@code-yeongyu/senpi 2026.9.7 → 2026.9.8

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

@@ -1 +1 @@
-{"version":3,"file":"overflow.js","sourceRoot":"","sources":["../../src/utils/overflow.ts"],"names":[],"mappings":"AAEA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAqCG;AACH,MAAM,iBAAiB,GAAG;IACzB,6BAA6B,EAAE,qEAAqE;IACpG,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;IAClF,0DAA0D,EAAE,4UAA4U;IACxY,qGAAqG,EAAE,+GAA+G;IACtN,0EAA0E,EAAE,+DAA+D;CAC3I,CAAC;AAEF;;;;;;;;GAQG;AACH,MAAM,qBAAqB,GAAG;IAC7B,2CAA2C,EAAE,oFAAoF;IACjI,aAAa,EAAE,wBAAwB;IACvC,oBAAoB,EAAE,yBAAyB;CAC/C,CAAC;AAEF;;;;;;;;GAQG;AACH,MAAM,0BAA0B,GAAG,sBAAsB,CAAC;AAE1D;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAqDG;AACH,MAAM,UAAU,iBAAiB,CAAC,OAAyB,EAAE,aAAsB;IAClF,uCAAuC;IACvC,IAAI,OAAO,CAAC,UAAU,KAAK,OAAO,IAAI,OAAO,CAAC,YAAY,EAAE,CAAC;QAC5D,oFAAoF;QACpF,MAAM,aAAa,GAAG,qBAAqB,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC,OAAO,CAAC,YAAa,CAAC,CAAC,CAAC;QACvF,IAAI,CAAC,aAAa,IAAI,iBAAiB,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC,OAAO,CAAC,YAAa,CAAC,CAAC,EAAE,CAAC;YACpF,OAAO,IAAI,CAAC;QACb,CAAC;QACD,MAAM,KAAK,GAAG,OAAO,CAAC,KAAK,CAAC;QAC5B,MAAM,gBAAgB,GACrB,CAAC,KAAK,CAAC,WAAW,IAAI,KAAK,CAAC,KAAK,GAAG,KAAK,CAAC,MAAM,GAAG,KAAK,CAAC,SAAS,GAAG,KAAK,CAAC,UAAU,CAAC,GAAG,CAAC,CAAC;QAC5F,IACC,CAAC,aAAa;YACd,gBAAgB;YAChB,0BAA0B,CAAC,IAAI,CAAC,OAAO,CAAC,YAAY,CAAC;YACrD,CAAC,aAAa,KAAK,SAAS,IAAI,aAAa,IAAI,CAAC,IAAI,oBAAoB,CAAC,OAAO,CAAC,IAAI,aAAa,GAAG,GAAG,CAAC,EAC1G,CAAC;YACF,OAAO,IAAI,CAAC;QACb,CAAC;IACF,CAAC;IAED,8EAA8E;IAC9E,IAAI,aAAa,IAAI,OAAO,CAAC,UAAU,KAAK,MAAM,EAAE,CAAC;QACpD,MAAM,WAAW,GAAG,OAAO,CAAC,KAAK,CAAC,KAAK,GAAG,OAAO,CAAC,KAAK,CAAC,SAAS,CAAC;QAClE,IAAI,WAAW,GAAG,aAAa,EAAE,CAAC;YACjC,OAAO,IAAI,CAAC;QACb,CAAC;IACF,CAAC;IAED,sFAAsF;IACtF,qFAAqF;IACrF,gEAAgE;IAChE,IAAI,aAAa,IAAI,OAAO,CAAC,UAAU,KAAK,QAAQ,IAAI,OAAO,CAAC,KAAK,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;QACpF,MAAM,WAAW,GAAG,OAAO,CAAC,KAAK,CAAC,KAAK,GAAG,OAAO,CAAC,KAAK,CAAC,SAAS,CAAC;QAClE,IAAI,WAAW,IAAI,aAAa,GAAG,IAAI,EAAE,CAAC;YACzC,OAAO,IAAI,CAAC;QACb,CAAC;IACF,CAAC;IAED,OAAO,KAAK,CAAC;AACd,CAAC;AACD;;;;;GAKG;AACH,MAAM,UAAU,mBAAmB,CAAC,OAAyB,EAAE,gBAAwB;IACtF,OAAO,OAAO,CAAC,UAAU,KAAK,QAAQ,IAAI,gBAAgB,GAAG,CAAC,IAAI,OAAO,CAAC,KAAK,CAAC,MAAM,GAAG,gBAAgB,CAAC;AAC3G,CAAC;AAED;;GAEG;AACH,MAAM,UAAU,mBAAmB;IAClC,OAAO,CAAC,GAAG,iBAAiB,CAAC,CAAC;AAC/B,CAAC;AAED,SAAS,oBAAoB,CAAC,OAE7B;IACA,MAAM,KAAK,GAAG,OAAO,CAAC,KAAK,CAAC;IAC5B,IAAI,CAAC,KAAK;QAAE,OAAO,CAAC,CAAC;IACrB,OAAO,CACN,KAAK,CAAC,WAAW,IAAI,CAAC,KAAK,CAAC,KAAK,IAAI,CAAC,CAAC,GAAG,CAAC,KAAK,CAAC,MAAM,IAAI,CAAC,CAAC,GAAG,CAAC,KAAK,CAAC,SAAS,IAAI,CAAC,CAAC,GAAG,CAAC,KAAK,CAAC,UAAU,IAAI,CAAC,CAAC,CAChH,CAAC;AACH,CAAC;AAED,MAAM,UAAU,gCAAgC,CAC/C,OAIC,EACD,eAAuB;IAEvB,IAAI,OAAO,CAAC,UAAU,KAAK,OAAO,IAAI,CAAC,sBAAsB,CAAC,IAAI,CAAC,OAAO,CAAC,YAAY,IAAI,EAAE,CAAC,EAAE,CAAC;QAChG,OAAO,KAAK,CAAC;IACd,CAAC;IACD,OAAO,CAAC,CAAC,oBAAoB,CAAC,OAAO,CAAC,GAAG,CAAC,CAAC,CAAC;AAC7C,CAAC;AAED;;;;GAIG;AACH,MAAM,UAAU,8BAA8B,CAC7C,OAIC,EACD,aAAqB;IAErB,MAAM,MAAM,GAAG,oBAAoB,CAAC,OAAO,CAAC,CAAC;IAC7C,OAAO,CACN,OAAO,CAAC,UAAU,KAAK,OAAO;QAC9B,0BAA0B,CAAC,IAAI,CAAC,OAAO,CAAC,YAAY,IAAI,EAAE,CAAC;QAC3D,aAAa,GAAG,CAAC;QACjB,MAAM,GAAG,CAAC;QACV,MAAM,GAAG,aAAa,GAAG,GAAG,CAC5B,CAAC;AACH,CAAC;AAED,MAAM,UAAU,kCAAkC,CAAC,OAIlD;IACA,IAAI,OAAO,CAAC,UAAU,KAAK,OAAO,IAAI,CAAC,sBAAsB,CAAC,IAAI,CAAC,OAAO,CAAC,YAAY,IAAI,EAAE,CAAC,EAAE,CAAC;QAChG,OAAO,KAAK,CAAC;IACd,CAAC;IACD,OAAO,CAAC,CAAC,oBAAoB,CAAC,OAAO,CAAC,GAAG,CAAC,CAAC,CAAC;AAC7C,CAAC;AAED,MAAM,UAAU,yCAAyC,CAAC,OAAkD;IAC3G,OAAO,OAAO,EAAE,eAAe,KAAK,IAAI,CAAC;AAC1C,CAAC;AAED,MAAM,UAAU,iCAAiC,CAAC,SAAkB,EAAE,YAAoB;IACzF,IAAI,SAAS;QAAE,OAAO,KAAK,CAAC;IAC5B,OAAO,qBAAqB,CAAC,IAAI,CAAC,YAAY,IAAI,EAAE,CAAC,CAAC;AACvD,CAAC;AAED,MAAM,UAAU,gCAAgC,CAC/C,QAAW,EACX,QAA4B,EAC5B,MAA0B;IAE1B,IAAI,MAAM,KAAK,UAAU;QAAE,OAAO,QAAQ,CAAC;IAC3C,IAAI,QAAQ,KAAK,QAAQ,IAAI,QAAQ,KAAK,kBAAkB;QAAE,OAAO,QAAQ,CAAC;IAC9E,OAAO,EAAE,GAAG,QAAQ,EAAE,gBAAgB,EAAE,CAAC,EAAE,kBAAkB,EAAE,KAAK,EAAE,CAAC;AACxE,CAAC","sourcesContent":["import type { AssistantMessage } from \"../types.ts\";\n\n/**\n * Regex patterns to detect context overflow errors from different providers.\n *\n * These patterns match error messages returned when the input exceeds\n * the model's context window.\n *\n * Provider-specific patterns (with example error messages):\n *\n * - Anthropic: \"prompt is too long: 213462 tokens > 200000 maximum\"\n * - Anthropic: \"413 {\\\"error\\\":{\\\"type\\\":\\\"request_too_large\\\",\\\"message\\\":\\\"Request exceeds the maximum size\\\"}}\"\n * - OpenAI: \"Your input exceeds the context window of this model\"\n * - OpenAI/LiteLLM: \"Requested token count exceeds the model's maximum context length of 131072 tokens\"\n * - OpenAI-compatible: \"Input length (265330) exceeds model's maximum context length (262144).\"\n * - Google: \"The input token count (1196265) exceeds the maximum number of tokens allowed (1048575)\"\n * - xAI: \"This model's maximum prompt length is 131072 but the request contains 537812 tokens\"\n * - Groq: \"Please reduce the length of the messages or completion\"\n * - OpenRouter: \"This endpoint's maximum context length is X tokens. However, you requested about Y tokens\"\n * - OpenRouter/Poolside: \"Input length X exceeds the maximum allowed input length of Y tokens.\"\n * - Together AI: \"The input (X tokens) is longer than the model's context length (Y tokens).\"\n * - llama.cpp: \"the request exceeds the available context size, try increasing it\"\n * - LM Studio: \"tokens to keep from the initial prompt is greater than the context length\"\n * - GitHub Copilot: \"prompt token count of X exceeds the limit of Y\"\n * - MiniMax: \"invalid params, context window exceeds limit\"\n * - Kimi For Coding: \"Your request exceeded model token limit: X (requested: Y)\"\n * - DS4: \"Prompt has X tokens, but the configured context size is Y tokens\"\n * - Cerebras: \"400/413 status code (no body)\"\n * - Gateways: \"413 Request body too large\" / \"Request Entity Too Large\" / \"Payload Too Large\" (byte-size overflow)\n * - kiro-lb gateways: \"Request payload is 1095225 bytes, over the 1085435 byte limit Kiro accepts.\" / \"Request payload is N tokens, over the M token limit Kiro accepts.\" (HTTP 400 local payload guard)\n * - Kiro upstream via kiro-lb: \"Model context limit reached. Conversation size exceeds model capacity.\" (CONTENT_LENGTH_EXCEEDS_THRESHOLD token overflow)\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 * - pi-ai pre-flight guard (api/context-room.ts): \"Context window exhausted: the conversation is estimated at X of Y tokens, ...\" - raised before any provider call\n */\nconst OVERFLOW_PATTERNS = [\n\t/^Context window exhausted: /, // pi-ai pre-flight guard: no answer room left, provider never called\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\t/(?:request[ _])?(?:body|entity|payload)[_ ]too[_ ]large/i, // Gateway HTTP 413 byte-size rejections (\"Request body too large\", \"Request Entity Too Large\", \"body_too_large\", \"Payload Too Large\"). Substring-anchored by design (JSON bodies lack an adjacent status code); a non-context size rejection (e.g. an oversized image) can over-match, which costs one bounded shrink-retry, never a wedge.\n\t/Request payload is \\d+ (?:bytes, over the \\d+ byte|tokens, over the \\d+ token) limit Kiro accepts\\./, // kiro-lb local byte/token payload guard (HTTP 400; Anthropic uses invalid_request_error, OpenAI uses detail).\n\t/Model context limit reached\\. Conversation size exceeds model capacity\\./, // kiro-lb enhancement of Kiro CONTENT_LENGTH_EXCEEDS_THRESHOLD\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 * Cursor's api2.cursor.sh surfaces a context overflow as a bare gRPC\n * `resource_exhausted` end-stream — the same wording its backend uses for\n * quota and poisoned-conversation rejections. Token evidence disambiguates:\n * a request that already streamed or billed tokens overflowed mid-flight,\n * while a zero-token rejection is a quota/conversation failure that must stay\n * on the rate-limit path (the cursor client rotates the conversation id for\n * those and retry handling supplies the backoff).\n */\nconst RESOURCE_EXHAUSTED_PATTERN = /resource.?exhausted/i;\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\tconst usage = message.usage;\n\t\tconst hasTokenEvidence =\n\t\t\t(usage.totalTokens || usage.input + usage.output + usage.cacheRead + usage.cacheWrite) > 0;\n\t\tif (\n\t\t\t!isNonOverflow &&\n\t\t\thasTokenEvidence &&\n\t\t\tRESOURCE_EXHAUSTED_PATTERN.test(message.errorMessage) &&\n\t\t\t(contextWindow === undefined || contextWindow <= 0 || cursorZeroTokenCount(message) >= contextWindow * 0.5)\n\t\t) {\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 * 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\nfunction cursorZeroTokenCount(message: {\n\tusage?: { input?: number; output?: number; cacheRead?: number; cacheWrite?: number; totalTokens?: number };\n}): number {\n\tconst usage = message.usage;\n\tif (!usage) return 0;\n\treturn (\n\t\tusage.totalTokens || (usage.input ?? 0) + (usage.output ?? 0) + (usage.cacheRead ?? 0) + (usage.cacheWrite ?? 0)\n\t);\n}\n\nexport function isCursorPayloadResourceExhausted(\n\tmessage: {\n\t\tstopReason?: string;\n\t\terrorMessage?: string;\n\t\tusage?: { input?: number; output?: number; cacheRead?: number; cacheWrite?: number; totalTokens?: number };\n\t},\n\t_estimateTokens: number,\n): boolean {\n\tif (message.stopReason !== \"error\" || !/resource.?exhausted/i.test(message.errorMessage || \"\")) {\n\t\treturn false;\n\t}\n\treturn !(cursorZeroTokenCount(message) > 0);\n}\n\n/**\n * Detects Cursor's verified usage-pool exhaustion signature: a token-bearing\n * `resource_exhausted` error while the conversation is well below the model\n * context window.\n */\nexport function isCursorQuotaResourceExhausted(\n\tmessage: {\n\t\tstopReason?: string;\n\t\terrorMessage?: string;\n\t\tusage?: { input?: number; output?: number; cacheRead?: number; cacheWrite?: number; totalTokens?: number };\n\t},\n\tcontextWindow: number,\n): boolean {\n\tconst tokens = cursorZeroTokenCount(message);\n\treturn (\n\t\tmessage.stopReason === \"error\" &&\n\t\tRESOURCE_EXHAUSTED_PATTERN.test(message.errorMessage || \"\") &&\n\t\tcontextWindow > 0 &&\n\t\ttokens > 0 &&\n\t\ttokens < contextWindow * 0.5\n\t);\n}\n\nexport function isCursorZeroTokenResourceExhausted(message: {\n\tstopReason?: string;\n\terrorMessage?: string;\n\tusage?: { input?: number; output?: number; cacheRead?: number; cacheWrite?: number; totalTokens?: number };\n}): boolean {\n\tif (message.stopReason !== \"error\" || !/resource.?exhausted/i.test(message.errorMessage || \"\")) {\n\t\treturn false;\n\t}\n\treturn !(cursorZeroTokenCount(message) > 0);\n}\n\nexport function shouldSkipProviderFallbackForCursorZeroRe(options: { sameModelRemint?: boolean } | undefined): boolean {\n\treturn options?.sameModelRemint === true;\n}\n\nexport function shouldRetryOverflowWithoutCompact(compacted: boolean, errorMessage: string): boolean {\n\tif (compacted) return false;\n\treturn /Nothing to compact/i.test(errorMessage || \"\");\n}\n\nexport function cursorOverflowCompactionSettings<T extends { keepRecentTokens?: number; restorationEnabled?: boolean }>(\n\tsettings: T,\n\tprovider: string | undefined,\n\treason: string | undefined,\n): T {\n\tif (reason !== \"overflow\") return settings;\n\tif (provider !== \"cursor\" && provider !== \"cursor-cli-oauth\") return settings;\n\treturn { ...settings, keepRecentTokens: 0, restorationEnabled: false };\n}\n"]}
+{"version":3,"file":"overflow.js","sourceRoot":"","sources":["../../src/utils/overflow.ts"],"names":[],"mappings":"AAEA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAsCG;AACH,MAAM,iBAAiB,GAAG;IACzB,6BAA6B,EAAE,qEAAqE;IACpG,qBAAqB,EAAE,2BAA2B;IAClD,oBAAoB,EAAE,kDAAkD;IACxE,wCAAwC,EAAE,iBAAiB;IAC3D,yDAAyD,EAAE,uCAAuC;IAClG,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;IAClF,0DAA0D,EAAE,4UAA4U;IACxY,qGAAqG,EAAE,+GAA+G;IACtN,0EAA0E,EAAE,+DAA+D;CAC3I,CAAC;AAEF;;;;;;;;;;GAUG;AACH,MAAM,qBAAqB,GAAG;IAC7B,2CAA2C,EAAE,oFAAoF;IACjI,aAAa,EAAE,wBAAwB;IACvC,oBAAoB,EAAE,yBAAyB;IAC/C,qCAAqC,EAAE,sDAAsD;IAC7F,UAAU,EAAE,iCAAiC;IAC7C,UAAU,EAAE,mCAAmC;IAC/C,iBAAiB,EAAE,qCAAqC;IACxD,wBAAwB,EAAE,iCAAiC;IAC3D,QAAQ,EAAE,kBAAkB;IAC5B,kBAAkB,EAAE,iCAAiC;IACrD,aAAa,EAAE,sCAAsC;CACrD,CAAC;AAEF;;;;;;;;GAQG;AACH,MAAM,0BAA0B,GAAG,sBAAsB,CAAC;AAE1D;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAqDG;AACH,MAAM,UAAU,iBAAiB,CAAC,OAAyB,EAAE,aAAsB;IAClF,uCAAuC;IACvC,IAAI,OAAO,CAAC,UAAU,KAAK,OAAO,IAAI,OAAO,CAAC,YAAY,EAAE,CAAC;QAC5D,oFAAoF;QACpF,MAAM,aAAa,GAAG,qBAAqB,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC,OAAO,CAAC,YAAa,CAAC,CAAC,CAAC;QACvF,IAAI,CAAC,aAAa,IAAI,iBAAiB,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC,OAAO,CAAC,YAAa,CAAC,CAAC,EAAE,CAAC;YACpF,OAAO,IAAI,CAAC;QACb,CAAC;QACD,MAAM,KAAK,GAAG,OAAO,CAAC,KAAK,CAAC;QAC5B,MAAM,gBAAgB,GACrB,CAAC,KAAK,CAAC,WAAW,IAAI,KAAK,CAAC,KAAK,GAAG,KAAK,CAAC,MAAM,GAAG,KAAK,CAAC,SAAS,GAAG,KAAK,CAAC,UAAU,CAAC,GAAG,CAAC,CAAC;QAC5F,IACC,CAAC,aAAa;YACd,gBAAgB;YAChB,0BAA0B,CAAC,IAAI,CAAC,OAAO,CAAC,YAAY,CAAC;YACrD,CAAC,aAAa,KAAK,SAAS,IAAI,aAAa,IAAI,CAAC,IAAI,oBAAoB,CAAC,OAAO,CAAC,IAAI,aAAa,GAAG,GAAG,CAAC,EAC1G,CAAC;YACF,OAAO,IAAI,CAAC;QACb,CAAC;IACF,CAAC;IAED,8EAA8E;IAC9E,IAAI,aAAa,IAAI,OAAO,CAAC,UAAU,KAAK,MAAM,EAAE,CAAC;QACpD,MAAM,WAAW,GAAG,OAAO,CAAC,KAAK,CAAC,KAAK,GAAG,OAAO,CAAC,KAAK,CAAC,SAAS,CAAC;QAClE,IAAI,WAAW,GAAG,aAAa,EAAE,CAAC;YACjC,OAAO,IAAI,CAAC;QACb,CAAC;IACF,CAAC;IAED,sFAAsF;IACtF,qFAAqF;IACrF,gEAAgE;IAChE,IAAI,aAAa,IAAI,OAAO,CAAC,UAAU,KAAK,QAAQ,IAAI,OAAO,CAAC,KAAK,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;QACpF,MAAM,WAAW,GAAG,OAAO,CAAC,KAAK,CAAC,KAAK,GAAG,OAAO,CAAC,KAAK,CAAC,SAAS,CAAC;QAClE,IAAI,WAAW,IAAI,aAAa,GAAG,IAAI,EAAE,CAAC;YACzC,OAAO,IAAI,CAAC;QACb,CAAC;IACF,CAAC;IAED,OAAO,KAAK,CAAC;AACd,CAAC;AACD;;;;;GAKG;AACH,MAAM,UAAU,mBAAmB,CAAC,OAAyB,EAAE,gBAAwB;IACtF,OAAO,OAAO,CAAC,UAAU,KAAK,QAAQ,IAAI,gBAAgB,GAAG,CAAC,IAAI,OAAO,CAAC,KAAK,CAAC,MAAM,GAAG,gBAAgB,CAAC;AAC3G,CAAC;AAED;;GAEG;AACH,MAAM,UAAU,mBAAmB;IAClC,OAAO,CAAC,GAAG,iBAAiB,CAAC,CAAC;AAC/B,CAAC;AAED,SAAS,oBAAoB,CAAC,OAE7B;IACA,MAAM,KAAK,GAAG,OAAO,CAAC,KAAK,CAAC;IAC5B,IAAI,CAAC,KAAK;QAAE,OAAO,CAAC,CAAC;IACrB,OAAO,CACN,KAAK,CAAC,WAAW,IAAI,CAAC,KAAK,CAAC,KAAK,IAAI,CAAC,CAAC,GAAG,CAAC,KAAK,CAAC,MAAM,IAAI,CAAC,CAAC,GAAG,CAAC,KAAK,CAAC,SAAS,IAAI,CAAC,CAAC,GAAG,CAAC,KAAK,CAAC,UAAU,IAAI,CAAC,CAAC,CAChH,CAAC;AACH,CAAC;AAED,MAAM,UAAU,gCAAgC,CAC/C,OAIC,EACD,eAAuB;IAEvB,IAAI,OAAO,CAAC,UAAU,KAAK,OAAO,IAAI,CAAC,sBAAsB,CAAC,IAAI,CAAC,OAAO,CAAC,YAAY,IAAI,EAAE,CAAC,EAAE,CAAC;QAChG,OAAO,KAAK,CAAC;IACd,CAAC;IACD,OAAO,CAAC,CAAC,oBAAoB,CAAC,OAAO,CAAC,GAAG,CAAC,CAAC,CAAC;AAC7C,CAAC;AAED;;;;GAIG;AACH,MAAM,UAAU,8BAA8B,CAC7C,OAIC,EACD,aAAqB;IAErB,MAAM,MAAM,GAAG,oBAAoB,CAAC,OAAO,CAAC,CAAC;IAC7C,OAAO,CACN,OAAO,CAAC,UAAU,KAAK,OAAO;QAC9B,0BAA0B,CAAC,IAAI,CAAC,OAAO,CAAC,YAAY,IAAI,EAAE,CAAC;QAC3D,aAAa,GAAG,CAAC;QACjB,MAAM,GAAG,CAAC;QACV,MAAM,GAAG,aAAa,GAAG,GAAG,CAC5B,CAAC;AACH,CAAC;AAED,MAAM,UAAU,kCAAkC,CAAC,OAIlD;IACA,IAAI,OAAO,CAAC,UAAU,KAAK,OAAO,IAAI,CAAC,sBAAsB,CAAC,IAAI,CAAC,OAAO,CAAC,YAAY,IAAI,EAAE,CAAC,EAAE,CAAC;QAChG,OAAO,KAAK,CAAC;IACd,CAAC;IACD,OAAO,CAAC,CAAC,oBAAoB,CAAC,OAAO,CAAC,GAAG,CAAC,CAAC,CAAC;AAC7C,CAAC;AAED,MAAM,UAAU,yCAAyC,CAAC,OAAkD;IAC3G,OAAO,OAAO,EAAE,eAAe,KAAK,IAAI,CAAC;AAC1C,CAAC;AAED,MAAM,UAAU,iCAAiC,CAAC,SAAkB,EAAE,YAAoB;IACzF,IAAI,SAAS;QAAE,OAAO,KAAK,CAAC;IAC5B,OAAO,qBAAqB,CAAC,IAAI,CAAC,YAAY,IAAI,EAAE,CAAC,CAAC;AACvD,CAAC;AAED,MAAM,UAAU,gCAAgC,CAC/C,QAAW,EACX,QAA4B,EAC5B,MAA0B;IAE1B,IAAI,MAAM,KAAK,UAAU;QAAE,OAAO,QAAQ,CAAC;IAC3C,IAAI,QAAQ,KAAK,QAAQ,IAAI,QAAQ,KAAK,kBAAkB;QAAE,OAAO,QAAQ,CAAC;IAC9E,OAAO,EAAE,GAAG,QAAQ,EAAE,gBAAgB,EAAE,CAAC,EAAE,kBAAkB,EAAE,KAAK,EAAE,CAAC;AACxE,CAAC","sourcesContent":["import type { AssistantMessage } from \"../types.ts\";\n\n/**\n * Regex patterns to detect context overflow errors from different providers.\n *\n * These patterns match error messages returned when the input exceeds\n * the model's context window.\n *\n * Provider-specific patterns (with example error messages):\n *\n * - Anthropic: \"prompt is too long: 213462 tokens > 200000 maximum\"\n * - Anthropic: \"413 {\\\"error\\\":{\\\"type\\\":\\\"request_too_large\\\",\\\"message\\\":\\\"Request exceeds the maximum size\\\"}}\"\n * - OpenAI: \"Your input exceeds the context window of this model\"\n * - OpenAI: \"Your input exceeds the model's context window\" / \"exceeds this model's context window\"\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 * - Gateways: \"413 Request body too large\" / \"Request Entity Too Large\" / \"Payload Too Large\" (byte-size overflow)\n * - kiro-lb gateways: \"Request payload is 1095225 bytes, over the 1085435 byte limit Kiro accepts.\" / \"Request payload is N tokens, over the M token limit Kiro accepts.\" (HTTP 400 local payload guard)\n * - Kiro upstream via kiro-lb: \"Model context limit reached. Conversation size exceeds model capacity.\" (CONTENT_LENGTH_EXCEEDS_THRESHOLD token overflow)\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 * - pi-ai pre-flight guard (api/context-room.ts): \"Context window exhausted: the conversation is estimated at X of Y tokens, ...\" - raised before any provider call\n */\nconst OVERFLOW_PATTERNS = [\n\t/^Context window exhausted: /, // pi-ai pre-flight guard: no answer room left, provider never called\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|this) )?(?:model'?s )?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\t/(?:request[ _])?(?:body|entity|payload)[_ ]too[_ ]large/i, // Gateway HTTP 413 byte-size rejections (\"Request body too large\", \"Request Entity Too Large\", \"body_too_large\", \"Payload Too Large\"). Substring-anchored by design (JSON bodies lack an adjacent status code); a non-context size rejection (e.g. an oversized image) can over-match, which costs one bounded shrink-retry, never a wedge.\n\t/Request payload is \\d+ (?:bytes, over the \\d+ byte|tokens, over the \\d+ token) limit Kiro accepts\\./, // kiro-lb local byte/token payload guard (HTTP 400; Anthropic uses invalid_request_error, OpenAI uses detail).\n\t/Model context limit reached\\. Conversation size exceeds model capacity\\./, // kiro-lb enhancement of Kiro CONTENT_LENGTH_EXCEEDS_THRESHOLD\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. Token-quota / TPM messages such as \"Too many tokens\n * per minute\" or \"This request exceeds the limit of 30000 tokens per minute\" similarly\n * match generic overflow fallbacks and must stay on the rate-limit path.\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\t/tokens per (?:min|minute|hour|day)/i, // Token-quota windows (TPM/TPH/TPD), not context size\n\t/\\bTPM\\b/i, // Tokens-per-minute abbreviation\n\t/\\bRPM\\b/i, // Requests-per-minute abbreviation\n\t/quota exceeded/i, // Provider quota, not context window\n\t/retry (?:after|in) \\d/i, // Retry-after rate-limit wording\n\t/^429\\b/, // HTTP 429 prefix\n\t/status code 429/i, // HTTP 429 mentioned mid-message\n\t/overloaded/i, // Provider capacity, not context size\n];\n\n/**\n * Cursor's api2.cursor.sh surfaces a context overflow as a bare gRPC\n * `resource_exhausted` end-stream — the same wording its backend uses for\n * quota and poisoned-conversation rejections. Token evidence disambiguates:\n * a request that already streamed or billed tokens overflowed mid-flight,\n * while a zero-token rejection is a quota/conversation failure that must stay\n * on the rate-limit path (the cursor client rotates the conversation id for\n * those and retry handling supplies the backoff).\n */\nconst RESOURCE_EXHAUSTED_PATTERN = /resource.?exhausted/i;\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 context window\", \"exceeds this model's 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\tconst usage = message.usage;\n\t\tconst hasTokenEvidence =\n\t\t\t(usage.totalTokens || usage.input + usage.output + usage.cacheRead + usage.cacheWrite) > 0;\n\t\tif (\n\t\t\t!isNonOverflow &&\n\t\t\thasTokenEvidence &&\n\t\t\tRESOURCE_EXHAUSTED_PATTERN.test(message.errorMessage) &&\n\t\t\t(contextWindow === undefined || contextWindow <= 0 || cursorZeroTokenCount(message) >= contextWindow * 0.5)\n\t\t) {\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 * 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\nfunction cursorZeroTokenCount(message: {\n\tusage?: { input?: number; output?: number; cacheRead?: number; cacheWrite?: number; totalTokens?: number };\n}): number {\n\tconst usage = message.usage;\n\tif (!usage) return 0;\n\treturn (\n\t\tusage.totalTokens || (usage.input ?? 0) + (usage.output ?? 0) + (usage.cacheRead ?? 0) + (usage.cacheWrite ?? 0)\n\t);\n}\n\nexport function isCursorPayloadResourceExhausted(\n\tmessage: {\n\t\tstopReason?: string;\n\t\terrorMessage?: string;\n\t\tusage?: { input?: number; output?: number; cacheRead?: number; cacheWrite?: number; totalTokens?: number };\n\t},\n\t_estimateTokens: number,\n): boolean {\n\tif (message.stopReason !== \"error\" || !/resource.?exhausted/i.test(message.errorMessage || \"\")) {\n\t\treturn false;\n\t}\n\treturn !(cursorZeroTokenCount(message) > 0);\n}\n\n/**\n * Detects Cursor's verified usage-pool exhaustion signature: a token-bearing\n * `resource_exhausted` error while the conversation is well below the model\n * context window.\n */\nexport function isCursorQuotaResourceExhausted(\n\tmessage: {\n\t\tstopReason?: string;\n\t\terrorMessage?: string;\n\t\tusage?: { input?: number; output?: number; cacheRead?: number; cacheWrite?: number; totalTokens?: number };\n\t},\n\tcontextWindow: number,\n): boolean {\n\tconst tokens = cursorZeroTokenCount(message);\n\treturn (\n\t\tmessage.stopReason === \"error\" &&\n\t\tRESOURCE_EXHAUSTED_PATTERN.test(message.errorMessage || \"\") &&\n\t\tcontextWindow > 0 &&\n\t\ttokens > 0 &&\n\t\ttokens < contextWindow * 0.5\n\t);\n}\n\nexport function isCursorZeroTokenResourceExhausted(message: {\n\tstopReason?: string;\n\terrorMessage?: string;\n\tusage?: { input?: number; output?: number; cacheRead?: number; cacheWrite?: number; totalTokens?: number };\n}): boolean {\n\tif (message.stopReason !== \"error\" || !/resource.?exhausted/i.test(message.errorMessage || \"\")) {\n\t\treturn false;\n\t}\n\treturn !(cursorZeroTokenCount(message) > 0);\n}\n\nexport function shouldSkipProviderFallbackForCursorZeroRe(options: { sameModelRemint?: boolean } | undefined): boolean {\n\treturn options?.sameModelRemint === true;\n}\n\nexport function shouldRetryOverflowWithoutCompact(compacted: boolean, errorMessage: string): boolean {\n\tif (compacted) return false;\n\treturn /Nothing to compact/i.test(errorMessage || \"\");\n}\n\nexport function cursorOverflowCompactionSettings<T extends { keepRecentTokens?: number; restorationEnabled?: boolean }>(\n\tsettings: T,\n\tprovider: string | undefined,\n\treason: string | undefined,\n): T {\n\tif (reason !== \"overflow\") return settings;\n\tif (provider !== \"cursor\" && provider !== \"cursor-cli-oauth\") return settings;\n\treturn { ...settings, keepRecentTokens: 0, restorationEnabled: false };\n}\n"]}

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

@@ -1,4 +1,4 @@
-import type { AssistantMessage } from "../types.ts";
+import type { AssistantMessage, Model } from "../types.ts";
 export declare const SERVER_FALLBACK_ABORTED_DIAGNOSTIC = "server_fallback_aborted";
 export declare const BILLING_INCOMPLETE_DIAGNOSTIC = "billing_incomplete_after_client_abort";
 export interface ServerFallbackReceipt {
@@ -20,6 +20,7 @@ export declare function parseServerFallbackReceipt(block: unknown): ServerFallba
  */
 export declare function parseStickyFallbackReceipt(usage: unknown, requestedModel: string, servedModel?: string): ServerFallbackReceipt | undefined;
 export declare function serverFallbackRefusalExplanation(receipt: ServerFallbackReceipt): string;
+export declare function applyServerFallbackContinuation(message: AssistantMessage, model: Model<"anthropic-messages">, receipt: ServerFallbackReceipt): Model<"anthropic-messages">;
 /**
  * Rewrites an aborted turn into the same shape a pre-output classifier refusal
  * has, so `isClassifierRefusal()` routes it through the caller's fallback chain.

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

@@ -1 +1 @@
-{"version":3,"file":"server-fallback-receipt.d.ts","sourceRoot":"","sources":["../../src/utils/server-fallback-receipt.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,gBAAgB,EAAE,MAAM,aAAa,CAAC;AAGpD,eAAO,MAAM,kCAAkC,4BAA4B,CAAC;AAC5E,eAAO,MAAM,6BAA6B,0CAA0C,CAAC;AAErF,MAAM,WAAW,qBAAqB;IACrC,QAAQ,CAAC,IAAI,EAAE,MAAM,CAAC;IACtB,QAAQ,CAAC,EAAE,EAAE,MAAM,CAAC;CACpB;AAWD;;;;GAIG;AACH,wBAAgB,0BAA0B,CAAC,KAAK,EAAE,OAAO,GAAG,qBAAqB,GAAG,SAAS,CAK5F;AAED;;;;;;GAMG;AACH,wBAAgB,0BAA0B,CACzC,KAAK,EAAE,OAAO,EACd,cAAc,EAAE,MAAM,EACtB,WAAW,CAAC,EAAE,MAAM,GAClB,qBAAqB,GAAG,SAAS,CAUnC;AAED,wBAAgB,gCAAgC,CAAC,OAAO,EAAE,qBAAqB,GAAG,MAAM,CAEvF;AAED;;;;;GAKG;AACH,wBAAgB,wBAAwB,CAAC,OAAO,EAAE,gBAAgB,EAAE,OAAO,EAAE,qBAAqB,GAAG,IAAI,CAgBxG"}
+{"version":3,"file":"server-fallback-receipt.d.ts","sourceRoot":"","sources":["../../src/utils/server-fallback-receipt.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,gBAAgB,EAAE,KAAK,EAAE,MAAM,aAAa,CAAC;AAG3D,eAAO,MAAM,kCAAkC,4BAA4B,CAAC;AAC5E,eAAO,MAAM,6BAA6B,0CAA0C,CAAC;AAErF,MAAM,WAAW,qBAAqB;IACrC,QAAQ,CAAC,IAAI,EAAE,MAAM,CAAC;IACtB,QAAQ,CAAC,EAAE,EAAE,MAAM,CAAC;CACpB;AAWD;;;;GAIG;AACH,wBAAgB,0BAA0B,CAAC,KAAK,EAAE,OAAO,GAAG,qBAAqB,GAAG,SAAS,CAK5F;AAED;;;;;;GAMG;AACH,wBAAgB,0BAA0B,CACzC,KAAK,EAAE,OAAO,EACd,cAAc,EAAE,MAAM,EACtB,WAAW,CAAC,EAAE,MAAM,GAClB,qBAAqB,GAAG,SAAS,CAUnC;AAED,wBAAgB,gCAAgC,CAAC,OAAO,EAAE,qBAAqB,GAAG,MAAM,CAEvF;AAED,wBAAgB,+BAA+B,CAC9C,OAAO,EAAE,gBAAgB,EACzB,KAAK,EAAE,KAAK,CAAC,oBAAoB,CAAC,EAClC,OAAO,EAAE,qBAAqB,GAC5B,KAAK,CAAC,oBAAoB,CAAC,CAkB7B;AAED;;;;;GAKG;AACH,wBAAgB,wBAAwB,CAAC,OAAO,EAAE,gBAAgB,EAAE,OAAO,EAAE,qBAAqB,GAAG,IAAI,CAgBxG"}

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

@@ -44,6 +44,22 @@ export function parseStickyFallbackReceipt(usage, requestedModel, servedModel) {
 export function serverFallbackRefusalExplanation(receipt) {
     return `Server-side fallback (${receipt.from} -> ${receipt.to}) aborted by client policy`;
 }
+export function applyServerFallbackContinuation(message, model, receipt) {
+    // Keep source indices stable for stream projections while making abandoned
+    // client calls audit-only before the agent can execute the completed message.
+    for (const [index, block] of message.content.entries()) {
+        if (block.type === "toolCall") {
+            message.content[index] = { type: "providerNative", subtype: "discarded_tool_call", raw: block };
+        }
+    }
+    message.model = receipt.to;
+    const fallback = model.compat?.allowedFallbackModels?.find((candidate) => typeof candidate !== "string" && candidate.provider === model.provider && candidate.model === receipt.to);
+    return {
+        ...model,
+        id: receipt.to,
+        cost: fallback && typeof fallback !== "string" ? fallback.cost : model.cost,
+    };
+}
 /**
  * Rewrites an aborted turn into the same shape a pre-output classifier refusal
  * has, so `isClassifierRefusal()` routes it through the caller's fallback chain.

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

@@ -1 +1 @@
-{"version":3,"file":"server-fallback-receipt.js","sourceRoot":"","sources":["../../src/utils/server-fallback-receipt.ts"],"names":[],"mappings":"AACA,OAAO,EAAE,gCAAgC,EAAE,MAAM,kBAAkB,CAAC;AAEpE,MAAM,CAAC,MAAM,kCAAkC,GAAG,yBAAyB,CAAC;AAC5E,MAAM,CAAC,MAAM,6BAA6B,GAAG,uCAAuC,CAAC;AAOrF,SAAS,QAAQ,CAAC,KAAc;IAC/B,OAAO,OAAO,KAAK,KAAK,QAAQ,IAAI,KAAK,KAAK,IAAI,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC,CAAC;AAC7E,CAAC;AAED,SAAS,SAAS,CAAC,KAAc;IAChC,IAAI,CAAC,QAAQ,CAAC,KAAK,CAAC;QAAE,OAAO,SAAS,CAAC;IACvC,OAAO,OAAO,KAAK,CAAC,KAAK,KAAK,QAAQ,IAAI,KAAK,CAAC,KAAK,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC,KAAK,CAAC,CAAC,CAAC,SAAS,CAAC;AAC5F,CAAC;AAED;;;;GAIG;AACH,MAAM,UAAU,0BAA0B,CAAC,KAAc;IACxD,IAAI,CAAC,QAAQ,CAAC,KAAK,CAAC,IAAI,KAAK,CAAC,IAAI,KAAK,UAAU;QAAE,OAAO,SAAS,CAAC;IACpE,MAAM,IAAI,GAAG,SAAS,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC;IACnC,MAAM,EAAE,GAAG,SAAS,CAAC,KAAK,CAAC,EAAE,CAAC,CAAC;IAC/B,OAAO,IAAI,KAAK,SAAS,IAAI,EAAE,KAAK,SAAS,CAAC,CAAC,CAAC,EAAE,IAAI,EAAE,EAAE,EAAE,CAAC,CAAC,CAAC,SAAS,CAAC;AAC1E,CAAC;AAED;;;;;;GAMG;AACH,MAAM,UAAU,0BAA0B,CACzC,KAAc,EACd,cAAsB,EACtB,WAAoB;IAEpB,IAAI,CAAC,QAAQ,CAAC,KAAK,CAAC,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC,UAAU,CAAC;QAAE,OAAO,SAAS,CAAC;IAC3E,KAAK,IAAI,KAAK,GAAG,KAAK,CAAC,UAAU,CAAC,MAAM,GAAG,CAAC,EAAE,KAAK,IAAI,CAAC,EAAE,KAAK,EAAE,EAAE,CAAC;QACnE,MAAM,KAAK,GAAY,KAAK,CAAC,UAAU,CAAC,KAAK,CAAC,CAAC;QAC/C,IAAI,CAAC,QAAQ,CAAC,KAAK,CAAC,IAAI,KAAK,CAAC,IAAI,KAAK,kBAAkB;YAAE,SAAS;QACpE,MAAM,UAAU,GAAG,OAAO,KAAK,CAAC,KAAK,KAAK,QAAQ,IAAI,KAAK,CAAC,KAAK,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC,KAAK,CAAC,CAAC,CAAC,SAAS,CAAC;QACvG,MAAM,EAAE,GAAG,UAAU,IAAI,WAAW,CAAC;QACrC,OAAO,EAAE,KAAK,SAAS,CAAC,CAAC,CAAC,EAAE,IAAI,EAAE,cAAc,EAAE,EAAE,EAAE,CAAC,CAAC,CAAC,SAAS,CAAC;IACpE,CAAC;IACD,OAAO,SAAS,CAAC;AAClB,CAAC;AAED,MAAM,UAAU,gCAAgC,CAAC,OAA8B;IAC9E,OAAO,yBAAyB,OAAO,CAAC,IAAI,OAAO,OAAO,CAAC,EAAE,4BAA4B,CAAC;AAC3F,CAAC;AAED;;;;;GAKG;AACH,MAAM,UAAU,wBAAwB,CAAC,OAAyB,EAAE,OAA8B;IACjG,MAAM,WAAW,GAAG,gCAAgC,CAAC,OAAO,CAAC,CAAC;IAC9D,OAAO,CAAC,OAAO,GAAG,EAAE,CAAC;IACrB,OAAO,CAAC,UAAU,GAAG,OAAO,CAAC;IAC7B,OAAO,CAAC,WAAW,GAAG,EAAE,IAAI,EAAE,SAAS,EAAE,WAAW,EAAE,CAAC;IACvD,OAAO,CAAC,YAAY,GAAG,WAAW,CAAC;IACnC,gCAAgC,CAAC,OAAO,EAAE;QACzC,IAAI,EAAE,kCAAkC;QACxC,SAAS,EAAE,IAAI,CAAC,GAAG,EAAE;QACrB,OAAO,EAAE,EAAE,IAAI,EAAE,OAAO,CAAC,IAAI,EAAE,EAAE,EAAE,OAAO,CAAC,EAAE,EAAE;KAC/C,CAAC,CAAC;IACH,gCAAgC,CAAC,OAAO,EAAE;QACzC,IAAI,EAAE,6BAA6B;QACnC,SAAS,EAAE,IAAI,CAAC,GAAG,EAAE;QACrB,OAAO,EAAE,EAAE,MAAM,EAAE,wDAAwD,EAAE;KAC7E,CAAC,CAAC;AACJ,CAAC","sourcesContent":["import type { AssistantMessage } from \"../types.ts\";\nimport { appendAssistantMessageDiagnostic } from \"./diagnostics.ts\";\n\nexport const SERVER_FALLBACK_ABORTED_DIAGNOSTIC = \"server_fallback_aborted\";\nexport const BILLING_INCOMPLETE_DIAGNOSTIC = \"billing_incomplete_after_client_abort\";\n\nexport interface ServerFallbackReceipt {\n\treadonly from: string;\n\treadonly to: string;\n}\n\nfunction isRecord(value: unknown): value is Record<string, unknown> {\n\treturn typeof value === \"object\" && value !== null && !Array.isArray(value);\n}\n\nfunction readModel(value: unknown): string | undefined {\n\tif (!isRecord(value)) return undefined;\n\treturn typeof value.model === \"string\" && value.model.length > 0 ? value.model : undefined;\n}\n\n/**\n * Anthropic marks a classifier handoff with `{type:\"fallback\", from:{model}, to:{model}}`\n * (`server-side-fallback-*` betas). Returns undefined for anything else, so an\n * unrelated or malformed provider-native block falls through to normal handling.\n */\nexport function parseServerFallbackReceipt(block: unknown): ServerFallbackReceipt | undefined {\n\tif (!isRecord(block) || block.type !== \"fallback\") return undefined;\n\tconst from = readModel(block.from);\n\tconst to = readModel(block.to);\n\treturn from !== undefined && to !== undefined ? { from, to } : undefined;\n}\n\n/**\n * Sticky routing serves later turns of a fallen-back conversation from the\n * substitute model with no `fallback` block at all; a `fallback_message` entry\n * in `usage.iterations` is the documented signal. A served-model string\n * comparison is deliberately NOT used: gateways and Bedrock-style endpoints\n * rewrite model ids, so a mismatch is not evidence of a fallback.\n */\nexport function parseStickyFallbackReceipt(\n\tusage: unknown,\n\trequestedModel: string,\n\tservedModel?: string,\n): ServerFallbackReceipt | undefined {\n\tif (!isRecord(usage) || !Array.isArray(usage.iterations)) return undefined;\n\tfor (let index = usage.iterations.length - 1; index >= 0; index--) {\n\t\tconst entry: unknown = usage.iterations[index];\n\t\tif (!isRecord(entry) || entry.type !== \"fallback_message\") continue;\n\t\tconst entryModel = typeof entry.model === \"string\" && entry.model.length > 0 ? entry.model : undefined;\n\t\tconst to = entryModel ?? servedModel;\n\t\treturn to !== undefined ? { from: requestedModel, to } : undefined;\n\t}\n\treturn undefined;\n}\n\nexport function serverFallbackRefusalExplanation(receipt: ServerFallbackReceipt): string {\n\treturn `Server-side fallback (${receipt.from} -> ${receipt.to}) aborted by client policy`;\n}\n\n/**\n * Rewrites an aborted turn into the same shape a pre-output classifier refusal\n * has, so `isClassifierRefusal()` routes it through the caller's fallback chain.\n * Content is dropped: the substitute model's partial output was never requested,\n * and replaying a half-turn containing a fallback marker breaks Anthropic replay.\n */\nexport function applyServerFallbackAbort(message: AssistantMessage, receipt: ServerFallbackReceipt): void {\n\tconst explanation = serverFallbackRefusalExplanation(receipt);\n\tmessage.content = [];\n\tmessage.stopReason = \"error\";\n\tmessage.stopDetails = { type: \"refusal\", explanation };\n\tmessage.errorMessage = explanation;\n\tappendAssistantMessageDiagnostic(message, {\n\t\ttype: SERVER_FALLBACK_ABORTED_DIAGNOSTIC,\n\t\ttimestamp: Date.now(),\n\t\tdetails: { from: receipt.from, to: receipt.to },\n\t});\n\tappendAssistantMessageDiagnostic(message, {\n\t\ttype: BILLING_INCOMPLETE_DIAGNOSTIC,\n\t\ttimestamp: Date.now(),\n\t\tdetails: { reason: \"per-attempt usage does not arrive after a client abort\" },\n\t});\n}\n"]}
+{"version":3,"file":"server-fallback-receipt.js","sourceRoot":"","sources":["../../src/utils/server-fallback-receipt.ts"],"names":[],"mappings":"AACA,OAAO,EAAE,gCAAgC,EAAE,MAAM,kBAAkB,CAAC;AAEpE,MAAM,CAAC,MAAM,kCAAkC,GAAG,yBAAyB,CAAC;AAC5E,MAAM,CAAC,MAAM,6BAA6B,GAAG,uCAAuC,CAAC;AAOrF,SAAS,QAAQ,CAAC,KAAc;IAC/B,OAAO,OAAO,KAAK,KAAK,QAAQ,IAAI,KAAK,KAAK,IAAI,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC,CAAC;AAC7E,CAAC;AAED,SAAS,SAAS,CAAC,KAAc;IAChC,IAAI,CAAC,QAAQ,CAAC,KAAK,CAAC;QAAE,OAAO,SAAS,CAAC;IACvC,OAAO,OAAO,KAAK,CAAC,KAAK,KAAK,QAAQ,IAAI,KAAK,CAAC,KAAK,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC,KAAK,CAAC,CAAC,CAAC,SAAS,CAAC;AAC5F,CAAC;AAED;;;;GAIG;AACH,MAAM,UAAU,0BAA0B,CAAC,KAAc;IACxD,IAAI,CAAC,QAAQ,CAAC,KAAK,CAAC,IAAI,KAAK,CAAC,IAAI,KAAK,UAAU;QAAE,OAAO,SAAS,CAAC;IACpE,MAAM,IAAI,GAAG,SAAS,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC;IACnC,MAAM,EAAE,GAAG,SAAS,CAAC,KAAK,CAAC,EAAE,CAAC,CAAC;IAC/B,OAAO,IAAI,KAAK,SAAS,IAAI,EAAE,KAAK,SAAS,CAAC,CAAC,CAAC,EAAE,IAAI,EAAE,EAAE,EAAE,CAAC,CAAC,CAAC,SAAS,CAAC;AAC1E,CAAC;AAED;;;;;;GAMG;AACH,MAAM,UAAU,0BAA0B,CACzC,KAAc,EACd,cAAsB,EACtB,WAAoB;IAEpB,IAAI,CAAC,QAAQ,CAAC,KAAK,CAAC,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC,UAAU,CAAC;QAAE,OAAO,SAAS,CAAC;IAC3E,KAAK,IAAI,KAAK,GAAG,KAAK,CAAC,UAAU,CAAC,MAAM,GAAG,CAAC,EAAE,KAAK,IAAI,CAAC,EAAE,KAAK,EAAE,EAAE,CAAC;QACnE,MAAM,KAAK,GAAY,KAAK,CAAC,UAAU,CAAC,KAAK,CAAC,CAAC;QAC/C,IAAI,CAAC,QAAQ,CAAC,KAAK,CAAC,IAAI,KAAK,CAAC,IAAI,KAAK,kBAAkB;YAAE,SAAS;QACpE,MAAM,UAAU,GAAG,OAAO,KAAK,CAAC,KAAK,KAAK,QAAQ,IAAI,KAAK,CAAC,KAAK,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC,KAAK,CAAC,CAAC,CAAC,SAAS,CAAC;QACvG,MAAM,EAAE,GAAG,UAAU,IAAI,WAAW,CAAC;QACrC,OAAO,EAAE,KAAK,SAAS,CAAC,CAAC,CAAC,EAAE,IAAI,EAAE,cAAc,EAAE,EAAE,EAAE,CAAC,CAAC,CAAC,SAAS,CAAC;IACpE,CAAC;IACD,OAAO,SAAS,CAAC;AAClB,CAAC;AAED,MAAM,UAAU,gCAAgC,CAAC,OAA8B;IAC9E,OAAO,yBAAyB,OAAO,CAAC,IAAI,OAAO,OAAO,CAAC,EAAE,4BAA4B,CAAC;AAC3F,CAAC;AAED,MAAM,UAAU,+BAA+B,CAC9C,OAAyB,EACzB,KAAkC,EAClC,OAA8B;IAE9B,2EAA2E;IAC3E,8EAA8E;IAC9E,KAAK,MAAM,CAAC,KAAK,EAAE,KAAK,CAAC,IAAI,OAAO,CAAC,OAAO,CAAC,OAAO,EAAE,EAAE,CAAC;QACxD,IAAI,KAAK,CAAC,IAAI,KAAK,UAAU,EAAE,CAAC;YAC/B,OAAO,CAAC,OAAO,CAAC,KAAK,CAAC,GAAG,EAAE,IAAI,EAAE,gBAAgB,EAAE,OAAO,EAAE,qBAAqB,EAAE,GAAG,EAAE,KAAK,EAAE,CAAC;QACjG,CAAC;IACF,CAAC;IACD,OAAO,CAAC,KAAK,GAAG,OAAO,CAAC,EAAE,CAAC;IAC3B,MAAM,QAAQ,GAAG,KAAK,CAAC,MAAM,EAAE,qBAAqB,EAAE,IAAI,CACzD,CAAC,SAAS,EAAE,EAAE,CACb,OAAO,SAAS,KAAK,QAAQ,IAAI,SAAS,CAAC,QAAQ,KAAK,KAAK,CAAC,QAAQ,IAAI,SAAS,CAAC,KAAK,KAAK,OAAO,CAAC,EAAE,CACzG,CAAC;IACF,OAAO;QACN,GAAG,KAAK;QACR,EAAE,EAAE,OAAO,CAAC,EAAE;QACd,IAAI,EAAE,QAAQ,IAAI,OAAO,QAAQ,KAAK,QAAQ,CAAC,CAAC,CAAC,QAAQ,CAAC,IAAI,CAAC,CAAC,CAAC,KAAK,CAAC,IAAI;KAC3E,CAAC;AACH,CAAC;AAED;;;;;GAKG;AACH,MAAM,UAAU,wBAAwB,CAAC,OAAyB,EAAE,OAA8B;IACjG,MAAM,WAAW,GAAG,gCAAgC,CAAC,OAAO,CAAC,CAAC;IAC9D,OAAO,CAAC,OAAO,GAAG,EAAE,CAAC;IACrB,OAAO,CAAC,UAAU,GAAG,OAAO,CAAC;IAC7B,OAAO,CAAC,WAAW,GAAG,EAAE,IAAI,EAAE,SAAS,EAAE,WAAW,EAAE,CAAC;IACvD,OAAO,CAAC,YAAY,GAAG,WAAW,CAAC;IACnC,gCAAgC,CAAC,OAAO,EAAE;QACzC,IAAI,EAAE,kCAAkC;QACxC,SAAS,EAAE,IAAI,CAAC,GAAG,EAAE;QACrB,OAAO,EAAE,EAAE,IAAI,EAAE,OAAO,CAAC,IAAI,EAAE,EAAE,EAAE,OAAO,CAAC,EAAE,EAAE;KAC/C,CAAC,CAAC;IACH,gCAAgC,CAAC,OAAO,EAAE;QACzC,IAAI,EAAE,6BAA6B;QACnC,SAAS,EAAE,IAAI,CAAC,GAAG,EAAE;QACrB,OAAO,EAAE,EAAE,MAAM,EAAE,wDAAwD,EAAE;KAC7E,CAAC,CAAC;AACJ,CAAC","sourcesContent":["import type { AssistantMessage, Model } from \"../types.ts\";\nimport { appendAssistantMessageDiagnostic } from \"./diagnostics.ts\";\n\nexport const SERVER_FALLBACK_ABORTED_DIAGNOSTIC = \"server_fallback_aborted\";\nexport const BILLING_INCOMPLETE_DIAGNOSTIC = \"billing_incomplete_after_client_abort\";\n\nexport interface ServerFallbackReceipt {\n\treadonly from: string;\n\treadonly to: string;\n}\n\nfunction isRecord(value: unknown): value is Record<string, unknown> {\n\treturn typeof value === \"object\" && value !== null && !Array.isArray(value);\n}\n\nfunction readModel(value: unknown): string | undefined {\n\tif (!isRecord(value)) return undefined;\n\treturn typeof value.model === \"string\" && value.model.length > 0 ? value.model : undefined;\n}\n\n/**\n * Anthropic marks a classifier handoff with `{type:\"fallback\", from:{model}, to:{model}}`\n * (`server-side-fallback-*` betas). Returns undefined for anything else, so an\n * unrelated or malformed provider-native block falls through to normal handling.\n */\nexport function parseServerFallbackReceipt(block: unknown): ServerFallbackReceipt | undefined {\n\tif (!isRecord(block) || block.type !== \"fallback\") return undefined;\n\tconst from = readModel(block.from);\n\tconst to = readModel(block.to);\n\treturn from !== undefined && to !== undefined ? { from, to } : undefined;\n}\n\n/**\n * Sticky routing serves later turns of a fallen-back conversation from the\n * substitute model with no `fallback` block at all; a `fallback_message` entry\n * in `usage.iterations` is the documented signal. A served-model string\n * comparison is deliberately NOT used: gateways and Bedrock-style endpoints\n * rewrite model ids, so a mismatch is not evidence of a fallback.\n */\nexport function parseStickyFallbackReceipt(\n\tusage: unknown,\n\trequestedModel: string,\n\tservedModel?: string,\n): ServerFallbackReceipt | undefined {\n\tif (!isRecord(usage) || !Array.isArray(usage.iterations)) return undefined;\n\tfor (let index = usage.iterations.length - 1; index >= 0; index--) {\n\t\tconst entry: unknown = usage.iterations[index];\n\t\tif (!isRecord(entry) || entry.type !== \"fallback_message\") continue;\n\t\tconst entryModel = typeof entry.model === \"string\" && entry.model.length > 0 ? entry.model : undefined;\n\t\tconst to = entryModel ?? servedModel;\n\t\treturn to !== undefined ? { from: requestedModel, to } : undefined;\n\t}\n\treturn undefined;\n}\n\nexport function serverFallbackRefusalExplanation(receipt: ServerFallbackReceipt): string {\n\treturn `Server-side fallback (${receipt.from} -> ${receipt.to}) aborted by client policy`;\n}\n\nexport function applyServerFallbackContinuation(\n\tmessage: AssistantMessage,\n\tmodel: Model<\"anthropic-messages\">,\n\treceipt: ServerFallbackReceipt,\n): Model<\"anthropic-messages\"> {\n\t// Keep source indices stable for stream projections while making abandoned\n\t// client calls audit-only before the agent can execute the completed message.\n\tfor (const [index, block] of message.content.entries()) {\n\t\tif (block.type === \"toolCall\") {\n\t\t\tmessage.content[index] = { type: \"providerNative\", subtype: \"discarded_tool_call\", raw: block };\n\t\t}\n\t}\n\tmessage.model = receipt.to;\n\tconst fallback = model.compat?.allowedFallbackModels?.find(\n\t\t(candidate) =>\n\t\t\ttypeof candidate !== \"string\" && candidate.provider === model.provider && candidate.model === receipt.to,\n\t);\n\treturn {\n\t\t...model,\n\t\tid: receipt.to,\n\t\tcost: fallback && typeof fallback !== \"string\" ? fallback.cost : model.cost,\n\t};\n}\n\n/**\n * Rewrites an aborted turn into the same shape a pre-output classifier refusal\n * has, so `isClassifierRefusal()` routes it through the caller's fallback chain.\n * Content is dropped: the substitute model's partial output was never requested,\n * and replaying a half-turn containing a fallback marker breaks Anthropic replay.\n */\nexport function applyServerFallbackAbort(message: AssistantMessage, receipt: ServerFallbackReceipt): void {\n\tconst explanation = serverFallbackRefusalExplanation(receipt);\n\tmessage.content = [];\n\tmessage.stopReason = \"error\";\n\tmessage.stopDetails = { type: \"refusal\", explanation };\n\tmessage.errorMessage = explanation;\n\tappendAssistantMessageDiagnostic(message, {\n\t\ttype: SERVER_FALLBACK_ABORTED_DIAGNOSTIC,\n\t\ttimestamp: Date.now(),\n\t\tdetails: { from: receipt.from, to: receipt.to },\n\t});\n\tappendAssistantMessageDiagnostic(message, {\n\t\ttype: BILLING_INCOMPLETE_DIAGNOSTIC,\n\t\ttimestamp: Date.now(),\n\t\tdetails: { reason: \"per-attempt usage does not arrive after a client abort\" },\n\t});\n}\n"]}

package/node_modules/@earendil-works/pi-ai/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@earendil-works/pi-ai",
-	"version": "2026.9.7",
+	"version": "2026.9.8",
 	"description": "Unified LLM API with automatic model discovery and provider configuration",
 	"type": "module",
 	"main": "./dist/index.js",
@@ -76,7 +76,7 @@
 	"dependencies": {
 		"@anthropic-ai/sdk": "0.123.0",
 		"@aws-sdk/client-bedrock-runtime": "3.1116.0",
-		"@earendil-works/pi-telemetry": "npm:@code-yeongyu/senpi-telemetry@2026.9.7",
+		"@earendil-works/pi-telemetry": "npm:@code-yeongyu/senpi-telemetry@2026.9.8",
 		"@google/genai": "2.18.0",
 		"@smithy/node-http-handler": "4.11.3",
 		"http-proxy-agent": "9.1.0",

package/node_modules/@earendil-works/pi-pty/package.json

@@ -1,7 +1,7 @@
 {
 	"name": "@earendil-works/pi-pty",
 	"private": true,
-	"version": "2026.9.7",
+	"version": "2026.9.8",
 	"description": "Typed PTY session loader for Senpi persistent terminals",
 	"type": "module",
 	"main": "./dist/index.js",

package/node_modules/@earendil-works/pi-telemetry/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@earendil-works/pi-telemetry",
-	"version": "2026.9.7",
+	"version": "2026.9.8",
 	"description": "Vendor-neutral telemetry contracts and typed schema utilities for pi",
 	"type": "module",
 	"main": "./dist/index.js",

package/node_modules/@earendil-works/pi-tui/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@earendil-works/pi-tui",
-	"version": "2026.9.7",
+	"version": "2026.9.8",
 	"description": "Terminal User Interface library with differential rendering for efficient text-based applications",
 	"type": "module",
 	"main": "dist/index.js",

package/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@code-yeongyu/senpi",
-	"version": "2026.9.7",
+	"version": "2026.9.8",
 	"description": "Coding agent CLI with read, bash, edit, write tools and session management",
 	"type": "module",
 	"piConfig": {
@@ -48,9 +48,9 @@
 		"qa:app-server": "node scripts/qa-app-server/run-all.mjs"
 	},
 	"dependencies": {
-		"@earendil-works/pi-agent-core": "npm:@code-yeongyu/senpi-agent-core@2026.9.7",
-		"@earendil-works/pi-ai": "npm:@code-yeongyu/senpi-ai@2026.9.7",
-		"@earendil-works/pi-tui": "npm:@code-yeongyu/senpi-tui@2026.9.7",
+		"@earendil-works/pi-agent-core": "npm:@code-yeongyu/senpi-agent-core@2026.9.8",
+		"@earendil-works/pi-ai": "npm:@code-yeongyu/senpi-ai@2026.9.8",
+		"@earendil-works/pi-tui": "npm:@code-yeongyu/senpi-tui@2026.9.8",
 		"@silvia-odwyer/photon-node": "0.3.4",
 		"chalk": "6.0.0",
 		"cross-spawn": "7.0.6",
@@ -70,8 +70,8 @@
 		"@anthropic-ai/sdk": "0.123.0",
 		"@aws-sdk/client-bedrock-runtime": "3.1116.0",
 		"@bufbuild/protobuf": "2.14.0",
-		"@code-yeongyu/senpi-codemode": "2026.9.7",
-		"@earendil-works/pi-pty": "npm:@code-yeongyu/senpi-pty@2026.9.7",
+		"@code-yeongyu/senpi-codemode": "2026.9.8",
+		"@earendil-works/pi-pty": "npm:@code-yeongyu/senpi-pty@2026.9.8",
 		"@modelcontextprotocol/sdk": "1.30.0",
 		"@mozilla/readability": "0.6.0",
 		"@opentelemetry/api": "1.9.1",
@@ -119,7 +119,7 @@
 		"@csstools/css-parser-algorithms": "4.0.0",
 		"@csstools/css-syntax-patches-for-csstree": "1.1.8",
 		"@csstools/css-tokenizer": "4.0.0",
-		"@earendil-works/pi-telemetry": "npm:@code-yeongyu/senpi-telemetry@2026.9.7",
+		"@earendil-works/pi-telemetry": "npm:@code-yeongyu/senpi-telemetry@2026.9.8",
 		"@exodus/bytes": "1.15.1",
 		"@google/genai": "2.18.0",
 		"@hono/node-server": "2.1.1",