@sammorrowdrums/mcpi-ai 0.70.9 → 0.85.0

package/dist/utils/node-http-proxy.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"node-http-proxy.js","sourceRoot":"","sources":["../../src/utils/node-http-proxy.ts"],"names":[],"mappings":"AACA,OAAO,EAAE,mBAAmB,EAAE,MAAM,mBAAmB,CAAC;AAExD,MAAM,mBAAmB,GAA2B;IACnD,GAAG,EAAE,EAAE;IACP,MAAM,EAAE,EAAE;IACV,IAAI,EAAE,EAAE;IACR,KAAK,EAAE,GAAG;IACV,EAAE,EAAE,EAAE;IACN,GAAG,EAAE,GAAG;CACR,CAAC;AAEF,SAAS,WAAW,CAAC,GAAW,EAAE,GAAiB,EAAU;IAC5D,MAAM,YAAY,GAAG,GAAG,CAAC,WAAW,EAAE,CAAC;IACvC,MAAM,YAAY,GAAG,GAAG,CAAC,WAAW,EAAE,CAAC;IACvC,OAAO,CACN,GAAG,EAAE,CAAC,YAAY,CAAC;QACnB,GAAG,EAAE,CAAC,YAAY,CAAC;QACnB,mBAAmB,CAAC,YAAY,CAAC;QACjC,mBAAmB,CAAC,YAAY,CAAC;QACjC,EAAE,CACF,CAAC;AAAA,CACF;AAED,SAAS,mBAAmB,CAAC,SAAuB,EAAmB;IACtE,IAAI,SAAS,YAAY,GAAG,EAAE,CAAC;QAC9B,OAAO,SAAS,CAAC;IAClB,CAAC;IAED,IAAI,CAAC;QACJ,OAAO,IAAI,GAAG,CAAC,SAAS,CAAC,CAAC;IAC3B,CAAC;IAAC,MAAM,CAAC;QACR,OAAO,SAAS,CAAC;IAClB,CAAC;AAAA,CACD;AAED,SAAS,mBAAmB,CAAC,QAAgB,EAAE,IAAY,EAAE,GAAiB,EAAW;IACxF,MAAM,OAAO,GAAG,WAAW,CAAC,UAAU,EAAE,GAAG,CAAC,CAAC,WAAW,EAAE,CAAC;IAC3D,IAAI,CAAC,OAAO,EAAE,CAAC;QACd,OAAO,IAAI,CAAC;IACb,CAAC;IACD,IAAI,OAAO,KAAK,GAAG,EAAE,CAAC;QACrB,OAAO,KAAK,CAAC;IACd,CAAC;IAED,OAAO,OAAO,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC,KAAK,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC;QAC9C,IAAI,CAAC,KAAK,EAAE,CAAC;YACZ,OAAO,IAAI,CAAC;QACb,CAAC;QAED,MAAM,WAAW,GAAG,KAAK,CAAC,KAAK,CAAC,cAAc,CAAC,CAAC;QAChD,IAAI,aAAa,GAAG,WAAW,CAAC,CAAC,CAAC,WAAW,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC;QACzD,MAAM,SAAS,GAAG,WAAW,CAAC,CAAC,CAAC,MAAM,CAAC,QAAQ,CAAC,WAAW,CAAC,CAAC,CAAE,EAAE,EAAE,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;QACzE,IAAI,SAAS,IAAI,SAAS,KAAK,IAAI,EAAE,CAAC;YACrC,OAAO,IAAI,CAAC;QACb,CAAC;QAED,IAAI,CAAC,OAAO,CAAC,IAAI,CAAC,aAAa,CAAC,EAAE,CAAC;YAClC,OAAO,QAAQ,KAAK,aAAa,CAAC;QACnC,CAAC;QAED,IAAI,aAAa,CAAC,UAAU,CAAC,GAAG,CAAC,EAAE,CAAC;YACnC,aAAa,GAAG,aAAa,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC;QACxC,CAAC;QACD,OAAO,CAAC,QAAQ,CAAC,QAAQ,CAAC,aAAa,CAAC,CAAC;IAAA,CACzC,CAAC,CAAC;AAAA,CACH;AAED,SAAS,cAAc,CAAC,SAAuB,EAAE,GAAiB,EAAU;IAC3E,MAAM,SAAS,GAAG,mBAAmB,CAAC,SAAS,CAAC,CAAC;IACjD,IAAI,CAAC,SAAS,EAAE,QAAQ,IAAI,CAAC,SAAS,CAAC,IAAI,EAAE,CAAC;QAC7C,OAAO,EAAE,CAAC;IACX,CAAC;IAED,MAAM,QAAQ,GAAG,SAAS,CAAC,QAAQ,CAAC,KAAK,CAAC,GAAG,EAAE,CAAC,CAAC,CAAC,CAAC,CAAE,CAAC;IACtD,MAAM,QAAQ,GAAG,SAAS,CAAC,IAAI,CAAC,OAAO,CAAC,OAAO,EAAE,EAAE,CAAC,CAAC;IACrD,MAAM,IAAI,GAAG,MAAM,CAAC,QAAQ,CAAC,SAAS,CAAC,IAAI,EAAE,EAAE,CAAC,IAAI,mBAAmB,CAAC,QAAQ,CAAC,IAAI,CAAC,CAAC;IACvF,IAAI,CAAC,mBAAmB,CAAC,QAAQ,EAAE,IAAI,EAAE,GAAG,CAAC,EAAE,CAAC;QAC/C,OAAO,EAAE,CAAC;IACX,CAAC;IAED,IAAI,KAAK,GAAG,WAAW,CAAC,GAAG,QAAQ,QAAQ,EAAE,GAAG,CAAC,IAAI,WAAW,CAAC,WAAW,EAAE,GAAG,CAAC,CAAC;IACnF,IAAI,KAAK,IAAI,CAAC,KAAK,CAAC,QAAQ,CAAC,KAAK,CAAC,EAAE,CAAC;QACrC,KAAK,GAAG,GAAG,QAAQ,MAAM,KAAK,EAAE,CAAC;IAClC,CAAC;IACD,OAAO,KAAK,CAAC;AAAA,CACb;AAED,MAAM,CAAC,MAAM,kCAAkC,GAC9C,yGAAyG,CAAC;AAE3G,MAAM,UAAU,4BAA4B,CAAC,SAAuB,EAAE,GAAiB,EAAmB;IACzG,MAAM,KAAK,GAAG,cAAc,CAAC,SAAS,EAAE,GAAG,CAAC,CAAC;IAC7C,IAAI,CAAC,KAAK,EAAE,CAAC;QACZ,OAAO,SAAS,CAAC;IAClB,CAAC;IAED,IAAI,QAAa,CAAC;IAClB,IAAI,CAAC;QACJ,QAAQ,GAAG,IAAI,GAAG,CAAC,KAAK,CAAC,CAAC;IAC3B,CAAC;IAAC,OAAO,KAAK,EAAE,CAAC;QAChB,MAAM,IAAI,KAAK,CACd,qBAAqB,IAAI,CAAC,SAAS,CAAC,KAAK,CAAC,KAAK,KAAK,YAAY,KAAK,CAAC,CAAC,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,MAAM,CAAC,KAAK,CAAC,EAAE,CACvG,CAAC;IACH,CAAC;IAED,IAAI,QAAQ,CAAC,QAAQ,KAAK,OAAO,IAAI,QAAQ,CAAC,QAAQ,KAAK,QAAQ,EAAE,CAAC;QACrE,MAAM,IAAI,KAAK,CAAC,GAAG,kCAAkC,QAAQ,QAAQ,CAAC,QAAQ,EAAE,CAAC,CAAC;IACnF,CAAC;IAED,OAAO,QAAQ,CAAC;AAAA,CAChB","sourcesContent":["import type { ProviderEnv } from \"../types.ts\";\nimport { getProviderEnvValue } from \"./provider-env.ts\";\n\nconst DEFAULT_PROXY_PORTS: Record<string, number> = {\n\tftp: 21,\n\tgopher: 70,\n\thttp: 80,\n\thttps: 443,\n\tws: 80,\n\twss: 443,\n};\n\nfunction getProxyEnv(key: string, env?: ProviderEnv): string {\n\tconst lowercaseKey = key.toLowerCase();\n\tconst uppercaseKey = key.toUpperCase();\n\treturn (\n\t\tenv?.[lowercaseKey] ||\n\t\tenv?.[uppercaseKey] ||\n\t\tgetProviderEnvValue(lowercaseKey) ||\n\t\tgetProviderEnvValue(uppercaseKey) ||\n\t\t\"\"\n\t);\n}\n\nfunction parseProxyTargetUrl(targetUrl: string | URL): URL | undefined {\n\tif (targetUrl instanceof URL) {\n\t\treturn targetUrl;\n\t}\n\n\ttry {\n\t\treturn new URL(targetUrl);\n\t} catch {\n\t\treturn undefined;\n\t}\n}\n\nfunction shouldProxyHostname(hostname: string, port: number, env?: ProviderEnv): boolean {\n\tconst noProxy = getProxyEnv(\"no_proxy\", env).toLowerCase();\n\tif (!noProxy) {\n\t\treturn true;\n\t}\n\tif (noProxy === \"*\") {\n\t\treturn false;\n\t}\n\n\treturn noProxy.split(/[,\\s]/).every((proxy) => {\n\t\tif (!proxy) {\n\t\t\treturn true;\n\t\t}\n\n\t\tconst parsedProxy = proxy.match(/^(.+):(\\d+)$/);\n\t\tlet proxyHostname = parsedProxy ? parsedProxy[1] : proxy;\n\t\tconst proxyPort = parsedProxy ? Number.parseInt(parsedProxy[2]!, 10) : 0;\n\t\tif (proxyPort && proxyPort !== port) {\n\t\t\treturn true;\n\t\t}\n\n\t\tif (!/^[.*]/.test(proxyHostname)) {\n\t\t\treturn hostname !== proxyHostname;\n\t\t}\n\n\t\tif (proxyHostname.startsWith(\"*\")) {\n\t\t\tproxyHostname = proxyHostname.slice(1);\n\t\t}\n\t\treturn !hostname.endsWith(proxyHostname);\n\t});\n}\n\nfunction getProxyForUrl(targetUrl: string | URL, env?: ProviderEnv): string {\n\tconst parsedUrl = parseProxyTargetUrl(targetUrl);\n\tif (!parsedUrl?.protocol || !parsedUrl.host) {\n\t\treturn \"\";\n\t}\n\n\tconst protocol = parsedUrl.protocol.split(\":\", 1)[0]!;\n\tconst hostname = parsedUrl.host.replace(/:\\d*$/, \"\");\n\tconst port = Number.parseInt(parsedUrl.port, 10) || DEFAULT_PROXY_PORTS[protocol] || 0;\n\tif (!shouldProxyHostname(hostname, port, env)) {\n\t\treturn \"\";\n\t}\n\n\tlet proxy = getProxyEnv(`${protocol}_proxy`, env) || getProxyEnv(\"all_proxy\", env);\n\tif (proxy && !proxy.includes(\"://\")) {\n\t\tproxy = `${protocol}://${proxy}`;\n\t}\n\treturn proxy;\n}\n\nexport const UNSUPPORTED_PROXY_PROTOCOL_MESSAGE =\n\t\"Unsupported proxy protocol. SOCKS and PAC proxy URLs are not supported; use an HTTP or HTTPS proxy URL.\";\n\nexport function resolveHttpProxyUrlForTarget(targetUrl: string | URL, env?: ProviderEnv): URL | undefined {\n\tconst proxy = getProxyForUrl(targetUrl, env);\n\tif (!proxy) {\n\t\treturn undefined;\n\t}\n\n\tlet proxyUrl: URL;\n\ttry {\n\t\tproxyUrl = new URL(proxy);\n\t} catch (error) {\n\t\tthrow new Error(\n\t\t\t`Invalid proxy URL ${JSON.stringify(proxy)}: ${error instanceof Error ? error.message : String(error)}`,\n\t\t);\n\t}\n\n\tif (proxyUrl.protocol !== \"http:\" && proxyUrl.protocol !== \"https:\") {\n\t\tthrow new Error(`${UNSUPPORTED_PROXY_PROTOCOL_MESSAGE} Got ${proxyUrl.protocol}`);\n\t}\n\n\treturn proxyUrl;\n}\n"]}

package/dist/utils/overflow.d.ts

@@ -1,31 +1,39 @@
-import type { AssistantMessage } from "../types.js";
+import type { AssistantMessage } from "../types.ts";
 /**
  * Check if an assistant message represents a context overflow error.
  *
- * This handles two cases:
+ * This handles three cases:
  * 1. Error-based overflow: Most providers return stopReason "error" with a
  *    specific error message pattern.
  * 2. Silent overflow: Some providers accept overflow requests and return
  *    successfully. For these, we check if usage.input exceeds the context window.
+ * 3. Length-stop overflow: Xiaomi MiMo can return "length" with zero output when
+ *    the input fills the context window.
  *
  * ## Reliability by Provider
  *
  * **Reliable detection (returns error with detectable message):**
  * - Anthropic: "prompt is too long: X tokens > Y maximum" or "request_too_large"
- * - OpenAI (Completions & Responses): "exceeds the context window"
+ * - OpenAI (Completions & Responses): "exceeds the context window", "exceeds the model's maximum context length of X tokens", or "exceeds model's maximum context length (X)"
  * - Google Gemini: "input token count exceeds the maximum"
  * - xAI (Grok): "maximum prompt length is X but request contains Y"
  * - Groq: "reduce the length of the messages"
  * - Cerebras: 400/413 status code (no body)
  * - Mistral: "Prompt contains X tokens ... too large for model with Y maximum context length"
- * - OpenRouter (all backends): "maximum context length is X tokens"
+ * - OpenRouter (most backends): "maximum context length is X tokens"
+ * - OpenRouter/Poolside: "Input length X exceeds the maximum allowed input length of Y tokens."
+ * - Together AI: "The input (X tokens) is longer than the model's context length (Y tokens)."
  * - llama.cpp: "exceeds the available context size"
  * - LM Studio: "greater than the context length"
  * - Kimi For Coding: "exceeded model token limit: X (requested: Y)"
+ * - DS4: "Prompt has X tokens, but the configured context size is Y tokens"
+ * - DashScope/Qwen: "Range of input length should be [1, X]"
  *
  * **Unreliable detection:**
  * - z.ai: Sometimes accepts overflow silently (detectable via usage.input > contextWindow),
  *   sometimes returns rate limit errors. Pass contextWindow param to detect silent overflow.
+ * - Xiaomi MiMo: Truncates input to fit contextWindow then returns stopReason "length" with
+ *   output=0. Pass contextWindow param to detect via the "filled context + zero output" signal.
  * - Ollama: May truncate input silently for some setups, but may also return explicit
  *   overflow errors that match the patterns above. Silent truncation still cannot be
  *   detected here because we do not know the expected token count.
@@ -46,6 +54,13 @@ import type { AssistantMessage } from "../types.js";
  * @returns true if the message indicates a context overflow
  */
 export declare function isContextOverflow(message: AssistantMessage, contextWindow?: number): boolean;
+/**
+ * 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 declare function isRecoverableLength(message: AssistantMessage, desiredMaxOutput: number): boolean;
 /**
  * Get the overflow patterns for testing purposes.
  */

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;AAiEpD;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA6CG;AACH,wBAAgB,iBAAiB,CAAC,OAAO,EAAE,gBAAgB,EAAE,aAAa,CAAC,EAAE,MAAM,GAAG,OAAO,CAmB5F;AAED;;GAEG;AACH,wBAAgB,mBAAmB,IAAI,MAAM,EAAE,CAE9C","sourcesContent":["import type { AssistantMessage } from \"../types.js\";\n\n/**\n * Regex patterns to detect context overflow errors from different providers.\n *\n * These patterns match error messages returned when the input exceeds\n * the model's context window.\n *\n * Provider-specific patterns (with example error messages):\n *\n * - Anthropic: \"prompt is too long: 213462 tokens > 200000 maximum\"\n * - Anthropic: \"413 {\\\"error\\\":{\\\"type\\\":\\\"request_too_large\\\",\\\"message\\\":\\\"Request exceeds the maximum size\\\"}}\"\n * - OpenAI: \"Your input exceeds the context window of this model\"\n * - Google: \"The input token count (1196265) exceeds the maximum number of tokens allowed (1048575)\"\n * - xAI: \"This model's maximum prompt length is 131072 but the request contains 537812 tokens\"\n * - Groq: \"Please reduce the length of the messages or completion\"\n * - OpenRouter: \"This endpoint's maximum context length is X tokens. However, you requested about Y tokens\"\n * - llama.cpp: \"the request exceeds the available context size, try increasing it\"\n * - LM Studio: \"tokens to keep from the initial prompt is greater than the context length\"\n * - GitHub Copilot: \"prompt token count of X exceeds the limit of Y\"\n * - MiniMax: \"invalid params, context window exceeds limit\"\n * - Kimi For Coding: \"Your request exceeded model token limit: X (requested: Y)\"\n * - Cerebras: \"400/413 status code (no body)\"\n * - Mistral: \"Prompt contains X tokens ... too large for model with Y maximum context length\"\n * - z.ai: Does NOT error, accepts overflow silently - handled via usage.input > contextWindow\n * - Ollama: Some deployments truncate silently, others return errors like \"prompt too long; exceeded max context length by X tokens\"\n */\nconst OVERFLOW_PATTERNS = [\n\t/prompt is too long/i, // Anthropic token overflow\n\t/request_too_large/i, // Anthropic request byte-size overflow (HTTP 413)\n\t/input is too long for requested model/i, // Amazon Bedrock\n\t/exceeds the context window/i, // OpenAI (Completions & Responses API)\n\t/input token count.*exceeds the maximum/i, // Google (Gemini)\n\t/maximum prompt length is \\d+/i, // xAI (Grok)\n\t/reduce the length of the messages/i, // Groq\n\t/maximum context length is \\d+ tokens/i, // OpenRouter (all backends)\n\t/exceeds the limit of \\d+/i, // GitHub Copilot\n\t/exceeds the available context size/i, // llama.cpp server\n\t/greater than the context length/i, // LM Studio\n\t/context window exceeds limit/i, // MiniMax\n\t/exceeded model token limit/i, // Kimi For Coding\n\t/too large for model with \\d+ maximum context length/i, // Mistral\n\t/model_context_window_exceeded/i, // z.ai non-standard finish_reason surfaced as error text\n\t/prompt too long; exceeded (?:max )?context length/i, // Ollama explicit overflow error\n\t/context[_ ]length[_ ]exceeded/i, // Generic fallback\n\t/too many tokens/i, // Generic fallback\n\t/token limit exceeded/i, // Generic fallback\n\t/^4(?:00|13)\\s*(?:status code)?\\s*\\(no body\\)/i, // Cerebras: 400/413 with no body\n];\n\n/**\n * Patterns that indicate non-overflow errors (e.g. rate limiting, server errors).\n * Error messages matching any of these are excluded from overflow detection\n * even if they also match an OVERFLOW_PATTERN.\n *\n * Example: Bedrock formats throttling errors as \"ThrottlingException: Too many tokens,\n * please wait before trying again.\" which would match the /too many tokens/i overflow\n * pattern without this exclusion.\n */\nconst NON_OVERFLOW_PATTERNS = [\n\t/^(Throttling error|Service unavailable):/i, // AWS Bedrock non-overflow errors (human-readable prefixes from formatBedrockError)\n\t/rate limit/i, // Generic rate limiting\n\t/too many requests/i, // Generic HTTP 429 style\n];\n\n/**\n * Check if an assistant message represents a context overflow error.\n *\n * This handles two cases:\n * 1. Error-based overflow: Most providers return stopReason \"error\" with a\n *    specific error message pattern.\n * 2. Silent overflow: Some providers accept overflow requests and return\n *    successfully. For these, we check if usage.input exceeds the context window.\n *\n * ## Reliability by Provider\n *\n * **Reliable detection (returns error with detectable message):**\n * - Anthropic: \"prompt is too long: X tokens > Y maximum\" or \"request_too_large\"\n * - OpenAI (Completions & Responses): \"exceeds the context window\"\n * - Google Gemini: \"input token count exceeds the maximum\"\n * - xAI (Grok): \"maximum prompt length is X but request contains Y\"\n * - Groq: \"reduce the length of the messages\"\n * - Cerebras: 400/413 status code (no body)\n * - Mistral: \"Prompt contains X tokens ... too large for model with Y maximum context length\"\n * - OpenRouter (all backends): \"maximum context length is X tokens\"\n * - llama.cpp: \"exceeds the available context size\"\n * - LM Studio: \"greater than the context length\"\n * - Kimi For Coding: \"exceeded model token limit: X (requested: Y)\"\n *\n * **Unreliable detection:**\n * - z.ai: Sometimes accepts overflow silently (detectable via usage.input > contextWindow),\n *   sometimes returns rate limit errors. Pass contextWindow param to detect silent overflow.\n * - 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\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;AA+EpD;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAqDG;AACH,wBAAgB,iBAAiB,CAAC,OAAO,EAAE,gBAAgB,EAAE,aAAa,CAAC,EAAE,MAAM,GAAG,OAAO,CA6B5F;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.ts\";\n\n/**\n * Regex patterns to detect context overflow errors from different providers.\n *\n * These patterns match error messages returned when the input exceeds\n * the model's context window.\n *\n * Provider-specific patterns (with example error messages):\n *\n * - Anthropic: \"prompt is too long: 213462 tokens > 200000 maximum\"\n * - Anthropic: \"413 {\\\"error\\\":{\\\"type\\\":\\\"request_too_large\\\",\\\"message\\\":\\\"Request exceeds the maximum size\\\"}}\"\n * - OpenAI: \"Your input exceeds the context window of this model\"\n * - OpenAI/LiteLLM: \"Requested token count exceeds the model's maximum context length of 131072 tokens\"\n * - OpenAI-compatible: \"Input length (265330) exceeds model's maximum context length (262144).\"\n * - Google: \"The input token count (1196265) exceeds the maximum number of tokens allowed (1048575)\"\n * - xAI: \"This model's maximum prompt length is 131072 but the request contains 537812 tokens\"\n * - Groq: \"Please reduce the length of the messages or completion\"\n * - OpenRouter: \"This endpoint's maximum context length is X tokens. However, you requested about Y tokens\"\n * - OpenRouter/Poolside: \"Input length X exceeds the maximum allowed input length of Y tokens.\"\n * - Together AI: \"The input (X tokens) is longer than the model's context length (Y tokens).\"\n * - llama.cpp: \"the request exceeds the available context size, try increasing it\"\n * - LM Studio: \"tokens to keep from the initial prompt is greater than the context length\"\n * - GitHub Copilot: \"prompt token count of X exceeds the limit of Y\"\n * - MiniMax: \"invalid params, context window exceeds limit\"\n * - Kimi For Coding: \"Your request exceeded model token limit: X (requested: Y)\"\n * - DS4: \"Prompt has X tokens, but the configured context size is Y tokens\"\n * - Cerebras: \"400/413 status code (no body)\"\n * - Mistral: \"Prompt contains X tokens ... too large for model with Y maximum context length\"\n * - z.ai: Does NOT error, accepts overflow silently - handled via usage.input > contextWindow\n * - Xiaomi MiMo: Truncates input to fill contextWindow exactly, then returns finish_reason \"length\"\n *   with output=0 (no room left to generate). Detected via stopReason \"length\" + zero output +\n *   input filling the context window.\n * - DashScope/Qwen: \"Range of input length should be [1, X]\" (HTTP 400 invalid_parameter_error)\n * - Ollama: Some deployments truncate silently, others return errors like \"prompt too long; exceeded max context length by X tokens\"\n */\nconst OVERFLOW_PATTERNS = [\n\t/prompt is too long/i, // Anthropic token overflow\n\t/request_too_large/i, // Anthropic request byte-size overflow (HTTP 413)\n\t/input is too long for requested model/i, // Amazon Bedrock\n\t/exceeds the context window/i, // OpenAI (Completions & Responses API)\n\t/exceeds (?:the )?(?:model'?s )?maximum context length(?: of [\\d,]+ tokens?|\\s*\\([\\d,]+\\))/i, // OpenAI-compatible proxies (LiteLLM)\n\t/input token count.*exceeds the maximum/i, // Google (Gemini)\n\t/maximum prompt length is \\d+/i, // xAI (Grok)\n\t/reduce the length of the messages/i, // Groq\n\t/maximum context length is \\d+ tokens/i, // OpenRouter (most backends)\n\t/exceeds (?:the )?maximum allowed input length of [\\d,]+ tokens?/i, // OpenRouter/Poolside\n\t/input \\(\\d+ tokens\\) is longer than the model'?s context length \\(\\d+ tokens\\)/i, // Together AI\n\t/exceeds the limit of \\d+/i, // GitHub Copilot\n\t/exceeds the available context size/i, // llama.cpp server\n\t/greater than the context length/i, // LM Studio\n\t/context window exceeds limit/i, // MiniMax\n\t/exceeded model token limit/i, // Kimi For Coding\n\t/too large for model with \\d+ maximum context length/i, // Mistral\n\t/prompt has [\\d,]+ tokens?, but the configured context size is [\\d,]+ tokens?/i, // DS4 server\n\t/model_context_window_exceeded/i, // z.ai non-standard finish_reason surfaced as error text\n\t/prompt too long; exceeded (?:max )?context length/i, // Ollama explicit overflow error\n\t/range of input length should be/i, // DashScope / Qwen Token Plan\n\t/context[_ ]length[_ ]exceeded/i, // Generic fallback\n\t/too many tokens/i, // Generic fallback\n\t/token limit exceeded/i, // Generic fallback\n\t/^4(?:00|13)\\s*(?:status code)?\\s*\\(no body\\)/i, // Cerebras: 400/413 with no body\n];\n\n/**\n * Patterns that indicate non-overflow errors (e.g. rate limiting, server errors).\n * Error messages matching any of these are excluded from overflow detection\n * even if they also match an OVERFLOW_PATTERN.\n *\n * Example: Bedrock formats throttling errors as \"ThrottlingException: Too many tokens,\n * please wait before trying again.\" which would match the /too many tokens/i overflow\n * pattern without this exclusion.\n */\nconst NON_OVERFLOW_PATTERNS = [\n\t/^(Throttling error|Service unavailable):/i, // AWS Bedrock non-overflow errors (human-readable prefixes from formatBedrockError)\n\t/rate limit/i, // Generic rate limiting\n\t/too many requests/i, // Generic HTTP 429 style\n];\n\n/**\n * Check if an assistant message represents a context overflow error.\n *\n * This handles three 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 * 3. Length-stop overflow: Xiaomi MiMo can return \"length\" with zero output when\n *    the input fills the context window.\n *\n * ## Reliability by Provider\n *\n * **Reliable detection (returns error with detectable message):**\n * - Anthropic: \"prompt is too long: X tokens > Y maximum\" or \"request_too_large\"\n * - OpenAI (Completions & Responses): \"exceeds the context window\", \"exceeds the model's maximum context length of X tokens\", or \"exceeds model's maximum context length (X)\"\n * - Google Gemini: \"input token count exceeds the maximum\"\n * - xAI (Grok): \"maximum prompt length is X but request contains Y\"\n * - Groq: \"reduce the length of the messages\"\n * - Cerebras: 400/413 status code (no body)\n * - Mistral: \"Prompt contains X tokens ... too large for model with Y maximum context length\"\n * - OpenRouter (most backends): \"maximum context length is X tokens\"\n * - OpenRouter/Poolside: \"Input length X exceeds the maximum allowed input length of Y tokens.\"\n * - Together AI: \"The input (X tokens) is longer than the model's context length (Y tokens).\"\n * - llama.cpp: \"exceeds the available context size\"\n * - LM Studio: \"greater than the context length\"\n * - Kimi For Coding: \"exceeded model token limit: X (requested: Y)\"\n * - DS4: \"Prompt has X tokens, but the configured context size is Y tokens\"\n * - DashScope/Qwen: \"Range of input length should be [1, X]\"\n *\n * **Unreliable detection:**\n * - z.ai: Sometimes accepts overflow silently (detectable via usage.input > contextWindow),\n *   sometimes returns rate limit errors. Pass contextWindow param to detect silent overflow.\n * - Xiaomi MiMo: Truncates input to fit contextWindow then returns stopReason \"length\" with\n *   output=0. Pass contextWindow param to detect via the \"filled context + zero output\" signal.\n * - Ollama: May truncate input silently for some setups, but may also return explicit\n *   overflow errors that match the patterns above. Silent truncation still cannot be\n *   detected here because we do not know the expected token count.\n *\n * ## Custom Providers\n *\n * If you've added custom models via settings.json, this function may not detect\n * overflow errors from those providers. To add support:\n *\n * 1. Send a request that exceeds the model's context window\n * 2. Check the errorMessage in the response\n * 3. Create a regex pattern that matches the error\n * 4. The pattern should be added to OVERFLOW_PATTERNS in this file, or\n *    check the errorMessage yourself before calling this function\n *\n * @param message - The assistant message to check\n * @param contextWindow - Optional context window size for detecting silent overflow (z.ai)\n * @returns true if the message indicates a context overflow\n */\nexport function isContextOverflow(message: AssistantMessage, contextWindow?: number): boolean {\n\t// Case 1: Check error message patterns\n\tif (message.stopReason === \"error\" && message.errorMessage) {\n\t\t// Skip messages matching known non-overflow patterns (e.g. throttling / rate-limit)\n\t\tconst isNonOverflow = NON_OVERFLOW_PATTERNS.some((p) => p.test(message.errorMessage!));\n\t\tif (!isNonOverflow && OVERFLOW_PATTERNS.some((p) => p.test(message.errorMessage!))) {\n\t\t\treturn true;\n\t\t}\n\t}\n\n\t// Case 2: Silent overflow (z.ai style) - successful but usage exceeds context\n\tif (contextWindow && message.stopReason === \"stop\") {\n\t\tconst inputTokens = message.usage.input + message.usage.cacheRead;\n\t\tif (inputTokens > contextWindow) {\n\t\t\treturn true;\n\t\t}\n\t}\n\n\t// Case 3: Length-stop overflow (Xiaomi MiMo style) - server truncates oversized input\n\t// to fit the context window, leaving no room for output. Returns stopReason \"length\"\n\t// with output=0 and input+cacheRead filling the context window.\n\tif (contextWindow && message.stopReason === \"length\" && message.usage.output === 0) {\n\t\tconst inputTokens = message.usage.input + message.usage.cacheRead;\n\t\tif (inputTokens >= contextWindow * 0.99) {\n\t\t\treturn true;\n\t\t}\n\t}\n\n\treturn false;\n}\n\n/**\n * 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

@@ -9,18 +9,27 @@
  * - Anthropic: "prompt is too long: 213462 tokens > 200000 maximum"
  * - Anthropic: "413 {\"error\":{\"type\":\"request_too_large\",\"message\":\"Request exceeds the maximum size\"}}"
  * - OpenAI: "Your input exceeds the context window of this model"
+ * - OpenAI/LiteLLM: "Requested token count exceeds the model's maximum context length of 131072 tokens"
+ * - OpenAI-compatible: "Input length (265330) exceeds model's maximum context length (262144)."
  * - Google: "The input token count (1196265) exceeds the maximum number of tokens allowed (1048575)"
  * - xAI: "This model's maximum prompt length is 131072 but the request contains 537812 tokens"
  * - Groq: "Please reduce the length of the messages or completion"
  * - OpenRouter: "This endpoint's maximum context length is X tokens. However, you requested about Y tokens"
+ * - OpenRouter/Poolside: "Input length X exceeds the maximum allowed input length of Y tokens."
+ * - Together AI: "The input (X tokens) is longer than the model's context length (Y tokens)."
  * - llama.cpp: "the request exceeds the available context size, try increasing it"
  * - LM Studio: "tokens to keep from the initial prompt is greater than the context length"
  * - GitHub Copilot: "prompt token count of X exceeds the limit of Y"
  * - MiniMax: "invalid params, context window exceeds limit"
  * - Kimi For Coding: "Your request exceeded model token limit: X (requested: Y)"
+ * - DS4: "Prompt has X tokens, but the configured context size is Y tokens"
  * - Cerebras: "400/413 status code (no body)"
  * - Mistral: "Prompt contains X tokens ... too large for model with Y maximum context length"
  * - z.ai: Does NOT error, accepts overflow silently - handled via usage.input > contextWindow
+ * - Xiaomi MiMo: Truncates input to fill contextWindow exactly, then returns finish_reason "length"
+ *   with output=0 (no room left to generate). Detected via stopReason "length" + zero output +
+ *   input filling the context window.
+ * - DashScope/Qwen: "Range of input length should be [1, X]" (HTTP 400 invalid_parameter_error)
  * - Ollama: Some deployments truncate silently, others return errors like "prompt too long; exceeded max context length by X tokens"
  */
 const OVERFLOW_PATTERNS = [
@@ -28,18 +37,23 @@ const OVERFLOW_PATTERNS = [
     /request_too_large/i, // Anthropic request byte-size overflow (HTTP 413)
     /input is too long for requested model/i, // Amazon Bedrock
     /exceeds the context window/i, // OpenAI (Completions & Responses API)
+    /exceeds (?:the )?(?:model'?s )?maximum context length(?: of [\d,]+ tokens?|\s*\([\d,]+\))/i, // OpenAI-compatible proxies (LiteLLM)
     /input token count.*exceeds the maximum/i, // Google (Gemini)
     /maximum prompt length is \d+/i, // xAI (Grok)
     /reduce the length of the messages/i, // Groq
-    /maximum context length is \d+ tokens/i, // OpenRouter (all backends)
+    /maximum context length is \d+ tokens/i, // OpenRouter (most backends)
+    /exceeds (?:the )?maximum allowed input length of [\d,]+ tokens?/i, // OpenRouter/Poolside
+    /input \(\d+ tokens\) is longer than the model'?s context length \(\d+ tokens\)/i, // Together AI
     /exceeds the limit of \d+/i, // GitHub Copilot
     /exceeds the available context size/i, // llama.cpp server
     /greater than the context length/i, // LM Studio
     /context window exceeds limit/i, // MiniMax
     /exceeded model token limit/i, // Kimi For Coding
     /too large for model with \d+ maximum context length/i, // Mistral
+    /prompt has [\d,]+ tokens?, but the configured context size is [\d,]+ tokens?/i, // DS4 server
     /model_context_window_exceeded/i, // z.ai non-standard finish_reason surfaced as error text
     /prompt too long; exceeded (?:max )?context length/i, // Ollama explicit overflow error
+    /range of input length should be/i, // DashScope / Qwen Token Plan
     /context[_ ]length[_ ]exceeded/i, // Generic fallback
     /too many tokens/i, // Generic fallback
     /token limit exceeded/i, // Generic fallback
@@ -62,30 +76,38 @@ const NON_OVERFLOW_PATTERNS = [
 /**
  * Check if an assistant message represents a context overflow error.
  *
- * This handles two cases:
+ * This handles three cases:
  * 1. Error-based overflow: Most providers return stopReason "error" with a
  *    specific error message pattern.
  * 2. Silent overflow: Some providers accept overflow requests and return
  *    successfully. For these, we check if usage.input exceeds the context window.
+ * 3. Length-stop overflow: Xiaomi MiMo can return "length" with zero output when
+ *    the input fills the context window.
  *
  * ## Reliability by Provider
  *
  * **Reliable detection (returns error with detectable message):**
  * - Anthropic: "prompt is too long: X tokens > Y maximum" or "request_too_large"
- * - OpenAI (Completions & Responses): "exceeds the context window"
+ * - OpenAI (Completions & Responses): "exceeds the context window", "exceeds the model's maximum context length of X tokens", or "exceeds model's maximum context length (X)"
  * - Google Gemini: "input token count exceeds the maximum"
  * - xAI (Grok): "maximum prompt length is X but request contains Y"
  * - Groq: "reduce the length of the messages"
  * - Cerebras: 400/413 status code (no body)
  * - Mistral: "Prompt contains X tokens ... too large for model with Y maximum context length"
- * - OpenRouter (all backends): "maximum context length is X tokens"
+ * - OpenRouter (most backends): "maximum context length is X tokens"
+ * - OpenRouter/Poolside: "Input length X exceeds the maximum allowed input length of Y tokens."
+ * - Together AI: "The input (X tokens) is longer than the model's context length (Y tokens)."
  * - llama.cpp: "exceeds the available context size"
  * - LM Studio: "greater than the context length"
  * - Kimi For Coding: "exceeded model token limit: X (requested: Y)"
+ * - DS4: "Prompt has X tokens, but the configured context size is Y tokens"
+ * - DashScope/Qwen: "Range of input length should be [1, X]"
  *
  * **Unreliable detection:**
  * - z.ai: Sometimes accepts overflow silently (detectable via usage.input > contextWindow),
  *   sometimes returns rate limit errors. Pass contextWindow param to detect silent overflow.
+ * - Xiaomi MiMo: Truncates input to fit contextWindow then returns stopReason "length" with
+ *   output=0. Pass contextWindow param to detect via the "filled context + zero output" signal.
  * - Ollama: May truncate input silently for some setups, but may also return explicit
  *   overflow errors that match the patterns above. Silent truncation still cannot be
  *   detected here because we do not know the expected token count.
@@ -121,8 +143,26 @@ export function isContextOverflow(message, contextWindow) {
             return true;
         }
     }
+    // Case 3: Length-stop overflow (Xiaomi MiMo style) - server truncates oversized input
+    // to fit the context window, leaving no room for output. Returns stopReason "length"
+    // with output=0 and input+cacheRead filling the context window.
+    if (contextWindow && message.stopReason === "length" && message.usage.output === 0) {
+        const inputTokens = message.usage.input + message.usage.cacheRead;
+        if (inputTokens >= contextWindow * 0.99) {
+            return true;
+        }
+    }
     return false;
 }
+/**
+ * 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;;;;;;;;;;;;;;;;;;;;;;;;GAwBG;AACH,MAAM,iBAAiB,GAAG;IACzB,qBAAqB,EAAE,2BAA2B;IAClD,oBAAoB,EAAE,kDAAkD;IACxE,wCAAwC,EAAE,iBAAiB;IAC3D,6BAA6B,EAAE,uCAAuC;IACtE,yCAAyC,EAAE,kBAAkB;IAC7D,+BAA+B,EAAE,aAAa;IAC9C,oCAAoC,EAAE,OAAO;IAC7C,uCAAuC,EAAE,4BAA4B;IACrE,2BAA2B,EAAE,iBAAiB;IAC9C,qCAAqC,EAAE,mBAAmB;IAC1D,kCAAkC,EAAE,YAAY;IAChD,+BAA+B,EAAE,UAAU;IAC3C,6BAA6B,EAAE,kBAAkB;IACjD,sDAAsD,EAAE,UAAU;IAClE,gCAAgC,EAAE,yDAAyD;IAC3F,oDAAoD,EAAE,iCAAiC;IACvF,gCAAgC,EAAE,mBAAmB;IACrD,kBAAkB,EAAE,mBAAmB;IACvC,uBAAuB,EAAE,mBAAmB;IAC5C,+CAA+C,EAAE,iCAAiC;CAClF,CAAC;AAEF;;;;;;;;GAQG;AACH,MAAM,qBAAqB,GAAG;IAC7B,2CAA2C,EAAE,oFAAoF;IACjI,aAAa,EAAE,wBAAwB;IACvC,oBAAoB,EAAE,yBAAyB;CAC/C,CAAC;AAEF;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA6CG;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,OAAO,KAAK,CAAC;AAAA,CACb;AAED;;GAEG;AACH,MAAM,UAAU,mBAAmB,GAAa;IAC/C,OAAO,CAAC,GAAG,iBAAiB,CAAC,CAAC;AAAA,CAC9B","sourcesContent":["import type { AssistantMessage } from \"../types.js\";\n\n/**\n * Regex patterns to detect context overflow errors from different providers.\n *\n * These patterns match error messages returned when the input exceeds\n * the model's context window.\n *\n * Provider-specific patterns (with example error messages):\n *\n * - Anthropic: \"prompt is too long: 213462 tokens > 200000 maximum\"\n * - Anthropic: \"413 {\\\"error\\\":{\\\"type\\\":\\\"request_too_large\\\",\\\"message\\\":\\\"Request exceeds the maximum size\\\"}}\"\n * - OpenAI: \"Your input exceeds the context window of this model\"\n * - Google: \"The input token count (1196265) exceeds the maximum number of tokens allowed (1048575)\"\n * - xAI: \"This model's maximum prompt length is 131072 but the request contains 537812 tokens\"\n * - Groq: \"Please reduce the length of the messages or completion\"\n * - OpenRouter: \"This endpoint's maximum context length is X tokens. However, you requested about Y tokens\"\n * - llama.cpp: \"the request exceeds the available context size, try increasing it\"\n * - LM Studio: \"tokens to keep from the initial prompt is greater than the context length\"\n * - GitHub Copilot: \"prompt token count of X exceeds the limit of Y\"\n * - MiniMax: \"invalid params, context window exceeds limit\"\n * - Kimi For Coding: \"Your request exceeded model token limit: X (requested: Y)\"\n * - Cerebras: \"400/413 status code (no body)\"\n * - Mistral: \"Prompt contains X tokens ... too large for model with Y maximum context length\"\n * - z.ai: Does NOT error, accepts overflow silently - handled via usage.input > contextWindow\n * - Ollama: Some deployments truncate silently, others return errors like \"prompt too long; exceeded max context length by X tokens\"\n */\nconst OVERFLOW_PATTERNS = [\n\t/prompt is too long/i, // Anthropic token overflow\n\t/request_too_large/i, // Anthropic request byte-size overflow (HTTP 413)\n\t/input is too long for requested model/i, // Amazon Bedrock\n\t/exceeds the context window/i, // OpenAI (Completions & Responses API)\n\t/input token count.*exceeds the maximum/i, // Google (Gemini)\n\t/maximum prompt length is \\d+/i, // xAI (Grok)\n\t/reduce the length of the messages/i, // Groq\n\t/maximum context length is \\d+ tokens/i, // OpenRouter (all backends)\n\t/exceeds the limit of \\d+/i, // GitHub Copilot\n\t/exceeds the available context size/i, // llama.cpp server\n\t/greater than the context length/i, // LM Studio\n\t/context window exceeds limit/i, // MiniMax\n\t/exceeded model token limit/i, // Kimi For Coding\n\t/too large for model with \\d+ maximum context length/i, // Mistral\n\t/model_context_window_exceeded/i, // z.ai non-standard finish_reason surfaced as error text\n\t/prompt too long; exceeded (?:max )?context length/i, // Ollama explicit overflow error\n\t/context[_ ]length[_ ]exceeded/i, // Generic fallback\n\t/too many tokens/i, // Generic fallback\n\t/token limit exceeded/i, // Generic fallback\n\t/^4(?:00|13)\\s*(?:status code)?\\s*\\(no body\\)/i, // Cerebras: 400/413 with no body\n];\n\n/**\n * Patterns that indicate non-overflow errors (e.g. rate limiting, server errors).\n * Error messages matching any of these are excluded from overflow detection\n * even if they also match an OVERFLOW_PATTERN.\n *\n * Example: Bedrock formats throttling errors as \"ThrottlingException: Too many tokens,\n * please wait before trying again.\" which would match the /too many tokens/i overflow\n * pattern without this exclusion.\n */\nconst NON_OVERFLOW_PATTERNS = [\n\t/^(Throttling error|Service unavailable):/i, // AWS Bedrock non-overflow errors (human-readable prefixes from formatBedrockError)\n\t/rate limit/i, // Generic rate limiting\n\t/too many requests/i, // Generic HTTP 429 style\n];\n\n/**\n * Check if an assistant message represents a context overflow error.\n *\n * This handles two cases:\n * 1. Error-based overflow: Most providers return stopReason \"error\" with a\n *    specific error message pattern.\n * 2. Silent overflow: Some providers accept overflow requests and return\n *    successfully. For these, we check if usage.input exceeds the context window.\n *\n * ## Reliability by Provider\n *\n * **Reliable detection (returns error with detectable message):**\n * - Anthropic: \"prompt is too long: X tokens > Y maximum\" or \"request_too_large\"\n * - OpenAI (Completions & Responses): \"exceeds the context window\"\n * - Google Gemini: \"input token count exceeds the maximum\"\n * - xAI (Grok): \"maximum prompt length is X but request contains Y\"\n * - Groq: \"reduce the length of the messages\"\n * - Cerebras: 400/413 status code (no body)\n * - Mistral: \"Prompt contains X tokens ... too large for model with Y maximum context length\"\n * - OpenRouter (all backends): \"maximum context length is X tokens\"\n * - llama.cpp: \"exceeds the available context size\"\n * - LM Studio: \"greater than the context length\"\n * - Kimi For Coding: \"exceeded model token limit: X (requested: Y)\"\n *\n * **Unreliable detection:**\n * - z.ai: Sometimes accepts overflow silently (detectable via usage.input > contextWindow),\n *   sometimes returns rate limit errors. Pass contextWindow param to detect silent overflow.\n * - 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\treturn false;\n}\n\n/**\n * Get the overflow patterns for testing purposes.\n */\nexport function getOverflowPatterns(): RegExp[] {\n\treturn [...OVERFLOW_PATTERNS];\n}\n"]}
+{"version":3,"file":"overflow.js","sourceRoot":"","sources":["../../src/utils/overflow.ts"],"names":[],"mappings":"AAEA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAiCG;AACH,MAAM,iBAAiB,GAAG;IACzB,qBAAqB,EAAE,2BAA2B;IAClD,oBAAoB,EAAE,kDAAkD;IACxE,wCAAwC,EAAE,iBAAiB;IAC3D,6BAA6B,EAAE,uCAAuC;IACtE,4FAA4F,EAAE,sCAAsC;IACpI,yCAAyC,EAAE,kBAAkB;IAC7D,+BAA+B,EAAE,aAAa;IAC9C,oCAAoC,EAAE,OAAO;IAC7C,uCAAuC,EAAE,6BAA6B;IACtE,kEAAkE,EAAE,sBAAsB;IAC1F,iFAAiF,EAAE,cAAc;IACjG,2BAA2B,EAAE,iBAAiB;IAC9C,qCAAqC,EAAE,mBAAmB;IAC1D,kCAAkC,EAAE,YAAY;IAChD,+BAA+B,EAAE,UAAU;IAC3C,6BAA6B,EAAE,kBAAkB;IACjD,sDAAsD,EAAE,UAAU;IAClE,+EAA+E,EAAE,aAAa;IAC9F,gCAAgC,EAAE,yDAAyD;IAC3F,oDAAoD,EAAE,iCAAiC;IACvF,kCAAkC,EAAE,8BAA8B;IAClE,gCAAgC,EAAE,mBAAmB;IACrD,kBAAkB,EAAE,mBAAmB;IACvC,uBAAuB,EAAE,mBAAmB;IAC5C,+CAA+C,EAAE,iCAAiC;CAClF,CAAC;AAEF;;;;;;;;GAQG;AACH,MAAM,qBAAqB,GAAG;IAC7B,2CAA2C,EAAE,oFAAoF;IACjI,aAAa,EAAE,wBAAwB;IACvC,oBAAoB,EAAE,yBAAyB;CAC/C,CAAC;AAEF;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAqDG;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;;;;;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.ts\";\n\n/**\n * Regex patterns to detect context overflow errors from different providers.\n *\n * These patterns match error messages returned when the input exceeds\n * the model's context window.\n *\n * Provider-specific patterns (with example error messages):\n *\n * - Anthropic: \"prompt is too long: 213462 tokens > 200000 maximum\"\n * - Anthropic: \"413 {\\\"error\\\":{\\\"type\\\":\\\"request_too_large\\\",\\\"message\\\":\\\"Request exceeds the maximum size\\\"}}\"\n * - OpenAI: \"Your input exceeds the context window of this model\"\n * - OpenAI/LiteLLM: \"Requested token count exceeds the model's maximum context length of 131072 tokens\"\n * - OpenAI-compatible: \"Input length (265330) exceeds model's maximum context length (262144).\"\n * - Google: \"The input token count (1196265) exceeds the maximum number of tokens allowed (1048575)\"\n * - xAI: \"This model's maximum prompt length is 131072 but the request contains 537812 tokens\"\n * - Groq: \"Please reduce the length of the messages or completion\"\n * - OpenRouter: \"This endpoint's maximum context length is X tokens. However, you requested about Y tokens\"\n * - OpenRouter/Poolside: \"Input length X exceeds the maximum allowed input length of Y tokens.\"\n * - Together AI: \"The input (X tokens) is longer than the model's context length (Y tokens).\"\n * - llama.cpp: \"the request exceeds the available context size, try increasing it\"\n * - LM Studio: \"tokens to keep from the initial prompt is greater than the context length\"\n * - GitHub Copilot: \"prompt token count of X exceeds the limit of Y\"\n * - MiniMax: \"invalid params, context window exceeds limit\"\n * - Kimi For Coding: \"Your request exceeded model token limit: X (requested: Y)\"\n * - DS4: \"Prompt has X tokens, but the configured context size is Y tokens\"\n * - Cerebras: \"400/413 status code (no body)\"\n * - Mistral: \"Prompt contains X tokens ... too large for model with Y maximum context length\"\n * - z.ai: Does NOT error, accepts overflow silently - handled via usage.input > contextWindow\n * - Xiaomi MiMo: Truncates input to fill contextWindow exactly, then returns finish_reason \"length\"\n *   with output=0 (no room left to generate). Detected via stopReason \"length\" + zero output +\n *   input filling the context window.\n * - DashScope/Qwen: \"Range of input length should be [1, X]\" (HTTP 400 invalid_parameter_error)\n * - Ollama: Some deployments truncate silently, others return errors like \"prompt too long; exceeded max context length by X tokens\"\n */\nconst OVERFLOW_PATTERNS = [\n\t/prompt is too long/i, // Anthropic token overflow\n\t/request_too_large/i, // Anthropic request byte-size overflow (HTTP 413)\n\t/input is too long for requested model/i, // Amazon Bedrock\n\t/exceeds the context window/i, // OpenAI (Completions & Responses API)\n\t/exceeds (?:the )?(?:model'?s )?maximum context length(?: of [\\d,]+ tokens?|\\s*\\([\\d,]+\\))/i, // OpenAI-compatible proxies (LiteLLM)\n\t/input token count.*exceeds the maximum/i, // Google (Gemini)\n\t/maximum prompt length is \\d+/i, // xAI (Grok)\n\t/reduce the length of the messages/i, // Groq\n\t/maximum context length is \\d+ tokens/i, // OpenRouter (most backends)\n\t/exceeds (?:the )?maximum allowed input length of [\\d,]+ tokens?/i, // OpenRouter/Poolside\n\t/input \\(\\d+ tokens\\) is longer than the model'?s context length \\(\\d+ tokens\\)/i, // Together AI\n\t/exceeds the limit of \\d+/i, // GitHub Copilot\n\t/exceeds the available context size/i, // llama.cpp server\n\t/greater than the context length/i, // LM Studio\n\t/context window exceeds limit/i, // MiniMax\n\t/exceeded model token limit/i, // Kimi For Coding\n\t/too large for model with \\d+ maximum context length/i, // Mistral\n\t/prompt has [\\d,]+ tokens?, but the configured context size is [\\d,]+ tokens?/i, // DS4 server\n\t/model_context_window_exceeded/i, // z.ai non-standard finish_reason surfaced as error text\n\t/prompt too long; exceeded (?:max )?context length/i, // Ollama explicit overflow error\n\t/range of input length should be/i, // DashScope / Qwen Token Plan\n\t/context[_ ]length[_ ]exceeded/i, // Generic fallback\n\t/too many tokens/i, // Generic fallback\n\t/token limit exceeded/i, // Generic fallback\n\t/^4(?:00|13)\\s*(?:status code)?\\s*\\(no body\\)/i, // Cerebras: 400/413 with no body\n];\n\n/**\n * Patterns that indicate non-overflow errors (e.g. rate limiting, server errors).\n * Error messages matching any of these are excluded from overflow detection\n * even if they also match an OVERFLOW_PATTERN.\n *\n * Example: Bedrock formats throttling errors as \"ThrottlingException: Too many tokens,\n * please wait before trying again.\" which would match the /too many tokens/i overflow\n * pattern without this exclusion.\n */\nconst NON_OVERFLOW_PATTERNS = [\n\t/^(Throttling error|Service unavailable):/i, // AWS Bedrock non-overflow errors (human-readable prefixes from formatBedrockError)\n\t/rate limit/i, // Generic rate limiting\n\t/too many requests/i, // Generic HTTP 429 style\n];\n\n/**\n * Check if an assistant message represents a context overflow error.\n *\n * This handles three 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 * 3. Length-stop overflow: Xiaomi MiMo can return \"length\" with zero output when\n *    the input fills the context window.\n *\n * ## Reliability by Provider\n *\n * **Reliable detection (returns error with detectable message):**\n * - Anthropic: \"prompt is too long: X tokens > Y maximum\" or \"request_too_large\"\n * - OpenAI (Completions & Responses): \"exceeds the context window\", \"exceeds the model's maximum context length of X tokens\", or \"exceeds model's maximum context length (X)\"\n * - Google Gemini: \"input token count exceeds the maximum\"\n * - xAI (Grok): \"maximum prompt length is X but request contains Y\"\n * - Groq: \"reduce the length of the messages\"\n * - Cerebras: 400/413 status code (no body)\n * - Mistral: \"Prompt contains X tokens ... too large for model with Y maximum context length\"\n * - OpenRouter (most backends): \"maximum context length is X tokens\"\n * - OpenRouter/Poolside: \"Input length X exceeds the maximum allowed input length of Y tokens.\"\n * - Together AI: \"The input (X tokens) is longer than the model's context length (Y tokens).\"\n * - llama.cpp: \"exceeds the available context size\"\n * - LM Studio: \"greater than the context length\"\n * - Kimi For Coding: \"exceeded model token limit: X (requested: Y)\"\n * - DS4: \"Prompt has X tokens, but the configured context size is Y tokens\"\n * - DashScope/Qwen: \"Range of input length should be [1, X]\"\n *\n * **Unreliable detection:**\n * - z.ai: Sometimes accepts overflow silently (detectable via usage.input > contextWindow),\n *   sometimes returns rate limit errors. Pass contextWindow param to detect silent overflow.\n * - Xiaomi MiMo: Truncates input to fit contextWindow then returns stopReason \"length\" with\n *   output=0. Pass contextWindow param to detect via the \"filled context + zero output\" signal.\n * - Ollama: May truncate input silently for some setups, but may also return explicit\n *   overflow errors that match the patterns above. Silent truncation still cannot be\n *   detected here because we do not know the expected token count.\n *\n * ## Custom Providers\n *\n * If you've added custom models via settings.json, this function may not detect\n * overflow errors from those providers. To add support:\n *\n * 1. Send a request that exceeds the model's context window\n * 2. Check the errorMessage in the response\n * 3. Create a regex pattern that matches the error\n * 4. The pattern should be added to OVERFLOW_PATTERNS in this file, or\n *    check the errorMessage yourself before calling this function\n *\n * @param message - The assistant message to check\n * @param contextWindow - Optional context window size for detecting silent overflow (z.ai)\n * @returns true if the message indicates a context overflow\n */\nexport function isContextOverflow(message: AssistantMessage, contextWindow?: number): boolean {\n\t// Case 1: Check error message patterns\n\tif (message.stopReason === \"error\" && message.errorMessage) {\n\t\t// Skip messages matching known non-overflow patterns (e.g. throttling / rate-limit)\n\t\tconst isNonOverflow = NON_OVERFLOW_PATTERNS.some((p) => p.test(message.errorMessage!));\n\t\tif (!isNonOverflow && OVERFLOW_PATTERNS.some((p) => p.test(message.errorMessage!))) {\n\t\t\treturn true;\n\t\t}\n\t}\n\n\t// Case 2: Silent overflow (z.ai style) - successful but usage exceeds context\n\tif (contextWindow && message.stopReason === \"stop\") {\n\t\tconst inputTokens = message.usage.input + message.usage.cacheRead;\n\t\tif (inputTokens > contextWindow) {\n\t\t\treturn true;\n\t\t}\n\t}\n\n\t// Case 3: Length-stop overflow (Xiaomi MiMo style) - server truncates oversized input\n\t// to fit the context window, leaving no room for output. Returns stopReason \"length\"\n\t// with output=0 and input+cacheRead filling the context window.\n\tif (contextWindow && message.stopReason === \"length\" && message.usage.output === 0) {\n\t\tconst inputTokens = message.usage.input + message.usage.cacheRead;\n\t\tif (inputTokens >= contextWindow * 0.99) {\n\t\t\treturn true;\n\t\t}\n\t}\n\n\treturn false;\n}\n\n/**\n * 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/pi-user-agent.d.ts

@@ -0,0 +1,2 @@
+export declare function getPiUserAgent(): string;
+//# sourceMappingURL=pi-user-agent.d.ts.map

package/dist/utils/pi-user-agent.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"pi-user-agent.d.ts","sourceRoot":"","sources":["../../src/utils/pi-user-agent.ts"],"names":[],"mappings":"AAgBA,wBAAgB,cAAc,IAAI,MAAM,CAEvC","sourcesContent":["import type * as NodeOs from \"node:os\";\n\ntype ProcessWithOsBuiltinModule = typeof process & {\n\tgetBuiltinModule?: (id: \"node:os\") => typeof NodeOs;\n};\n\nfunction loadNodeOs(): typeof NodeOs | null {\n\tif (typeof process === \"undefined\" || !(process.versions?.node || process.versions?.bun)) {\n\t\treturn null;\n\t}\n\treturn (process as ProcessWithOsBuiltinModule).getBuiltinModule?.(\"node:os\") ?? null;\n}\n\n// Keep runtime OS loading browser-safe. A top-level runtime import of node:os breaks browser/Vite builds.\nconst nodeOs = loadNodeOs();\n\nexport function getPiUserAgent(): string {\n\treturn nodeOs ? `mcpi (${nodeOs.platform()} ${nodeOs.release()}; ${nodeOs.arch()})` : \"mcpi (browser)\";\n}\n"]}

package/dist/utils/pi-user-agent.js

@@ -0,0 +1,12 @@
+function loadNodeOs() {
+    if (typeof process === "undefined" || !(process.versions?.node || process.versions?.bun)) {
+        return null;
+    }
+    return process.getBuiltinModule?.("node:os") ?? null;
+}
+// Keep runtime OS loading browser-safe. A top-level runtime import of node:os breaks browser/Vite builds.
+const nodeOs = loadNodeOs();
+export function getPiUserAgent() {
+    return nodeOs ? `mcpi (${nodeOs.platform()} ${nodeOs.release()}; ${nodeOs.arch()})` : "mcpi (browser)";
+}
+//# sourceMappingURL=pi-user-agent.js.map

package/dist/utils/pi-user-agent.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"pi-user-agent.js","sourceRoot":"","sources":["../../src/utils/pi-user-agent.ts"],"names":[],"mappings":"AAMA,SAAS,UAAU,GAAyB;IAC3C,IAAI,OAAO,OAAO,KAAK,WAAW,IAAI,CAAC,CAAC,OAAO,CAAC,QAAQ,EAAE,IAAI,IAAI,OAAO,CAAC,QAAQ,EAAE,GAAG,CAAC,EAAE,CAAC;QAC1F,OAAO,IAAI,CAAC;IACb,CAAC;IACD,OAAQ,OAAsC,CAAC,gBAAgB,EAAE,CAAC,SAAS,CAAC,IAAI,IAAI,CAAC;AAAA,CACrF;AAED,0GAA0G;AAC1G,MAAM,MAAM,GAAG,UAAU,EAAE,CAAC;AAE5B,MAAM,UAAU,cAAc,GAAW;IACxC,OAAO,MAAM,CAAC,CAAC,CAAC,SAAS,MAAM,CAAC,QAAQ,EAAE,IAAI,MAAM,CAAC,OAAO,EAAE,KAAK,MAAM,CAAC,IAAI,EAAE,GAAG,CAAC,CAAC,CAAC,gBAAgB,CAAC;AAAA,CACvG","sourcesContent":["import type * as NodeOs from \"node:os\";\n\ntype ProcessWithOsBuiltinModule = typeof process & {\n\tgetBuiltinModule?: (id: \"node:os\") => typeof NodeOs;\n};\n\nfunction loadNodeOs(): typeof NodeOs | null {\n\tif (typeof process === \"undefined\" || !(process.versions?.node || process.versions?.bun)) {\n\t\treturn null;\n\t}\n\treturn (process as ProcessWithOsBuiltinModule).getBuiltinModule?.(\"node:os\") ?? null;\n}\n\n// Keep runtime OS loading browser-safe. A top-level runtime import of node:os breaks browser/Vite builds.\nconst nodeOs = loadNodeOs();\n\nexport function getPiUserAgent(): string {\n\treturn nodeOs ? `mcpi (${nodeOs.platform()} ${nodeOs.release()}; ${nodeOs.arch()})` : \"mcpi (browser)\";\n}\n"]}

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

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

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

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

package/dist/utils/provider-env.js

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

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

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

package/dist/utils/provider-retry.d.ts

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

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

@@ -0,0 +1 @@
+{"version":3,"file":"provider-retry.d.ts","sourceRoot":"","sources":["../../src/utils/provider-retry.ts"],"names":[],"mappings":"AAEA,UAAU,oBAAoB;IAC7B,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,eAAe,CAAC,EAAE,MAAM,CAAC;IACzB,MAAM,CAAC,EAAE,WAAW,CAAC;CACrB;AA0FD;;;;;;;GAOG;AACH,wBAAsB,oBAAoB,CAAC,CAAC,EAC3C,OAAO,EAAE,MAAM,OAAO,CAAC,CAAC,CAAC,EACzB,OAAO,GAAE,oBAAyB,GAChC,OAAO,CAAC,CAAC,CAAC,CAiBZ","sourcesContent":["const DEFAULT_MAX_RETRY_DELAY_MS = 60_000;\n\ninterface ProviderRetryOptions {\n\tmaxRetries?: number;\n\tmaxRetryDelayMs?: number;\n\tsignal?: AbortSignal;\n}\n\ninterface ProviderError extends Error {\n\tstatus: number | undefined;\n\theaders: Headers | undefined;\n}\n\nfunction isProviderError(error: unknown): error is ProviderError {\n\tif (!(error instanceof Error) || !(\"status\" in error) || !(\"headers\" in error)) return false;\n\treturn (\n\t\t(error.status === undefined || typeof error.status === \"number\") &&\n\t\t(error.headers === undefined || error.headers instanceof Headers)\n\t);\n}\n\n/** Mirrors the pinned OpenAI/Anthropic SDK retry policy; review when either SDK is upgraded. */\nfunction isRetryableProviderError(error: ProviderError): boolean {\n\tconst shouldRetry = error.headers?.get(\"x-should-retry\");\n\tif (shouldRetry === \"true\") return true;\n\tif (shouldRetry === \"false\") return false;\n\n\tif (error.status === undefined) return true;\n\treturn (\n\t\terror.status === 408 ||\n\t\terror.status === 409 ||\n\t\terror.status === 429 ||\n\t\t(typeof error.status === \"number\" && error.status >= 500)\n\t);\n}\n\nfunction validateServerRetryDelayMs(\n\tdelayMs: number,\n\tmaxRetryDelayMs: number | undefined,\n\tproviderErrorMessage: string,\n): number {\n\tconst maxDelayMs = maxRetryDelayMs ?? DEFAULT_MAX_RETRY_DELAY_MS;\n\tif (maxDelayMs > 0 && delayMs > maxDelayMs) {\n\t\tthrow new Error(\n\t\t\t`Server requested ${Math.ceil(delayMs / 1000)}s retry delay (max: ${Math.ceil(maxDelayMs / 1000)}s). ${providerErrorMessage}`,\n\t\t);\n\t}\n\treturn delayMs;\n}\n\nfunction getRetryDelayMs(error: ProviderError, retryIndex: number, maxRetryDelayMs: number | undefined): number {\n\tconst retryAfterMs = error.headers?.get(\"retry-after-ms\");\n\tif (retryAfterMs) {\n\t\tconst value = Number.parseFloat(retryAfterMs);\n\t\tif (!Number.isNaN(value)) return validateServerRetryDelayMs(value, maxRetryDelayMs, error.message);\n\t}\n\n\tconst retryAfter = error.headers?.get(\"retry-after\");\n\tif (retryAfter) {\n\t\tconst seconds = Number.parseFloat(retryAfter);\n\t\tconst delayMs = Number.isNaN(seconds) ? Date.parse(retryAfter) - Date.now() : seconds * 1000;\n\t\treturn validateServerRetryDelayMs(delayMs, maxRetryDelayMs, error.message);\n\t}\n\n\tconst exponentialDelay = Math.min(0.5 * 2 ** retryIndex, 8) * 1000;\n\treturn exponentialDelay * (1 - Math.random() * 0.25);\n}\n\nfunction createAbortError(): Error {\n\tconst error = new Error(\"Request aborted\");\n\terror.name = \"AbortError\";\n\treturn error;\n}\n\nfunction abortableSleep(ms: number, signal?: AbortSignal): Promise<void> {\n\treturn new Promise((resolve, reject) => {\n\t\tif (signal?.aborted) {\n\t\t\treject(createAbortError());\n\t\t\treturn;\n\t\t}\n\n\t\tconst onAbort = () => {\n\t\t\tclearTimeout(timeout);\n\t\t\treject(createAbortError());\n\t\t};\n\t\tconst timeout = setTimeout(\n\t\t\t() => {\n\t\t\t\tsignal?.removeEventListener(\"abort\", onAbort);\n\t\t\t\tresolve();\n\t\t\t},\n\t\t\tMath.max(0, ms),\n\t\t);\n\t\tsignal?.addEventListener(\"abort\", onAbort, { once: true });\n\t});\n}\n\n/**\n * Reproduce the retry behavior used by the OpenAI and Anthropic SDKs while making\n * their backoff sleep interruptible. Their built-in retry timers ignore the\n * request AbortSignal, so callers must invoke the SDK with `maxRetries: 0` and\n * wrap the request with this helper. Provider-requested delays above\n * `maxRetryDelayMs` fail immediately (60 seconds by default); set it to zero to\n * disable the limit.\n */\nexport async function retryProviderRequest<T>(\n\trequest: () => Promise<T>,\n\toptions: ProviderRetryOptions = {},\n): Promise<T> {\n\tconst maxRetries = options.maxRetries ?? 0;\n\tlet retriesRemaining = maxRetries;\n\n\tfor (;;) {\n\t\ttry {\n\t\t\t// Each retry is a fresh SDK request, so X-Stainless-Retry-Count remains zero.\n\t\t\treturn await request();\n\t\t} catch (error) {\n\t\t\tif (options.signal?.aborted) throw createAbortError();\n\t\t\tif (retriesRemaining <= 0 || !isProviderError(error) || !isRetryableProviderError(error)) throw error;\n\n\t\t\tconst retryIndex = maxRetries - retriesRemaining;\n\t\t\tretriesRemaining--;\n\t\t\tawait abortableSleep(getRetryDelayMs(error, retryIndex, options.maxRetryDelayMs), options.signal);\n\t\t}\n\t}\n}\n"]}

package/dist/utils/provider-retry.js

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

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

@@ -0,0 +1 @@
+{"version":3,"file":"provider-retry.js","sourceRoot":"","sources":["../../src/utils/provider-retry.ts"],"names":[],"mappings":"AAAA,MAAM,0BAA0B,GAAG,MAAM,CAAC;AAa1C,SAAS,eAAe,CAAC,KAAc,EAA0B;IAChE,IAAI,CAAC,CAAC,KAAK,YAAY,KAAK,CAAC,IAAI,CAAC,CAAC,QAAQ,IAAI,KAAK,CAAC,IAAI,CAAC,CAAC,SAAS,IAAI,KAAK,CAAC;QAAE,OAAO,KAAK,CAAC;IAC7F,OAAO,CACN,CAAC,KAAK,CAAC,MAAM,KAAK,SAAS,IAAI,OAAO,KAAK,CAAC,MAAM,KAAK,QAAQ,CAAC;QAChE,CAAC,KAAK,CAAC,OAAO,KAAK,SAAS,IAAI,KAAK,CAAC,OAAO,YAAY,OAAO,CAAC,CACjE,CAAC;AAAA,CACF;AAED,gGAAgG;AAChG,SAAS,wBAAwB,CAAC,KAAoB,EAAW;IAChE,MAAM,WAAW,GAAG,KAAK,CAAC,OAAO,EAAE,GAAG,CAAC,gBAAgB,CAAC,CAAC;IACzD,IAAI,WAAW,KAAK,MAAM;QAAE,OAAO,IAAI,CAAC;IACxC,IAAI,WAAW,KAAK,OAAO;QAAE,OAAO,KAAK,CAAC;IAE1C,IAAI,KAAK,CAAC,MAAM,KAAK,SAAS;QAAE,OAAO,IAAI,CAAC;IAC5C,OAAO,CACN,KAAK,CAAC,MAAM,KAAK,GAAG;QACpB,KAAK,CAAC,MAAM,KAAK,GAAG;QACpB,KAAK,CAAC,MAAM,KAAK,GAAG;QACpB,CAAC,OAAO,KAAK,CAAC,MAAM,KAAK,QAAQ,IAAI,KAAK,CAAC,MAAM,IAAI,GAAG,CAAC,CACzD,CAAC;AAAA,CACF;AAED,SAAS,0BAA0B,CAClC,OAAe,EACf,eAAmC,EACnC,oBAA4B,EACnB;IACT,MAAM,UAAU,GAAG,eAAe,IAAI,0BAA0B,CAAC;IACjE,IAAI,UAAU,GAAG,CAAC,IAAI,OAAO,GAAG,UAAU,EAAE,CAAC;QAC5C,MAAM,IAAI,KAAK,CACd,oBAAoB,IAAI,CAAC,IAAI,CAAC,OAAO,GAAG,IAAI,CAAC,uBAAuB,IAAI,CAAC,IAAI,CAAC,UAAU,GAAG,IAAI,CAAC,OAAO,oBAAoB,EAAE,CAC7H,CAAC;IACH,CAAC;IACD,OAAO,OAAO,CAAC;AAAA,CACf;AAED,SAAS,eAAe,CAAC,KAAoB,EAAE,UAAkB,EAAE,eAAmC,EAAU;IAC/G,MAAM,YAAY,GAAG,KAAK,CAAC,OAAO,EAAE,GAAG,CAAC,gBAAgB,CAAC,CAAC;IAC1D,IAAI,YAAY,EAAE,CAAC;QAClB,MAAM,KAAK,GAAG,MAAM,CAAC,UAAU,CAAC,YAAY,CAAC,CAAC;QAC9C,IAAI,CAAC,MAAM,CAAC,KAAK,CAAC,KAAK,CAAC;YAAE,OAAO,0BAA0B,CAAC,KAAK,EAAE,eAAe,EAAE,KAAK,CAAC,OAAO,CAAC,CAAC;IACpG,CAAC;IAED,MAAM,UAAU,GAAG,KAAK,CAAC,OAAO,EAAE,GAAG,CAAC,aAAa,CAAC,CAAC;IACrD,IAAI,UAAU,EAAE,CAAC;QAChB,MAAM,OAAO,GAAG,MAAM,CAAC,UAAU,CAAC,UAAU,CAAC,CAAC;QAC9C,MAAM,OAAO,GAAG,MAAM,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,KAAK,CAAC,UAAU,CAAC,GAAG,IAAI,CAAC,GAAG,EAAE,CAAC,CAAC,CAAC,OAAO,GAAG,IAAI,CAAC;QAC7F,OAAO,0BAA0B,CAAC,OAAO,EAAE,eAAe,EAAE,KAAK,CAAC,OAAO,CAAC,CAAC;IAC5E,CAAC;IAED,MAAM,gBAAgB,GAAG,IAAI,CAAC,GAAG,CAAC,GAAG,GAAG,CAAC,IAAI,UAAU,EAAE,CAAC,CAAC,GAAG,IAAI,CAAC;IACnE,OAAO,gBAAgB,GAAG,CAAC,CAAC,GAAG,IAAI,CAAC,MAAM,EAAE,GAAG,IAAI,CAAC,CAAC;AAAA,CACrD;AAED,SAAS,gBAAgB,GAAU;IAClC,MAAM,KAAK,GAAG,IAAI,KAAK,CAAC,iBAAiB,CAAC,CAAC;IAC3C,KAAK,CAAC,IAAI,GAAG,YAAY,CAAC;IAC1B,OAAO,KAAK,CAAC;AAAA,CACb;AAED,SAAS,cAAc,CAAC,EAAU,EAAE,MAAoB,EAAiB;IACxE,OAAO,IAAI,OAAO,CAAC,CAAC,OAAO,EAAE,MAAM,EAAE,EAAE,CAAC;QACvC,IAAI,MAAM,EAAE,OAAO,EAAE,CAAC;YACrB,MAAM,CAAC,gBAAgB,EAAE,CAAC,CAAC;YAC3B,OAAO;QACR,CAAC;QAED,MAAM,OAAO,GAAG,GAAG,EAAE,CAAC;YACrB,YAAY,CAAC,OAAO,CAAC,CAAC;YACtB,MAAM,CAAC,gBAAgB,EAAE,CAAC,CAAC;QAAA,CAC3B,CAAC;QACF,MAAM,OAAO,GAAG,UAAU,CACzB,GAAG,EAAE,CAAC;YACL,MAAM,EAAE,mBAAmB,CAAC,OAAO,EAAE,OAAO,CAAC,CAAC;YAC9C,OAAO,EAAE,CAAC;QAAA,CACV,EACD,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,EAAE,CAAC,CACf,CAAC;QACF,MAAM,EAAE,gBAAgB,CAAC,OAAO,EAAE,OAAO,EAAE,EAAE,IAAI,EAAE,IAAI,EAAE,CAAC,CAAC;IAAA,CAC3D,CAAC,CAAC;AAAA,CACH;AAED;;;;;;;GAOG;AACH,MAAM,CAAC,KAAK,UAAU,oBAAoB,CACzC,OAAyB,EACzB,OAAO,GAAyB,EAAE,EACrB;IACb,MAAM,UAAU,GAAG,OAAO,CAAC,UAAU,IAAI,CAAC,CAAC;IAC3C,IAAI,gBAAgB,GAAG,UAAU,CAAC;IAElC,SAAS,CAAC;QACT,IAAI,CAAC;YACJ,8EAA8E;YAC9E,OAAO,MAAM,OAAO,EAAE,CAAC;QACxB,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YAChB,IAAI,OAAO,CAAC,MAAM,EAAE,OAAO;gBAAE,MAAM,gBAAgB,EAAE,CAAC;YACtD,IAAI,gBAAgB,IAAI,CAAC,IAAI,CAAC,eAAe,CAAC,KAAK,CAAC,IAAI,CAAC,wBAAwB,CAAC,KAAK,CAAC;gBAAE,MAAM,KAAK,CAAC;YAEtG,MAAM,UAAU,GAAG,UAAU,GAAG,gBAAgB,CAAC;YACjD,gBAAgB,EAAE,CAAC;YACnB,MAAM,cAAc,CAAC,eAAe,CAAC,KAAK,EAAE,UAAU,EAAE,OAAO,CAAC,eAAe,CAAC,EAAE,OAAO,CAAC,MAAM,CAAC,CAAC;QACnG,CAAC;IACF,CAAC;AAAA,CACD","sourcesContent":["const DEFAULT_MAX_RETRY_DELAY_MS = 60_000;\n\ninterface ProviderRetryOptions {\n\tmaxRetries?: number;\n\tmaxRetryDelayMs?: number;\n\tsignal?: AbortSignal;\n}\n\ninterface ProviderError extends Error {\n\tstatus: number | undefined;\n\theaders: Headers | undefined;\n}\n\nfunction isProviderError(error: unknown): error is ProviderError {\n\tif (!(error instanceof Error) || !(\"status\" in error) || !(\"headers\" in error)) return false;\n\treturn (\n\t\t(error.status === undefined || typeof error.status === \"number\") &&\n\t\t(error.headers === undefined || error.headers instanceof Headers)\n\t);\n}\n\n/** Mirrors the pinned OpenAI/Anthropic SDK retry policy; review when either SDK is upgraded. */\nfunction isRetryableProviderError(error: ProviderError): boolean {\n\tconst shouldRetry = error.headers?.get(\"x-should-retry\");\n\tif (shouldRetry === \"true\") return true;\n\tif (shouldRetry === \"false\") return false;\n\n\tif (error.status === undefined) return true;\n\treturn (\n\t\terror.status === 408 ||\n\t\terror.status === 409 ||\n\t\terror.status === 429 ||\n\t\t(typeof error.status === \"number\" && error.status >= 500)\n\t);\n}\n\nfunction validateServerRetryDelayMs(\n\tdelayMs: number,\n\tmaxRetryDelayMs: number | undefined,\n\tproviderErrorMessage: string,\n): number {\n\tconst maxDelayMs = maxRetryDelayMs ?? DEFAULT_MAX_RETRY_DELAY_MS;\n\tif (maxDelayMs > 0 && delayMs > maxDelayMs) {\n\t\tthrow new Error(\n\t\t\t`Server requested ${Math.ceil(delayMs / 1000)}s retry delay (max: ${Math.ceil(maxDelayMs / 1000)}s). ${providerErrorMessage}`,\n\t\t);\n\t}\n\treturn delayMs;\n}\n\nfunction getRetryDelayMs(error: ProviderError, retryIndex: number, maxRetryDelayMs: number | undefined): number {\n\tconst retryAfterMs = error.headers?.get(\"retry-after-ms\");\n\tif (retryAfterMs) {\n\t\tconst value = Number.parseFloat(retryAfterMs);\n\t\tif (!Number.isNaN(value)) return validateServerRetryDelayMs(value, maxRetryDelayMs, error.message);\n\t}\n\n\tconst retryAfter = error.headers?.get(\"retry-after\");\n\tif (retryAfter) {\n\t\tconst seconds = Number.parseFloat(retryAfter);\n\t\tconst delayMs = Number.isNaN(seconds) ? Date.parse(retryAfter) - Date.now() : seconds * 1000;\n\t\treturn validateServerRetryDelayMs(delayMs, maxRetryDelayMs, error.message);\n\t}\n\n\tconst exponentialDelay = Math.min(0.5 * 2 ** retryIndex, 8) * 1000;\n\treturn exponentialDelay * (1 - Math.random() * 0.25);\n}\n\nfunction createAbortError(): Error {\n\tconst error = new Error(\"Request aborted\");\n\terror.name = \"AbortError\";\n\treturn error;\n}\n\nfunction abortableSleep(ms: number, signal?: AbortSignal): Promise<void> {\n\treturn new Promise((resolve, reject) => {\n\t\tif (signal?.aborted) {\n\t\t\treject(createAbortError());\n\t\t\treturn;\n\t\t}\n\n\t\tconst onAbort = () => {\n\t\t\tclearTimeout(timeout);\n\t\t\treject(createAbortError());\n\t\t};\n\t\tconst timeout = setTimeout(\n\t\t\t() => {\n\t\t\t\tsignal?.removeEventListener(\"abort\", onAbort);\n\t\t\t\tresolve();\n\t\t\t},\n\t\t\tMath.max(0, ms),\n\t\t);\n\t\tsignal?.addEventListener(\"abort\", onAbort, { once: true });\n\t});\n}\n\n/**\n * Reproduce the retry behavior used by the OpenAI and Anthropic SDKs while making\n * their backoff sleep interruptible. Their built-in retry timers ignore the\n * request AbortSignal, so callers must invoke the SDK with `maxRetries: 0` and\n * wrap the request with this helper. Provider-requested delays above\n * `maxRetryDelayMs` fail immediately (60 seconds by default); set it to zero to\n * disable the limit.\n */\nexport async function retryProviderRequest<T>(\n\trequest: () => Promise<T>,\n\toptions: ProviderRetryOptions = {},\n): Promise<T> {\n\tconst maxRetries = options.maxRetries ?? 0;\n\tlet retriesRemaining = maxRetries;\n\n\tfor (;;) {\n\t\ttry {\n\t\t\t// Each retry is a fresh SDK request, so X-Stainless-Retry-Count remains zero.\n\t\t\treturn await request();\n\t\t} catch (error) {\n\t\t\tif (options.signal?.aborted) throw createAbortError();\n\t\t\tif (retriesRemaining <= 0 || !isProviderError(error) || !isRetryableProviderError(error)) throw error;\n\n\t\t\tconst retryIndex = maxRetries - retriesRemaining;\n\t\t\tretriesRemaining--;\n\t\t\tawait abortableSleep(getRetryDelayMs(error, retryIndex, options.maxRetryDelayMs), options.signal);\n\t\t}\n\t}\n}\n"]}

package/dist/utils/retry.d.ts

@@ -0,0 +1,53 @@
+import type { AssistantMessage } from "../types.ts";
+/**
+ * Retry policy: bounded attempts with exponential backoff (`baseDelayMs * 2^(attempt-1)`).
+ * Matches `settings.retry` (`enabled`, `maxRetries`, `baseDelayMs`) in coding-agent; kept
+ * here so the classifier and the policy-driven retry loop live together and stay reusable
+ * by the SDK and other callers.
+ */
+export interface RetryPolicy {
+    enabled: boolean;
+    /** Max retry attempts (0 = no retries). The initial call never counts as a retry. */
+    maxRetries: number;
+    /** Base delay in ms. Per-attempt delay is `baseDelayMs * 2^(attempt-1)` before jitter. */
+    baseDelayMs: number;
+}
+/** Optional callbacks emitted by {@link retryAssistantCall} around each retry. */
+export interface RetryCallbacks {
+    /** Emitted before the backoff sleep of each retry attempt (1-indexed). */
+    onRetryScheduled?: (attempt: number, maxAttempts: number, delayMs: number, errorMessage: string) => void | Promise<void>;
+    /** Emitted after the backoff sleep, immediately before the retried call starts. */
+    onRetryAttemptStart?: () => void | Promise<void>;
+    /** Emitted once when the loop ends: success if a later call completed normally. */
+    onRetryFinished?: (success: boolean, attempt: number, finalError?: string) => void | Promise<void>;
+}
+/**
+ * Run a single assistant-producing call with bounded retry on transient errors.
+ *
+ * Behavior:
+ * - A successful response is returned immediately. Aborts are terminal and never
+ *   retried, but reported as unsuccessful if they happen after a retry was scheduled.
+ *   Aborts during the backoff sleep are normalized to an aborted `AssistantMessage`
+ *   too, so callers do not need to care when cancellation happened.
+ * - A non-retryable error (per {@link isRetryableAssistantError}, including quota/
+ *   billing exhaustion) is returned immediately so deterministic errors fail fast.
+ * - Otherwise retries up to `maxRetries` times with exponential backoff, emitting
+ *   `onRetryScheduled` before each sleep, `onRetryAttemptStart` after each sleep before
+ *   the retried call starts, and `onRetryFinished` once at the end (whether the loop
+ *   ends in success, exhausted retries, or an aborted backoff).
+ *
+ * When `policy` is undefined or disabled, the first response is returned unchanged
+ * (equivalent to calling `produce()` directly).
+ */
+export declare function retryAssistantCall(produce: () => Promise<AssistantMessage>, policy: RetryPolicy | undefined, signal: AbortSignal | undefined, callbacks?: RetryCallbacks): Promise<AssistantMessage>;
+/**
+ * Classifies whether a failed assistant message looks like a transient provider
+ * or transport error, so callers can decide if the last assistant turn should be
+ * restarted.
+ *
+ * This does not implement retry policy. Callers should first handle context
+ * overflow separately, then apply their own retry budget, backoff, and reporting
+ * before restarting the assistant turn.
+ */
+export declare function isRetryableAssistantError(message: AssistantMessage): boolean;
+//# sourceMappingURL=retry.d.ts.map