@earendil-works/pi-ai 1.3.0 → 1.5.0

package/dist/utils/openrouter-cost.d.ts

@@ -0,0 +1,24 @@
+import type { Usage } from "../types.ts";
+export interface OpenRouterUsageAccounting {
+    /** Amount charged to the user's OpenRouter account. */
+    cost: number | null;
+    /** Whether OpenRouter used the user's upstream provider key. */
+    is_byok: boolean;
+    cost_details: {
+        /** Total inference cost billed through the user's upstream provider account, when available. */
+        upstream_inference_cost: number | null;
+        /** Upstream cost for prompt processing, including any provider-side cache pricing. */
+        upstream_inference_prompt_cost: number;
+        /** Upstream cost for completion generation. */
+        upstream_inference_completions_cost: number;
+        /** Metered server-tool cost already included in OpenRouter's charged `cost`. */
+        server_tool_cost?: number | null;
+    } | null;
+}
+export declare function isOpenRouterModel(model: {
+    provider: string;
+    baseUrl: string;
+}): boolean;
+/** Replace the catalog total with OpenRouter's billed cost when the response contains complete, valid accounting. */
+export declare function applyOpenRouterReportedCost(usage: Usage, reported: OpenRouterUsageAccounting): void;
+//# sourceMappingURL=openrouter-cost.d.ts.map

package/dist/utils/openrouter-cost.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"openrouter-cost.d.ts","sourceRoot":"","sources":["../../src/utils/openrouter-cost.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,KAAK,EAAE,MAAM,aAAa,CAAC;AAEzC,MAAM,WAAW,yBAAyB;IACzC,uDAAuD;IACvD,IAAI,EAAE,MAAM,GAAG,IAAI,CAAC;IACpB,gEAAgE;IAChE,OAAO,EAAE,OAAO,CAAC;IACjB,YAAY,EAAE;QACb,gGAAgG;QAChG,uBAAuB,EAAE,MAAM,GAAG,IAAI,CAAC;QACvC,sFAAsF;QACtF,8BAA8B,EAAE,MAAM,CAAC;QACvC,+CAA+C;QAC/C,mCAAmC,EAAE,MAAM,CAAC;QAC5C,gFAAgF;QAChF,gBAAgB,CAAC,EAAE,MAAM,GAAG,IAAI,CAAC;KACjC,GAAG,IAAI,CAAC;CACT;AAMD,wBAAgB,iBAAiB,CAAC,KAAK,EAAE;IAAE,QAAQ,EAAE,MAAM,CAAC;IAAC,OAAO,EAAE,MAAM,CAAA;CAAE,GAAG,OAAO,CAEvF;AAED,qHAAqH;AACrH,wBAAgB,2BAA2B,CAAC,KAAK,EAAE,KAAK,EAAE,QAAQ,EAAE,yBAAyB,GAAG,IAAI,CA0BnG"}

package/dist/utils/openrouter-cost.js

@@ -0,0 +1,34 @@
+function isValidCost(value) {
+    return typeof value === "number" && Number.isFinite(value) && value >= 0;
+}
+export function isOpenRouterModel(model) {
+    return model.provider === "openrouter" || model.baseUrl.includes("openrouter.ai");
+}
+/** Replace the catalog total with OpenRouter's billed cost when the response contains complete, valid accounting. */
+export function applyOpenRouterReportedCost(usage, reported) {
+    if (!isValidCost(reported.cost))
+        return;
+    if (reported.is_byok === false) {
+        usage.cost.total = reported.cost;
+        usage.cost.totalFromProvider = true;
+        return;
+    }
+    if (reported.is_byok === true) {
+        const details = reported.cost_details;
+        if (!details)
+            return;
+        let upstreamCost = details.upstream_inference_cost;
+        if (upstreamCost === null) {
+            if (!isValidCost(details.upstream_inference_prompt_cost) ||
+                !isValidCost(details.upstream_inference_completions_cost)) {
+                return;
+            }
+            upstreamCost = details.upstream_inference_prompt_cost + details.upstream_inference_completions_cost;
+        }
+        if (!isValidCost(upstreamCost))
+            return;
+        usage.cost.total = reported.cost + upstreamCost;
+        usage.cost.totalFromProvider = true;
+    }
+}
+//# sourceMappingURL=openrouter-cost.js.map

package/dist/utils/openrouter-cost.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"openrouter-cost.js","sourceRoot":"","sources":["../../src/utils/openrouter-cost.ts"],"names":[],"mappings":"AAmBA,SAAS,WAAW,CAAC,KAAc;IAClC,OAAO,OAAO,KAAK,KAAK,QAAQ,IAAI,MAAM,CAAC,QAAQ,CAAC,KAAK,CAAC,IAAI,KAAK,IAAI,CAAC,CAAC;AAC1E,CAAC;AAED,MAAM,UAAU,iBAAiB,CAAC,KAA4C;IAC7E,OAAO,KAAK,CAAC,QAAQ,KAAK,YAAY,IAAI,KAAK,CAAC,OAAO,CAAC,QAAQ,CAAC,eAAe,CAAC,CAAC;AACnF,CAAC;AAED,qHAAqH;AACrH,MAAM,UAAU,2BAA2B,CAAC,KAAY,EAAE,QAAmC;IAC5F,IAAI,CAAC,WAAW,CAAC,QAAQ,CAAC,IAAI,CAAC;QAAE,OAAO;IAExC,IAAI,QAAQ,CAAC,OAAO,KAAK,KAAK,EAAE,CAAC;QAChC,KAAK,CAAC,IAAI,CAAC,KAAK,GAAG,QAAQ,CAAC,IAAI,CAAC;QACjC,KAAK,CAAC,IAAI,CAAC,iBAAiB,GAAG,IAAI,CAAC;QACpC,OAAO;IACR,CAAC;IAED,IAAI,QAAQ,CAAC,OAAO,KAAK,IAAI,EAAE,CAAC;QAC/B,MAAM,OAAO,GAAG,QAAQ,CAAC,YAAY,CAAC;QACtC,IAAI,CAAC,OAAO;YAAE,OAAO;QACrB,IAAI,YAAY,GAAY,OAAO,CAAC,uBAAuB,CAAC;QAC5D,IAAI,YAAY,KAAK,IAAI,EAAE,CAAC;YAC3B,IACC,CAAC,WAAW,CAAC,OAAO,CAAC,8BAA8B,CAAC;gBACpD,CAAC,WAAW,CAAC,OAAO,CAAC,mCAAmC,CAAC,EACxD,CAAC;gBACF,OAAO;YACR,CAAC;YACD,YAAY,GAAG,OAAO,CAAC,8BAA8B,GAAG,OAAO,CAAC,mCAAmC,CAAC;QACrG,CAAC;QACD,IAAI,CAAC,WAAW,CAAC,YAAY,CAAC;YAAE,OAAO;QACvC,KAAK,CAAC,IAAI,CAAC,KAAK,GAAG,QAAQ,CAAC,IAAI,GAAG,YAAY,CAAC;QAChD,KAAK,CAAC,IAAI,CAAC,iBAAiB,GAAG,IAAI,CAAC;IACrC,CAAC;AACF,CAAC","sourcesContent":["import type { Usage } from \"../types.ts\";\n\nexport interface OpenRouterUsageAccounting {\n\t/** Amount charged to the user's OpenRouter account. */\n\tcost: number | null;\n\t/** Whether OpenRouter used the user's upstream provider key. */\n\tis_byok: boolean;\n\tcost_details: {\n\t\t/** Total inference cost billed through the user's upstream provider account, when available. */\n\t\tupstream_inference_cost: number | null;\n\t\t/** Upstream cost for prompt processing, including any provider-side cache pricing. */\n\t\tupstream_inference_prompt_cost: number;\n\t\t/** Upstream cost for completion generation. */\n\t\tupstream_inference_completions_cost: number;\n\t\t/** Metered server-tool cost already included in OpenRouter's charged `cost`. */\n\t\tserver_tool_cost?: number | null;\n\t} | null;\n}\n\nfunction isValidCost(value: unknown): value is number {\n\treturn typeof value === \"number\" && Number.isFinite(value) && value >= 0;\n}\n\nexport function isOpenRouterModel(model: { provider: string; baseUrl: string }): boolean {\n\treturn model.provider === \"openrouter\" || model.baseUrl.includes(\"openrouter.ai\");\n}\n\n/** Replace the catalog total with OpenRouter's billed cost when the response contains complete, valid accounting. */\nexport function applyOpenRouterReportedCost(usage: Usage, reported: OpenRouterUsageAccounting): void {\n\tif (!isValidCost(reported.cost)) return;\n\n\tif (reported.is_byok === false) {\n\t\tusage.cost.total = reported.cost;\n\t\tusage.cost.totalFromProvider = true;\n\t\treturn;\n\t}\n\n\tif (reported.is_byok === true) {\n\t\tconst details = reported.cost_details;\n\t\tif (!details) return;\n\t\tlet upstreamCost: unknown = details.upstream_inference_cost;\n\t\tif (upstreamCost === null) {\n\t\t\tif (\n\t\t\t\t!isValidCost(details.upstream_inference_prompt_cost) ||\n\t\t\t\t!isValidCost(details.upstream_inference_completions_cost)\n\t\t\t) {\n\t\t\t\treturn;\n\t\t\t}\n\t\t\tupstreamCost = details.upstream_inference_prompt_cost + details.upstream_inference_completions_cost;\n\t\t}\n\t\tif (!isValidCost(upstreamCost)) return;\n\t\tusage.cost.total = reported.cost + upstreamCost;\n\t\tusage.cost.totalFromProvider = true;\n\t}\n}\n"]}

package/dist/utils/overflow.d.ts

@@ -18,6 +18,7 @@ import type { AssistantMessage } from "../types.ts";
  * - Google Gemini: "input token count exceeds the maximum"
  * - xAI (Grok): "maximum prompt length is X but request contains Y"
  * - Groq: "reduce the length of the messages"
+ * - GitHub Copilot: HTTP 413 "failed to parse request" (request byte-size overflow)
  * - Cerebras: 400/413 status code (no body)
  * - Mistral: "Prompt contains X tokens ... too large for model with Y maximum context length"
  * - OpenRouter (most backends): "maximum context length is X tokens"

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;AAiFpD;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAsDG;AACH,wBAAgB,iBAAiB,CAAC,OAAO,EAAE,gBAAgB,EAAE,aAAa,CAAC,EAAE,MAAM,GAAG,OAAO,CAkC5F;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"}
+{"version":3,"file":"overflow.d.ts","sourceRoot":"","sources":["../../src/utils/overflow.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,gBAAgB,EAAE,MAAM,aAAa,CAAC;AAuFpD;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAuDG;AACH,wBAAgB,iBAAiB,CAAC,OAAO,EAAE,gBAAgB,EAAE,aAAa,CAAC,EAAE,MAAM,GAAG,OAAO,CAkC5F;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"}

package/dist/utils/overflow.js

@@ -2,7 +2,7 @@
  * Regex patterns to detect context overflow errors from different providers.
  *
  * These patterns match error messages returned when the input exceeds
- * the model's context window.
+ * the model's context window or the provider's request byte-size limit.
  *
  * Provider-specific patterns (with example error messages):
  *
@@ -10,6 +10,7 @@
  * - Anthropic: "413 {\"error\":{\"type\":\"request_too_large\",\"message\":\"Request exceeds the maximum size\"}}"
  * - OpenAI: "Your input exceeds the context window of this model"
  * - OpenAI/LiteLLM: "Requested token count exceeds the model's maximum context length of 131072 tokens"
+ * - LiteLLM pre-call check: "litellm.ContextWindowExceededError: ... Context Window exceeded for given call."
  * - 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"
@@ -20,6 +21,7 @@
  * - llama.cpp: "the request exceeds the available context size, try increasing it"
  * - LM Studio: "tokens to keep from the initial prompt is greater than the context length"
  * - GitHub Copilot: "prompt token count of X exceeds the limit of Y"
+ * - GitHub Copilot: 'github-copilot API error (413): {"message":"failed to parse request","code":""}'
  * - MiniMax: "invalid params, context window exceeds limit"
  * - Kimi For Coding: "Your request exceeded model token limit: X (requested: Y)"
  * - DS4: "Prompt has X tokens, but the configured context size is Y tokens"
@@ -39,6 +41,7 @@ const OVERFLOW_PATTERNS = [
     /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)
+    /context window exceeded/i, // LiteLLM pre-call check (ContextWindowExceededError)
     /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
@@ -58,6 +61,9 @@ const OVERFLOW_PATTERNS = [
     /context[_ ]length[_ ]exceeded/i, // Generic fallback
     /too many tokens/i, // Generic fallback
     /token limit exceeded/i, // Generic fallback
+    // HTTP 413 means request-size overflow regardless of provider or body text.
+    // Match raw SDK errors and formatProviderError's prefix, not a status mentioned in the body.
+    /^(?:[\w -]+ API error \(413\):|413(?:\s|:|$))/i,
 ];
 const CEREBRAS_BODYLESS_OVERFLOW_PATTERN = /^4(?:00|13)\s*(?:status code)?\s*\(no body\)/i;
 /**
@@ -93,6 +99,7 @@ const NON_OVERFLOW_PATTERNS = [
  * - Google Gemini: "input token count exceeds the maximum"
  * - xAI (Grok): "maximum prompt length is X but request contains Y"
  * - Groq: "reduce the length of the messages"
+ * - GitHub Copilot: HTTP 413 "failed to parse request" (request byte-size overflow)
  * - Cerebras: 400/413 status code (no body)
  * - Mistral: "Prompt contains X tokens ... too large for model with Y maximum context length"
  * - OpenRouter (most backends): "maximum context length is X tokens"

package/dist/utils/overflow.js.map

@@ -1 +1 @@
-{"version":3,"file":"overflow.js","sourceRoot":"","sources":["../../src/utils/overflow.ts"],"names":[],"mappings":"AAEA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAiCG;AACH,MAAM,iBAAiB,GAAG;IACzB,0BAA0B,EAAE,oCAAoC;IAChE,4BAA4B,EAAE,kCAAkC;IAChE,oBAAoB,EAAE,kDAAkD;IACxE,wCAAwC,EAAE,iBAAiB;IAC3D,6BAA6B,EAAE,uCAAuC;IACtE,4FAA4F,EAAE,sCAAsC;IACpI,yCAAyC,EAAE,kBAAkB;IAC7D,+BAA+B,EAAE,aAAa;IAC9C,oCAAoC,EAAE,OAAO;IAC7C,uCAAuC,EAAE,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;CAC5C,CAAC;AAEF,MAAM,kCAAkC,GAAG,+CAA+C,CAAC;AAE3F;;;;;;;;GAQG;AACH,MAAM,qBAAqB,GAAG;IAC7B,2CAA2C,EAAE,oFAAoF;IACjI,aAAa,EAAE,wBAAwB;IACvC,oBAAoB,EAAE,yBAAyB;CAC/C,CAAC;AAEF;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAsDG;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,EAAE,CAAC;YACpB,IAAI,iBAAiB,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC,OAAO,CAAC,YAAa,CAAC,CAAC,EAAE,CAAC;gBAClE,OAAO,IAAI,CAAC;YACb,CAAC;YACD,IAAI,OAAO,CAAC,QAAQ,KAAK,UAAU,IAAI,kCAAkC,CAAC,IAAI,CAAC,OAAO,CAAC,YAAY,CAAC,EAAE,CAAC;gBACtG,OAAO,IAAI,CAAC;YACb,CAAC;QACF,CAAC;IACF,CAAC;IAED,8EAA8E;IAC9E,IAAI,aAAa,IAAI,OAAO,CAAC,UAAU,KAAK,MAAM,EAAE,CAAC;QACpD,MAAM,WAAW,GAAG,OAAO,CAAC,KAAK,CAAC,KAAK,GAAG,OAAO,CAAC,KAAK,CAAC,SAAS,CAAC;QAClE,IAAI,WAAW,GAAG,aAAa,EAAE,CAAC;YACjC,OAAO,IAAI,CAAC;QACb,CAAC;IACF,CAAC;IAED,sFAAsF;IACtF,qFAAqF;IACrF,gEAAgE;IAChE,IAAI,aAAa,IAAI,OAAO,CAAC,UAAU,KAAK,QAAQ,IAAI,OAAO,CAAC,KAAK,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;QACpF,MAAM,WAAW,GAAG,OAAO,CAAC,KAAK,CAAC,KAAK,GAAG,OAAO,CAAC,KAAK,CAAC,SAAS,CAAC;QAClE,IAAI,WAAW,IAAI,aAAa,GAAG,IAAI,EAAE,CAAC;YACzC,OAAO,IAAI,CAAC;QACb,CAAC;IACF,CAAC;IAED,OAAO,KAAK,CAAC;AACd,CAAC;AAED;;;;;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","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: `{\"code\":\"1261\",\"message\":\"Prompt too long\"}`, `{\"code\":\"1261\",\"message\":\"Prompt exceeds max length\"}` (CN endpoint), or silent overflow via usage.input > contextWindow\n * - Xiaomi MiMo: Truncates input to fill contextWindow exactly, then returns finish_reason \"length\"\n *   with output=0 (no room left to generate). Detected via stopReason \"length\" + zero output +\n *   input filling the context window.\n * - 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 and z.ai token overflow\n\t/prompt exceeds max length/i, // z.ai CN endpoint token overflow\n\t/request_too_large/i, // Anthropic request byte-size overflow (HTTP 413)\n\t/input is too long for requested model/i, // Amazon Bedrock\n\t/exceeds the context window/i, // OpenAI (Completions & Responses API)\n\t/exceeds (?:the )?(?:model'?s )?maximum context length(?: of [\\d,]+ tokens?|\\s*\\([\\d,]+\\))/i, // OpenAI-compatible proxies (LiteLLM)\n\t/input token count.*exceeds the maximum/i, // Google (Gemini)\n\t/maximum prompt length is \\d+/i, // xAI (Grok)\n\t/reduce the length of the messages/i, // Groq\n\t/maximum context length is \\d+ tokens/i, // OpenRouter (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];\n\nconst CEREBRAS_BODYLESS_OVERFLOW_PATTERN = /^4(?:00|13)\\s*(?:status code)?\\s*\\(no body\\)/i;\n\n/**\n * Patterns that indicate non-overflow errors (e.g. rate limiting, server errors).\n * Error messages matching any of these are excluded from overflow detection\n * even if they also match an OVERFLOW_PATTERN.\n *\n * Example: Bedrock formats throttling errors as \"ThrottlingException: Too many tokens,\n * please wait before trying again.\" which would match the /too many tokens/i overflow\n * pattern without this exclusion.\n */\nconst NON_OVERFLOW_PATTERNS = [\n\t/^(Throttling error|Service unavailable):/i, // AWS Bedrock non-overflow errors (human-readable prefixes from formatBedrockError)\n\t/rate limit/i, // Generic rate limiting\n\t/too many requests/i, // Generic HTTP 429 style\n];\n\n/**\n * Check if an assistant message represents a context overflow error.\n *\n * This handles 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 * - z.ai: \"Prompt too long\"\n *\n * **Unreliable detection:**\n * - z.ai: Sometimes accepts overflow silently (detectable via usage.input > contextWindow),\n *   sometimes returns rate limit errors. Pass contextWindow param to detect silent overflow.\n * - Xiaomi MiMo: Truncates input to fit contextWindow then returns stopReason \"length\" with\n *   output=0. Pass contextWindow param to detect via the \"filled context + zero output\" signal.\n * - Ollama: May truncate input silently for some setups, but may also return explicit\n *   overflow errors that match the patterns above. Silent truncation still cannot be\n *   detected here because we do not know the expected token count.\n *\n * ## Custom Providers\n *\n * If you've added custom models via settings.json, this function may not detect\n * overflow errors from those providers. To add support:\n *\n * 1. Send a request that exceeds the model's context window\n * 2. Check the errorMessage in the response\n * 3. Create a regex pattern that matches the error\n * 4. The pattern should be added to OVERFLOW_PATTERNS in this file, or\n *    check the errorMessage yourself before calling this function\n *\n * @param message - The assistant message to check\n * @param contextWindow - Optional context window size for detecting silent overflow (z.ai)\n * @returns true if the message indicates a context overflow\n */\nexport function isContextOverflow(message: AssistantMessage, contextWindow?: number): boolean {\n\t// Case 1: Check error message patterns\n\tif (message.stopReason === \"error\" && message.errorMessage) {\n\t\t// Skip messages matching known non-overflow patterns (e.g. throttling / rate-limit)\n\t\tconst isNonOverflow = NON_OVERFLOW_PATTERNS.some((p) => p.test(message.errorMessage!));\n\t\tif (!isNonOverflow) {\n\t\t\tif (OVERFLOW_PATTERNS.some((p) => p.test(message.errorMessage!))) {\n\t\t\t\treturn true;\n\t\t\t}\n\t\t\tif (message.provider === \"cerebras\" && CEREBRAS_BODYLESS_OVERFLOW_PATTERN.test(message.errorMessage)) {\n\t\t\t\treturn true;\n\t\t\t}\n\t\t}\n\t}\n\n\t// Case 2: Silent overflow (z.ai style) - successful but usage exceeds context\n\tif (contextWindow && message.stopReason === \"stop\") {\n\t\tconst inputTokens = message.usage.input + message.usage.cacheRead;\n\t\tif (inputTokens > contextWindow) {\n\t\t\treturn true;\n\t\t}\n\t}\n\n\t// Case 3: Length-stop overflow (Xiaomi MiMo style) - server truncates oversized input\n\t// to fit the context window, leaving no room for output. Returns stopReason \"length\"\n\t// with output=0 and input+cacheRead filling the context window.\n\tif (contextWindow && message.stopReason === \"length\" && message.usage.output === 0) {\n\t\tconst inputTokens = message.usage.input + message.usage.cacheRead;\n\t\tif (inputTokens >= contextWindow * 0.99) {\n\t\t\treturn true;\n\t\t}\n\t}\n\n\treturn false;\n}\n\n/**\n * Check whether a length stop ended below the caller or model's intended output limit.\n * Such responses may be caused by context pressure or provider-side truncation, so callers\n * can make one bounded compact-and-retry attempt. `desiredMaxOutput` must be the original\n * limit before any context-based clamping.\n */\nexport function isRecoverableLength(message: AssistantMessage, desiredMaxOutput: number): boolean {\n\treturn message.stopReason === \"length\" && desiredMaxOutput > 0 && message.usage.output < desiredMaxOutput;\n}\n\n/**\n * Get the overflow patterns for testing purposes.\n */\nexport function getOverflowPatterns(): RegExp[] {\n\treturn [...OVERFLOW_PATTERNS];\n}\n"]}
+{"version":3,"file":"overflow.js","sourceRoot":"","sources":["../../src/utils/overflow.ts"],"names":[],"mappings":"AAEA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAmCG;AACH,MAAM,iBAAiB,GAAG;IACzB,0BAA0B,EAAE,oCAAoC;IAChE,4BAA4B,EAAE,kCAAkC;IAChE,oBAAoB,EAAE,kDAAkD;IACxE,wCAAwC,EAAE,iBAAiB;IAC3D,6BAA6B,EAAE,uCAAuC;IACtE,4FAA4F,EAAE,sCAAsC;IACpI,0BAA0B,EAAE,sDAAsD;IAClF,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,4EAA4E;IAC5E,6FAA6F;IAC7F,gDAAgD;CAChD,CAAC;AAEF,MAAM,kCAAkC,GAAG,+CAA+C,CAAC;AAE3F;;;;;;;;GAQG;AACH,MAAM,qBAAqB,GAAG;IAC7B,2CAA2C,EAAE,oFAAoF;IACjI,aAAa,EAAE,wBAAwB;IACvC,oBAAoB,EAAE,yBAAyB;CAC/C,CAAC;AAEF;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAuDG;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,EAAE,CAAC;YACpB,IAAI,iBAAiB,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC,OAAO,CAAC,YAAa,CAAC,CAAC,EAAE,CAAC;gBAClE,OAAO,IAAI,CAAC;YACb,CAAC;YACD,IAAI,OAAO,CAAC,QAAQ,KAAK,UAAU,IAAI,kCAAkC,CAAC,IAAI,CAAC,OAAO,CAAC,YAAY,CAAC,EAAE,CAAC;gBACtG,OAAO,IAAI,CAAC;YACb,CAAC;QACF,CAAC;IACF,CAAC;IAED,8EAA8E;IAC9E,IAAI,aAAa,IAAI,OAAO,CAAC,UAAU,KAAK,MAAM,EAAE,CAAC;QACpD,MAAM,WAAW,GAAG,OAAO,CAAC,KAAK,CAAC,KAAK,GAAG,OAAO,CAAC,KAAK,CAAC,SAAS,CAAC;QAClE,IAAI,WAAW,GAAG,aAAa,EAAE,CAAC;YACjC,OAAO,IAAI,CAAC;QACb,CAAC;IACF,CAAC;IAED,sFAAsF;IACtF,qFAAqF;IACrF,gEAAgE;IAChE,IAAI,aAAa,IAAI,OAAO,CAAC,UAAU,KAAK,QAAQ,IAAI,OAAO,CAAC,KAAK,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;QACpF,MAAM,WAAW,GAAG,OAAO,CAAC,KAAK,CAAC,KAAK,GAAG,OAAO,CAAC,KAAK,CAAC,SAAS,CAAC;QAClE,IAAI,WAAW,IAAI,aAAa,GAAG,IAAI,EAAE,CAAC;YACzC,OAAO,IAAI,CAAC;QACb,CAAC;IACF,CAAC;IAED,OAAO,KAAK,CAAC;AACd,CAAC;AAED;;;;;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","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 or the provider's request byte-size limit.\n *\n * Provider-specific patterns (with example error messages):\n *\n * - Anthropic: \"prompt is too long: 213462 tokens > 200000 maximum\"\n * - Anthropic: \"413 {\\\"error\\\":{\\\"type\\\":\\\"request_too_large\\\",\\\"message\\\":\\\"Request exceeds the maximum size\\\"}}\"\n * - OpenAI: \"Your input exceeds the context window of this model\"\n * - OpenAI/LiteLLM: \"Requested token count exceeds the model's maximum context length of 131072 tokens\"\n * - LiteLLM pre-call check: \"litellm.ContextWindowExceededError: ... Context Window exceeded for given call.\"\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 * - GitHub Copilot: 'github-copilot API error (413): {\"message\":\"failed to parse request\",\"code\":\"\"}'\n * - MiniMax: \"invalid params, context window exceeds limit\"\n * - Kimi For Coding: \"Your request exceeded model token limit: X (requested: Y)\"\n * - DS4: \"Prompt has X tokens, but the configured context size is Y tokens\"\n * - Cerebras: \"400/413 status code (no body)\"\n * - Mistral: \"Prompt contains X tokens ... too large for model with Y maximum context length\"\n * - z.ai: `{\"code\":\"1261\",\"message\":\"Prompt too long\"}`, `{\"code\":\"1261\",\"message\":\"Prompt exceeds max length\"}` (CN endpoint), or silent overflow via usage.input > contextWindow\n * - Xiaomi MiMo: Truncates input to fill contextWindow exactly, then returns finish_reason \"length\"\n *   with output=0 (no room left to generate). Detected via stopReason \"length\" + zero output +\n *   input filling the context window.\n * - 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 and z.ai token overflow\n\t/prompt exceeds max length/i, // z.ai CN endpoint token overflow\n\t/request_too_large/i, // Anthropic request byte-size overflow (HTTP 413)\n\t/input is too long for requested model/i, // Amazon Bedrock\n\t/exceeds the context window/i, // OpenAI (Completions & Responses API)\n\t/exceeds (?:the )?(?:model'?s )?maximum context length(?: of [\\d,]+ tokens?|\\s*\\([\\d,]+\\))/i, // OpenAI-compatible proxies (LiteLLM)\n\t/context window exceeded/i, // LiteLLM pre-call check (ContextWindowExceededError)\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// HTTP 413 means request-size overflow regardless of provider or body text.\n\t// Match raw SDK errors and formatProviderError's prefix, not a status mentioned in the body.\n\t/^(?:[\\w -]+ API error \\(413\\):|413(?:\\s|:|$))/i,\n];\n\nconst CEREBRAS_BODYLESS_OVERFLOW_PATTERN = /^4(?:00|13)\\s*(?:status code)?\\s*\\(no body\\)/i;\n\n/**\n * Patterns that indicate non-overflow errors (e.g. rate limiting, server errors).\n * Error messages matching any of these are excluded from overflow detection\n * even if they also match an OVERFLOW_PATTERN.\n *\n * Example: Bedrock formats throttling errors as \"ThrottlingException: Too many tokens,\n * please wait before trying again.\" which would match the /too many tokens/i overflow\n * pattern without this exclusion.\n */\nconst NON_OVERFLOW_PATTERNS = [\n\t/^(Throttling error|Service unavailable):/i, // AWS Bedrock non-overflow errors (human-readable prefixes from formatBedrockError)\n\t/rate limit/i, // Generic rate limiting\n\t/too many requests/i, // Generic HTTP 429 style\n];\n\n/**\n * Check if an assistant message represents a context overflow error.\n *\n * This handles 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 * - GitHub Copilot: HTTP 413 \"failed to parse request\" (request byte-size overflow)\n * - Cerebras: 400/413 status code (no body)\n * - Mistral: \"Prompt contains X tokens ... too large for model with Y maximum context length\"\n * - OpenRouter (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 * - z.ai: \"Prompt too long\"\n *\n * **Unreliable detection:**\n * - z.ai: Sometimes accepts overflow silently (detectable via usage.input > contextWindow),\n *   sometimes returns rate limit errors. Pass contextWindow param to detect silent overflow.\n * - Xiaomi MiMo: Truncates input to fit contextWindow then returns stopReason \"length\" with\n *   output=0. Pass contextWindow param to detect via the \"filled context + zero output\" signal.\n * - Ollama: May truncate input silently for some setups, but may also return explicit\n *   overflow errors that match the patterns above. Silent truncation still cannot be\n *   detected here because we do not know the expected token count.\n *\n * ## Custom Providers\n *\n * If you've added custom models via settings.json, this function may not detect\n * overflow errors from those providers. To add support:\n *\n * 1. Send a request that exceeds the model's context window\n * 2. Check the errorMessage in the response\n * 3. Create a regex pattern that matches the error\n * 4. The pattern should be added to OVERFLOW_PATTERNS in this file, or\n *    check the errorMessage yourself before calling this function\n *\n * @param message - The assistant message to check\n * @param contextWindow - Optional context window size for detecting silent overflow (z.ai)\n * @returns true if the message indicates a context overflow\n */\nexport function isContextOverflow(message: AssistantMessage, contextWindow?: number): boolean {\n\t// Case 1: Check error message patterns\n\tif (message.stopReason === \"error\" && message.errorMessage) {\n\t\t// Skip messages matching known non-overflow patterns (e.g. throttling / rate-limit)\n\t\tconst isNonOverflow = NON_OVERFLOW_PATTERNS.some((p) => p.test(message.errorMessage!));\n\t\tif (!isNonOverflow) {\n\t\t\tif (OVERFLOW_PATTERNS.some((p) => p.test(message.errorMessage!))) {\n\t\t\t\treturn true;\n\t\t\t}\n\t\t\tif (message.provider === \"cerebras\" && CEREBRAS_BODYLESS_OVERFLOW_PATTERN.test(message.errorMessage)) {\n\t\t\t\treturn true;\n\t\t\t}\n\t\t}\n\t}\n\n\t// Case 2: Silent overflow (z.ai style) - successful but usage exceeds context\n\tif (contextWindow && message.stopReason === \"stop\") {\n\t\tconst inputTokens = message.usage.input + message.usage.cacheRead;\n\t\tif (inputTokens > contextWindow) {\n\t\t\treturn true;\n\t\t}\n\t}\n\n\t// Case 3: Length-stop overflow (Xiaomi MiMo style) - server truncates oversized input\n\t// to fit the context window, leaving no room for output. Returns stopReason \"length\"\n\t// with output=0 and input+cacheRead filling the context window.\n\tif (contextWindow && message.stopReason === \"length\" && message.usage.output === 0) {\n\t\tconst inputTokens = message.usage.input + message.usage.cacheRead;\n\t\tif (inputTokens >= contextWindow * 0.99) {\n\t\t\treturn true;\n\t\t}\n\t}\n\n\treturn false;\n}\n\n/**\n * Check whether a length stop ended below the caller or model's intended output limit.\n * Such responses may be caused by context pressure or provider-side truncation, so callers\n * can make one bounded compact-and-retry attempt. `desiredMaxOutput` must be the original\n * limit before any context-based clamping.\n */\nexport function isRecoverableLength(message: AssistantMessage, desiredMaxOutput: number): boolean {\n\treturn message.stopReason === \"length\" && desiredMaxOutput > 0 && message.usage.output < desiredMaxOutput;\n}\n\n/**\n * Get the overflow patterns for testing purposes.\n */\nexport function getOverflowPatterns(): RegExp[] {\n\treturn [...OVERFLOW_PATTERNS];\n}\n"]}

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

@@ -1 +1 @@
-{"version":3,"file":"retry.d.ts","sourceRoot":"","sources":["../../src/utils/retry.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,gBAAgB,EAAE,MAAM,aAAa,CAAC;AAiHpD;;;;;;GAMG;AACH,MAAM,WAAW,WAAW;IAC3B,OAAO,EAAE,OAAO,CAAC;IACjB,qFAAqF;IACrF,UAAU,EAAE,MAAM,CAAC;IACnB,0FAA0F;IAC1F,WAAW,EAAE,MAAM,CAAC;IACpB,+EAA+E;IAC/E,eAAe,CAAC,EAAE,MAAM,CAAC;CACzB;AAED,eAAO,MAAM,gCAAgC,QAAS,CAAC;AAEvD,wBAAgB,YAAY,CAAC,MAAM,EAAE,IAAI,CAAC,WAAW,EAAE,aAAa,GAAG,iBAAiB,CAAC,EAAE,OAAO,EAAE,MAAM,GAAG,MAAM,CAIlH;AAED,kFAAkF;AAClF,MAAM,WAAW,cAAc;IAC9B,0EAA0E;IAC1E,gBAAgB,CAAC,EAAE,CAClB,OAAO,EAAE,MAAM,EACf,WAAW,EAAE,MAAM,EACnB,OAAO,EAAE,MAAM,EACf,YAAY,EAAE,MAAM,KAChB,IAAI,GAAG,OAAO,CAAC,IAAI,CAAC,CAAC;IAC1B,mFAAmF;IACnF,mBAAmB,CAAC,EAAE,MAAM,IAAI,GAAG,OAAO,CAAC,IAAI,CAAC,CAAC;IACjD,mFAAmF;IACnF,eAAe,CAAC,EAAE,CAAC,OAAO,EAAE,OAAO,EAAE,OAAO,EAAE,MAAM,EAAE,UAAU,CAAC,EAAE,MAAM,KAAK,IAAI,GAAG,OAAO,CAAC,IAAI,CAAC,CAAC;CACnG;AA0BD;;;;;;;;;;;;;;;;;GAiBG;AACH,wBAAsB,kBAAkB,CACvC,OAAO,EAAE,MAAM,OAAO,CAAC,gBAAgB,CAAC,EACxC,MAAM,EAAE,WAAW,GAAG,SAAS,EAC/B,MAAM,EAAE,WAAW,GAAG,SAAS,EAC/B,SAAS,CAAC,EAAE,cAAc,GACxB,OAAO,CAAC,gBAAgB,CAAC,CA6C3B;AAED;;;;;;;;GAQG;AACH,wBAAgB,yBAAyB,CAAC,OAAO,EAAE,gBAAgB,GAAG,OAAO,CAK5E"}
+{"version":3,"file":"retry.d.ts","sourceRoot":"","sources":["../../src/utils/retry.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,gBAAgB,EAAE,MAAM,aAAa,CAAC;AA8HpD;;;;;;GAMG;AACH,MAAM,WAAW,WAAW;IAC3B,OAAO,EAAE,OAAO,CAAC;IACjB,qFAAqF;IACrF,UAAU,EAAE,MAAM,CAAC;IACnB,0FAA0F;IAC1F,WAAW,EAAE,MAAM,CAAC;IACpB,+EAA+E;IAC/E,eAAe,CAAC,EAAE,MAAM,CAAC;CACzB;AAED,eAAO,MAAM,gCAAgC,QAAS,CAAC;AAEvD,wBAAgB,YAAY,CAAC,MAAM,EAAE,IAAI,CAAC,WAAW,EAAE,aAAa,GAAG,iBAAiB,CAAC,EAAE,OAAO,EAAE,MAAM,GAAG,MAAM,CAIlH;AAED,kFAAkF;AAClF,MAAM,WAAW,cAAc;IAC9B,0EAA0E;IAC1E,gBAAgB,CAAC,EAAE,CAClB,OAAO,EAAE,MAAM,EACf,WAAW,EAAE,MAAM,EACnB,OAAO,EAAE,MAAM,EACf,YAAY,EAAE,MAAM,KAChB,IAAI,GAAG,OAAO,CAAC,IAAI,CAAC,CAAC;IAC1B,mFAAmF;IACnF,mBAAmB,CAAC,EAAE,MAAM,IAAI,GAAG,OAAO,CAAC,IAAI,CAAC,CAAC;IACjD,mFAAmF;IACnF,eAAe,CAAC,EAAE,CAAC,OAAO,EAAE,OAAO,EAAE,OAAO,EAAE,MAAM,EAAE,UAAU,CAAC,EAAE,MAAM,KAAK,IAAI,GAAG,OAAO,CAAC,IAAI,CAAC,CAAC;CACnG;AA0BD;;;;;;;;;;;;;;;;;GAiBG;AACH,wBAAsB,kBAAkB,CACvC,OAAO,EAAE,MAAM,OAAO,CAAC,gBAAgB,CAAC,EACxC,MAAM,EAAE,WAAW,GAAG,SAAS,EAC/B,MAAM,EAAE,WAAW,GAAG,SAAS,EAC/B,SAAS,CAAC,EAAE,cAAc,GACxB,OAAO,CAAC,gBAAgB,CAAC,CA6C3B;AAED;;;;;;;;GAQG;AACH,wBAAgB,yBAAyB,CAAC,OAAO,EAAE,gBAAgB,GAAG,OAAO,CAK5E"}

package/dist/utils/retry.js

@@ -19,6 +19,12 @@ const NON_RETRYABLE_PROVIDER_LIMIT_ERROR_PATTERN = buildProviderErrorPattern([
     // Sign in with ChatGPT: the subscription's shared usage limit, which resets
     // after hours rather than seconds.
     "subscription_sharing_usage_limit_exceeded",
+    // Sign in with ChatGPT on a ChatGPT Business plan: a usage limit set by the
+    // workspace admin. Only the admin can raise it.
+    "subscription_sharing_workplace_individual_usage_limit_exceeded",
+    // Zhipu/Z.ai (open.bigmodel.cn) error code 1308: the plan's usage limit is
+    // reached, sent as HTTP 429. It resets after hours, not seconds (#11537).
+    '"code"\\s*:\\s*"?1308(?![0-9])',
 ]);
 const RETRYABLE_PROVIDER_ERROR_PATTERN = buildProviderErrorPattern([
     // Generic provider load, HTTP status, and server-side transient failures.
@@ -29,13 +35,8 @@ const RETRYABLE_PROVIDER_ERROR_PATTERN = buildProviderErrorPattern([
     "model is at capacity",
     "rate.?limit",
     "too many requests",
-    "429",
-    "500",
-    "502",
-    "503",
-    "504",
-    "520",
-    "524",
+    // Status codes must not be part of a longer word or number, e.g. a request id or byte count (#11057).
+    "(?<![a-z0-9])(?:429|500|502|503|504|520|524)(?![a-z0-9])",
     // Also matches "The service is temporarily unavailable" (#10965).
     "service.?(?:is.?)?(?:temporarily.?)?unavailable",
     "server.?error",
@@ -48,6 +49,11 @@ const RETRYABLE_PROVIDER_ERROR_PATTERN = buildProviderErrorPattern([
     // Gateway failure relaying the upstream response stream, sent as an OpenAI
     // Responses `response.failed` event without an error code (#10920).
     "unable to forward upstream",
+    // Other gateway upstream failures relayed as `response.failed` error codes (#10920).
+    "upstream.?request.?failed",
+    "upstream_stream_read_error",
+    "stream was interrupted",
+    "http/?2 stream failed",
     // Network, proxy, and fetch transport failures. This includes OpenAI Codex
     // raw-fetch failures such as "upstream connect", "connection refused", and
     // "reset before headers" (#733), plus OpenRouter connection drops (#3317).
@@ -57,6 +63,9 @@ const RETRYABLE_PROVIDER_ERROR_PATTERN = buildProviderErrorPattern([
     "connection.?lost",
     "other side closed",
     "fetch failed",
+    // node-fetch system errors, e.g. the Google ADC OAuth token refresh used by
+    // google-vertex. The reason is often empty (#10996).
+    "request to \\S+ failed, reason:",
     "getaddrinfo",
     "ENOTFOUND",
     "EAI_AGAIN",
@@ -77,6 +86,9 @@ const RETRYABLE_PROVIDER_ERROR_PATTERN = buildProviderErrorPattern([
     "stream ended before message_stop",
     "stream ended before a terminal response event",
     "http2 request did not get a response",
+    // Codex CLI mid-stream disconnect, relayed by proxies that wrap the Codex CLI,
+    // e.g. "stream disconnected before completion: stream closed before response.completed" (#11057).
+    "stream disconnected before completion",
     // Node ERR_HTTP2_STREAM_CANCEL: the HTTP/2 session died before the request was
     // sent, e.g. after the Bedrock SDK's 5-minute session timeout (#10379).
     "pending stream has been canceled",

package/dist/utils/retry.js.map

@@ -1 +1 @@
-{"version":3,"file":"retry.js","sourceRoot":"","sources":["../../src/utils/retry.ts"],"names":[],"mappings":"AAEA,SAAS,yBAAyB,CAAC,QAA2B;IAC7D,OAAO,IAAI,MAAM,CAAC,QAAQ,CAAC,IAAI,CAAC,GAAG,CAAC,EAAE,GAAG,CAAC,CAAC;AAC5C,CAAC;AAED,MAAM,0CAA0C,GAAG,yBAAyB,CAAC;IAC5E,8EAA8E;IAC9E,2EAA2E;IAC3E,mBAAmB;IACnB,qBAAqB;IAErB,6EAA6E;IAC7E,yDAAyD;IACzD,6BAA6B;IAC7B,mBAAmB;IAEnB,4EAA4E;IAC5E,4EAA4E;IAC5E,oBAAoB;IACpB,eAAe;IACf,gBAAgB;IAChB,SAAS;IAET,4EAA4E;IAC5E,mCAAmC;IACnC,2CAA2C;CAC3C,CAAC,CAAC;AAEH,MAAM,gCAAgC,GAAG,yBAAyB,CAAC;IAClE,0EAA0E;IAC1E,YAAY;IACZ,aAAa;IACb,4BAA4B;IAC5B,oCAAoC;IACpC,sBAAsB;IACtB,aAAa;IACb,mBAAmB;IACnB,KAAK;IACL,KAAK;IACL,KAAK;IACL,KAAK;IACL,KAAK;IACL,KAAK;IACL,KAAK;IACL,kEAAkE;IAClE,iDAAiD;IACjD,eAAe;IACf,iBAAiB;IAEjB,8EAA8E;IAC9E,0EAA0E;IAC1E,sDAAsD;IACtD,oCAAoC;IACpC,uDAAuD;IACvD,2EAA2E;IAC3E,oEAAoE;IACpE,4BAA4B;IAE5B,2EAA2E;IAC3E,2EAA2E;IAC3E,2EAA2E;IAC3E,gBAAgB;IAChB,mBAAmB;IACnB,qBAAqB;IACrB,kBAAkB;IAClB,mBAAmB;IACnB,cAAc;IACd,aAAa;IACb,WAAW;IACX,WAAW;IACX,mBAAmB;IACnB,sBAAsB;IACtB,gBAAgB;IAChB,8BAA8B;IAC9B,YAAY;IACZ,SAAS;IACT,YAAY;IAEZ,+EAA+E;IAC/E,mBAAmB;IACnB,kBAAkB;IAElB,yEAAyE;IACzE,8EAA8E;IAC9E,yEAAyE;IACzE,eAAe;IACf,kCAAkC;IAClC,+CAA+C;IAC/C,sCAAsC;IACtC,+EAA+E;IAC/E,wEAAwE;IACxE,kCAAkC;IAElC,4EAA4E;IAC5E,iEAAiE;IACjE,aAAa;IAEb,6EAA6E;IAC7E,6BAA6B;IAC7B,4BAA4B;IAC5B,wBAAwB;IACxB,2BAA2B;IAE3B,yCAAyC;IACzC,mBAAmB;IAEnB,0EAA0E;IAC1E,qEAAqE;IACrE,wCAAwC;IACxC,uCAAuC;CACvC,CAAC,CAAC;AAmBH,MAAM,CAAC,MAAM,gCAAgC,GAAG,MAAM,CAAC;AAEvD,MAAM,UAAU,YAAY,CAAC,MAA4D,EAAE,OAAe;IACzG,MAAM,KAAK,GAAG,MAAM,CAAC,WAAW,GAAG,CAAC,IAAI,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,OAAO,GAAG,CAAC,CAAC,CAAC;IACjE,MAAM,SAAS,GAAG,MAAM,CAAC,aAAa,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,MAAM,CAAC,gBAAgB,CAAC;IAChF,OAAO,IAAI,CAAC,GAAG,CAAC,SAAS,EAAE,MAAM,CAAC,eAAe,IAAI,gCAAgC,CAAC,CAAC;AACxF,CAAC;AAiBD,MAAM,oBAAqB,SAAQ,KAAK;IACvC;QACC,KAAK,CAAC,SAAS,CAAC,CAAC;IAClB,CAAC;CACD;AAED,SAAS,KAAK,CAAC,EAAU,EAAE,MAAoB;IAC9C,OAAO,IAAI,OAAO,CAAC,CAAC,OAAO,EAAE,MAAM,EAAE,EAAE;QACtC,IAAI,MAAM,EAAE,OAAO,EAAE,CAAC;YACrB,MAAM,CAAC,IAAI,oBAAoB,EAAE,CAAC,CAAC;YACnC,OAAO;QACR,CAAC;QACD,MAAM,OAAO,GAAG,UAAU,CAAC,OAAO,EAAE,EAAE,CAAC,CAAC;QACxC,MAAM,EAAE,gBAAgB,CACvB,OAAO,EACP,GAAG,EAAE;YACJ,YAAY,CAAC,OAAO,CAAC,CAAC;YACtB,MAAM,CAAC,IAAI,oBAAoB,EAAE,CAAC,CAAC;QACpC,CAAC,EACD,EAAE,IAAI,EAAE,IAAI,EAAE,CACd,CAAC;IACH,CAAC,CAAC,CAAC;AACJ,CAAC;AAED;;;;;;;;;;;;;;;;;GAiBG;AACH,MAAM,CAAC,KAAK,UAAU,kBAAkB,CACvC,OAAwC,EACxC,MAA+B,EAC/B,MAA+B,EAC/B,SAA0B;IAE1B,MAAM,WAAW,GAAG,MAAM,EAAE,OAAO,CAAC,CAAC,CAAC,MAAM,CAAC,UAAU,CAAC,CAAC,CAAC,CAAC,CAAC;IAE5D,IAAI,OAAO,GAAG,CAAC,CAAC;IAChB,IAAI,SAAgE,CAAC;IACrE,SAAS,CAAC;QACT,MAAM,QAAQ,GAAG,MAAM,OAAO,EAAE,CAAC;QAEjC,sEAAsE;QACtE,IAAI,QAAQ,CAAC,UAAU,KAAK,SAAS,EAAE,CAAC;YACvC,IAAI,SAAS;gBAAE,MAAM,SAAS,EAAE,eAAe,EAAE,CAAC,KAAK,EAAE,SAAS,CAAC,OAAO,CAAC,CAAC;YAC5E,OAAO,QAAQ,CAAC;QACjB,CAAC;QAED,wDAAwD;QACxD,IAAI,QAAQ,CAAC,UAAU,KAAK,OAAO,EAAE,CAAC;YACrC,IAAI,SAAS;gBAAE,MAAM,SAAS,EAAE,eAAe,EAAE,CAAC,IAAI,EAAE,SAAS,CAAC,OAAO,CAAC,CAAC;YAC3E,OAAO,QAAQ,CAAC;QACjB,CAAC;QAED,sEAAsE;QACtE,IAAI,OAAO,IAAI,WAAW,IAAI,CAAC,yBAAyB,CAAC,QAAQ,CAAC,EAAE,CAAC;YACpE,IAAI,SAAS;gBAAE,MAAM,SAAS,EAAE,eAAe,EAAE,CAAC,KAAK,EAAE,SAAS,CAAC,OAAO,EAAE,QAAQ,CAAC,YAAY,CAAC,CAAC;YACnG,OAAO,QAAQ,CAAC;QACjB,CAAC;QAED,OAAO,EAAE,CAAC;QACV,SAAS,GAAG,EAAE,OAAO,EAAE,YAAY,EAAE,QAAQ,CAAC,YAAY,IAAI,eAAe,EAAE,CAAC;QAChF,MAAM,OAAO,GAAG,YAAY,CAAC,MAAO,EAAE,OAAO,CAAC,CAAC;QAC/C,MAAM,SAAS,EAAE,gBAAgB,EAAE,CAAC,OAAO,EAAE,WAAW,EAAE,OAAO,EAAE,SAAS,CAAC,YAAY,CAAC,CAAC;QAE3F,8EAA8E;QAC9E,qFAAqF;QACrF,IAAI,CAAC;YACJ,MAAM,KAAK,CAAC,OAAO,EAAE,MAAM,CAAC,CAAC;QAC9B,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YAChB,MAAM,SAAS,EAAE,eAAe,EAAE,CAAC,KAAK,EAAE,OAAO,EAAE,SAAS,CAAC,YAAY,CAAC,CAAC;YAC3E,IAAI,KAAK,YAAY,oBAAoB,EAAE,CAAC;gBAC3C,MAAM,EAAE,YAAY,EAAE,aAAa,EAAE,GAAG,IAAI,EAAE,GAAG,QAAQ,CAAC;gBAC1D,OAAO,EAAE,GAAG,IAAI,EAAE,UAAU,EAAE,SAAS,EAAE,CAAC;YAC3C,CAAC;YACD,MAAM,KAAK,CAAC;QACb,CAAC;QACD,MAAM,SAAS,EAAE,mBAAmB,EAAE,EAAE,CAAC;IAC1C,CAAC;AACF,CAAC;AAED;;;;;;;;GAQG;AACH,MAAM,UAAU,yBAAyB,CAAC,OAAyB;IAClE,IAAI,OAAO,CAAC,UAAU,KAAK,OAAO,IAAI,CAAC,OAAO,CAAC,YAAY;QAAE,OAAO,KAAK,CAAC;IAC1E,MAAM,YAAY,GAAG,OAAO,CAAC,YAAY,CAAC;IAC1C,IAAI,0CAA0C,CAAC,IAAI,CAAC,YAAY,CAAC;QAAE,OAAO,KAAK,CAAC;IAChF,OAAO,gCAAgC,CAAC,IAAI,CAAC,YAAY,CAAC,CAAC;AAC5D,CAAC","sourcesContent":["import type { AssistantMessage } from \"../types.ts\";\n\nfunction buildProviderErrorPattern(patterns: readonly string[]): RegExp {\n\treturn new RegExp(patterns.join(\"|\"), \"i\");\n}\n\nconst NON_RETRYABLE_PROVIDER_LIMIT_ERROR_PATTERN = buildProviderErrorPattern([\n\t// OpenCode Go/free-tier limits returned as 429 JSON error types by OpenCode's\n\t// Zen API. These are subscription/account limits, not transient throttles.\n\t\"GoUsageLimitError\",\n\t\"FreeUsageLimitError\",\n\n\t// OpenCode Go subscription-limit text asks users to enable available-balance\n\t// usage after rolling/weekly/monthly limits are reached.\n\t\"Monthly usage limit reached\",\n\t\"available balance\",\n\n\t// Generic quota/budget/billing exhaustion. `insufficient_quota` is OpenAI's\n\t// quota/billing error code; the other strings cover common gateway wording.\n\t\"insufficient_quota\",\n\t\"out of budget\",\n\t\"quota exceeded\",\n\t\"billing\",\n\n\t// Sign in with ChatGPT: the subscription's shared usage limit, which resets\n\t// after hours rather than seconds.\n\t\"subscription_sharing_usage_limit_exceeded\",\n]);\n\nconst RETRYABLE_PROVIDER_ERROR_PATTERN = buildProviderErrorPattern([\n\t// Generic provider load, HTTP status, and server-side transient failures.\n\t\"overloaded\",\n\t\"server_busy\",\n\t\"servers are currently busy\",\n\t\"currently experiencing high demand\",\n\t\"model is at capacity\",\n\t\"rate.?limit\",\n\t\"too many requests\",\n\t\"429\",\n\t\"500\",\n\t\"502\",\n\t\"503\",\n\t\"504\",\n\t\"520\",\n\t\"524\",\n\t// Also matches \"The service is temporarily unavailable\" (#10965).\n\t\"service.?(?:is.?)?(?:temporarily.?)?unavailable\",\n\t\"server.?error\",\n\t\"internal.?error\",\n\n\t// Wrapper/provider text for transient upstream failures, including OpenRouter\n\t// \"Provider returned error\" responses (#2264) and gateway wording such as\n\t// \"The upstream provider returned an error\" (#10842).\n\t\"provider.?returned.?(?:an.?)?error\",\n\t\"exceeded request buffer limit while retrying upstream\",\n\t// Gateway failure relaying the upstream response stream, sent as an OpenAI\n\t// Responses `response.failed` event without an error code (#10920).\n\t\"unable to forward upstream\",\n\n\t// Network, proxy, and fetch transport failures. This includes OpenAI Codex\n\t// raw-fetch failures such as \"upstream connect\", \"connection refused\", and\n\t// \"reset before headers\" (#733), plus OpenRouter connection drops (#3317).\n\t\"network.?error\",\n\t\"connection.?error\",\n\t\"connection.?refused\",\n\t\"connection.?lost\",\n\t\"other side closed\",\n\t\"fetch failed\",\n\t\"getaddrinfo\",\n\t\"ENOTFOUND\",\n\t\"EAI_AGAIN\",\n\t\"upstream.?connect\",\n\t\"reset before headers\",\n\t\"socket hang up\",\n\t\"socket connection was closed\",\n\t\"timed? out\",\n\t\"timeout\",\n\t\"terminated\",\n\n\t// WebSocket transports can report close/error text instead of HTTP/fetch text.\n\t\"websocket.?closed\",\n\t\"websocket.?error\",\n\n\t// Premature stream endings from SDKs and transports. Anthropic can throw\n\t// \"stream ended without ...\" and \"Anthropic stream ended before message_stop\"\n\t// (#4433); Bedrock/Smithy can throw an HTTP/2 no-response error (#3594).\n\t\"ended without\",\n\t\"stream ended before message_stop\",\n\t\"stream ended before a terminal response event\",\n\t\"http2 request did not get a response\",\n\t// Node ERR_HTTP2_STREAM_CANCEL: the HTTP/2 session died before the request was\n\t// sent, e.g. after the Bedrock SDK's 5-minute session timeout (#10379).\n\t\"pending stream has been canceled\",\n\n\t// Provider-requested retry delay cap failures should flow through the outer\n\t// retry policy so callers can surface/abort the backoff (#1123).\n\t\"retry delay\",\n\n\t// Explicit retry guidance emitted mid-stream by OpenAI Responses and Bedrock\n\t// stream exceptions (#6019).\n\t\"you can retry your request\",\n\t\"try your request again\",\n\t\"please retry your request\",\n\n\t// gRPC based providers (e.g. NVIDIA NIM)\n\t\"ResourceExhausted\",\n\n\t// Sign in with ChatGPT: usage or user data temporarily unavailable. Usage\n\t// failures can arrive mid-stream without an HTTP 503 in the message.\n\t\"subscription_sharing_usage_unavailable\",\n\t\"subscription_sharing_user_unavailable\",\n]);\n\n/**\n * Retry policy: bounded attempts with exponential backoff (`baseDelayMs * 2^(attempt-1)`).\n * `maxAgentDelayMs` caps each computed delay and defaults to 60 seconds.\n * Matches `settings.retry` (`enabled`, `maxRetries`, `baseDelayMs`, `maxAgentDelayMs`) in coding-agent; kept\n * here so the classifier and the policy-driven retry loop live together and stay reusable\n * by the SDK and other callers.\n */\nexport interface RetryPolicy {\n\tenabled: boolean;\n\t/** Max retry attempts (0 = no retries). The initial call never counts as a retry. */\n\tmaxRetries: number;\n\t/** Base delay in ms. Per-attempt delay is `baseDelayMs * 2^(attempt-1)` before jitter. */\n\tbaseDelayMs: number;\n\t/** Optional cap for agent-level retry delays in ms. Defaults to 60 seconds. */\n\tmaxAgentDelayMs?: number;\n}\n\nexport const DEFAULT_MAX_AGENT_RETRY_DELAY_MS = 60_000;\n\nexport function retryDelayMs(policy: Pick<RetryPolicy, \"baseDelayMs\" | \"maxAgentDelayMs\">, attempt: number): number {\n\tconst delay = policy.baseDelayMs * 2 ** Math.max(0, attempt - 1);\n\tconst safeDelay = Number.isSafeInteger(delay) ? delay : Number.MAX_SAFE_INTEGER;\n\treturn Math.min(safeDelay, policy.maxAgentDelayMs ?? DEFAULT_MAX_AGENT_RETRY_DELAY_MS);\n}\n\n/** Optional callbacks emitted by {@link retryAssistantCall} around each retry. */\nexport interface RetryCallbacks {\n\t/** Emitted before the backoff sleep of each retry attempt (1-indexed). */\n\tonRetryScheduled?: (\n\t\tattempt: number,\n\t\tmaxAttempts: number,\n\t\tdelayMs: number,\n\t\terrorMessage: string,\n\t) => void | Promise<void>;\n\t/** Emitted after the backoff sleep, immediately before the retried call starts. */\n\tonRetryAttemptStart?: () => void | Promise<void>;\n\t/** Emitted once when the loop ends: success if a later call completed normally. */\n\tonRetryFinished?: (success: boolean, attempt: number, finalError?: string) => void | Promise<void>;\n}\n\nclass RetrySleepAbortError extends Error {\n\tconstructor() {\n\t\tsuper(\"Aborted\");\n\t}\n}\n\nfunction sleep(ms: number, signal?: AbortSignal): Promise<void> {\n\treturn new Promise((resolve, reject) => {\n\t\tif (signal?.aborted) {\n\t\t\treject(new RetrySleepAbortError());\n\t\t\treturn;\n\t\t}\n\t\tconst timeout = setTimeout(resolve, ms);\n\t\tsignal?.addEventListener(\n\t\t\t\"abort\",\n\t\t\t() => {\n\t\t\t\tclearTimeout(timeout);\n\t\t\t\treject(new RetrySleepAbortError());\n\t\t\t},\n\t\t\t{ once: true },\n\t\t);\n\t});\n}\n\n/**\n * Run a single assistant-producing call with bounded retry on transient errors.\n *\n * Behavior:\n * - A successful response is returned immediately. Aborts are terminal and never\n *   retried, but reported as unsuccessful if they happen after a retry was scheduled.\n *   Aborts during the backoff sleep are normalized to an aborted `AssistantMessage`\n *   too, so callers do not need to care when cancellation happened.\n * - A non-retryable error (per {@link isRetryableAssistantError}, including quota/\n *   billing exhaustion) is returned immediately so deterministic errors fail fast.\n * - Otherwise retries up to `maxRetries` times with exponential backoff, emitting\n *   `onRetryScheduled` before each sleep, `onRetryAttemptStart` after each sleep before\n *   the retried call starts, and `onRetryFinished` once at the end (whether the loop\n *   ends in success, exhausted retries, or an aborted backoff).\n *\n * When `policy` is undefined or disabled, the first response is returned unchanged\n * (equivalent to calling `produce()` directly).\n */\nexport async function retryAssistantCall(\n\tproduce: () => Promise<AssistantMessage>,\n\tpolicy: RetryPolicy | undefined,\n\tsignal: AbortSignal | undefined,\n\tcallbacks?: RetryCallbacks,\n): Promise<AssistantMessage> {\n\tconst maxAttempts = policy?.enabled ? policy.maxRetries : 0;\n\n\tlet attempt = 0;\n\tlet lastRetry: { attempt: number; errorMessage: string } | undefined;\n\tfor (;;) {\n\t\tconst response = await produce();\n\n\t\t// Abort: terminal but not successful. Never retry an aborted message.\n\t\tif (response.stopReason === \"aborted\") {\n\t\t\tif (lastRetry) await callbacks?.onRetryFinished?.(false, lastRetry.attempt);\n\t\t\treturn response;\n\t\t}\n\n\t\t// Success: non-error, non-abort responses return as-is.\n\t\tif (response.stopReason !== \"error\") {\n\t\t\tif (lastRetry) await callbacks?.onRetryFinished?.(true, lastRetry.attempt);\n\t\t\treturn response;\n\t\t}\n\n\t\t// Non-retryable, or budget exhausted: return the final error message.\n\t\tif (attempt >= maxAttempts || !isRetryableAssistantError(response)) {\n\t\t\tif (lastRetry) await callbacks?.onRetryFinished?.(false, lastRetry.attempt, response.errorMessage);\n\t\t\treturn response;\n\t\t}\n\n\t\tattempt++;\n\t\tlastRetry = { attempt, errorMessage: response.errorMessage || \"Unknown error\" };\n\t\tconst delayMs = retryDelayMs(policy!, attempt);\n\t\tawait callbacks?.onRetryScheduled?.(attempt, maxAttempts, delayMs, lastRetry.errorMessage);\n\n\t\t// Normalize aborts during retry backoff to the same AssistantMessage shape as\n\t\t// provider stream aborts, so callers do not need to care when cancellation happened.\n\t\ttry {\n\t\t\tawait sleep(delayMs, signal);\n\t\t} catch (error) {\n\t\t\tawait callbacks?.onRetryFinished?.(false, attempt, lastRetry.errorMessage);\n\t\t\tif (error instanceof RetrySleepAbortError) {\n\t\t\t\tconst { errorMessage: _errorMessage, ...rest } = response;\n\t\t\t\treturn { ...rest, stopReason: \"aborted\" };\n\t\t\t}\n\t\t\tthrow error;\n\t\t}\n\t\tawait callbacks?.onRetryAttemptStart?.();\n\t}\n}\n\n/**\n * Classifies whether a failed assistant message looks like a transient provider\n * or transport error, so callers can decide if the last assistant turn should be\n * restarted.\n *\n * This does not implement retry policy. Callers should first handle context\n * overflow separately, then apply their own retry budget, backoff, and reporting\n * before restarting the assistant turn.\n */\nexport function isRetryableAssistantError(message: AssistantMessage): boolean {\n\tif (message.stopReason !== \"error\" || !message.errorMessage) return false;\n\tconst errorMessage = message.errorMessage;\n\tif (NON_RETRYABLE_PROVIDER_LIMIT_ERROR_PATTERN.test(errorMessage)) return false;\n\treturn RETRYABLE_PROVIDER_ERROR_PATTERN.test(errorMessage);\n}\n"]}
+{"version":3,"file":"retry.js","sourceRoot":"","sources":["../../src/utils/retry.ts"],"names":[],"mappings":"AAEA,SAAS,yBAAyB,CAAC,QAA2B;IAC7D,OAAO,IAAI,MAAM,CAAC,QAAQ,CAAC,IAAI,CAAC,GAAG,CAAC,EAAE,GAAG,CAAC,CAAC;AAC5C,CAAC;AAED,MAAM,0CAA0C,GAAG,yBAAyB,CAAC;IAC5E,8EAA8E;IAC9E,2EAA2E;IAC3E,mBAAmB;IACnB,qBAAqB;IAErB,6EAA6E;IAC7E,yDAAyD;IACzD,6BAA6B;IAC7B,mBAAmB;IAEnB,4EAA4E;IAC5E,4EAA4E;IAC5E,oBAAoB;IACpB,eAAe;IACf,gBAAgB;IAChB,SAAS;IAET,4EAA4E;IAC5E,mCAAmC;IACnC,2CAA2C;IAC3C,4EAA4E;IAC5E,gDAAgD;IAChD,gEAAgE;IAEhE,2EAA2E;IAC3E,0EAA0E;IAC1E,gCAAgC;CAChC,CAAC,CAAC;AAEH,MAAM,gCAAgC,GAAG,yBAAyB,CAAC;IAClE,0EAA0E;IAC1E,YAAY;IACZ,aAAa;IACb,4BAA4B;IAC5B,oCAAoC;IACpC,sBAAsB;IACtB,aAAa;IACb,mBAAmB;IACnB,sGAAsG;IACtG,0DAA0D;IAC1D,kEAAkE;IAClE,iDAAiD;IACjD,eAAe;IACf,iBAAiB;IAEjB,8EAA8E;IAC9E,0EAA0E;IAC1E,sDAAsD;IACtD,oCAAoC;IACpC,uDAAuD;IACvD,2EAA2E;IAC3E,oEAAoE;IACpE,4BAA4B;IAC5B,qFAAqF;IACrF,2BAA2B;IAC3B,4BAA4B;IAC5B,wBAAwB;IACxB,uBAAuB;IAEvB,2EAA2E;IAC3E,2EAA2E;IAC3E,2EAA2E;IAC3E,gBAAgB;IAChB,mBAAmB;IACnB,qBAAqB;IACrB,kBAAkB;IAClB,mBAAmB;IACnB,cAAc;IACd,4EAA4E;IAC5E,qDAAqD;IACrD,iCAAiC;IACjC,aAAa;IACb,WAAW;IACX,WAAW;IACX,mBAAmB;IACnB,sBAAsB;IACtB,gBAAgB;IAChB,8BAA8B;IAC9B,YAAY;IACZ,SAAS;IACT,YAAY;IAEZ,+EAA+E;IAC/E,mBAAmB;IACnB,kBAAkB;IAElB,yEAAyE;IACzE,8EAA8E;IAC9E,yEAAyE;IACzE,eAAe;IACf,kCAAkC;IAClC,+CAA+C;IAC/C,sCAAsC;IACtC,+EAA+E;IAC/E,kGAAkG;IAClG,uCAAuC;IACvC,+EAA+E;IAC/E,wEAAwE;IACxE,kCAAkC;IAElC,4EAA4E;IAC5E,iEAAiE;IACjE,aAAa;IAEb,6EAA6E;IAC7E,6BAA6B;IAC7B,4BAA4B;IAC5B,wBAAwB;IACxB,2BAA2B;IAE3B,yCAAyC;IACzC,mBAAmB;IAEnB,0EAA0E;IAC1E,qEAAqE;IACrE,wCAAwC;IACxC,uCAAuC;CACvC,CAAC,CAAC;AAmBH,MAAM,CAAC,MAAM,gCAAgC,GAAG,MAAM,CAAC;AAEvD,MAAM,UAAU,YAAY,CAAC,MAA4D,EAAE,OAAe;IACzG,MAAM,KAAK,GAAG,MAAM,CAAC,WAAW,GAAG,CAAC,IAAI,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,OAAO,GAAG,CAAC,CAAC,CAAC;IACjE,MAAM,SAAS,GAAG,MAAM,CAAC,aAAa,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,MAAM,CAAC,gBAAgB,CAAC;IAChF,OAAO,IAAI,CAAC,GAAG,CAAC,SAAS,EAAE,MAAM,CAAC,eAAe,IAAI,gCAAgC,CAAC,CAAC;AACxF,CAAC;AAiBD,MAAM,oBAAqB,SAAQ,KAAK;IACvC;QACC,KAAK,CAAC,SAAS,CAAC,CAAC;IAClB,CAAC;CACD;AAED,SAAS,KAAK,CAAC,EAAU,EAAE,MAAoB;IAC9C,OAAO,IAAI,OAAO,CAAC,CAAC,OAAO,EAAE,MAAM,EAAE,EAAE;QACtC,IAAI,MAAM,EAAE,OAAO,EAAE,CAAC;YACrB,MAAM,CAAC,IAAI,oBAAoB,EAAE,CAAC,CAAC;YACnC,OAAO;QACR,CAAC;QACD,MAAM,OAAO,GAAG,UAAU,CAAC,OAAO,EAAE,EAAE,CAAC,CAAC;QACxC,MAAM,EAAE,gBAAgB,CACvB,OAAO,EACP,GAAG,EAAE;YACJ,YAAY,CAAC,OAAO,CAAC,CAAC;YACtB,MAAM,CAAC,IAAI,oBAAoB,EAAE,CAAC,CAAC;QACpC,CAAC,EACD,EAAE,IAAI,EAAE,IAAI,EAAE,CACd,CAAC;IACH,CAAC,CAAC,CAAC;AACJ,CAAC;AAED;;;;;;;;;;;;;;;;;GAiBG;AACH,MAAM,CAAC,KAAK,UAAU,kBAAkB,CACvC,OAAwC,EACxC,MAA+B,EAC/B,MAA+B,EAC/B,SAA0B;IAE1B,MAAM,WAAW,GAAG,MAAM,EAAE,OAAO,CAAC,CAAC,CAAC,MAAM,CAAC,UAAU,CAAC,CAAC,CAAC,CAAC,CAAC;IAE5D,IAAI,OAAO,GAAG,CAAC,CAAC;IAChB,IAAI,SAAgE,CAAC;IACrE,SAAS,CAAC;QACT,MAAM,QAAQ,GAAG,MAAM,OAAO,EAAE,CAAC;QAEjC,sEAAsE;QACtE,IAAI,QAAQ,CAAC,UAAU,KAAK,SAAS,EAAE,CAAC;YACvC,IAAI,SAAS;gBAAE,MAAM,SAAS,EAAE,eAAe,EAAE,CAAC,KAAK,EAAE,SAAS,CAAC,OAAO,CAAC,CAAC;YAC5E,OAAO,QAAQ,CAAC;QACjB,CAAC;QAED,wDAAwD;QACxD,IAAI,QAAQ,CAAC,UAAU,KAAK,OAAO,EAAE,CAAC;YACrC,IAAI,SAAS;gBAAE,MAAM,SAAS,EAAE,eAAe,EAAE,CAAC,IAAI,EAAE,SAAS,CAAC,OAAO,CAAC,CAAC;YAC3E,OAAO,QAAQ,CAAC;QACjB,CAAC;QAED,sEAAsE;QACtE,IAAI,OAAO,IAAI,WAAW,IAAI,CAAC,yBAAyB,CAAC,QAAQ,CAAC,EAAE,CAAC;YACpE,IAAI,SAAS;gBAAE,MAAM,SAAS,EAAE,eAAe,EAAE,CAAC,KAAK,EAAE,SAAS,CAAC,OAAO,EAAE,QAAQ,CAAC,YAAY,CAAC,CAAC;YACnG,OAAO,QAAQ,CAAC;QACjB,CAAC;QAED,OAAO,EAAE,CAAC;QACV,SAAS,GAAG,EAAE,OAAO,EAAE,YAAY,EAAE,QAAQ,CAAC,YAAY,IAAI,eAAe,EAAE,CAAC;QAChF,MAAM,OAAO,GAAG,YAAY,CAAC,MAAO,EAAE,OAAO,CAAC,CAAC;QAC/C,MAAM,SAAS,EAAE,gBAAgB,EAAE,CAAC,OAAO,EAAE,WAAW,EAAE,OAAO,EAAE,SAAS,CAAC,YAAY,CAAC,CAAC;QAE3F,8EAA8E;QAC9E,qFAAqF;QACrF,IAAI,CAAC;YACJ,MAAM,KAAK,CAAC,OAAO,EAAE,MAAM,CAAC,CAAC;QAC9B,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YAChB,MAAM,SAAS,EAAE,eAAe,EAAE,CAAC,KAAK,EAAE,OAAO,EAAE,SAAS,CAAC,YAAY,CAAC,CAAC;YAC3E,IAAI,KAAK,YAAY,oBAAoB,EAAE,CAAC;gBAC3C,MAAM,EAAE,YAAY,EAAE,aAAa,EAAE,GAAG,IAAI,EAAE,GAAG,QAAQ,CAAC;gBAC1D,OAAO,EAAE,GAAG,IAAI,EAAE,UAAU,EAAE,SAAS,EAAE,CAAC;YAC3C,CAAC;YACD,MAAM,KAAK,CAAC;QACb,CAAC;QACD,MAAM,SAAS,EAAE,mBAAmB,EAAE,EAAE,CAAC;IAC1C,CAAC;AACF,CAAC;AAED;;;;;;;;GAQG;AACH,MAAM,UAAU,yBAAyB,CAAC,OAAyB;IAClE,IAAI,OAAO,CAAC,UAAU,KAAK,OAAO,IAAI,CAAC,OAAO,CAAC,YAAY;QAAE,OAAO,KAAK,CAAC;IAC1E,MAAM,YAAY,GAAG,OAAO,CAAC,YAAY,CAAC;IAC1C,IAAI,0CAA0C,CAAC,IAAI,CAAC,YAAY,CAAC;QAAE,OAAO,KAAK,CAAC;IAChF,OAAO,gCAAgC,CAAC,IAAI,CAAC,YAAY,CAAC,CAAC;AAC5D,CAAC","sourcesContent":["import type { AssistantMessage } from \"../types.ts\";\n\nfunction buildProviderErrorPattern(patterns: readonly string[]): RegExp {\n\treturn new RegExp(patterns.join(\"|\"), \"i\");\n}\n\nconst NON_RETRYABLE_PROVIDER_LIMIT_ERROR_PATTERN = buildProviderErrorPattern([\n\t// OpenCode Go/free-tier limits returned as 429 JSON error types by OpenCode's\n\t// Zen API. These are subscription/account limits, not transient throttles.\n\t\"GoUsageLimitError\",\n\t\"FreeUsageLimitError\",\n\n\t// OpenCode Go subscription-limit text asks users to enable available-balance\n\t// usage after rolling/weekly/monthly limits are reached.\n\t\"Monthly usage limit reached\",\n\t\"available balance\",\n\n\t// Generic quota/budget/billing exhaustion. `insufficient_quota` is OpenAI's\n\t// quota/billing error code; the other strings cover common gateway wording.\n\t\"insufficient_quota\",\n\t\"out of budget\",\n\t\"quota exceeded\",\n\t\"billing\",\n\n\t// Sign in with ChatGPT: the subscription's shared usage limit, which resets\n\t// after hours rather than seconds.\n\t\"subscription_sharing_usage_limit_exceeded\",\n\t// Sign in with ChatGPT on a ChatGPT Business plan: a usage limit set by the\n\t// workspace admin. Only the admin can raise it.\n\t\"subscription_sharing_workplace_individual_usage_limit_exceeded\",\n\n\t// Zhipu/Z.ai (open.bigmodel.cn) error code 1308: the plan's usage limit is\n\t// reached, sent as HTTP 429. It resets after hours, not seconds (#11537).\n\t'\"code\"\\\\s*:\\\\s*\"?1308(?![0-9])',\n]);\n\nconst RETRYABLE_PROVIDER_ERROR_PATTERN = buildProviderErrorPattern([\n\t// Generic provider load, HTTP status, and server-side transient failures.\n\t\"overloaded\",\n\t\"server_busy\",\n\t\"servers are currently busy\",\n\t\"currently experiencing high demand\",\n\t\"model is at capacity\",\n\t\"rate.?limit\",\n\t\"too many requests\",\n\t// Status codes must not be part of a longer word or number, e.g. a request id or byte count (#11057).\n\t\"(?<![a-z0-9])(?:429|500|502|503|504|520|524)(?![a-z0-9])\",\n\t// Also matches \"The service is temporarily unavailable\" (#10965).\n\t\"service.?(?:is.?)?(?:temporarily.?)?unavailable\",\n\t\"server.?error\",\n\t\"internal.?error\",\n\n\t// Wrapper/provider text for transient upstream failures, including OpenRouter\n\t// \"Provider returned error\" responses (#2264) and gateway wording such as\n\t// \"The upstream provider returned an error\" (#10842).\n\t\"provider.?returned.?(?:an.?)?error\",\n\t\"exceeded request buffer limit while retrying upstream\",\n\t// Gateway failure relaying the upstream response stream, sent as an OpenAI\n\t// Responses `response.failed` event without an error code (#10920).\n\t\"unable to forward upstream\",\n\t// Other gateway upstream failures relayed as `response.failed` error codes (#10920).\n\t\"upstream.?request.?failed\",\n\t\"upstream_stream_read_error\",\n\t\"stream was interrupted\",\n\t\"http/?2 stream failed\",\n\n\t// Network, proxy, and fetch transport failures. This includes OpenAI Codex\n\t// raw-fetch failures such as \"upstream connect\", \"connection refused\", and\n\t// \"reset before headers\" (#733), plus OpenRouter connection drops (#3317).\n\t\"network.?error\",\n\t\"connection.?error\",\n\t\"connection.?refused\",\n\t\"connection.?lost\",\n\t\"other side closed\",\n\t\"fetch failed\",\n\t// node-fetch system errors, e.g. the Google ADC OAuth token refresh used by\n\t// google-vertex. The reason is often empty (#10996).\n\t\"request to \\\\S+ failed, reason:\",\n\t\"getaddrinfo\",\n\t\"ENOTFOUND\",\n\t\"EAI_AGAIN\",\n\t\"upstream.?connect\",\n\t\"reset before headers\",\n\t\"socket hang up\",\n\t\"socket connection was closed\",\n\t\"timed? out\",\n\t\"timeout\",\n\t\"terminated\",\n\n\t// WebSocket transports can report close/error text instead of HTTP/fetch text.\n\t\"websocket.?closed\",\n\t\"websocket.?error\",\n\n\t// Premature stream endings from SDKs and transports. Anthropic can throw\n\t// \"stream ended without ...\" and \"Anthropic stream ended before message_stop\"\n\t// (#4433); Bedrock/Smithy can throw an HTTP/2 no-response error (#3594).\n\t\"ended without\",\n\t\"stream ended before message_stop\",\n\t\"stream ended before a terminal response event\",\n\t\"http2 request did not get a response\",\n\t// Codex CLI mid-stream disconnect, relayed by proxies that wrap the Codex CLI,\n\t// e.g. \"stream disconnected before completion: stream closed before response.completed\" (#11057).\n\t\"stream disconnected before completion\",\n\t// Node ERR_HTTP2_STREAM_CANCEL: the HTTP/2 session died before the request was\n\t// sent, e.g. after the Bedrock SDK's 5-minute session timeout (#10379).\n\t\"pending stream has been canceled\",\n\n\t// Provider-requested retry delay cap failures should flow through the outer\n\t// retry policy so callers can surface/abort the backoff (#1123).\n\t\"retry delay\",\n\n\t// Explicit retry guidance emitted mid-stream by OpenAI Responses and Bedrock\n\t// stream exceptions (#6019).\n\t\"you can retry your request\",\n\t\"try your request again\",\n\t\"please retry your request\",\n\n\t// gRPC based providers (e.g. NVIDIA NIM)\n\t\"ResourceExhausted\",\n\n\t// Sign in with ChatGPT: usage or user data temporarily unavailable. Usage\n\t// failures can arrive mid-stream without an HTTP 503 in the message.\n\t\"subscription_sharing_usage_unavailable\",\n\t\"subscription_sharing_user_unavailable\",\n]);\n\n/**\n * Retry policy: bounded attempts with exponential backoff (`baseDelayMs * 2^(attempt-1)`).\n * `maxAgentDelayMs` caps each computed delay and defaults to 60 seconds.\n * Matches `settings.retry` (`enabled`, `maxRetries`, `baseDelayMs`, `maxAgentDelayMs`) in coding-agent; kept\n * here so the classifier and the policy-driven retry loop live together and stay reusable\n * by the SDK and other callers.\n */\nexport interface RetryPolicy {\n\tenabled: boolean;\n\t/** Max retry attempts (0 = no retries). The initial call never counts as a retry. */\n\tmaxRetries: number;\n\t/** Base delay in ms. Per-attempt delay is `baseDelayMs * 2^(attempt-1)` before jitter. */\n\tbaseDelayMs: number;\n\t/** Optional cap for agent-level retry delays in ms. Defaults to 60 seconds. */\n\tmaxAgentDelayMs?: number;\n}\n\nexport const DEFAULT_MAX_AGENT_RETRY_DELAY_MS = 60_000;\n\nexport function retryDelayMs(policy: Pick<RetryPolicy, \"baseDelayMs\" | \"maxAgentDelayMs\">, attempt: number): number {\n\tconst delay = policy.baseDelayMs * 2 ** Math.max(0, attempt - 1);\n\tconst safeDelay = Number.isSafeInteger(delay) ? delay : Number.MAX_SAFE_INTEGER;\n\treturn Math.min(safeDelay, policy.maxAgentDelayMs ?? DEFAULT_MAX_AGENT_RETRY_DELAY_MS);\n}\n\n/** Optional callbacks emitted by {@link retryAssistantCall} around each retry. */\nexport interface RetryCallbacks {\n\t/** Emitted before the backoff sleep of each retry attempt (1-indexed). */\n\tonRetryScheduled?: (\n\t\tattempt: number,\n\t\tmaxAttempts: number,\n\t\tdelayMs: number,\n\t\terrorMessage: string,\n\t) => void | Promise<void>;\n\t/** Emitted after the backoff sleep, immediately before the retried call starts. */\n\tonRetryAttemptStart?: () => void | Promise<void>;\n\t/** Emitted once when the loop ends: success if a later call completed normally. */\n\tonRetryFinished?: (success: boolean, attempt: number, finalError?: string) => void | Promise<void>;\n}\n\nclass RetrySleepAbortError extends Error {\n\tconstructor() {\n\t\tsuper(\"Aborted\");\n\t}\n}\n\nfunction sleep(ms: number, signal?: AbortSignal): Promise<void> {\n\treturn new Promise((resolve, reject) => {\n\t\tif (signal?.aborted) {\n\t\t\treject(new RetrySleepAbortError());\n\t\t\treturn;\n\t\t}\n\t\tconst timeout = setTimeout(resolve, ms);\n\t\tsignal?.addEventListener(\n\t\t\t\"abort\",\n\t\t\t() => {\n\t\t\t\tclearTimeout(timeout);\n\t\t\t\treject(new RetrySleepAbortError());\n\t\t\t},\n\t\t\t{ once: true },\n\t\t);\n\t});\n}\n\n/**\n * Run a single assistant-producing call with bounded retry on transient errors.\n *\n * Behavior:\n * - A successful response is returned immediately. Aborts are terminal and never\n *   retried, but reported as unsuccessful if they happen after a retry was scheduled.\n *   Aborts during the backoff sleep are normalized to an aborted `AssistantMessage`\n *   too, so callers do not need to care when cancellation happened.\n * - A non-retryable error (per {@link isRetryableAssistantError}, including quota/\n *   billing exhaustion) is returned immediately so deterministic errors fail fast.\n * - Otherwise retries up to `maxRetries` times with exponential backoff, emitting\n *   `onRetryScheduled` before each sleep, `onRetryAttemptStart` after each sleep before\n *   the retried call starts, and `onRetryFinished` once at the end (whether the loop\n *   ends in success, exhausted retries, or an aborted backoff).\n *\n * When `policy` is undefined or disabled, the first response is returned unchanged\n * (equivalent to calling `produce()` directly).\n */\nexport async function retryAssistantCall(\n\tproduce: () => Promise<AssistantMessage>,\n\tpolicy: RetryPolicy | undefined,\n\tsignal: AbortSignal | undefined,\n\tcallbacks?: RetryCallbacks,\n): Promise<AssistantMessage> {\n\tconst maxAttempts = policy?.enabled ? policy.maxRetries : 0;\n\n\tlet attempt = 0;\n\tlet lastRetry: { attempt: number; errorMessage: string } | undefined;\n\tfor (;;) {\n\t\tconst response = await produce();\n\n\t\t// Abort: terminal but not successful. Never retry an aborted message.\n\t\tif (response.stopReason === \"aborted\") {\n\t\t\tif (lastRetry) await callbacks?.onRetryFinished?.(false, lastRetry.attempt);\n\t\t\treturn response;\n\t\t}\n\n\t\t// Success: non-error, non-abort responses return as-is.\n\t\tif (response.stopReason !== \"error\") {\n\t\t\tif (lastRetry) await callbacks?.onRetryFinished?.(true, lastRetry.attempt);\n\t\t\treturn response;\n\t\t}\n\n\t\t// Non-retryable, or budget exhausted: return the final error message.\n\t\tif (attempt >= maxAttempts || !isRetryableAssistantError(response)) {\n\t\t\tif (lastRetry) await callbacks?.onRetryFinished?.(false, lastRetry.attempt, response.errorMessage);\n\t\t\treturn response;\n\t\t}\n\n\t\tattempt++;\n\t\tlastRetry = { attempt, errorMessage: response.errorMessage || \"Unknown error\" };\n\t\tconst delayMs = retryDelayMs(policy!, attempt);\n\t\tawait callbacks?.onRetryScheduled?.(attempt, maxAttempts, delayMs, lastRetry.errorMessage);\n\n\t\t// Normalize aborts during retry backoff to the same AssistantMessage shape as\n\t\t// provider stream aborts, so callers do not need to care when cancellation happened.\n\t\ttry {\n\t\t\tawait sleep(delayMs, signal);\n\t\t} catch (error) {\n\t\t\tawait callbacks?.onRetryFinished?.(false, attempt, lastRetry.errorMessage);\n\t\t\tif (error instanceof RetrySleepAbortError) {\n\t\t\t\tconst { errorMessage: _errorMessage, ...rest } = response;\n\t\t\t\treturn { ...rest, stopReason: \"aborted\" };\n\t\t\t}\n\t\t\tthrow error;\n\t\t}\n\t\tawait callbacks?.onRetryAttemptStart?.();\n\t}\n}\n\n/**\n * Classifies whether a failed assistant message looks like a transient provider\n * or transport error, so callers can decide if the last assistant turn should be\n * restarted.\n *\n * This does not implement retry policy. Callers should first handle context\n * overflow separately, then apply their own retry budget, backoff, and reporting\n * before restarting the assistant turn.\n */\nexport function isRetryableAssistantError(message: AssistantMessage): boolean {\n\tif (message.stopReason !== \"error\" || !message.errorMessage) return false;\n\tconst errorMessage = message.errorMessage;\n\tif (NON_RETRYABLE_PROVIDER_LIMIT_ERROR_PATTERN.test(errorMessage)) return false;\n\treturn RETRYABLE_PROVIDER_ERROR_PATTERN.test(errorMessage);\n}\n"]}

package/dist/utils/streaming-json.d.ts

@@ -0,0 +1,84 @@
+/**
+ * Incremental, tolerant parsing of JSON text that arrives in pieces, such as streamed tool-call arguments.
+ *
+ * `StreamingJsonParser.append()` reads each piece once, so a stream costs time and memory linear in its length, and
+ * `snapshot()` returns the value of the text read so far. Snapshots are plain JSON values and never change once
+ * returned: the parser copies the containers that are still open (the "open path") into every new snapshot and shares
+ * completed containers and strings. The rules for incomplete text follow `parseStreamingJson()`:
+ *
+ * - An unterminated string shows the characters read so far; an unterminated escape is held back until complete.
+ * - An unterminated container shows the members and elements read so far.
+ * - A member whose key or value has not started is omitted.
+ * - An unterminated literal shows its value once it is a prefix of `true`, `false` or `null`; an unterminated number
+ *   shows its value when the digits read so far form a number (an exponent without digits is ignored).
+ *
+ * Malformed text is read leniently, the way `repairJson()` and `parseStreamingJson()` treat it: raw control characters
+ * in strings are kept, an invalid escape such as `\d` stays literal text, missing commas and colons are tolerated,
+ * values in invalid positions and unknown words are dropped, and text after the root value is ignored. The parser never
+ * throws. Callers that need the exact value of the finished text parse it once, e.g. with `parseToolCallArguments()`.
+ *
+ * Differences from `parseStreamingJson()` on incomplete text: trailing whitespace inside an unterminated string is
+ * kept (it was trimmed), a string that contains a raw control character or an invalid escape stays visible while it
+ * streams (it was dropped until the document was complete), and `NaN` and `Infinity` words are dropped.
+ *
+ * `append()` also returns the changes the piece made, for a writer that keeps its own copy of the value and applies
+ * them instead of re-assigning snapshots: `set` when a key, an array item or a finished value appears (a new string is
+ * set once with its first text, a container empty), `append` when the piece extended an open string. Unterminated
+ * numbers and literals are not reported until they end. A container nested more than 32 levels deep is reported as one
+ * `set` of its whole value when it closes, so paths stay short.
+ *
+ * Cost: copying the open path costs its member and element count. Each `append()` earns copy credit proportional to its
+ * length plus a constant; `snapshot()` spends it and returns the previous snapshot while the open path costs more than
+ * the credit. So a document that keeps a very large container open (thousands of elements) shows a preview that lags
+ * a little behind the text, and no input makes the total work grow faster than the text.
+ */
+import type { JsonValue, ToolCall } from "../types.ts";
+/** A path from the root value: member keys and array indices. */
+export type JsonPath = readonly (string | number)[];
+/** One change a piece of text made to the value; see the module comment. */
+export type JsonChange = {
+    readonly kind: "set";
+    readonly path: JsonPath;
+    readonly value: JsonValue;
+} | {
+    readonly kind: "append";
+    readonly path: JsonPath;
+    readonly text: string;
+};
+/** Incremental, tolerant JSON parser; see the module comment for its rules and costs. */
+export declare class StreamingJsonParser {
+    #private;
+    /**
+     * Read the next piece of text; returns the changes it made, in order. Costs time linear in `text.length`.
+     */
+    append(text: string): readonly JsonChange[];
+    /**
+     * The value of the text read so far: `undefined` before a value starts, otherwise a plain JSON value that never
+     * changes afterwards. Returns the previous snapshot while the text read since did not change the value, or while a
+     * fresh copy of the open path is not yet paid for (see the module comment).
+     */
+    snapshot(): JsonValue | undefined;
+}
+/**
+ * The arguments of one streaming tool call: the raw text, for the one full parse when the call ends, and a
+ * `StreamingJsonParser` for the partial arguments in between. Providers keep one per call in their own stream state,
+ * never on the content block, so the raw text never reaches partial messages or transcripts.
+ */
+export declare class ToolCallArgumentsStream {
+    #private;
+    /** The raw argument text received so far. */
+    get text(): string;
+    /** Append a piece of argument text; returns the partial arguments, a new object whenever they changed. */
+    append(delta: string): ToolCall["arguments"];
+    /**
+     * Take the provider's complete argument text. When it continues the text so far, returns the missing piece as
+     * `delta`; otherwise starts over with `text`. Either way returns the partial arguments of `text`.
+     */
+    replace(text: string): {
+        readonly delta: string | undefined;
+        readonly arguments: ToolCall["arguments"];
+    };
+    /** The final arguments: one full parse of the complete text, which reports malformed arguments. */
+    finish(): Pick<ToolCall, "arguments" | "argumentsError">;
+}
+//# sourceMappingURL=streaming-json.d.ts.map

package/dist/utils/streaming-json.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"streaming-json.d.ts","sourceRoot":"","sources":["../../src/utils/streaming-json.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAiCG;AAEH,OAAO,KAAK,EAAE,SAAS,EAAE,QAAQ,EAAE,MAAM,aAAa,CAAC;AAWvD,iEAAiE;AACjE,MAAM,MAAM,QAAQ,GAAG,SAAS,CAAC,MAAM,GAAG,MAAM,CAAC,EAAE,CAAC;AAEpD,4EAA4E;AAC5E,MAAM,MAAM,UAAU,GACnB;IAAE,QAAQ,CAAC,IAAI,EAAE,KAAK,CAAC;IAAC,QAAQ,CAAC,IAAI,EAAE,QAAQ,CAAC;IAAC,QAAQ,CAAC,KAAK,EAAE,SAAS,CAAA;CAAE,GAC5E;IAAE,QAAQ,CAAC,IAAI,EAAE,QAAQ,CAAC;IAAC,QAAQ,CAAC,IAAI,EAAE,QAAQ,CAAC;IAAC,QAAQ,CAAC,IAAI,EAAE,MAAM,CAAA;CAAE,CAAC;AAuD/E,yFAAyF;AACzF,qBAAa,mBAAmB;;IAe/B;;OAEG;IACH,MAAM,CAAC,IAAI,EAAE,MAAM,GAAG,SAAS,UAAU,EAAE,CAY1C;IAED;;;;OAIG;IACH,QAAQ,IAAI,SAAS,GAAG,SAAS,CAShC;CAkSD;AAED;;;;GAIG;AACH,qBAAa,uBAAuB;;IAMnC,6CAA6C;IAC7C,IAAI,IAAI,IAAI,MAAM,CAEjB;IAED,0GAA0G;IAC1G,MAAM,CAAC,KAAK,EAAE,MAAM,GAAG,QAAQ,CAAC,WAAW,CAAC,CAI3C;IAED;;;OAGG;IACH,OAAO,CAAC,IAAI,EAAE,MAAM,GAAG;QAAE,QAAQ,CAAC,KAAK,EAAE,MAAM,GAAG,SAAS,CAAC;QAAC,QAAQ,CAAC,SAAS,EAAE,QAAQ,CAAC,WAAW,CAAC,CAAA;KAAE,CAQvG;IAED,mGAAmG;IACnG,MAAM,IAAI,IAAI,CAAC,QAAQ,EAAE,WAAW,GAAG,gBAAgB,CAAC,CAEvD;CAaD"}