (CN endpoint), or silent overflow via usage.input > contextWindow
  * - Xiaomi MiMo: Truncates input to fill contextWindow exactly, then returns finish_reason "length"
  *   with output=0 (no room left to generate). Detected via stopReason "length" + zero output +
  *   input filling the context window.
  * - 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
+    /prompt (?:is )?too long/i, // Anthropic and z.ai token overflow
+    /prompt exceeds max length/i, // z.ai CN endpoint token overflow
     /request_too_large/i, // Anthropic request byte-size overflow (HTTP 413)
     /input is too long for requested model/i, // Amazon Bedrock
     /exceeds the context window/i, // OpenAI (Completions & Responses API)
@@ -51,8 +54,28 @@ const OVERFLOW_PATTERNS = [
     /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
+    // HTTP 413 means request-size overflow regardless of provider or body text.
+    // Bare SDK status; formatted provider headers are parsed separately below.
+    /^413(?:\s|:|$)/i,
 ];
+/** Match the first single-line API-error header with linear work and unrestricted provider names. */
+function hasRequestSizeOverflowHeader(message) {
+    const header = " api error (";
+    for (let index = 0; index < message.length; index++) {
+        const code = message.charCodeAt(index);
+        if (code === 10 || code === 13)
+            return false;
+        if (code !== 32 || index === 0)
+            continue;
+        // The fixed-size comparison cannot backtrack over the provider prefix.
+        if (message.slice(index, index + header.length).toLowerCase() === header) {
+            // Stop at the first header so a different status's body cannot supply a 413 header.
+            return message.startsWith("413):", index + header.length);
+        }
+    }
+    return false;
+}
+const CEREBRAS_BODYLESS_OVERFLOW_PATTERN = /^4(?:00|13)\s*(?:status code)?\s*\(no body\)/i;
 /**
  * Patterns that indicate non-overflow errors (e.g. rate limiting, server errors).
  * Error messages matching any of these are excluded from overflow detection
@@ -80,10 +103,11 @@ const NON_OVERFLOW_PATTERNS = [
  *
  * **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" or "exceeds the model's maximum context length of X tokens"
+ * - 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"
+ * - GitHub Copilot: HTTP 413 "failed to parse request" (request byte-size overflow)
  * - 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"
@@ -91,6 +115,7 @@ const NON_OVERFLOW_PATTERNS = [
  * - 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"
+ * - z.ai: "Prompt too long"
  *
  * **Unreliable detection:**
  * - z.ai: Sometimes accepts overflow silently (detectable via usage.input > contextWindow),
@@ -121,8 +146,14 @@ 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 (!isNonOverflow) {
+            if (hasRequestSizeOverflowHeader(message.errorMessage) ||
+                OVERFLOW_PATTERNS.some((p) => p.test(message.errorMessage))) {
+                return true;
+            }
+            if (message.provider === "cerebras" && CEREBRAS_BODYLESS_OVERFLOW_PATTERN.test(message.errorMessage)) {
+                return true;
+            }
         }
     }
     // Case 2: Silent overflow (z.ai style) - successful but usage exceeds context
@@ -143,6 +174,15 @@ export function isContextOverflow(message, contextWindow) {
     }
     return false;
 }
+/**
+ * Check whether a length stop ended below the caller or model's intended output limit.
+ * Such responses may be caused by context pressure or provider-side truncation, so callers
+ * can make one bounded compact-and-retry attempt. `desiredMaxOutput` must be the original
+ * limit before any context-based clamping.
+ */
+export function isRecoverableLength(message, desiredMaxOutput) {
+    return message.stopReason === "length" && desiredMaxOutput > 0 && message.usage.output < desiredMaxOutput;
+}
 /**
  * Get the overflow patterns for testing purposes.
  */

package/dist/utils/overflow.js.map

@@ -1 +1 @@
-{"version":3,"file":"overflow.js","sourceRoot":"","sources":["../../src/utils/overflow.ts"],"names":[],"mappings":"AAEA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA6BG;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,4BAA4B;IACrE,kEAAkE,EAAE,sBAAsB;IAC1F,2BAA2B,EAAE,iBAAiB;IAC9C,qCAAqC,EAAE,mBAAmB;IAC1D,kCAAkC,EAAE,YAAY;IAChD,+BAA+B,EAAE,UAAU;IAC3C,6BAA6B,EAAE,kBAAkB;IACjD,sDAAsD,EAAE,UAAU;IAClE,+EAA+E,EAAE,aAAa;IAC9F,gCAAgC,EAAE,yDAAyD;IAC3F,oDAAoD,EAAE,iCAAiC;IACvF,gCAAgC,EAAE,mBAAmB;IACrD,kBAAkB,EAAE,mBAAmB;IACvC,uBAAuB,EAAE,mBAAmB;IAC5C,+CAA+C,EAAE,iCAAiC;CAClF,CAAC;AAEF;;;;;;;;GAQG;AACH,MAAM,qBAAqB,GAAG;IAC7B,2CAA2C,EAAE,oFAAoF;IACjI,aAAa,EAAE,wBAAwB;IACvC,oBAAoB,EAAE,yBAAyB;CAC/C,CAAC;AAEF;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAgDG;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.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 * - OpenAI/LiteLLM: \"Requested token count exceeds the model's maximum context length of 131072 tokens\"\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 * - DS4: \"Prompt has X tokens, but the configured context size is Y tokens\"\n * - Cerebras: \"400/413 status code (no body)\"\n * - Mistral: \"Prompt contains X tokens ... too large for model with Y maximum context length\"\n * - z.ai: Does NOT error, accepts overflow silently - handled via usage.input > contextWindow\n * - Xiaomi MiMo: Truncates input to fill contextWindow exactly, then returns finish_reason \"length\"\n *   with output=0 (no room left to generate). Detected via stopReason \"length\" + zero output +\n *   input filling the context window.\n * - Ollama: Some deployments truncate silently, others return errors like \"prompt too long; exceeded max context length by X tokens\"\n */\nconst OVERFLOW_PATTERNS = [\n\t/prompt is too long/i, // Anthropic token overflow\n\t/request_too_large/i, // Anthropic request byte-size overflow (HTTP 413)\n\t/input is too long for requested model/i, // Amazon Bedrock\n\t/exceeds the context window/i, // OpenAI (Completions & Responses API)\n\t/exceeds (?:the )?(?:model'?s )?maximum context length(?: of [\\d,]+ tokens?|\\s*\\([\\d,]+\\))/i, // OpenAI-compatible proxies (LiteLLM)\n\t/input token count.*exceeds the maximum/i, // Google (Gemini)\n\t/maximum prompt length is \\d+/i, // xAI (Grok)\n\t/reduce the length of the messages/i, // Groq\n\t/maximum context length is \\d+ tokens/i, // OpenRouter (all backends)\n\t/exceeds (?:the )?maximum allowed input length of [\\d,]+ tokens?/i, // OpenRouter/Poolside\n\t/exceeds the limit of \\d+/i, // GitHub Copilot\n\t/exceeds the available context size/i, // llama.cpp server\n\t/greater than the context length/i, // LM Studio\n\t/context window exceeds limit/i, // MiniMax\n\t/exceeded model token limit/i, // Kimi For Coding\n\t/too large for model with \\d+ maximum context length/i, // Mistral\n\t/prompt has [\\d,]+ tokens?, but the configured context size is [\\d,]+ tokens?/i, // DS4 server\n\t/model_context_window_exceeded/i, // z.ai non-standard finish_reason surfaced as error text\n\t/prompt too long; exceeded (?:max )?context length/i, // Ollama explicit overflow error\n\t/context[_ ]length[_ ]exceeded/i, // Generic fallback\n\t/too many tokens/i, // Generic fallback\n\t/token limit exceeded/i, // Generic fallback\n\t/^4(?:00|13)\\s*(?:status code)?\\s*\\(no body\\)/i, // Cerebras: 400/413 with no body\n];\n\n/**\n * Patterns that indicate non-overflow errors (e.g. rate limiting, server errors).\n * Error messages matching any of these are excluded from overflow detection\n * even if they also match an OVERFLOW_PATTERN.\n *\n * Example: Bedrock formats throttling errors as \"ThrottlingException: Too many tokens,\n * please wait before trying again.\" which would match the /too many tokens/i overflow\n * pattern without this exclusion.\n */\nconst NON_OVERFLOW_PATTERNS = [\n\t/^(Throttling error|Service unavailable):/i, // AWS Bedrock non-overflow errors (human-readable prefixes from formatBedrockError)\n\t/rate limit/i, // Generic rate limiting\n\t/too many requests/i, // Generic HTTP 429 style\n];\n\n/**\n * Check if an assistant message represents a context overflow error.\n *\n * This handles two cases:\n * 1. Error-based overflow: Most providers return stopReason \"error\" with a\n *    specific error message pattern.\n * 2. Silent overflow: Some providers accept overflow requests and return\n *    successfully. For these, we check if usage.input exceeds the context window.\n *\n * ## Reliability by Provider\n *\n * **Reliable detection (returns error with detectable message):**\n * - Anthropic: \"prompt is too long: X tokens > Y maximum\" or \"request_too_large\"\n * - OpenAI (Completions & Responses): \"exceeds the context window\" or \"exceeds the model's maximum context length of X tokens\"\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 * - DS4: \"Prompt has X tokens, but the configured context size is Y tokens\"\n *\n * **Unreliable detection:**\n * - z.ai: Sometimes accepts overflow silently (detectable via usage.input > contextWindow),\n *   sometimes returns rate limit errors. Pass contextWindow param to detect silent overflow.\n * - Xiaomi MiMo: Truncates input to fit contextWindow then returns stopReason \"length\" with\n *   output=0. Pass contextWindow param to detect via the \"filled context + zero output\" signal.\n * - Ollama: May truncate input silently for some setups, but may also return explicit\n *   overflow errors that match the patterns above. Silent truncation still cannot be\n *   detected here because we do not know the expected token count.\n *\n * ## Custom Providers\n *\n * If you've added custom models via settings.json, this function may not detect\n * overflow errors from those providers. To add support:\n *\n * 1. Send a request that exceeds the model's context window\n * 2. Check the errorMessage in the response\n * 3. Create a regex pattern that matches the error\n * 4. The pattern should be added to OVERFLOW_PATTERNS in this file, or\n *    check the errorMessage yourself before calling this function\n *\n * @param message - The assistant message to check\n * @param contextWindow - Optional context window size for detecting silent overflow (z.ai)\n * @returns true if the message indicates a context overflow\n */\nexport function isContextOverflow(message: AssistantMessage, contextWindow?: number): boolean {\n\t// Case 1: Check error message patterns\n\tif (message.stopReason === \"error\" && message.errorMessage) {\n\t\t// Skip messages matching known non-overflow patterns (e.g. throttling / rate-limit)\n\t\tconst isNonOverflow = NON_OVERFLOW_PATTERNS.some((p) => p.test(message.errorMessage!));\n\t\tif (!isNonOverflow && OVERFLOW_PATTERNS.some((p) => p.test(message.errorMessage!))) {\n\t\t\treturn true;\n\t\t}\n\t}\n\n\t// Case 2: Silent overflow (z.ai style) - successful but usage exceeds context\n\tif (contextWindow && message.stopReason === \"stop\") {\n\t\tconst inputTokens = message.usage.input + message.usage.cacheRead;\n\t\tif (inputTokens > contextWindow) {\n\t\t\treturn true;\n\t\t}\n\t}\n\n\t// Case 3: Length-stop overflow (Xiaomi MiMo style) - server truncates oversized input\n\t// to fit the context window, leaving no room for output. Returns stopReason \"length\"\n\t// with output=0 and input+cacheRead filling the context window.\n\tif (contextWindow && message.stopReason === \"length\" && message.usage.output === 0) {\n\t\tconst inputTokens = message.usage.input + message.usage.cacheRead;\n\t\tif (inputTokens >= contextWindow * 0.99) {\n\t\t\treturn true;\n\t\t}\n\t}\n\n\treturn false;\n}\n\n/**\n * Get the overflow patterns for testing purposes.\n */\nexport function getOverflowPatterns(): RegExp[] {\n\treturn [...OVERFLOW_PATTERNS];\n}\n"]}
+{"version":3,"file":"overflow.js","sourceRoot":"","sources":["../../src/utils/overflow.ts"],"names":[],"mappings":"AAEA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA+BG;AACH,MAAM,iBAAiB,GAAG;IACzB,0BAA0B,EAAE,oCAAoC;IAChE,4BAA4B,EAAE,kCAAkC;IAChE,oBAAoB,EAAE,kDAAkD;IACxE,wCAAwC,EAAE,iBAAiB;IAC3D,6BAA6B,EAAE,uCAAuC;IACtE,4FAA4F,EAAE,sCAAsC;IACpI,yCAAyC,EAAE,kBAAkB;IAC7D,+BAA+B,EAAE,aAAa;IAC9C,oCAAoC,EAAE,OAAO;IAC7C,uCAAuC,EAAE,4BAA4B;IACrE,kEAAkE,EAAE,sBAAsB;IAC1F,2BAA2B,EAAE,iBAAiB;IAC9C,qCAAqC,EAAE,mBAAmB;IAC1D,kCAAkC,EAAE,YAAY;IAChD,+BAA+B,EAAE,UAAU;IAC3C,6BAA6B,EAAE,kBAAkB;IACjD,sDAAsD,EAAE,UAAU;IAClE,+EAA+E,EAAE,aAAa;IAC9F,gCAAgC,EAAE,yDAAyD;IAC3F,oDAAoD,EAAE,iCAAiC;IACvF,gCAAgC,EAAE,mBAAmB;IACrD,kBAAkB,EAAE,mBAAmB;IACvC,uBAAuB,EAAE,mBAAmB;IAC5C,4EAA4E;IAC5E,2EAA2E;IAC3E,iBAAiB;CACjB,CAAC;AAEF,qGAAqG;AACrG,SAAS,4BAA4B,CAAC,OAAe,EAAW;IAC/D,MAAM,MAAM,GAAG,cAAc,CAAC;IAC9B,KAAK,IAAI,KAAK,GAAG,CAAC,EAAE,KAAK,GAAG,OAAO,CAAC,MAAM,EAAE,KAAK,EAAE,EAAE,CAAC;QACrD,MAAM,IAAI,GAAG,OAAO,CAAC,UAAU,CAAC,KAAK,CAAC,CAAC;QACvC,IAAI,IAAI,KAAK,EAAE,IAAI,IAAI,KAAK,EAAE;YAAE,OAAO,KAAK,CAAC;QAC7C,IAAI,IAAI,KAAK,EAAE,IAAI,KAAK,KAAK,CAAC;YAAE,SAAS;QACzC,uEAAuE;QACvE,IAAI,OAAO,CAAC,KAAK,CAAC,KAAK,EAAE,KAAK,GAAG,MAAM,CAAC,MAAM,CAAC,CAAC,WAAW,EAAE,KAAK,MAAM,EAAE,CAAC;YAC1E,oFAAoF;YACpF,OAAO,OAAO,CAAC,UAAU,CAAC,OAAO,EAAE,KAAK,GAAG,MAAM,CAAC,MAAM,CAAC,CAAC;QAC3D,CAAC;IACF,CAAC;IACD,OAAO,KAAK,CAAC;AAAA,CACb;AAED,MAAM,kCAAkC,GAAG,+CAA+C,CAAC;AAE3F;;;;;;;;GAQG;AACH,MAAM,qBAAqB,GAAG;IAC7B,2CAA2C,EAAE,oFAAoF;IACjI,aAAa,EAAE,wBAAwB;IACvC,oBAAoB,EAAE,yBAAyB;CAC/C,CAAC;AAEF;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAkDG;AACH,MAAM,UAAU,iBAAiB,CAAC,OAAyB,EAAE,aAAsB,EAAW;IAC7F,uCAAuC;IACvC,IAAI,OAAO,CAAC,UAAU,KAAK,OAAO,IAAI,OAAO,CAAC,YAAY,EAAE,CAAC;QAC5D,oFAAoF;QACpF,MAAM,aAAa,GAAG,qBAAqB,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC,OAAO,CAAC,YAAa,CAAC,CAAC,CAAC;QACvF,IAAI,CAAC,aAAa,EAAE,CAAC;YACpB,IACC,4BAA4B,CAAC,OAAO,CAAC,YAAY,CAAC;gBAClD,iBAAiB,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC,OAAO,CAAC,YAAa,CAAC,CAAC,EAC3D,CAAC;gBACF,OAAO,IAAI,CAAC;YACb,CAAC;YACD,IAAI,OAAO,CAAC,QAAQ,KAAK,UAAU,IAAI,kCAAkC,CAAC,IAAI,CAAC,OAAO,CAAC,YAAY,CAAC,EAAE,CAAC;gBACtG,OAAO,IAAI,CAAC;YACb,CAAC;QACF,CAAC;IACF,CAAC;IAED,8EAA8E;IAC9E,IAAI,aAAa,IAAI,OAAO,CAAC,UAAU,KAAK,MAAM,EAAE,CAAC;QACpD,MAAM,WAAW,GAAG,OAAO,CAAC,KAAK,CAAC,KAAK,GAAG,OAAO,CAAC,KAAK,CAAC,SAAS,CAAC;QAClE,IAAI,WAAW,GAAG,aAAa,EAAE,CAAC;YACjC,OAAO,IAAI,CAAC;QACb,CAAC;IACF,CAAC;IAED,sFAAsF;IACtF,qFAAqF;IACrF,gEAAgE;IAChE,IAAI,aAAa,IAAI,OAAO,CAAC,UAAU,KAAK,QAAQ,IAAI,OAAO,CAAC,KAAK,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;QACpF,MAAM,WAAW,GAAG,OAAO,CAAC,KAAK,CAAC,KAAK,GAAG,OAAO,CAAC,KAAK,CAAC,SAAS,CAAC;QAClE,IAAI,WAAW,IAAI,aAAa,GAAG,IAAI,EAAE,CAAC;YACzC,OAAO,IAAI,CAAC;QACb,CAAC;IACF,CAAC;IAED,OAAO,KAAK,CAAC;AAAA,CACb;AAED;;;;;GAKG;AACH,MAAM,UAAU,mBAAmB,CAAC,OAAyB,EAAE,gBAAwB,EAAW;IACjG,OAAO,OAAO,CAAC,UAAU,KAAK,QAAQ,IAAI,gBAAgB,GAAG,CAAC,IAAI,OAAO,CAAC,KAAK,CAAC,MAAM,GAAG,gBAAgB,CAAC;AAAA,CAC1G;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 or the provider's request byte-size limit.\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 * - 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 * - GitHub Copilot: 'github-copilot API error (413): {\"message\":\"failed to parse request\",\"code\":\"\"}'\n * - MiniMax: \"invalid params, context window exceeds limit\"\n * - Kimi For Coding: \"Your request exceeded model token limit: X (requested: Y)\"\n * - DS4: \"Prompt has X tokens, but the configured context size is Y tokens\"\n * - Cerebras: \"400/413 status code (no body)\"\n * - Mistral: \"Prompt contains X tokens ... too large for model with Y maximum context length\"\n * - z.ai: `{\"code\":\"1261\",\"message\":\"Prompt too long\"}`, `{\"code\":\"1261\",\"message\":\"Prompt exceeds max length\"}` (CN endpoint), or silent overflow via usage.input > contextWindow\n * - Xiaomi MiMo: Truncates input to fill contextWindow exactly, then returns finish_reason \"length\"\n *   with output=0 (no room left to generate). Detected via stopReason \"length\" + zero output +\n *   input filling the context window.\n * - Ollama: Some deployments truncate silently, others return errors like \"prompt too long; exceeded max context length by X tokens\"\n */\nconst OVERFLOW_PATTERNS = [\n\t/prompt (?:is )?too long/i, // Anthropic and z.ai token overflow\n\t/prompt exceeds max length/i, // z.ai CN endpoint token overflow\n\t/request_too_large/i, // Anthropic request byte-size overflow (HTTP 413)\n\t/input is too long for requested model/i, // Amazon Bedrock\n\t/exceeds the context window/i, // OpenAI (Completions & Responses API)\n\t/exceeds (?:the )?(?:model'?s )?maximum context length(?: of [\\d,]+ tokens?|\\s*\\([\\d,]+\\))/i, // OpenAI-compatible proxies (LiteLLM)\n\t/input token count.*exceeds the maximum/i, // Google (Gemini)\n\t/maximum prompt length is \\d+/i, // xAI (Grok)\n\t/reduce the length of the messages/i, // Groq\n\t/maximum context length is \\d+ tokens/i, // OpenRouter (all backends)\n\t/exceeds (?:the )?maximum allowed input length of [\\d,]+ tokens?/i, // OpenRouter/Poolside\n\t/exceeds the limit of \\d+/i, // GitHub Copilot\n\t/exceeds the available context size/i, // llama.cpp server\n\t/greater than the context length/i, // LM Studio\n\t/context window exceeds limit/i, // MiniMax\n\t/exceeded model token limit/i, // Kimi For Coding\n\t/too large for model with \\d+ maximum context length/i, // Mistral\n\t/prompt has [\\d,]+ tokens?, but the configured context size is [\\d,]+ tokens?/i, // DS4 server\n\t/model_context_window_exceeded/i, // z.ai non-standard finish_reason surfaced as error text\n\t/prompt too long; exceeded (?:max )?context length/i, // Ollama explicit overflow error\n\t/context[_ ]length[_ ]exceeded/i, // Generic fallback\n\t/too many tokens/i, // Generic fallback\n\t/token limit exceeded/i, // Generic fallback\n\t// HTTP 413 means request-size overflow regardless of provider or body text.\n\t// Bare SDK status; formatted provider headers are parsed separately below.\n\t/^413(?:\\s|:|$)/i,\n];\n\n/** Match the first single-line API-error header with linear work and unrestricted provider names. */\nfunction hasRequestSizeOverflowHeader(message: string): boolean {\n\tconst header = \" api error (\";\n\tfor (let index = 0; index < message.length; index++) {\n\t\tconst code = message.charCodeAt(index);\n\t\tif (code === 10 || code === 13) return false;\n\t\tif (code !== 32 || index === 0) continue;\n\t\t// The fixed-size comparison cannot backtrack over the provider prefix.\n\t\tif (message.slice(index, index + header.length).toLowerCase() === header) {\n\t\t\t// Stop at the first header so a different status's body cannot supply a 413 header.\n\t\t\treturn message.startsWith(\"413):\", index + header.length);\n\t\t}\n\t}\n\treturn false;\n}\n\nconst CEREBRAS_BODYLESS_OVERFLOW_PATTERN = /^4(?:00|13)\\s*(?:status code)?\\s*\\(no body\\)/i;\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 * - GitHub Copilot: HTTP 413 \"failed to parse request\" (request byte-size overflow)\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 * - DS4: \"Prompt has X tokens, but the configured context size is Y tokens\"\n * - z.ai: \"Prompt too long\"\n *\n * **Unreliable detection:**\n * - z.ai: Sometimes accepts overflow silently (detectable via usage.input > contextWindow),\n *   sometimes returns rate limit errors. Pass contextWindow param to detect silent overflow.\n * - Xiaomi MiMo: Truncates input to fit contextWindow then returns stopReason \"length\" with\n *   output=0. Pass contextWindow param to detect via the \"filled context + zero output\" signal.\n * - Ollama: May truncate input silently for some setups, but may also return explicit\n *   overflow errors that match the patterns above. Silent truncation still cannot be\n *   detected here because we do not know the expected token count.\n *\n * ## Custom Providers\n *\n * If you've added custom models via settings.json, this function may not detect\n * overflow errors from those providers. To add support:\n *\n * 1. Send a request that exceeds the model's context window\n * 2. Check the errorMessage in the response\n * 3. Create a regex pattern that matches the error\n * 4. The pattern should be added to OVERFLOW_PATTERNS in this file, or\n *    check the errorMessage yourself before calling this function\n *\n * @param message - The assistant message to check\n * @param contextWindow - Optional context window size for detecting silent overflow (z.ai)\n * @returns true if the message indicates a context overflow\n */\nexport function isContextOverflow(message: AssistantMessage, contextWindow?: number): boolean {\n\t// Case 1: Check error message patterns\n\tif (message.stopReason === \"error\" && message.errorMessage) {\n\t\t// Skip messages matching known non-overflow patterns (e.g. throttling / rate-limit)\n\t\tconst isNonOverflow = NON_OVERFLOW_PATTERNS.some((p) => p.test(message.errorMessage!));\n\t\tif (!isNonOverflow) {\n\t\t\tif (\n\t\t\t\thasRequestSizeOverflowHeader(message.errorMessage) ||\n\t\t\t\tOVERFLOW_PATTERNS.some((p) => p.test(message.errorMessage!))\n\t\t\t) {\n\t\t\t\treturn true;\n\t\t\t}\n\t\t\tif (message.provider === \"cerebras\" && CEREBRAS_BODYLESS_OVERFLOW_PATTERN.test(message.errorMessage)) {\n\t\t\t\treturn true;\n\t\t\t}\n\t\t}\n\t}\n\n\t// Case 2: Silent overflow (z.ai style) - successful but usage exceeds context\n\tif (contextWindow && message.stopReason === \"stop\") {\n\t\tconst inputTokens = message.usage.input + message.usage.cacheRead;\n\t\tif (inputTokens > contextWindow) {\n\t\t\treturn true;\n\t\t}\n\t}\n\n\t// Case 3: Length-stop overflow (Xiaomi MiMo style) - server truncates oversized input\n\t// to fit the context window, leaving no room for output. Returns stopReason \"length\"\n\t// with output=0 and input+cacheRead filling the context window.\n\tif (contextWindow && message.stopReason === \"length\" && message.usage.output === 0) {\n\t\tconst inputTokens = message.usage.input + message.usage.cacheRead;\n\t\tif (inputTokens >= contextWindow * 0.99) {\n\t\t\treturn true;\n\t\t}\n\t}\n\n\treturn false;\n}\n\n/**\n * Check whether a length stop ended below the caller or model's intended output limit.\n * Such responses may be caused by context pressure or provider-side truncation, so callers\n * can make one bounded compact-and-retry attempt. `desiredMaxOutput` must be the original\n * limit before any context-based clamping.\n */\nexport function isRecoverableLength(message: AssistantMessage, desiredMaxOutput: number): boolean {\n\treturn message.stopReason === \"length\" && desiredMaxOutput > 0 && message.usage.output < desiredMaxOutput;\n}\n\n/**\n * Get the overflow patterns for testing purposes.\n */\nexport function getOverflowPatterns(): RegExp[] {\n\treturn [...OVERFLOW_PATTERNS];\n}\n"]}

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

@@ -1 +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;AAI/C;;;GAGG;AACH,wBAAgB,mBAAmB,CAAC,IAAI,EAAE,MAAM,EAAE,GAAG,CAAC,EAAE,WAAW,GAAG,MAAM,GAAG,SAAS,CAIvF;AAED,8EAA8E;AAC9E,wBAAgB,uBAAuB,CAAC,SAAS,EAAE,MAAM,EAAE,GAAG,EAAE,WAAW,GAAG,MAAM,CAanF","sourcesContent":["import type { ProviderEnv } from \"../types.js\";\n\nlet procEnvCache: Map<string, string> | null = null;\n\n/**\n * Resolve a provider environment value from a scoped override, process.env,\n * or the Bun sandbox fallback used by direct AI package consumers.\n */\nexport function getProviderEnvValue(name: string, env?: ProviderEnv): string | undefined {\n\tif (env && Object.hasOwn(env, name)) return env[name];\n\tif (typeof process !== \"undefined\" && process.env[name] !== undefined) return process.env[name];\n\treturn getBunSandboxEnvValue(name);\n}\n\n/** Resolve a WebSocket/HTTP proxy using the supplied provider environment. */\nexport function resolveProviderProxyUrl(targetUrl: string, env: ProviderEnv): string {\n\tconst target = new URL(targetUrl);\n\tconst protocol = target.protocol.replace(\":\", \"\");\n\tconst port = Number(target.port) || (protocol === \"wss\" || protocol === \"https\" ? 443 : 80);\n\tconst noProxy = getScopedAlias(env, [\"NO_PROXY\", \"no_proxy\"]);\n\tif (noProxy && matchesNoProxy(target.hostname, port, noProxy)) return \"\";\n\n\tconst proxy =\n\t\tprotocol === \"wss\" || protocol === \"https\"\n\t\t\t? getScopedAlias(env, [\"HTTPS_PROXY\", \"https_proxy\", \"HTTP_PROXY\", \"http_proxy\"])\n\t\t\t: getScopedAlias(env, [\"HTTP_PROXY\", \"http_proxy\"]);\n\tif (!proxy) return \"\";\n\treturn proxy.includes(\"://\") ? proxy : `${protocol === \"wss\" ? \"https\" : \"http\"}://${proxy}`;\n}\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\tfor (const entry of readFileSync(\"/proc/self/environ\", \"utf-8\").split(\"\\0\")) {\n\t\t\t\tconst index = entry.indexOf(\"=\");\n\t\t\t\tif (index > 0) procEnvCache.set(entry.slice(0, index), entry.slice(index + 1));\n\t\t\t}\n\t\t} catch {\n\t\t\t// /proc/self/environ is optional in restricted runtimes.\n\t\t}\n\t}\n\n\treturn procEnvCache.get(name);\n}\n\nfunction getScopedAlias(env: ProviderEnv, names: string[]): string | undefined {\n\tconst hasScopedAlias = names.some((name) => Object.hasOwn(env, name));\n\tif (hasScopedAlias) {\n\t\tfor (const name of names) {\n\t\t\tif (Object.hasOwn(env, name)) return env[name];\n\t\t}\n\t\treturn undefined;\n\t}\n\tfor (const name of names) {\n\t\tconst value = getProviderEnvValue(name);\n\t\tif (value !== undefined) return value;\n\t}\n\treturn undefined;\n}\n\nfunction matchesNoProxy(hostname: string, port: number, value: string): boolean {\n\tif (value.trim() === \"*\") return true;\n\treturn value.split(/[\\s,]+/).some((token) => {\n\t\tif (!token) return false;\n\t\tconst match = token.match(/^(.+):(\\d+)$/);\n\t\tconst host = (match?.[1] ?? token).toLowerCase();\n\t\tif (match && Number(match[2]) !== port) return false;\n\t\tconst normalizedHostname = hostname.toLowerCase();\n\t\treturn host.startsWith(\"*\")\n\t\t\t? normalizedHostname.endsWith(host.slice(1))\n\t\t\t: host.startsWith(\".\")\n\t\t\t\t? normalizedHostname.endsWith(host)\n\t\t\t\t: normalizedHostname === host;\n\t});\n}\n"]}
+{"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;AAI/C;;;GAGG;AACH,wBAAgB,mBAAmB,CAAC,IAAI,EAAE,MAAM,EAAE,GAAG,CAAC,EAAE,WAAW,GAAG,MAAM,GAAG,SAAS,CAIvF;AAED,8EAA8E;AAC9E,wBAAgB,uBAAuB,CAAC,SAAS,EAAE,MAAM,EAAE,GAAG,EAAE,WAAW,GAAG,MAAM,CAanF","sourcesContent":["import type { ProviderEnv } from \"../types.js\";\n\nlet procEnvCache: Map<string, string> | null = null;\n\n/**\n * Resolve a provider environment value from a scoped override, process.env,\n * or the Bun sandbox fallback used by direct AI package consumers.\n */\nexport function getProviderEnvValue(name: string, env?: ProviderEnv): string | undefined {\n\tif (env && Object.hasOwn(env, name)) return env[name];\n\tif (typeof process !== \"undefined\" && process.env[name] !== undefined) return process.env[name];\n\treturn getBunSandboxEnvValue(name);\n}\n\n/** Resolve a WebSocket/HTTP proxy using the supplied provider environment. */\nexport function resolveProviderProxyUrl(targetUrl: string, env: ProviderEnv): string {\n\tconst target = new URL(targetUrl);\n\tconst protocol = target.protocol.replace(\":\", \"\");\n\tconst port = Number(target.port) || (protocol === \"wss\" || protocol === \"https\" ? 443 : 80);\n\tconst noProxy = getScopedAlias(env, [\"NO_PROXY\", \"no_proxy\"]);\n\tif (noProxy && matchesNoProxy(target.hostname, port, noProxy)) return \"\";\n\n\tconst proxy =\n\t\tprotocol === \"wss\" || protocol === \"https\"\n\t\t\t? getScopedAlias(env, [\"HTTPS_PROXY\", \"https_proxy\", \"HTTP_PROXY\", \"http_proxy\"])\n\t\t\t: getScopedAlias(env, [\"HTTP_PROXY\", \"http_proxy\"]);\n\tif (!proxy) return \"\";\n\treturn proxy.includes(\"://\") ? proxy : `${protocol === \"wss\" ? \"https\" : \"http\"}://${proxy}`;\n}\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\tfor (const entry of readFileSync(\"/proc/self/environ\", \"utf-8\").split(\"\\0\")) {\n\t\t\t\tconst index = entry.indexOf(\"=\");\n\t\t\t\tif (index > 0) procEnvCache.set(entry.slice(0, index), entry.slice(index + 1));\n\t\t\t}\n\t\t} catch {\n\t\t\t// /proc/self/environ is optional in restricted runtimes.\n\t\t}\n\t}\n\n\treturn procEnvCache.get(name);\n}\n\nfunction getScopedAlias(env: ProviderEnv, names: string[]): string | undefined {\n\tconst hasScopedAlias = names.some((name) => Object.hasOwn(env, name));\n\tif (hasScopedAlias) {\n\t\tfor (const name of names) {\n\t\t\tif (Object.hasOwn(env, name)) return env[name];\n\t\t}\n\t\treturn undefined;\n\t}\n\tfor (const name of names) {\n\t\tconst value = getProviderEnvValue(name);\n\t\tif (value !== undefined) return value;\n\t}\n\treturn undefined;\n}\n\nfunction stripBrackets(host: string): string {\n\treturn host.startsWith(\"[\") && host.endsWith(\"]\") ? host.slice(1, -1) : host;\n}\n\nfunction parseNoProxyEntry(entry: string): { host: string; port: number } | undefined {\n\tconst trimmed = entry.trim().toLowerCase();\n\tif (!trimmed) return undefined;\n\n\tif (trimmed.startsWith(\"[\")) {\n\t\tconst closingBracket = trimmed.indexOf(\"]\");\n\t\tif (closingBracket !== -1) {\n\t\t\tconst host = trimmed.slice(1, closingBracket);\n\t\t\tconst rest = trimmed.slice(closingBracket + 1);\n\t\t\tif (rest.startsWith(\":\")) {\n\t\t\t\tconst port = Number.parseInt(rest.slice(1), 10);\n\t\t\t\treturn { host, port: Number.isNaN(port) ? 0 : port };\n\t\t\t}\n\t\t\treturn { host, port: 0 };\n\t\t}\n\t}\n\n\t// A bare IPv6 literal has several colons and no port.\n\tif (trimmed.split(\":\").length > 2) return { host: trimmed, port: 0 };\n\n\tconst colonIndex = trimmed.lastIndexOf(\":\");\n\tif (colonIndex !== -1) {\n\t\tconst port = Number.parseInt(trimmed.slice(colonIndex + 1), 10);\n\t\tif (!Number.isNaN(port)) return { host: trimmed.slice(0, colonIndex), port };\n\t}\n\n\treturn { host: trimmed, port: 0 };\n}\n\nfunction matchesNoProxy(hostname: string, port: number, value: string): boolean {\n\tif (value.trim() === \"*\") return true;\n\tconst targetHost = stripBrackets(hostname.toLowerCase());\n\treturn value.split(/[\\s,]+/).some((token) => {\n\t\tconst parsed = parseNoProxyEntry(token);\n\t\tif (!parsed) return false;\n\t\tif (parsed.port && parsed.port !== port) return false;\n\n\t\tlet domain = stripBrackets(parsed.host);\n\t\tif (domain.startsWith(\"*.\")) {\n\t\t\tdomain = domain.slice(2);\n\t\t} else if (domain.startsWith(\".\") || domain.startsWith(\"*\")) {\n\t\t\tdomain = domain.slice(1);\n\t\t}\n\t\tif (!domain) return false;\n\n\t\t// A root-domain entry covers the domain itself and all of its subdomains.\n\t\treturn targetHost === domain || targetHost.endsWith(`.${domain}`);\n\t});\n}\n"]}

package/dist/utils/provider-env.js

@@ -61,22 +61,57 @@ function getScopedAlias(env, names) {
     }
     return undefined;
 }
+function stripBrackets(host) {
+    return host.startsWith("[") && host.endsWith("]") ? host.slice(1, -1) : host;
+}
+function parseNoProxyEntry(entry) {
+    const trimmed = entry.trim().toLowerCase();
+    if (!trimmed)
+        return undefined;
+    if (trimmed.startsWith("[")) {
+        const closingBracket = trimmed.indexOf("]");
+        if (closingBracket !== -1) {
+            const host = trimmed.slice(1, closingBracket);
+            const rest = trimmed.slice(closingBracket + 1);
+            if (rest.startsWith(":")) {
+                const port = Number.parseInt(rest.slice(1), 10);
+                return { host, port: Number.isNaN(port) ? 0 : port };
+            }
+            return { host, port: 0 };
+        }
+    }
+    // A bare IPv6 literal has several colons and no port.
+    if (trimmed.split(":").length > 2)
+        return { host: trimmed, port: 0 };
+    const colonIndex = trimmed.lastIndexOf(":");
+    if (colonIndex !== -1) {
+        const port = Number.parseInt(trimmed.slice(colonIndex + 1), 10);
+        if (!Number.isNaN(port))
+            return { host: trimmed.slice(0, colonIndex), port };
+    }
+    return { host: trimmed, port: 0 };
+}
 function matchesNoProxy(hostname, port, value) {
     if (value.trim() === "*")
         return true;
+    const targetHost = stripBrackets(hostname.toLowerCase());
     return value.split(/[\s,]+/).some((token) => {
-        if (!token)
+        const parsed = parseNoProxyEntry(token);
+        if (!parsed)
+            return false;
+        if (parsed.port && parsed.port !== port)
             return false;
-        const match = token.match(/^(.+):(\d+)$/);
-        const host = (match?.[1] ?? token).toLowerCase();
-        if (match && Number(match[2]) !== port)
+        let domain = stripBrackets(parsed.host);
+        if (domain.startsWith("*.")) {
+            domain = domain.slice(2);
+        }
+        else if (domain.startsWith(".") || domain.startsWith("*")) {
+            domain = domain.slice(1);
+        }
+        if (!domain)
             return false;
-        const normalizedHostname = hostname.toLowerCase();
-        return host.startsWith("*")
-            ? normalizedHostname.endsWith(host.slice(1))
-            : host.startsWith(".")
-                ? normalizedHostname.endsWith(host)
-                : normalizedHostname === host;
+        // A root-domain entry covers the domain itself and all of its subdomains.
+        return targetHost === domain || targetHost.endsWith(`.${domain}`);
     });
 }
 //# sourceMappingURL=provider-env.js.map

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

@@ -1 +1 @@
-{"version":3,"file":"provider-env.js","sourceRoot":"","sources":["../../src/utils/provider-env.ts"],"names":[],"mappings":"AAEA,IAAI,YAAY,GAA+B,IAAI,CAAC;AAEpD;;;GAGG;AACH,MAAM,UAAU,mBAAmB,CAAC,IAAY,EAAE,GAAiB,EAAsB;IACxF,IAAI,GAAG,IAAI,MAAM,CAAC,MAAM,CAAC,GAAG,EAAE,IAAI,CAAC;QAAE,OAAO,GAAG,CAAC,IAAI,CAAC,CAAC;IACtD,IAAI,OAAO,OAAO,KAAK,WAAW,IAAI,OAAO,CAAC,GAAG,CAAC,IAAI,CAAC,KAAK,SAAS;QAAE,OAAO,OAAO,CAAC,GAAG,CAAC,IAAI,CAAC,CAAC;IAChG,OAAO,qBAAqB,CAAC,IAAI,CAAC,CAAC;AAAA,CACnC;AAED,8EAA8E;AAC9E,MAAM,UAAU,uBAAuB,CAAC,SAAiB,EAAE,GAAgB,EAAU;IACpF,MAAM,MAAM,GAAG,IAAI,GAAG,CAAC,SAAS,CAAC,CAAC;IAClC,MAAM,QAAQ,GAAG,MAAM,CAAC,QAAQ,CAAC,OAAO,CAAC,GAAG,EAAE,EAAE,CAAC,CAAC;IAClD,MAAM,IAAI,GAAG,MAAM,CAAC,MAAM,CAAC,IAAI,CAAC,IAAI,CAAC,QAAQ,KAAK,KAAK,IAAI,QAAQ,KAAK,OAAO,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC;IAC5F,MAAM,OAAO,GAAG,cAAc,CAAC,GAAG,EAAE,CAAC,UAAU,EAAE,UAAU,CAAC,CAAC,CAAC;IAC9D,IAAI,OAAO,IAAI,cAAc,CAAC,MAAM,CAAC,QAAQ,EAAE,IAAI,EAAE,OAAO,CAAC;QAAE,OAAO,EAAE,CAAC;IAEzE,MAAM,KAAK,GACV,QAAQ,KAAK,KAAK,IAAI,QAAQ,KAAK,OAAO;QACzC,CAAC,CAAC,cAAc,CAAC,GAAG,EAAE,CAAC,aAAa,EAAE,aAAa,EAAE,YAAY,EAAE,YAAY,CAAC,CAAC;QACjF,CAAC,CAAC,cAAc,CAAC,GAAG,EAAE,CAAC,YAAY,EAAE,YAAY,CAAC,CAAC,CAAC;IACtD,IAAI,CAAC,KAAK;QAAE,OAAO,EAAE,CAAC;IACtB,OAAO,KAAK,CAAC,QAAQ,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,GAAG,QAAQ,KAAK,KAAK,CAAC,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC,MAAM,MAAM,KAAK,EAAE,CAAC;AAAA,CAC7F;AAED,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,KAAK,MAAM,KAAK,IAAI,YAAY,CAAC,oBAAoB,EAAE,OAAO,CAAC,CAAC,KAAK,CAAC,IAAI,CAAC,EAAE,CAAC;gBAC7E,MAAM,KAAK,GAAG,KAAK,CAAC,OAAO,CAAC,GAAG,CAAC,CAAC;gBACjC,IAAI,KAAK,GAAG,CAAC;oBAAE,YAAY,CAAC,GAAG,CAAC,KAAK,CAAC,KAAK,CAAC,CAAC,EAAE,KAAK,CAAC,EAAE,KAAK,CAAC,KAAK,CAAC,KAAK,GAAG,CAAC,CAAC,CAAC,CAAC;YAChF,CAAC;QACF,CAAC;QAAC,MAAM,CAAC;YACR,yDAAyD;QAC1D,CAAC;IACF,CAAC;IAED,OAAO,YAAY,CAAC,GAAG,CAAC,IAAI,CAAC,CAAC;AAAA,CAC9B;AAED,SAAS,cAAc,CAAC,GAAgB,EAAE,KAAe,EAAsB;IAC9E,MAAM,cAAc,GAAG,KAAK,CAAC,IAAI,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,MAAM,CAAC,MAAM,CAAC,GAAG,EAAE,IAAI,CAAC,CAAC,CAAC;IACtE,IAAI,cAAc,EAAE,CAAC;QACpB,KAAK,MAAM,IAAI,IAAI,KAAK,EAAE,CAAC;YAC1B,IAAI,MAAM,CAAC,MAAM,CAAC,GAAG,EAAE,IAAI,CAAC;gBAAE,OAAO,GAAG,CAAC,IAAI,CAAC,CAAC;QAChD,CAAC;QACD,OAAO,SAAS,CAAC;IAClB,CAAC;IACD,KAAK,MAAM,IAAI,IAAI,KAAK,EAAE,CAAC;QAC1B,MAAM,KAAK,GAAG,mBAAmB,CAAC,IAAI,CAAC,CAAC;QACxC,IAAI,KAAK,KAAK,SAAS;YAAE,OAAO,KAAK,CAAC;IACvC,CAAC;IACD,OAAO,SAAS,CAAC;AAAA,CACjB;AAED,SAAS,cAAc,CAAC,QAAgB,EAAE,IAAY,EAAE,KAAa,EAAW;IAC/E,IAAI,KAAK,CAAC,IAAI,EAAE,KAAK,GAAG;QAAE,OAAO,IAAI,CAAC;IACtC,OAAO,KAAK,CAAC,KAAK,CAAC,QAAQ,CAAC,CAAC,IAAI,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC;QAC5C,IAAI,CAAC,KAAK;YAAE,OAAO,KAAK,CAAC;QACzB,MAAM,KAAK,GAAG,KAAK,CAAC,KAAK,CAAC,cAAc,CAAC,CAAC;QAC1C,MAAM,IAAI,GAAG,CAAC,KAAK,EAAE,CAAC,CAAC,CAAC,IAAI,KAAK,CAAC,CAAC,WAAW,EAAE,CAAC;QACjD,IAAI,KAAK,IAAI,MAAM,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,KAAK,IAAI;YAAE,OAAO,KAAK,CAAC;QACrD,MAAM,kBAAkB,GAAG,QAAQ,CAAC,WAAW,EAAE,CAAC;QAClD,OAAO,IAAI,CAAC,UAAU,CAAC,GAAG,CAAC;YAC1B,CAAC,CAAC,kBAAkB,CAAC,QAAQ,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC;YAC5C,CAAC,CAAC,IAAI,CAAC,UAAU,CAAC,GAAG,CAAC;gBACrB,CAAC,CAAC,kBAAkB,CAAC,QAAQ,CAAC,IAAI,CAAC;gBACnC,CAAC,CAAC,kBAAkB,KAAK,IAAI,CAAC;IAAA,CAChC,CAAC,CAAC;AAAA,CACH","sourcesContent":["import type { ProviderEnv } from \"../types.js\";\n\nlet procEnvCache: Map<string, string> | null = null;\n\n/**\n * Resolve a provider environment value from a scoped override, process.env,\n * or the Bun sandbox fallback used by direct AI package consumers.\n */\nexport function getProviderEnvValue(name: string, env?: ProviderEnv): string | undefined {\n\tif (env && Object.hasOwn(env, name)) return env[name];\n\tif (typeof process !== \"undefined\" && process.env[name] !== undefined) return process.env[name];\n\treturn getBunSandboxEnvValue(name);\n}\n\n/** Resolve a WebSocket/HTTP proxy using the supplied provider environment. */\nexport function resolveProviderProxyUrl(targetUrl: string, env: ProviderEnv): string {\n\tconst target = new URL(targetUrl);\n\tconst protocol = target.protocol.replace(\":\", \"\");\n\tconst port = Number(target.port) || (protocol === \"wss\" || protocol === \"https\" ? 443 : 80);\n\tconst noProxy = getScopedAlias(env, [\"NO_PROXY\", \"no_proxy\"]);\n\tif (noProxy && matchesNoProxy(target.hostname, port, noProxy)) return \"\";\n\n\tconst proxy =\n\t\tprotocol === \"wss\" || protocol === \"https\"\n\t\t\t? getScopedAlias(env, [\"HTTPS_PROXY\", \"https_proxy\", \"HTTP_PROXY\", \"http_proxy\"])\n\t\t\t: getScopedAlias(env, [\"HTTP_PROXY\", \"http_proxy\"]);\n\tif (!proxy) return \"\";\n\treturn proxy.includes(\"://\") ? proxy : `${protocol === \"wss\" ? \"https\" : \"http\"}://${proxy}`;\n}\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\tfor (const entry of readFileSync(\"/proc/self/environ\", \"utf-8\").split(\"\\0\")) {\n\t\t\t\tconst index = entry.indexOf(\"=\");\n\t\t\t\tif (index > 0) procEnvCache.set(entry.slice(0, index), entry.slice(index + 1));\n\t\t\t}\n\t\t} catch {\n\t\t\t// /proc/self/environ is optional in restricted runtimes.\n\t\t}\n\t}\n\n\treturn procEnvCache.get(name);\n}\n\nfunction getScopedAlias(env: ProviderEnv, names: string[]): string | undefined {\n\tconst hasScopedAlias = names.some((name) => Object.hasOwn(env, name));\n\tif (hasScopedAlias) {\n\t\tfor (const name of names) {\n\t\t\tif (Object.hasOwn(env, name)) return env[name];\n\t\t}\n\t\treturn undefined;\n\t}\n\tfor (const name of names) {\n\t\tconst value = getProviderEnvValue(name);\n\t\tif (value !== undefined) return value;\n\t}\n\treturn undefined;\n}\n\nfunction matchesNoProxy(hostname: string, port: number, value: string): boolean {\n\tif (value.trim() === \"*\") return true;\n\treturn value.split(/[\\s,]+/).some((token) => {\n\t\tif (!token) return false;\n\t\tconst match = token.match(/^(.+):(\\d+)$/);\n\t\tconst host = (match?.[1] ?? token).toLowerCase();\n\t\tif (match && Number(match[2]) !== port) return false;\n\t\tconst normalizedHostname = hostname.toLowerCase();\n\t\treturn host.startsWith(\"*\")\n\t\t\t? normalizedHostname.endsWith(host.slice(1))\n\t\t\t: host.startsWith(\".\")\n\t\t\t\t? normalizedHostname.endsWith(host)\n\t\t\t\t: normalizedHostname === host;\n\t});\n}\n"]}
+{"version":3,"file":"provider-env.js","sourceRoot":"","sources":["../../src/utils/provider-env.ts"],"names":[],"mappings":"AAEA,IAAI,YAAY,GAA+B,IAAI,CAAC;AAEpD;;;GAGG;AACH,MAAM,UAAU,mBAAmB,CAAC,IAAY,EAAE,GAAiB,EAAsB;IACxF,IAAI,GAAG,IAAI,MAAM,CAAC,MAAM,CAAC,GAAG,EAAE,IAAI,CAAC;QAAE,OAAO,GAAG,CAAC,IAAI,CAAC,CAAC;IACtD,IAAI,OAAO,OAAO,KAAK,WAAW,IAAI,OAAO,CAAC,GAAG,CAAC,IAAI,CAAC,KAAK,SAAS;QAAE,OAAO,OAAO,CAAC,GAAG,CAAC,IAAI,CAAC,CAAC;IAChG,OAAO,qBAAqB,CAAC,IAAI,CAAC,CAAC;AAAA,CACnC;AAED,8EAA8E;AAC9E,MAAM,UAAU,uBAAuB,CAAC,SAAiB,EAAE,GAAgB,EAAU;IACpF,MAAM,MAAM,GAAG,IAAI,GAAG,CAAC,SAAS,CAAC,CAAC;IAClC,MAAM,QAAQ,GAAG,MAAM,CAAC,QAAQ,CAAC,OAAO,CAAC,GAAG,EAAE,EAAE,CAAC,CAAC;IAClD,MAAM,IAAI,GAAG,MAAM,CAAC,MAAM,CAAC,IAAI,CAAC,IAAI,CAAC,QAAQ,KAAK,KAAK,IAAI,QAAQ,KAAK,OAAO,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC;IAC5F,MAAM,OAAO,GAAG,cAAc,CAAC,GAAG,EAAE,CAAC,UAAU,EAAE,UAAU,CAAC,CAAC,CAAC;IAC9D,IAAI,OAAO,IAAI,cAAc,CAAC,MAAM,CAAC,QAAQ,EAAE,IAAI,EAAE,OAAO,CAAC;QAAE,OAAO,EAAE,CAAC;IAEzE,MAAM,KAAK,GACV,QAAQ,KAAK,KAAK,IAAI,QAAQ,KAAK,OAAO;QACzC,CAAC,CAAC,cAAc,CAAC,GAAG,EAAE,CAAC,aAAa,EAAE,aAAa,EAAE,YAAY,EAAE,YAAY,CAAC,CAAC;QACjF,CAAC,CAAC,cAAc,CAAC,GAAG,EAAE,CAAC,YAAY,EAAE,YAAY,CAAC,CAAC,CAAC;IACtD,IAAI,CAAC,KAAK;QAAE,OAAO,EAAE,CAAC;IACtB,OAAO,KAAK,CAAC,QAAQ,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,GAAG,QAAQ,KAAK,KAAK,CAAC,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC,MAAM,MAAM,KAAK,EAAE,CAAC;AAAA,CAC7F;AAED,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,KAAK,MAAM,KAAK,IAAI,YAAY,CAAC,oBAAoB,EAAE,OAAO,CAAC,CAAC,KAAK,CAAC,IAAI,CAAC,EAAE,CAAC;gBAC7E,MAAM,KAAK,GAAG,KAAK,CAAC,OAAO,CAAC,GAAG,CAAC,CAAC;gBACjC,IAAI,KAAK,GAAG,CAAC;oBAAE,YAAY,CAAC,GAAG,CAAC,KAAK,CAAC,KAAK,CAAC,CAAC,EAAE,KAAK,CAAC,EAAE,KAAK,CAAC,KAAK,CAAC,KAAK,GAAG,CAAC,CAAC,CAAC,CAAC;YAChF,CAAC;QACF,CAAC;QAAC,MAAM,CAAC;YACR,yDAAyD;QAC1D,CAAC;IACF,CAAC;IAED,OAAO,YAAY,CAAC,GAAG,CAAC,IAAI,CAAC,CAAC;AAAA,CAC9B;AAED,SAAS,cAAc,CAAC,GAAgB,EAAE,KAAe,EAAsB;IAC9E,MAAM,cAAc,GAAG,KAAK,CAAC,IAAI,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,MAAM,CAAC,MAAM,CAAC,GAAG,EAAE,IAAI,CAAC,CAAC,CAAC;IACtE,IAAI,cAAc,EAAE,CAAC;QACpB,KAAK,MAAM,IAAI,IAAI,KAAK,EAAE,CAAC;YAC1B,IAAI,MAAM,CAAC,MAAM,CAAC,GAAG,EAAE,IAAI,CAAC;gBAAE,OAAO,GAAG,CAAC,IAAI,CAAC,CAAC;QAChD,CAAC;QACD,OAAO,SAAS,CAAC;IAClB,CAAC;IACD,KAAK,MAAM,IAAI,IAAI,KAAK,EAAE,CAAC;QAC1B,MAAM,KAAK,GAAG,mBAAmB,CAAC,IAAI,CAAC,CAAC;QACxC,IAAI,KAAK,KAAK,SAAS;YAAE,OAAO,KAAK,CAAC;IACvC,CAAC;IACD,OAAO,SAAS,CAAC;AAAA,CACjB;AAED,SAAS,aAAa,CAAC,IAAY,EAAU;IAC5C,OAAO,IAAI,CAAC,UAAU,CAAC,GAAG,CAAC,IAAI,IAAI,CAAC,QAAQ,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC;AAAA,CAC7E;AAED,SAAS,iBAAiB,CAAC,KAAa,EAA8C;IACrF,MAAM,OAAO,GAAG,KAAK,CAAC,IAAI,EAAE,CAAC,WAAW,EAAE,CAAC;IAC3C,IAAI,CAAC,OAAO;QAAE,OAAO,SAAS,CAAC;IAE/B,IAAI,OAAO,CAAC,UAAU,CAAC,GAAG,CAAC,EAAE,CAAC;QAC7B,MAAM,cAAc,GAAG,OAAO,CAAC,OAAO,CAAC,GAAG,CAAC,CAAC;QAC5C,IAAI,cAAc,KAAK,CAAC,CAAC,EAAE,CAAC;YAC3B,MAAM,IAAI,GAAG,OAAO,CAAC,KAAK,CAAC,CAAC,EAAE,cAAc,CAAC,CAAC;YAC9C,MAAM,IAAI,GAAG,OAAO,CAAC,KAAK,CAAC,cAAc,GAAG,CAAC,CAAC,CAAC;YAC/C,IAAI,IAAI,CAAC,UAAU,CAAC,GAAG,CAAC,EAAE,CAAC;gBAC1B,MAAM,IAAI,GAAG,MAAM,CAAC,QAAQ,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC;gBAChD,OAAO,EAAE,IAAI,EAAE,IAAI,EAAE,MAAM,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,IAAI,EAAE,CAAC;YACtD,CAAC;YACD,OAAO,EAAE,IAAI,EAAE,IAAI,EAAE,CAAC,EAAE,CAAC;QAC1B,CAAC;IACF,CAAC;IAED,sDAAsD;IACtD,IAAI,OAAO,CAAC,KAAK,CAAC,GAAG,CAAC,CAAC,MAAM,GAAG,CAAC;QAAE,OAAO,EAAE,IAAI,EAAE,OAAO,EAAE,IAAI,EAAE,CAAC,EAAE,CAAC;IAErE,MAAM,UAAU,GAAG,OAAO,CAAC,WAAW,CAAC,GAAG,CAAC,CAAC;IAC5C,IAAI,UAAU,KAAK,CAAC,CAAC,EAAE,CAAC;QACvB,MAAM,IAAI,GAAG,MAAM,CAAC,QAAQ,CAAC,OAAO,CAAC,KAAK,CAAC,UAAU,GAAG,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC;QAChE,IAAI,CAAC,MAAM,CAAC,KAAK,CAAC,IAAI,CAAC;YAAE,OAAO,EAAE,IAAI,EAAE,OAAO,CAAC,KAAK,CAAC,CAAC,EAAE,UAAU,CAAC,EAAE,IAAI,EAAE,CAAC;IAC9E,CAAC;IAED,OAAO,EAAE,IAAI,EAAE,OAAO,EAAE,IAAI,EAAE,CAAC,EAAE,CAAC;AAAA,CAClC;AAED,SAAS,cAAc,CAAC,QAAgB,EAAE,IAAY,EAAE,KAAa,EAAW;IAC/E,IAAI,KAAK,CAAC,IAAI,EAAE,KAAK,GAAG;QAAE,OAAO,IAAI,CAAC;IACtC,MAAM,UAAU,GAAG,aAAa,CAAC,QAAQ,CAAC,WAAW,EAAE,CAAC,CAAC;IACzD,OAAO,KAAK,CAAC,KAAK,CAAC,QAAQ,CAAC,CAAC,IAAI,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC;QAC5C,MAAM,MAAM,GAAG,iBAAiB,CAAC,KAAK,CAAC,CAAC;QACxC,IAAI,CAAC,MAAM;YAAE,OAAO,KAAK,CAAC;QAC1B,IAAI,MAAM,CAAC,IAAI,IAAI,MAAM,CAAC,IAAI,KAAK,IAAI;YAAE,OAAO,KAAK,CAAC;QAEtD,IAAI,MAAM,GAAG,aAAa,CAAC,MAAM,CAAC,IAAI,CAAC,CAAC;QACxC,IAAI,MAAM,CAAC,UAAU,CAAC,IAAI,CAAC,EAAE,CAAC;YAC7B,MAAM,GAAG,MAAM,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC;QAC1B,CAAC;aAAM,IAAI,MAAM,CAAC,UAAU,CAAC,GAAG,CAAC,IAAI,MAAM,CAAC,UAAU,CAAC,GAAG,CAAC,EAAE,CAAC;YAC7D,MAAM,GAAG,MAAM,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC;QAC1B,CAAC;QACD,IAAI,CAAC,MAAM;YAAE,OAAO,KAAK,CAAC;QAE1B,0EAA0E;QAC1E,OAAO,UAAU,KAAK,MAAM,IAAI,UAAU,CAAC,QAAQ,CAAC,IAAI,MAAM,EAAE,CAAC,CAAC;IAAA,CAClE,CAAC,CAAC;AAAA,CACH","sourcesContent":["import type { ProviderEnv } from \"../types.js\";\n\nlet procEnvCache: Map<string, string> | null = null;\n\n/**\n * Resolve a provider environment value from a scoped override, process.env,\n * or the Bun sandbox fallback used by direct AI package consumers.\n */\nexport function getProviderEnvValue(name: string, env?: ProviderEnv): string | undefined {\n\tif (env && Object.hasOwn(env, name)) return env[name];\n\tif (typeof process !== \"undefined\" && process.env[name] !== undefined) return process.env[name];\n\treturn getBunSandboxEnvValue(name);\n}\n\n/** Resolve a WebSocket/HTTP proxy using the supplied provider environment. */\nexport function resolveProviderProxyUrl(targetUrl: string, env: ProviderEnv): string {\n\tconst target = new URL(targetUrl);\n\tconst protocol = target.protocol.replace(\":\", \"\");\n\tconst port = Number(target.port) || (protocol === \"wss\" || protocol === \"https\" ? 443 : 80);\n\tconst noProxy = getScopedAlias(env, [\"NO_PROXY\", \"no_proxy\"]);\n\tif (noProxy && matchesNoProxy(target.hostname, port, noProxy)) return \"\";\n\n\tconst proxy =\n\t\tprotocol === \"wss\" || protocol === \"https\"\n\t\t\t? getScopedAlias(env, [\"HTTPS_PROXY\", \"https_proxy\", \"HTTP_PROXY\", \"http_proxy\"])\n\t\t\t: getScopedAlias(env, [\"HTTP_PROXY\", \"http_proxy\"]);\n\tif (!proxy) return \"\";\n\treturn proxy.includes(\"://\") ? proxy : `${protocol === \"wss\" ? \"https\" : \"http\"}://${proxy}`;\n}\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\tfor (const entry of readFileSync(\"/proc/self/environ\", \"utf-8\").split(\"\\0\")) {\n\t\t\t\tconst index = entry.indexOf(\"=\");\n\t\t\t\tif (index > 0) procEnvCache.set(entry.slice(0, index), entry.slice(index + 1));\n\t\t\t}\n\t\t} catch {\n\t\t\t// /proc/self/environ is optional in restricted runtimes.\n\t\t}\n\t}\n\n\treturn procEnvCache.get(name);\n}\n\nfunction getScopedAlias(env: ProviderEnv, names: string[]): string | undefined {\n\tconst hasScopedAlias = names.some((name) => Object.hasOwn(env, name));\n\tif (hasScopedAlias) {\n\t\tfor (const name of names) {\n\t\t\tif (Object.hasOwn(env, name)) return env[name];\n\t\t}\n\t\treturn undefined;\n\t}\n\tfor (const name of names) {\n\t\tconst value = getProviderEnvValue(name);\n\t\tif (value !== undefined) return value;\n\t}\n\treturn undefined;\n}\n\nfunction stripBrackets(host: string): string {\n\treturn host.startsWith(\"[\") && host.endsWith(\"]\") ? host.slice(1, -1) : host;\n}\n\nfunction parseNoProxyEntry(entry: string): { host: string; port: number } | undefined {\n\tconst trimmed = entry.trim().toLowerCase();\n\tif (!trimmed) return undefined;\n\n\tif (trimmed.startsWith(\"[\")) {\n\t\tconst closingBracket = trimmed.indexOf(\"]\");\n\t\tif (closingBracket !== -1) {\n\t\t\tconst host = trimmed.slice(1, closingBracket);\n\t\t\tconst rest = trimmed.slice(closingBracket + 1);\n\t\t\tif (rest.startsWith(\":\")) {\n\t\t\t\tconst port = Number.parseInt(rest.slice(1), 10);\n\t\t\t\treturn { host, port: Number.isNaN(port) ? 0 : port };\n\t\t\t}\n\t\t\treturn { host, port: 0 };\n\t\t}\n\t}\n\n\t// A bare IPv6 literal has several colons and no port.\n\tif (trimmed.split(\":\").length > 2) return { host: trimmed, port: 0 };\n\n\tconst colonIndex = trimmed.lastIndexOf(\":\");\n\tif (colonIndex !== -1) {\n\t\tconst port = Number.parseInt(trimmed.slice(colonIndex + 1), 10);\n\t\tif (!Number.isNaN(port)) return { host: trimmed.slice(0, colonIndex), port };\n\t}\n\n\treturn { host: trimmed, port: 0 };\n}\n\nfunction matchesNoProxy(hostname: string, port: number, value: string): boolean {\n\tif (value.trim() === \"*\") return true;\n\tconst targetHost = stripBrackets(hostname.toLowerCase());\n\treturn value.split(/[\\s,]+/).some((token) => {\n\t\tconst parsed = parseNoProxyEntry(token);\n\t\tif (!parsed) return false;\n\t\tif (parsed.port && parsed.port !== port) return false;\n\n\t\tlet domain = stripBrackets(parsed.host);\n\t\tif (domain.startsWith(\"*.\")) {\n\t\t\tdomain = domain.slice(2);\n\t\t} else if (domain.startsWith(\".\") || domain.startsWith(\"*\")) {\n\t\t\tdomain = domain.slice(1);\n\t\t}\n\t\tif (!domain) return false;\n\n\t\t// A root-domain entry covers the domain itself and all of its subdomains.\n\t\treturn targetHost === domain || targetHost.endsWith(`.${domain}`);\n\t});\n}\n"]}

package/dist/utils/tool-call-id.d.ts

@@ -0,0 +1,11 @@
+import type { Api, AssistantMessage, Model } from "../types.js";
+type TargetModel = Pick<Model<Api>, "provider" | "api" | "id">;
+type SourceModel = Pick<AssistantMessage, "provider" | "api">;
+/** Normalize Responses correlation and item IDs identically for execution and replay. */
+export declare function normalizeResponsesToolCallId(id: string, source: SourceModel, target: SourceModel): string;
+/** Identify Google API model families that carry explicit function-call IDs. */
+export declare function requiresToolCallId(modelId: string): boolean;
+/** Create the provider's replay ID transformation; reuse one instance per request. */
+export declare function createProviderToolCallIdNormalizer(model: TargetModel, allowedResponsesProviders?: ReadonlySet<string>): (id: string, source?: SourceModel) => string;
+export {};
+//# sourceMappingURL=tool-call-id.d.ts.map

package/dist/utils/tool-call-id.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"tool-call-id.d.ts","sourceRoot":"","sources":["../../src/utils/tool-call-id.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,GAAG,EAAE,gBAAgB,EAAE,KAAK,EAAE,MAAM,aAAa,CAAC;AAGhE,KAAK,WAAW,GAAG,IAAI,CAAC,KAAK,CAAC,GAAG,CAAC,EAAE,UAAU,GAAG,KAAK,GAAG,IAAI,CAAC,CAAC;AAC/D,KAAK,WAAW,GAAG,IAAI,CAAC,gBAAgB,EAAE,UAAU,GAAG,KAAK,CAAC,CAAC;AAS9D,yFAAyF;AACzF,wBAAgB,4BAA4B,CAAC,EAAE,EAAE,MAAM,EAAE,MAAM,EAAE,WAAW,EAAE,MAAM,EAAE,WAAW,GAAG,MAAM,CASzG;AAED,gFAAgF;AAChF,wBAAgB,kBAAkB,CAAC,OAAO,EAAE,MAAM,GAAG,OAAO,CAO3D;AAED,sFAAsF;AACtF,wBAAgB,kCAAkC,CACjD,KAAK,EAAE,WAAW,EAClB,yBAAyB,CAAC,EAAE,WAAW,CAAC,MAAM,CAAC,GAC7C,CAAC,EAAE,EAAE,MAAM,EAAE,MAAM,CAAC,EAAE,WAAW,KAAK,MAAM,CA2B9C","sourcesContent":["import type { Api, AssistantMessage, Model } from \"../types.js\";\nimport { shortHash } from \"./hash.js\";\n\ntype TargetModel = Pick<Model<Api>, \"provider\" | \"api\" | \"id\">;\ntype SourceModel = Pick<AssistantMessage, \"provider\" | \"api\">;\n\nfunction normalizeResponsesPart(part: string): string {\n\treturn part\n\t\t.replace(/[^a-zA-Z0-9_-]/g, \"_\")\n\t\t.slice(0, 64)\n\t\t.replace(/_+$/, \"\");\n}\n\n/** Normalize Responses correlation and item IDs identically for execution and replay. */\nexport function normalizeResponsesToolCallId(id: string, source: SourceModel, target: SourceModel): string {\n\tconst [callId, itemId] = id.split(\"|\");\n\tconst call = normalizeResponsesPart(callId) || `call_${shortHash(callId)}`;\n\tif (itemId === undefined) return call;\n\tconst foreign = source.provider !== target.provider || source.api !== target.api;\n\tlet item = foreign ? `fc_${shortHash(itemId)}` : normalizeResponsesPart(itemId);\n\tif (!item.startsWith(\"fc_\")) item = normalizeResponsesPart(`fc_${item}`);\n\tif (!item.startsWith(\"fc_\") || item.length === 3) item = `fc_${shortHash(itemId)}`;\n\treturn `${call}|${item}`;\n}\n\n/** Identify Google API model families that carry explicit function-call IDs. */\nexport function requiresToolCallId(modelId: string): boolean {\n\tconst match = modelId.toLowerCase().match(/^gemini(?:-live)?-(\\d+)/);\n\treturn (\n\t\tmodelId.startsWith(\"claude-\") ||\n\t\tmodelId.startsWith(\"gpt-oss-\") ||\n\t\t(match !== null && Number.parseInt(match[1], 10) >= 3)\n\t);\n}\n\n/** Create the provider's replay ID transformation; reuse one instance per request. */\nexport function createProviderToolCallIdNormalizer(\n\tmodel: TargetModel,\n\tallowedResponsesProviders?: ReadonlySet<string>,\n): (id: string, source?: SourceModel) => string {\n\tif (model.api === \"mistral-conversations\") return createMistralNormalizer();\n\treturn (id: string, source: SourceModel = model): string => {\n\t\tswitch (model.api) {\n\t\t\tcase \"openai-responses\":\n\t\t\tcase \"openai-codex-responses\":\n\t\t\tcase \"azure-openai-responses\": {\n\t\t\t\tconst composite =\n\t\t\t\t\tallowedResponsesProviders?.has(model.provider) ??\n\t\t\t\t\t([\"openai\", \"openai-codex\", \"opencode\"].includes(model.provider) ||\n\t\t\t\t\t\t(model.api === \"azure-openai-responses\" && model.provider === \"azure-openai-responses\"));\n\t\t\t\treturn composite\n\t\t\t\t\t? normalizeResponsesToolCallId(id, source, model)\n\t\t\t\t\t: normalizeResponsesPart(id) || `call_${shortHash(id)}`;\n\t\t\t}\n\t\t\tcase \"openai-completions\":\n\t\t\t\treturn normalizeChatId(id, model.provider);\n\t\t\tcase \"anthropic-messages\":\n\t\t\tcase \"bedrock-converse-stream\":\n\t\t\t\treturn id.replace(/[^a-zA-Z0-9_-]/g, \"_\").slice(0, 64);\n\t\t\tcase \"google-generative-ai\":\n\t\t\tcase \"google-vertex\":\n\t\t\t\treturn requiresToolCallId(model.id) ? id.replace(/[^a-zA-Z0-9_-]/g, \"_\").slice(0, 64) : id;\n\t\t\tdefault:\n\t\t\t\treturn id;\n\t\t}\n\t};\n}\n\nfunction normalizeChatId(id: string, provider: string): string {\n\tif (!id.includes(\"|\")) return provider === \"openai\" ? id.slice(0, 40) : id;\n\tconst separatorIndex = id.indexOf(\"|\");\n\tconst callId = id.slice(0, separatorIndex).replace(/[^a-zA-Z0-9_-]/g, \"_\");\n\tconst itemId = id.slice(separatorIndex + 1).replace(/[^a-zA-Z0-9_-]/g, \"_\");\n\tconst combined = itemId.length > 0 ? `${callId}_${itemId}` : callId;\n\tif (combined.length <= 40) return combined;\n\tconst hash = shortHash(id).slice(0, 8);\n\treturn `${callId.slice(0, Math.max(1, 40 - hash.length - 1))}_${hash}`;\n}\n\nfunction createMistralNormalizer(): (id: string) => string {\n\tconst ids = new Map<string, string>();\n\tconst owners = new Map<string, string>();\n\treturn (id) => {\n\t\tconst existing = ids.get(id);\n\t\tif (existing) return existing;\n\t\tconst normalized = id.replace(/[^a-zA-Z0-9]/g, \"\");\n\t\tconst seed = normalized || id;\n\t\tlet attempt = 0;\n\t\twhile (true) {\n\t\t\tconst candidate =\n\t\t\t\tattempt === 0 && normalized.length === 9\n\t\t\t\t\t? normalized\n\t\t\t\t\t: shortHash(attempt === 0 ? seed : `${seed}:${attempt}`)\n\t\t\t\t\t\t\t.replace(/[^a-zA-Z0-9]/g, \"\")\n\t\t\t\t\t\t\t.slice(0, 9);\n\t\t\tconst owner = owners.get(candidate);\n\t\t\tif (!owner || owner === id) {\n\t\t\t\tids.set(id, candidate);\n\t\t\t\towners.set(candidate, id);\n\t\t\t\treturn candidate;\n\t\t\t}\n\t\t\tattempt++;\n\t\t}\n\t};\n}\n"]}

package/dist/utils/tool-call-id.js

@@ -0,0 +1,96 @@
+import { shortHash } from "./hash.js";
+function normalizeResponsesPart(part) {
+    return part
+        .replace(/[^a-zA-Z0-9_-]/g, "_")
+        .slice(0, 64)
+        .replace(/_+$/, "");
+}
+/** Normalize Responses correlation and item IDs identically for execution and replay. */
+export function normalizeResponsesToolCallId(id, source, target) {
+    const [callId, itemId] = id.split("|");
+    const call = normalizeResponsesPart(callId) || `call_${shortHash(callId)}`;
+    if (itemId === undefined)
+        return call;
+    const foreign = source.provider !== target.provider || source.api !== target.api;
+    let item = foreign ? `fc_${shortHash(itemId)}` : normalizeResponsesPart(itemId);
+    if (!item.startsWith("fc_"))
+        item = normalizeResponsesPart(`fc_${item}`);
+    if (!item.startsWith("fc_") || item.length === 3)
+        item = `fc_${shortHash(itemId)}`;
+    return `${call}|${item}`;
+}
+/** Identify Google API model families that carry explicit function-call IDs. */
+export function requiresToolCallId(modelId) {
+    const match = modelId.toLowerCase().match(/^gemini(?:-live)?-(\d+)/);
+    return (modelId.startsWith("claude-") ||
+        modelId.startsWith("gpt-oss-") ||
+        (match !== null && Number.parseInt(match[1], 10) >= 3));
+}
+/** Create the provider's replay ID transformation; reuse one instance per request. */
+export function createProviderToolCallIdNormalizer(model, allowedResponsesProviders) {
+    if (model.api === "mistral-conversations")
+        return createMistralNormalizer();
+    return (id, source = model) => {
+        switch (model.api) {
+            case "openai-responses":
+            case "openai-codex-responses":
+            case "azure-openai-responses": {
+                const composite = allowedResponsesProviders?.has(model.provider) ??
+                    (["openai", "openai-codex", "opencode"].includes(model.provider) ||
+                        (model.api === "azure-openai-responses" && model.provider === "azure-openai-responses"));
+                return composite
+                    ? normalizeResponsesToolCallId(id, source, model)
+                    : normalizeResponsesPart(id) || `call_${shortHash(id)}`;
+            }
+            case "openai-completions":
+                return normalizeChatId(id, model.provider);
+            case "anthropic-messages":
+            case "bedrock-converse-stream":
+                return id.replace(/[^a-zA-Z0-9_-]/g, "_").slice(0, 64);
+            case "google-generative-ai":
+            case "google-vertex":
+                return requiresToolCallId(model.id) ? id.replace(/[^a-zA-Z0-9_-]/g, "_").slice(0, 64) : id;
+            default:
+                return id;
+        }
+    };
+}
+function normalizeChatId(id, provider) {
+    if (!id.includes("|"))
+        return provider === "openai" ? id.slice(0, 40) : id;
+    const separatorIndex = id.indexOf("|");
+    const callId = id.slice(0, separatorIndex).replace(/[^a-zA-Z0-9_-]/g, "_");
+    const itemId = id.slice(separatorIndex + 1).replace(/[^a-zA-Z0-9_-]/g, "_");
+    const combined = itemId.length > 0 ? `${callId}_${itemId}` : callId;
+    if (combined.length <= 40)
+        return combined;
+    const hash = shortHash(id).slice(0, 8);
+    return `${callId.slice(0, Math.max(1, 40 - hash.length - 1))}_${hash}`;
+}
+function createMistralNormalizer() {
+    const ids = new Map();
+    const owners = new Map();
+    return (id) => {
+        const existing = ids.get(id);
+        if (existing)
+            return existing;
+        const normalized = id.replace(/[^a-zA-Z0-9]/g, "");
+        const seed = normalized || id;
+        let attempt = 0;
+        while (true) {
+            const candidate = attempt === 0 && normalized.length === 9
+                ? normalized
+                : shortHash(attempt === 0 ? seed : `${seed}:${attempt}`)
+                    .replace(/[^a-zA-Z0-9]/g, "")
+                    .slice(0, 9);
+            const owner = owners.get(candidate);
+            if (!owner || owner === id) {
+                ids.set(id, candidate);
+                owners.set(candidate, id);
+                return candidate;
+            }
+            attempt++;
+        }
+    };
+}
+//# sourceMappingURL=tool-call-id.js.map

package/dist/utils/tool-call-id.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"tool-call-id.js","sourceRoot":"","sources":["../../src/utils/tool-call-id.ts"],"names":[],"mappings":"AACA,OAAO,EAAE,SAAS,EAAE,MAAM,WAAW,CAAC;AAKtC,SAAS,sBAAsB,CAAC,IAAY,EAAU;IACrD,OAAO,IAAI;SACT,OAAO,CAAC,iBAAiB,EAAE,GAAG,CAAC;SAC/B,KAAK,CAAC,CAAC,EAAE,EAAE,CAAC;SACZ,OAAO,CAAC,KAAK,EAAE,EAAE,CAAC,CAAC;AAAA,CACrB;AAED,yFAAyF;AACzF,MAAM,UAAU,4BAA4B,CAAC,EAAU,EAAE,MAAmB,EAAE,MAAmB,EAAU;IAC1G,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,GAAG,EAAE,CAAC,KAAK,CAAC,GAAG,CAAC,CAAC;IACvC,MAAM,IAAI,GAAG,sBAAsB,CAAC,MAAM,CAAC,IAAI,QAAQ,SAAS,CAAC,MAAM,CAAC,EAAE,CAAC;IAC3E,IAAI,MAAM,KAAK,SAAS;QAAE,OAAO,IAAI,CAAC;IACtC,MAAM,OAAO,GAAG,MAAM,CAAC,QAAQ,KAAK,MAAM,CAAC,QAAQ,IAAI,MAAM,CAAC,GAAG,KAAK,MAAM,CAAC,GAAG,CAAC;IACjF,IAAI,IAAI,GAAG,OAAO,CAAC,CAAC,CAAC,MAAM,SAAS,CAAC,MAAM,CAAC,EAAE,CAAC,CAAC,CAAC,sBAAsB,CAAC,MAAM,CAAC,CAAC;IAChF,IAAI,CAAC,IAAI,CAAC,UAAU,CAAC,KAAK,CAAC;QAAE,IAAI,GAAG,sBAAsB,CAAC,MAAM,IAAI,EAAE,CAAC,CAAC;IACzE,IAAI,CAAC,IAAI,CAAC,UAAU,CAAC,KAAK,CAAC,IAAI,IAAI,CAAC,MAAM,KAAK,CAAC;QAAE,IAAI,GAAG,MAAM,SAAS,CAAC,MAAM,CAAC,EAAE,CAAC;IACnF,OAAO,GAAG,IAAI,IAAI,IAAI,EAAE,CAAC;AAAA,CACzB;AAED,gFAAgF;AAChF,MAAM,UAAU,kBAAkB,CAAC,OAAe,EAAW;IAC5D,MAAM,KAAK,GAAG,OAAO,CAAC,WAAW,EAAE,CAAC,KAAK,CAAC,yBAAyB,CAAC,CAAC;IACrE,OAAO,CACN,OAAO,CAAC,UAAU,CAAC,SAAS,CAAC;QAC7B,OAAO,CAAC,UAAU,CAAC,UAAU,CAAC;QAC9B,CAAC,KAAK,KAAK,IAAI,IAAI,MAAM,CAAC,QAAQ,CAAC,KAAK,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,IAAI,CAAC,CAAC,CACtD,CAAC;AAAA,CACF;AAED,sFAAsF;AACtF,MAAM,UAAU,kCAAkC,CACjD,KAAkB,EAClB,yBAA+C,EACA;IAC/C,IAAI,KAAK,CAAC,GAAG,KAAK,uBAAuB;QAAE,OAAO,uBAAuB,EAAE,CAAC;IAC5E,OAAO,CAAC,EAAU,EAAE,MAAM,GAAgB,KAAK,EAAU,EAAE,CAAC;QAC3D,QAAQ,KAAK,CAAC,GAAG,EAAE,CAAC;YACnB,KAAK,kBAAkB,CAAC;YACxB,KAAK,wBAAwB,CAAC;YAC9B,KAAK,wBAAwB,EAAE,CAAC;gBAC/B,MAAM,SAAS,GACd,yBAAyB,EAAE,GAAG,CAAC,KAAK,CAAC,QAAQ,CAAC;oBAC9C,CAAC,CAAC,QAAQ,EAAE,cAAc,EAAE,UAAU,CAAC,CAAC,QAAQ,CAAC,KAAK,CAAC,QAAQ,CAAC;wBAC/D,CAAC,KAAK,CAAC,GAAG,KAAK,wBAAwB,IAAI,KAAK,CAAC,QAAQ,KAAK,wBAAwB,CAAC,CAAC,CAAC;gBAC3F,OAAO,SAAS;oBACf,CAAC,CAAC,4BAA4B,CAAC,EAAE,EAAE,MAAM,EAAE,KAAK,CAAC;oBACjD,CAAC,CAAC,sBAAsB,CAAC,EAAE,CAAC,IAAI,QAAQ,SAAS,CAAC,EAAE,CAAC,EAAE,CAAC;YAC1D,CAAC;YACD,KAAK,oBAAoB;gBACxB,OAAO,eAAe,CAAC,EAAE,EAAE,KAAK,CAAC,QAAQ,CAAC,CAAC;YAC5C,KAAK,oBAAoB,CAAC;YAC1B,KAAK,yBAAyB;gBAC7B,OAAO,EAAE,CAAC,OAAO,CAAC,iBAAiB,EAAE,GAAG,CAAC,CAAC,KAAK,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC;YACxD,KAAK,sBAAsB,CAAC;YAC5B,KAAK,eAAe;gBACnB,OAAO,kBAAkB,CAAC,KAAK,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC,OAAO,CAAC,iBAAiB,EAAE,GAAG,CAAC,CAAC,KAAK,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC;YAC5F;gBACC,OAAO,EAAE,CAAC;QACZ,CAAC;IAAA,CACD,CAAC;AAAA,CACF;AAED,SAAS,eAAe,CAAC,EAAU,EAAE,QAAgB,EAAU;IAC9D,IAAI,CAAC,EAAE,CAAC,QAAQ,CAAC,GAAG,CAAC;QAAE,OAAO,QAAQ,KAAK,QAAQ,CAAC,CAAC,CAAC,EAAE,CAAC,KAAK,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC;IAC3E,MAAM,cAAc,GAAG,EAAE,CAAC,OAAO,CAAC,GAAG,CAAC,CAAC;IACvC,MAAM,MAAM,GAAG,EAAE,CAAC,KAAK,CAAC,CAAC,EAAE,cAAc,CAAC,CAAC,OAAO,CAAC,iBAAiB,EAAE,GAAG,CAAC,CAAC;IAC3E,MAAM,MAAM,GAAG,EAAE,CAAC,KAAK,CAAC,cAAc,GAAG,CAAC,CAAC,CAAC,OAAO,CAAC,iBAAiB,EAAE,GAAG,CAAC,CAAC;IAC5E,MAAM,QAAQ,GAAG,MAAM,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,CAAC,GAAG,MAAM,IAAI,MAAM,EAAE,CAAC,CAAC,CAAC,MAAM,CAAC;IACpE,IAAI,QAAQ,CAAC,MAAM,IAAI,EAAE;QAAE,OAAO,QAAQ,CAAC;IAC3C,MAAM,IAAI,GAAG,SAAS,CAAC,EAAE,CAAC,CAAC,KAAK,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC;IACvC,OAAO,GAAG,MAAM,CAAC,KAAK,CAAC,CAAC,EAAE,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,EAAE,GAAG,IAAI,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,IAAI,IAAI,EAAE,CAAC;AAAA,CACvE;AAED,SAAS,uBAAuB,GAA2B;IAC1D,MAAM,GAAG,GAAG,IAAI,GAAG,EAAkB,CAAC;IACtC,MAAM,MAAM,GAAG,IAAI,GAAG,EAAkB,CAAC;IACzC,OAAO,CAAC,EAAE,EAAE,EAAE,CAAC;QACd,MAAM,QAAQ,GAAG,GAAG,CAAC,GAAG,CAAC,EAAE,CAAC,CAAC;QAC7B,IAAI,QAAQ;YAAE,OAAO,QAAQ,CAAC;QAC9B,MAAM,UAAU,GAAG,EAAE,CAAC,OAAO,CAAC,eAAe,EAAE,EAAE,CAAC,CAAC;QACnD,MAAM,IAAI,GAAG,UAAU,IAAI,EAAE,CAAC;QAC9B,IAAI,OAAO,GAAG,CAAC,CAAC;QAChB,OAAO,IAAI,EAAE,CAAC;YACb,MAAM,SAAS,GACd,OAAO,KAAK,CAAC,IAAI,UAAU,CAAC,MAAM,KAAK,CAAC;gBACvC,CAAC,CAAC,UAAU;gBACZ,CAAC,CAAC,SAAS,CAAC,OAAO,KAAK,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,GAAG,IAAI,IAAI,OAAO,EAAE,CAAC;qBACrD,OAAO,CAAC,eAAe,EAAE,EAAE,CAAC;qBAC5B,KAAK,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC;YACjB,MAAM,KAAK,GAAG,MAAM,CAAC,GAAG,CAAC,SAAS,CAAC,CAAC;YACpC,IAAI,CAAC,KAAK,IAAI,KAAK,KAAK,EAAE,EAAE,CAAC;gBAC5B,GAAG,CAAC,GAAG,CAAC,EAAE,EAAE,SAAS,CAAC,CAAC;gBACvB,MAAM,CAAC,GAAG,CAAC,SAAS,EAAE,EAAE,CAAC,CAAC;gBAC1B,OAAO,SAAS,CAAC;YAClB,CAAC;YACD,OAAO,EAAE,CAAC;QACX,CAAC;IAAA,CACD,CAAC;AAAA,CACF","sourcesContent":["import type { Api, AssistantMessage, Model } from \"../types.js\";\nimport { shortHash } from \"./hash.js\";\n\ntype TargetModel = Pick<Model<Api>, \"provider\" | \"api\" | \"id\">;\ntype SourceModel = Pick<AssistantMessage, \"provider\" | \"api\">;\n\nfunction normalizeResponsesPart(part: string): string {\n\treturn part\n\t\t.replace(/[^a-zA-Z0-9_-]/g, \"_\")\n\t\t.slice(0, 64)\n\t\t.replace(/_+$/, \"\");\n}\n\n/** Normalize Responses correlation and item IDs identically for execution and replay. */\nexport function normalizeResponsesToolCallId(id: string, source: SourceModel, target: SourceModel): string {\n\tconst [callId, itemId] = id.split(\"|\");\n\tconst call = normalizeResponsesPart(callId) || `call_${shortHash(callId)}`;\n\tif (itemId === undefined) return call;\n\tconst foreign = source.provider !== target.provider || source.api !== target.api;\n\tlet item = foreign ? `fc_${shortHash(itemId)}` : normalizeResponsesPart(itemId);\n\tif (!item.startsWith(\"fc_\")) item = normalizeResponsesPart(`fc_${item}`);\n\tif (!item.startsWith(\"fc_\") || item.length === 3) item = `fc_${shortHash(itemId)}`;\n\treturn `${call}|${item}`;\n}\n\n/** Identify Google API model families that carry explicit function-call IDs. */\nexport function requiresToolCallId(modelId: string): boolean {\n\tconst match = modelId.toLowerCase().match(/^gemini(?:-live)?-(\\d+)/);\n\treturn (\n\t\tmodelId.startsWith(\"claude-\") ||\n\t\tmodelId.startsWith(\"gpt-oss-\") ||\n\t\t(match !== null && Number.parseInt(match[1], 10) >= 3)\n\t);\n}\n\n/** Create the provider's replay ID transformation; reuse one instance per request. */\nexport function createProviderToolCallIdNormalizer(\n\tmodel: TargetModel,\n\tallowedResponsesProviders?: ReadonlySet<string>,\n): (id: string, source?: SourceModel) => string {\n\tif (model.api === \"mistral-conversations\") return createMistralNormalizer();\n\treturn (id: string, source: SourceModel = model): string => {\n\t\tswitch (model.api) {\n\t\t\tcase \"openai-responses\":\n\t\t\tcase \"openai-codex-responses\":\n\t\t\tcase \"azure-openai-responses\": {\n\t\t\t\tconst composite =\n\t\t\t\t\tallowedResponsesProviders?.has(model.provider) ??\n\t\t\t\t\t([\"openai\", \"openai-codex\", \"opencode\"].includes(model.provider) ||\n\t\t\t\t\t\t(model.api === \"azure-openai-responses\" && model.provider === \"azure-openai-responses\"));\n\t\t\t\treturn composite\n\t\t\t\t\t? normalizeResponsesToolCallId(id, source, model)\n\t\t\t\t\t: normalizeResponsesPart(id) || `call_${shortHash(id)}`;\n\t\t\t}\n\t\t\tcase \"openai-completions\":\n\t\t\t\treturn normalizeChatId(id, model.provider);\n\t\t\tcase \"anthropic-messages\":\n\t\t\tcase \"bedrock-converse-stream\":\n\t\t\t\treturn id.replace(/[^a-zA-Z0-9_-]/g, \"_\").slice(0, 64);\n\t\t\tcase \"google-generative-ai\":\n\t\t\tcase \"google-vertex\":\n\t\t\t\treturn requiresToolCallId(model.id) ? id.replace(/[^a-zA-Z0-9_-]/g, \"_\").slice(0, 64) : id;\n\t\t\tdefault:\n\t\t\t\treturn id;\n\t\t}\n\t};\n}\n\nfunction normalizeChatId(id: string, provider: string): string {\n\tif (!id.includes(\"|\")) return provider === \"openai\" ? id.slice(0, 40) : id;\n\tconst separatorIndex = id.indexOf(\"|\");\n\tconst callId = id.slice(0, separatorIndex).replace(/[^a-zA-Z0-9_-]/g, \"_\");\n\tconst itemId = id.slice(separatorIndex + 1).replace(/[^a-zA-Z0-9_-]/g, \"_\");\n\tconst combined = itemId.length > 0 ? `${callId}_${itemId}` : callId;\n\tif (combined.length <= 40) return combined;\n\tconst hash = shortHash(id).slice(0, 8);\n\treturn `${callId.slice(0, Math.max(1, 40 - hash.length - 1))}_${hash}`;\n}\n\nfunction createMistralNormalizer(): (id: string) => string {\n\tconst ids = new Map<string, string>();\n\tconst owners = new Map<string, string>();\n\treturn (id) => {\n\t\tconst existing = ids.get(id);\n\t\tif (existing) return existing;\n\t\tconst normalized = id.replace(/[^a-zA-Z0-9]/g, \"\");\n\t\tconst seed = normalized || id;\n\t\tlet attempt = 0;\n\t\twhile (true) {\n\t\t\tconst candidate =\n\t\t\t\tattempt === 0 && normalized.length === 9\n\t\t\t\t\t? normalized\n\t\t\t\t\t: shortHash(attempt === 0 ? seed : `${seed}:${attempt}`)\n\t\t\t\t\t\t\t.replace(/[^a-zA-Z0-9]/g, \"\")\n\t\t\t\t\t\t\t.slice(0, 9);\n\t\t\tconst owner = owners.get(candidate);\n\t\t\tif (!owner || owner === id) {\n\t\t\t\tids.set(id, candidate);\n\t\t\t\towners.set(candidate, id);\n\t\t\t\treturn candidate;\n\t\t\t}\n\t\t\tattempt++;\n\t\t}\n\t};\n}\n"]}

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

@@ -1 +1 @@
-{"version":3,"file":"validation.d.ts","sourceRoot":"","sources":["../../src/utils/validation.ts"],"names":[],"mappings":"AAGA,OAAO,KAAK,EAAE,IAAI,EAAE,QAAQ,EAAE,MAAM,aAAa,CAAC;AA0QlD;;;;;;GAMG;AACH,wBAAgB,gBAAgB,CAAC,KAAK,EAAE,IAAI,EAAE,EAAE,QAAQ,EAAE,QAAQ,GAAG,GAAG,CAMvE;AAED;;;;;;GAMG;AACH,wBAAgB,qBAAqB,CAAC,IAAI,EAAE,IAAI,EAAE,QAAQ,EAAE,QAAQ,GAAG,GAAG,CAgCzE","sourcesContent":["import { Compile } from \"typebox/compile\";\nimport type { TLocalizedValidationError } from \"typebox/error\";\nimport { Value } from \"typebox/value\";\nimport type { Tool, ToolCall } from \"../types.js\";\n\nconst validatorCache = new WeakMap<object, ReturnType<typeof Compile>>();\nconst TYPEBOX_KIND = Symbol.for(\"TypeBox.Kind\");\n\ninterface JsonSchemaObject {\n\ttype?: string | string[];\n\tproperties?: Record<string, JsonSchemaObject>;\n\titems?: JsonSchemaObject | JsonSchemaObject[];\n\tadditionalProperties?: boolean | JsonSchemaObject;\n\tallOf?: JsonSchemaObject[];\n\tanyOf?: JsonSchemaObject[];\n\toneOf?: JsonSchemaObject[];\n}\n\nfunction isRecord(value: unknown): value is Record<string, unknown> {\n\treturn typeof value === \"object\" && value !== null;\n}\n\nfunction isJsonSchemaObject(value: unknown): value is JsonSchemaObject {\n\treturn isRecord(value);\n}\n\nfunction hasTypeBoxMetadata(schema: unknown): boolean {\n\treturn isRecord(schema) && Object.getOwnPropertySymbols(schema).includes(TYPEBOX_KIND);\n}\n\nfunction getSchemaTypes(schema: JsonSchemaObject): string[] {\n\tif (typeof schema.type === \"string\") {\n\t\treturn [schema.type];\n\t}\n\tif (Array.isArray(schema.type)) {\n\t\treturn schema.type.filter((type): type is string => typeof type === \"string\");\n\t}\n\treturn [];\n}\n\nfunction matchesJsonType(value: unknown, type: string): boolean {\n\tswitch (type) {\n\t\tcase \"number\":\n\t\t\treturn typeof value === \"number\";\n\t\tcase \"integer\":\n\t\t\treturn typeof value === \"number\" && Number.isInteger(value);\n\t\tcase \"boolean\":\n\t\t\treturn typeof value === \"boolean\";\n\t\tcase \"string\":\n\t\t\treturn typeof value === \"string\";\n\t\tcase \"null\":\n\t\t\treturn value === null;\n\t\tcase \"array\":\n\t\t\treturn Array.isArray(value);\n\t\tcase \"object\":\n\t\t\treturn isRecord(value) && !Array.isArray(value);\n\t\tdefault:\n\t\t\treturn false;\n\t}\n}\n\nfunction isValidatorSchema(value: unknown): value is Tool[\"parameters\"] {\n\treturn isRecord(value);\n}\n\nfunction getSubSchemaValidator(schema: JsonSchemaObject): ReturnType<typeof Compile> | undefined {\n\tif (!isValidatorSchema(schema)) {\n\t\treturn undefined;\n\t}\n\ttry {\n\t\treturn getValidator(schema);\n\t} catch {\n\t\treturn undefined;\n\t}\n}\n\nfunction coercePrimitiveByType(value: unknown, type: string): unknown {\n\tswitch (type) {\n\t\tcase \"number\": {\n\t\t\tif (value === null) {\n\t\t\t\treturn 0;\n\t\t\t}\n\t\t\tif (typeof value === \"string\" && value.trim() !== \"\") {\n\t\t\t\tconst parsed = Number(value);\n\t\t\t\tif (Number.isFinite(parsed)) {\n\t\t\t\t\treturn parsed;\n\t\t\t\t}\n\t\t\t}\n\t\t\tif (typeof value === \"boolean\") {\n\t\t\t\treturn value ? 1 : 0;\n\t\t\t}\n\t\t\treturn value;\n\t\t}\n\t\tcase \"integer\": {\n\t\t\tif (value === null) {\n\t\t\t\treturn 0;\n\t\t\t}\n\t\t\tif (typeof value === \"string\" && value.trim() !== \"\") {\n\t\t\t\tconst parsed = Number(value);\n\t\t\t\tif (Number.isInteger(parsed)) {\n\t\t\t\t\treturn parsed;\n\t\t\t\t}\n\t\t\t}\n\t\t\tif (typeof value === \"boolean\") {\n\t\t\t\treturn value ? 1 : 0;\n\t\t\t}\n\t\t\treturn value;\n\t\t}\n\t\tcase \"boolean\": {\n\t\t\tif (value === null) {\n\t\t\t\treturn false;\n\t\t\t}\n\t\t\tif (typeof value === \"string\") {\n\t\t\t\tif (value === \"true\") {\n\t\t\t\t\treturn true;\n\t\t\t\t}\n\t\t\t\tif (value === \"false\") {\n\t\t\t\t\treturn false;\n\t\t\t\t}\n\t\t\t}\n\t\t\tif (typeof value === \"number\") {\n\t\t\t\tif (value === 1) {\n\t\t\t\t\treturn true;\n\t\t\t\t}\n\t\t\t\tif (value === 0) {\n\t\t\t\t\treturn false;\n\t\t\t\t}\n\t\t\t}\n\t\t\treturn value;\n\t\t}\n\t\tcase \"string\": {\n\t\t\tif (value === null) {\n\t\t\t\treturn \"\";\n\t\t\t}\n\t\t\tif (typeof value === \"number\" || typeof value === \"boolean\") {\n\t\t\t\treturn String(value);\n\t\t\t}\n\t\t\treturn value;\n\t\t}\n\t\tcase \"null\": {\n\t\t\tif (value === \"\" || value === 0 || value === false) {\n\t\t\t\treturn null;\n\t\t\t}\n\t\t\treturn value;\n\t\t}\n\t\tdefault:\n\t\t\treturn value;\n\t}\n}\n\nfunction applySchemaObjectCoercion(value: Record<string, unknown>, schema: JsonSchemaObject): void {\n\tconst properties = schema.properties;\n\tconst definedKeys = new Set<string>(properties ? Object.keys(properties) : []);\n\n\tif (properties) {\n\t\tfor (const [key, propertySchema] of Object.entries(properties)) {\n\t\t\tif (!(key in value)) {\n\t\t\t\tcontinue;\n\t\t\t}\n\t\t\tvalue[key] = coerceWithJsonSchema(value[key], propertySchema);\n\t\t}\n\t}\n\n\tif (schema.additionalProperties && isJsonSchemaObject(schema.additionalProperties)) {\n\t\tfor (const [key, propertyValue] of Object.entries(value)) {\n\t\t\tif (definedKeys.has(key)) {\n\t\t\t\tcontinue;\n\t\t\t}\n\t\t\tvalue[key] = coerceWithJsonSchema(propertyValue, schema.additionalProperties);\n\t\t}\n\t}\n}\n\nfunction applySchemaArrayCoercion(value: unknown[], schema: JsonSchemaObject): void {\n\tif (Array.isArray(schema.items)) {\n\t\tfor (let index = 0; index < value.length; index++) {\n\t\t\tconst itemSchema = schema.items[index];\n\t\t\tif (!itemSchema) {\n\t\t\t\tcontinue;\n\t\t\t}\n\t\t\tvalue[index] = coerceWithJsonSchema(value[index], itemSchema);\n\t\t}\n\t\treturn;\n\t}\n\n\tif (isJsonSchemaObject(schema.items)) {\n\t\tfor (let index = 0; index < value.length; index++) {\n\t\t\tvalue[index] = coerceWithJsonSchema(value[index], schema.items);\n\t\t}\n\t}\n}\n\nfunction coerceWithUnionSchema(value: unknown, schemas: JsonSchemaObject[]): unknown {\n\tfor (const schema of schemas) {\n\t\tconst candidate = structuredClone(value);\n\t\tconst coerced = coerceWithJsonSchema(candidate, schema);\n\t\tconst validator = getSubSchemaValidator(schema);\n\t\tif (validator?.Check(coerced)) {\n\t\t\treturn coerced;\n\t\t}\n\t}\n\treturn value;\n}\n\nfunction coerceWithJsonSchema(value: unknown, schema: JsonSchemaObject): unknown {\n\tlet nextValue = value;\n\n\tif (Array.isArray(schema.allOf)) {\n\t\tfor (const nested of schema.allOf) {\n\t\t\tnextValue = coerceWithJsonSchema(nextValue, nested);\n\t\t}\n\t}\n\n\tif (Array.isArray(schema.anyOf)) {\n\t\tnextValue = coerceWithUnionSchema(nextValue, schema.anyOf);\n\t}\n\n\tif (Array.isArray(schema.oneOf)) {\n\t\tnextValue = coerceWithUnionSchema(nextValue, schema.oneOf);\n\t}\n\n\tconst schemaTypes = getSchemaTypes(schema);\n\tconst matchesUnionMember =\n\t\tschemaTypes.length > 1 && schemaTypes.some((schemaType) => matchesJsonType(nextValue, schemaType));\n\tif (schemaTypes.length > 0 && !matchesUnionMember) {\n\t\tfor (const schemaType of schemaTypes) {\n\t\t\tconst candidate = coercePrimitiveByType(nextValue, schemaType);\n\t\t\tif (candidate !== nextValue) {\n\t\t\t\tnextValue = candidate;\n\t\t\t\tbreak;\n\t\t\t}\n\t\t}\n\t}\n\n\tif (schemaTypes.includes(\"object\") && isRecord(nextValue) && !Array.isArray(nextValue)) {\n\t\tapplySchemaObjectCoercion(nextValue, schema);\n\t}\n\n\tif (schemaTypes.includes(\"array\") && Array.isArray(nextValue)) {\n\t\tapplySchemaArrayCoercion(nextValue, schema);\n\t}\n\n\treturn nextValue;\n}\n\nfunction getValidator(schema: Tool[\"parameters\"]): ReturnType<typeof Compile> {\n\tconst key = schema as object;\n\tconst cached = validatorCache.get(key);\n\tif (cached) {\n\t\treturn cached;\n\t}\n\tconst validator = Compile(schema);\n\tvalidatorCache.set(key, validator);\n\treturn validator;\n}\n\nfunction formatValidationPath(error: TLocalizedValidationError): string {\n\tif (error.keyword === \"required\") {\n\t\tconst requiredProperties = (error.params as { requiredProperties?: string[] }).requiredProperties;\n\t\tconst requiredProperty = requiredProperties?.[0];\n\t\tif (requiredProperty) {\n\t\t\tconst basePath = error.instancePath.replace(/^\\//, \"\").replace(/\\//g, \".\");\n\t\t\treturn basePath ? `${basePath}.${requiredProperty}` : requiredProperty;\n\t\t}\n\t}\n\tconst path = error.instancePath.replace(/^\\//, \"\").replace(/\\//g, \".\");\n\treturn path || \"root\";\n}\n\n/**\n * Finds a tool by name and validates the tool call arguments against its TypeBox schema\n * @param tools Array of tool definitions\n * @param toolCall The tool call from the LLM\n * @returns The validated arguments\n * @throws Error if tool is not found or validation fails\n */\nexport function validateToolCall(tools: Tool[], toolCall: ToolCall): any {\n\tconst tool = tools.find((t) => t.name === toolCall.name);\n\tif (!tool) {\n\t\tthrow new Error(`Tool \"${toolCall.name}\" not found`);\n\t}\n\treturn validateToolArguments(tool, toolCall);\n}\n\n/**\n * Validates tool call arguments against the tool's TypeBox schema\n * @param tool The tool definition with TypeBox schema\n * @param toolCall The tool call from the LLM\n * @returns The validated (and potentially coerced) arguments\n * @throws Error with formatted message if validation fails\n */\nexport function validateToolArguments(tool: Tool, toolCall: ToolCall): any {\n\tconst args = structuredClone(toolCall.arguments);\n\tValue.Convert(tool.parameters, args);\n\n\tconst validator = getValidator(tool.parameters);\n\tif (!hasTypeBoxMetadata(tool.parameters) && isJsonSchemaObject(tool.parameters)) {\n\t\tconst coerced = coerceWithJsonSchema(args, tool.parameters);\n\t\tif (coerced !== args) {\n\t\t\tif (isRecord(args) && isRecord(coerced)) {\n\t\t\t\tfor (const key of Object.keys(args)) {\n\t\t\t\t\tdelete args[key];\n\t\t\t\t}\n\t\t\t\tObject.assign(args, coerced);\n\t\t\t} else {\n\t\t\t\treturn validator.Check(coerced) ? coerced : args;\n\t\t\t}\n\t\t}\n\t}\n\n\tif (validator.Check(args)) {\n\t\treturn args;\n\t}\n\n\tconst errors =\n\t\tvalidator\n\t\t\t.Errors(args)\n\t\t\t.map((error) => `  - ${formatValidationPath(error)}: ${error.message}`)\n\t\t\t.join(\"\\n\") || \"Unknown validation error\";\n\n\tconst errorMessage = `Validation failed for tool \"${toolCall.name}\":\\n${errors}\\n\\nReceived arguments:\\n${JSON.stringify(toolCall.arguments, null, 2)}`;\n\n\tthrow new Error(errorMessage);\n}\n"]}
+{"version":3,"file":"validation.d.ts","sourceRoot":"","sources":["../../src/utils/validation.ts"],"names":[],"mappings":"AAGA,OAAO,KAAK,EAAE,IAAI,EAAE,QAAQ,EAAE,MAAM,aAAa,CAAC;AAiRlD;;;;;;GAMG;AACH,wBAAgB,gBAAgB,CAAC,KAAK,EAAE,IAAI,EAAE,EAAE,QAAQ,EAAE,QAAQ,GAAG,GAAG,CAMvE;AAED;;;;;;GAMG;AACH,wBAAgB,qBAAqB,CAAC,IAAI,EAAE,IAAI,EAAE,QAAQ,EAAE,QAAQ,GAAG,GAAG,CAmCzE","sourcesContent":["import { Compile } from \"typebox/compile\";\nimport type { TLocalizedValidationError } from \"typebox/error\";\nimport { Value } from \"typebox/value\";\nimport type { Tool, ToolCall } from \"../types.js\";\n\nconst validatorCache = new WeakMap<object, ReturnType<typeof Compile>>();\nconst TYPEBOX_KIND = Symbol.for(\"TypeBox.Kind\");\n\ninterface JsonSchemaObject {\n\ttype?: string | string[];\n\tproperties?: Record<string, JsonSchemaObject>;\n\titems?: JsonSchemaObject | JsonSchemaObject[];\n\tadditionalProperties?: boolean | JsonSchemaObject;\n\tallOf?: JsonSchemaObject[];\n\tanyOf?: JsonSchemaObject[];\n\toneOf?: JsonSchemaObject[];\n}\n\nfunction isRecord(value: unknown): value is Record<string, unknown> {\n\treturn typeof value === \"object\" && value !== null;\n}\n\nfunction isJsonSchemaObject(value: unknown): value is JsonSchemaObject {\n\treturn isRecord(value);\n}\n\nfunction hasTypeBoxMetadata(schema: unknown): boolean {\n\treturn isRecord(schema) && Object.getOwnPropertySymbols(schema).includes(TYPEBOX_KIND);\n}\n\nfunction getSchemaTypes(schema: JsonSchemaObject): string[] {\n\tif (typeof schema.type === \"string\") {\n\t\treturn [schema.type];\n\t}\n\tif (Array.isArray(schema.type)) {\n\t\treturn schema.type.filter((type): type is string => typeof type === \"string\");\n\t}\n\treturn [];\n}\n\nfunction matchesJsonType(value: unknown, type: string): boolean {\n\tswitch (type) {\n\t\tcase \"number\":\n\t\t\treturn typeof value === \"number\";\n\t\tcase \"integer\":\n\t\t\treturn typeof value === \"number\" && Number.isInteger(value);\n\t\tcase \"boolean\":\n\t\t\treturn typeof value === \"boolean\";\n\t\tcase \"string\":\n\t\t\treturn typeof value === \"string\";\n\t\tcase \"null\":\n\t\t\treturn value === null;\n\t\tcase \"array\":\n\t\t\treturn Array.isArray(value);\n\t\tcase \"object\":\n\t\t\treturn isRecord(value) && !Array.isArray(value);\n\t\tdefault:\n\t\t\treturn false;\n\t}\n}\n\nfunction isValidatorSchema(value: unknown): value is Tool[\"parameters\"] {\n\treturn isRecord(value);\n}\n\nfunction getSubSchemaValidator(schema: JsonSchemaObject): ReturnType<typeof Compile> | undefined {\n\tif (!isValidatorSchema(schema)) {\n\t\treturn undefined;\n\t}\n\ttry {\n\t\treturn getValidator(schema);\n\t} catch {\n\t\treturn undefined;\n\t}\n}\n\nfunction coercePrimitiveByType(value: unknown, type: string): unknown {\n\tswitch (type) {\n\t\tcase \"number\": {\n\t\t\tif (value === null) {\n\t\t\t\treturn 0;\n\t\t\t}\n\t\t\tif (typeof value === \"string\" && value.trim() !== \"\") {\n\t\t\t\tconst parsed = Number(value);\n\t\t\t\tif (Number.isFinite(parsed)) {\n\t\t\t\t\treturn parsed;\n\t\t\t\t}\n\t\t\t}\n\t\t\tif (typeof value === \"boolean\") {\n\t\t\t\treturn value ? 1 : 0;\n\t\t\t}\n\t\t\treturn value;\n\t\t}\n\t\tcase \"integer\": {\n\t\t\tif (value === null) {\n\t\t\t\treturn 0;\n\t\t\t}\n\t\t\tif (typeof value === \"string\" && value.trim() !== \"\") {\n\t\t\t\tconst parsed = Number(value);\n\t\t\t\tif (Number.isInteger(parsed)) {\n\t\t\t\t\treturn parsed;\n\t\t\t\t}\n\t\t\t}\n\t\t\tif (typeof value === \"boolean\") {\n\t\t\t\treturn value ? 1 : 0;\n\t\t\t}\n\t\t\treturn value;\n\t\t}\n\t\tcase \"boolean\": {\n\t\t\tif (value === null) {\n\t\t\t\treturn false;\n\t\t\t}\n\t\t\tif (typeof value === \"string\") {\n\t\t\t\tif (value === \"true\") {\n\t\t\t\t\treturn true;\n\t\t\t\t}\n\t\t\t\tif (value === \"false\") {\n\t\t\t\t\treturn false;\n\t\t\t\t}\n\t\t\t}\n\t\t\tif (typeof value === \"number\") {\n\t\t\t\tif (value === 1) {\n\t\t\t\t\treturn true;\n\t\t\t\t}\n\t\t\t\tif (value === 0) {\n\t\t\t\t\treturn false;\n\t\t\t\t}\n\t\t\t}\n\t\t\treturn value;\n\t\t}\n\t\tcase \"string\": {\n\t\t\tif (value === null) {\n\t\t\t\treturn \"\";\n\t\t\t}\n\t\t\tif (typeof value === \"number\" || typeof value === \"boolean\") {\n\t\t\t\treturn String(value);\n\t\t\t}\n\t\t\treturn value;\n\t\t}\n\t\tcase \"null\": {\n\t\t\tif (value === \"\" || value === 0 || value === false) {\n\t\t\t\treturn null;\n\t\t\t}\n\t\t\treturn value;\n\t\t}\n\t\tdefault:\n\t\t\treturn value;\n\t}\n}\n\nfunction applySchemaObjectCoercion(value: Record<string, unknown>, schema: JsonSchemaObject): void {\n\tconst properties = schema.properties;\n\tconst definedKeys = new Set<string>(properties ? Object.keys(properties) : []);\n\n\tif (properties) {\n\t\tfor (const [key, propertySchema] of Object.entries(properties)) {\n\t\t\tif (!(key in value)) {\n\t\t\t\tcontinue;\n\t\t\t}\n\t\t\tvalue[key] = coerceWithJsonSchema(value[key], propertySchema);\n\t\t}\n\t}\n\n\tif (schema.additionalProperties && isJsonSchemaObject(schema.additionalProperties)) {\n\t\tfor (const [key, propertyValue] of Object.entries(value)) {\n\t\t\tif (definedKeys.has(key)) {\n\t\t\t\tcontinue;\n\t\t\t}\n\t\t\tvalue[key] = coerceWithJsonSchema(propertyValue, schema.additionalProperties);\n\t\t}\n\t}\n}\n\nfunction applySchemaArrayCoercion(value: unknown[], schema: JsonSchemaObject): void {\n\tif (Array.isArray(schema.items)) {\n\t\tfor (let index = 0; index < value.length; index++) {\n\t\t\tconst itemSchema = schema.items[index];\n\t\t\tif (!itemSchema) {\n\t\t\t\tcontinue;\n\t\t\t}\n\t\t\tvalue[index] = coerceWithJsonSchema(value[index], itemSchema);\n\t\t}\n\t\treturn;\n\t}\n\n\tif (isJsonSchemaObject(schema.items)) {\n\t\tfor (let index = 0; index < value.length; index++) {\n\t\t\tvalue[index] = coerceWithJsonSchema(value[index], schema.items);\n\t\t}\n\t}\n}\n\nfunction coerceWithUnionSchema(value: unknown, schemas: JsonSchemaObject[]): unknown {\n\tfor (const schema of schemas) {\n\t\tconst validator = getSubSchemaValidator(schema);\n\t\tif (validator?.Check(value)) {\n\t\t\treturn value;\n\t\t}\n\t}\n\n\tfor (const schema of schemas) {\n\t\tconst candidate = structuredClone(value);\n\t\tconst coerced = coerceWithJsonSchema(candidate, schema);\n\t\tconst validator = getSubSchemaValidator(schema);\n\t\tif (validator?.Check(coerced)) {\n\t\t\treturn coerced;\n\t\t}\n\t}\n\treturn value;\n}\n\nfunction coerceWithJsonSchema(value: unknown, schema: JsonSchemaObject): unknown {\n\tlet nextValue = value;\n\n\tif (Array.isArray(schema.allOf)) {\n\t\tfor (const nested of schema.allOf) {\n\t\t\tnextValue = coerceWithJsonSchema(nextValue, nested);\n\t\t}\n\t}\n\n\tif (Array.isArray(schema.anyOf)) {\n\t\tnextValue = coerceWithUnionSchema(nextValue, schema.anyOf);\n\t}\n\n\tif (Array.isArray(schema.oneOf)) {\n\t\tnextValue = coerceWithUnionSchema(nextValue, schema.oneOf);\n\t}\n\n\tconst schemaTypes = getSchemaTypes(schema);\n\tconst matchesUnionMember =\n\t\tschemaTypes.length > 1 && schemaTypes.some((schemaType) => matchesJsonType(nextValue, schemaType));\n\tif (schemaTypes.length > 0 && !matchesUnionMember) {\n\t\tfor (const schemaType of schemaTypes) {\n\t\t\tconst candidate = coercePrimitiveByType(nextValue, schemaType);\n\t\t\tif (candidate !== nextValue) {\n\t\t\t\tnextValue = candidate;\n\t\t\t\tbreak;\n\t\t\t}\n\t\t}\n\t}\n\n\tif (schemaTypes.includes(\"object\") && isRecord(nextValue) && !Array.isArray(nextValue)) {\n\t\tapplySchemaObjectCoercion(nextValue, schema);\n\t}\n\n\tif (schemaTypes.includes(\"array\") && Array.isArray(nextValue)) {\n\t\tapplySchemaArrayCoercion(nextValue, schema);\n\t}\n\n\treturn nextValue;\n}\n\nfunction getValidator(schema: Tool[\"parameters\"]): ReturnType<typeof Compile> {\n\tconst key = schema as object;\n\tconst cached = validatorCache.get(key);\n\tif (cached) {\n\t\treturn cached;\n\t}\n\tconst validator = Compile(schema);\n\tvalidatorCache.set(key, validator);\n\treturn validator;\n}\n\nfunction formatValidationPath(error: TLocalizedValidationError): string {\n\tif (error.keyword === \"required\") {\n\t\tconst requiredProperties = (error.params as { requiredProperties?: string[] }).requiredProperties;\n\t\tconst requiredProperty = requiredProperties?.[0];\n\t\tif (requiredProperty) {\n\t\t\tconst basePath = error.instancePath.replace(/^\\//, \"\").replace(/\\//g, \".\");\n\t\t\treturn basePath ? `${basePath}.${requiredProperty}` : requiredProperty;\n\t\t}\n\t}\n\tconst path = error.instancePath.replace(/^\\//, \"\").replace(/\\//g, \".\");\n\treturn path || \"root\";\n}\n\n/**\n * Finds a tool by name and validates the tool call arguments against its TypeBox schema\n * @param tools Array of tool definitions\n * @param toolCall The tool call from the LLM\n * @returns The validated arguments\n * @throws Error if tool is not found or validation fails\n */\nexport function validateToolCall(tools: Tool[], toolCall: ToolCall): any {\n\tconst tool = tools.find((t) => t.name === toolCall.name);\n\tif (!tool) {\n\t\tthrow new Error(`Tool \"${toolCall.name}\" not found`);\n\t}\n\treturn validateToolArguments(tool, toolCall);\n}\n\n/**\n * Validates tool call arguments against the tool's TypeBox schema\n * @param tool The tool definition with TypeBox schema\n * @param toolCall The tool call from the LLM\n * @returns The validated (and potentially coerced) arguments\n * @throws Error with formatted message if validation fails\n */\nexport function validateToolArguments(tool: Tool, toolCall: ToolCall): any {\n\tif (toolCall.argumentsError !== undefined) {\n\t\tthrow new Error(toolCall.argumentsError);\n\t}\n\tconst args = structuredClone(toolCall.arguments);\n\tValue.Convert(tool.parameters, args);\n\n\tconst validator = getValidator(tool.parameters);\n\tif (!hasTypeBoxMetadata(tool.parameters) && isJsonSchemaObject(tool.parameters)) {\n\t\tconst coerced = coerceWithJsonSchema(args, tool.parameters);\n\t\tif (coerced !== args) {\n\t\t\tif (isRecord(args) && isRecord(coerced)) {\n\t\t\t\tfor (const key of Object.keys(args)) {\n\t\t\t\t\tdelete args[key];\n\t\t\t\t}\n\t\t\t\tObject.assign(args, coerced);\n\t\t\t} else {\n\t\t\t\treturn validator.Check(coerced) ? coerced : args;\n\t\t\t}\n\t\t}\n\t}\n\n\tif (validator.Check(args)) {\n\t\treturn args;\n\t}\n\n\tconst errors =\n\t\tvalidator\n\t\t\t.Errors(args)\n\t\t\t.map((error) => `  - ${formatValidationPath(error)}: ${error.message}`)\n\t\t\t.join(\"\\n\") || \"Unknown validation error\";\n\n\tconst errorMessage = `Validation failed for tool \"${toolCall.name}\":\\n${errors}\\n\\nReceived arguments:\\n${JSON.stringify(toolCall.arguments, null, 2)}`;\n\n\tthrow new Error(errorMessage);\n}\n"]}