gsd-pi 2.73.1 → 2.74.0-dev.0306a2e

package/packages/pi-ai/dist/utils/overflow.js.map

@@ -1 +1 @@
-{"version":3,"file":"overflow.js","sourceRoot":"","sources":["../../src/utils/overflow.ts"],"names":[],"mappings":"AAEA;;;;;;;;;;;;;;;;;;;;;;;GAuBG;AACH,MAAM,iBAAiB,GAAG;IACzB,qBAAqB,EAAE,YAAY;IACnC,wCAAwC,EAAE,iBAAiB;IAC3D,6BAA6B,EAAE,uCAAuC;IACtE,yCAAyC,EAAE,kBAAkB;IAC7D,+BAA+B,EAAE,aAAa;IAC9C,oCAAoC,EAAE,OAAO;IAC7C,uCAAuC,EAAE,4BAA4B;IACrE,2BAA2B,EAAE,iBAAiB;IAC9C,qCAAqC,EAAE,mBAAmB;IAC1D,kCAAkC,EAAE,YAAY;IAChD,+BAA+B,EAAE,UAAU;IAC3C,6BAA6B,EAAE,kBAAkB;IACjD,sDAAsD,EAAE,UAAU;IAClE,gCAAgC,EAAE,yDAAyD;IAC3F,gCAAgC,EAAE,mBAAmB;IACrD,kBAAkB,EAAE,mBAAmB;IACvC,uBAAuB,EAAE,mBAAmB;CAC5C,CAAC;AAEF;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA4CG;AACH,MAAM,UAAU,iBAAiB,CAAC,OAAyB,EAAE,aAAsB;IAClF,uCAAuC;IACvC,IAAI,OAAO,CAAC,UAAU,KAAK,OAAO,IAAI,OAAO,CAAC,YAAY,EAAE,CAAC;QAC5D,uBAAuB;QACvB,IAAI,iBAAiB,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC,OAAO,CAAC,YAAa,CAAC,CAAC,EAAE,CAAC;YAClE,OAAO,IAAI,CAAC;QACb,CAAC;QAED,6DAA6D;QAC7D,8EAA8E;QAC9E,IAAI,2CAA2C,CAAC,IAAI,CAAC,OAAO,CAAC,YAAY,CAAC,EAAE,CAAC;YAC5E,OAAO,IAAI,CAAC;QACb,CAAC;IACF,CAAC;IAED,8EAA8E;IAC9E,IAAI,aAAa,IAAI,OAAO,CAAC,UAAU,KAAK,MAAM,EAAE,CAAC;QACpD,MAAM,WAAW,GAAG,OAAO,CAAC,KAAK,CAAC,KAAK,GAAG,OAAO,CAAC,KAAK,CAAC,SAAS,CAAC;QAClE,IAAI,WAAW,GAAG,aAAa,EAAE,CAAC;YACjC,OAAO,IAAI,CAAC;QACb,CAAC;IACF,CAAC;IAED,OAAO,KAAK,CAAC;AACd,CAAC","sourcesContent":["import type { AssistantMessage } from \"../types.js\";\n\n/**\n * Regex patterns to detect context overflow errors from different providers.\n *\n * These patterns match error messages returned when the input exceeds\n * the model's context window.\n *\n * Provider-specific patterns (with example error messages):\n *\n * - Anthropic: \"prompt is too long: 213462 tokens > 200000 maximum\"\n * - OpenAI: \"Your input exceeds the context window of this model\"\n * - Google: \"The input token count (1196265) exceeds the maximum number of tokens allowed (1048575)\"\n * - xAI: \"This model's maximum prompt length is 131072 but the request contains 537812 tokens\"\n * - Groq: \"Please reduce the length of the messages or completion\"\n * - OpenRouter: \"This endpoint's maximum context length is X tokens. However, you requested about Y tokens\"\n * - llama.cpp: \"the request exceeds the available context size, try increasing it\"\n * - LM Studio: \"tokens to keep from the initial prompt is greater than the context length\"\n * - GitHub Copilot: \"prompt token count of X exceeds the limit of Y\"\n * - MiniMax: \"invalid params, context window exceeds limit\"\n * - Kimi For Coding: \"Your request exceeded model token limit: X (requested: Y)\"\n * - Cerebras: Returns \"400/413 status code (no body)\" - handled separately below\n * - Mistral: \"Prompt contains X tokens ... too large for model with Y maximum context length\"\n * - z.ai: Does NOT error, accepts overflow silently - handled via usage.input > contextWindow\n * - Ollama: Silently truncates input - not detectable via error message\n */\nconst OVERFLOW_PATTERNS = [\n\t/prompt is too long/i, // Anthropic\n\t/input is too long for requested model/i, // Amazon Bedrock\n\t/exceeds the context window/i, // OpenAI (Completions & Responses API)\n\t/input token count.*exceeds the maximum/i, // Google (Gemini)\n\t/maximum prompt length is \\d+/i, // xAI (Grok)\n\t/reduce the length of the messages/i, // Groq\n\t/maximum context length is \\d+ tokens/i, // OpenRouter (all backends)\n\t/exceeds the limit of \\d+/i, // GitHub Copilot\n\t/exceeds the available context size/i, // llama.cpp server\n\t/greater than the context length/i, // LM Studio\n\t/context window exceeds limit/i, // MiniMax\n\t/exceeded model token limit/i, // Kimi For Coding\n\t/too large for model with \\d+ maximum context length/i, // Mistral\n\t/model_context_window_exceeded/i, // z.ai non-standard finish_reason surfaced as error text\n\t/context[_ ]length[_ ]exceeded/i, // Generic fallback\n\t/too many tokens/i, // Generic fallback\n\t/token limit exceeded/i, // Generic fallback\n];\n\n/**\n * Check if an assistant message represents a context overflow error.\n *\n * This handles two cases:\n * 1. Error-based overflow: Most providers return stopReason \"error\" with a\n *    specific error message pattern.\n * 2. Silent overflow: Some providers accept overflow requests and return\n *    successfully. For these, we check if usage.input exceeds the context window.\n *\n * ## Reliability by Provider\n *\n * **Reliable detection (returns error with detectable message):**\n * - Anthropic: \"prompt is too long: X tokens > Y maximum\"\n * - OpenAI (Completions & Responses): \"exceeds the context window\"\n * - Google Gemini: \"input token count exceeds the maximum\"\n * - xAI (Grok): \"maximum prompt length is X but request contains Y\"\n * - Groq: \"reduce the length of the messages\"\n * - Cerebras: 400/413 status code (no body)\n * - Mistral: \"Prompt contains X tokens ... too large for model with Y maximum context length\"\n * - OpenRouter (all backends): \"maximum context length is X tokens\"\n * - llama.cpp: \"exceeds the available context size\"\n * - LM Studio: \"greater than the context length\"\n * - Kimi For Coding: \"exceeded model token limit: X (requested: Y)\"\n *\n * **Unreliable detection:**\n * - z.ai: Sometimes accepts overflow silently (detectable via usage.input > contextWindow),\n *   sometimes returns rate limit errors. Pass contextWindow param to detect silent overflow.\n * - Ollama: Silently truncates input without error. Cannot be detected via this function.\n *   The response will have usage.input < expected, but we don't know the expected value.\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// Check known patterns\n\t\tif (OVERFLOW_PATTERNS.some((p) => p.test(message.errorMessage!))) {\n\t\t\treturn true;\n\t\t}\n\n\t\t// Cerebras returns 400/413 with no body for context overflow\n\t\t// Note: 429 is rate limiting (requests/tokens per time), NOT context overflow\n\t\tif (/^4(00|13)\\s*(status code)?\\s*\\(no body\\)/i.test(message.errorMessage)) {\n\t\t\treturn true;\n\t\t}\n\t}\n\n\t// Case 2: Silent overflow (z.ai style) - successful but usage exceeds context\n\tif (contextWindow && message.stopReason === \"stop\") {\n\t\tconst inputTokens = message.usage.input + message.usage.cacheRead;\n\t\tif (inputTokens > contextWindow) {\n\t\t\treturn true;\n\t\t}\n\t}\n\n\treturn false;\n}"]}
+{"version":3,"file":"overflow.js","sourceRoot":"","sources":["../../src/utils/overflow.ts"],"names":[],"mappings":"AAEA;;;;;;;;;;;;;;;;;;;;;;;GAuBG;AACH,MAAM,iBAAiB,GAAG;IACzB,qBAAqB,EAAE,YAAY;IACnC,wCAAwC,EAAE,iBAAiB;IAC3D,6BAA6B,EAAE,uCAAuC;IACtE,yCAAyC,EAAE,kBAAkB;IAC7D,+BAA+B,EAAE,aAAa;IAC9C,oCAAoC,EAAE,OAAO;IAC7C,uCAAuC,EAAE,4BAA4B;IACrE,2BAA2B,EAAE,iBAAiB;IAC9C,qCAAqC,EAAE,mBAAmB;IAC1D,kCAAkC,EAAE,YAAY;IAChD,+BAA+B,EAAE,UAAU;IAC3C,6BAA6B,EAAE,kBAAkB;IACjD,sDAAsD,EAAE,UAAU;IAClE,gCAAgC,EAAE,yDAAyD;IAC3F,gCAAgC,EAAE,mBAAmB;IACrD,kBAAkB,EAAE,mBAAmB;IACvC,uBAAuB,EAAE,mBAAmB;CAC5C,CAAC;AAEF;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA4CG;AACH,MAAM,UAAU,iBAAiB,CAAC,OAAyB,EAAE,aAAsB;IAClF,uCAAuC;IACvC,IAAI,OAAO,CAAC,UAAU,KAAK,OAAO,IAAI,OAAO,CAAC,YAAY,EAAE,CAAC;QAC5D,uBAAuB;QACvB,IAAI,iBAAiB,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC,OAAO,CAAC,YAAa,CAAC,CAAC,EAAE,CAAC;YAClE,OAAO,IAAI,CAAC;QACb,CAAC;QAED,6DAA6D;QAC7D,8EAA8E;QAC9E,IAAI,2CAA2C,CAAC,IAAI,CAAC,OAAO,CAAC,YAAY,CAAC,EAAE,CAAC;YAC5E,OAAO,IAAI,CAAC;QACb,CAAC;IACF,CAAC;IAED,4EAA4E;IAC5E,8EAA8E;IAC9E,iEAAiE;IACjE,IAAI,OAAO,CAAC,UAAU,KAAK,OAAO,EAAE,CAAC;QACpC,MAAM,aAAa,GAAG,OAAO,CAAC,OAAO;aACnC,MAAM,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC,KAAK,CAAC,IAAI,KAAK,MAAM,CAAC;aACxC,GAAG,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC,KAAK,CAAC,IAAI,CAAC;aAC1B,IAAI,CAAC,IAAI,CAAC,CAAC;QACb,IAAI,aAAa,IAAI,iBAAiB,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC,aAAa,CAAC,CAAC,EAAE,CAAC;YAC3E,OAAO,IAAI,CAAC;QACb,CAAC;IACF,CAAC;IAED,8EAA8E;IAC9E,IAAI,aAAa,IAAI,OAAO,CAAC,UAAU,KAAK,MAAM,EAAE,CAAC;QACpD,MAAM,WAAW,GAAG,OAAO,CAAC,KAAK,CAAC,KAAK,GAAG,OAAO,CAAC,KAAK,CAAC,SAAS,CAAC;QAClE,IAAI,WAAW,GAAG,aAAa,EAAE,CAAC;YACjC,OAAO,IAAI,CAAC;QACb,CAAC;IACF,CAAC;IAED,OAAO,KAAK,CAAC;AACd,CAAC","sourcesContent":["import type { AssistantMessage } from \"../types.js\";\n\n/**\n * Regex patterns to detect context overflow errors from different providers.\n *\n * These patterns match error messages returned when the input exceeds\n * the model's context window.\n *\n * Provider-specific patterns (with example error messages):\n *\n * - Anthropic: \"prompt is too long: 213462 tokens > 200000 maximum\"\n * - OpenAI: \"Your input exceeds the context window of this model\"\n * - Google: \"The input token count (1196265) exceeds the maximum number of tokens allowed (1048575)\"\n * - xAI: \"This model's maximum prompt length is 131072 but the request contains 537812 tokens\"\n * - Groq: \"Please reduce the length of the messages or completion\"\n * - OpenRouter: \"This endpoint's maximum context length is X tokens. However, you requested about Y tokens\"\n * - llama.cpp: \"the request exceeds the available context size, try increasing it\"\n * - LM Studio: \"tokens to keep from the initial prompt is greater than the context length\"\n * - GitHub Copilot: \"prompt token count of X exceeds the limit of Y\"\n * - MiniMax: \"invalid params, context window exceeds limit\"\n * - Kimi For Coding: \"Your request exceeded model token limit: X (requested: Y)\"\n * - Cerebras: Returns \"400/413 status code (no body)\" - handled separately below\n * - Mistral: \"Prompt contains X tokens ... too large for model with Y maximum context length\"\n * - z.ai: Does NOT error, accepts overflow silently - handled via usage.input > contextWindow\n * - Ollama: Silently truncates input - not detectable via error message\n */\nconst OVERFLOW_PATTERNS = [\n\t/prompt is too long/i, // Anthropic\n\t/input is too long for requested model/i, // Amazon Bedrock\n\t/exceeds the context window/i, // OpenAI (Completions & Responses API)\n\t/input token count.*exceeds the maximum/i, // Google (Gemini)\n\t/maximum prompt length is \\d+/i, // xAI (Grok)\n\t/reduce the length of the messages/i, // Groq\n\t/maximum context length is \\d+ tokens/i, // OpenRouter (all backends)\n\t/exceeds the limit of \\d+/i, // GitHub Copilot\n\t/exceeds the available context size/i, // llama.cpp server\n\t/greater than the context length/i, // LM Studio\n\t/context window exceeds limit/i, // MiniMax\n\t/exceeded model token limit/i, // Kimi For Coding\n\t/too large for model with \\d+ maximum context length/i, // Mistral\n\t/model_context_window_exceeded/i, // z.ai non-standard finish_reason surfaced as error text\n\t/context[_ ]length[_ ]exceeded/i, // Generic fallback\n\t/too many tokens/i, // Generic fallback\n\t/token limit exceeded/i, // Generic fallback\n];\n\n/**\n * Check if an assistant message represents a context overflow error.\n *\n * This handles two cases:\n * 1. Error-based overflow: Most providers return stopReason \"error\" with a\n *    specific error message pattern.\n * 2. Silent overflow: Some providers accept overflow requests and return\n *    successfully. For these, we check if usage.input exceeds the context window.\n *\n * ## Reliability by Provider\n *\n * **Reliable detection (returns error with detectable message):**\n * - Anthropic: \"prompt is too long: X tokens > Y maximum\"\n * - OpenAI (Completions & Responses): \"exceeds the context window\"\n * - Google Gemini: \"input token count exceeds the maximum\"\n * - xAI (Grok): \"maximum prompt length is X but request contains Y\"\n * - Groq: \"reduce the length of the messages\"\n * - Cerebras: 400/413 status code (no body)\n * - Mistral: \"Prompt contains X tokens ... too large for model with Y maximum context length\"\n * - OpenRouter (all backends): \"maximum context length is X tokens\"\n * - llama.cpp: \"exceeds the available context size\"\n * - LM Studio: \"greater than the context length\"\n * - Kimi For Coding: \"exceeded model token limit: X (requested: Y)\"\n *\n * **Unreliable detection:**\n * - z.ai: Sometimes accepts overflow silently (detectable via usage.input > contextWindow),\n *   sometimes returns rate limit errors. Pass contextWindow param to detect silent overflow.\n * - Ollama: Silently truncates input without error. Cannot be detected via this function.\n *   The response will have usage.input < expected, but we don't know the expected value.\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// Check known patterns\n\t\tif (OVERFLOW_PATTERNS.some((p) => p.test(message.errorMessage!))) {\n\t\t\treturn true;\n\t\t}\n\n\t\t// Cerebras returns 400/413 with no body for context overflow\n\t\t// Note: 429 is rate limiting (requests/tokens per time), NOT context overflow\n\t\tif (/^4(00|13)\\s*(status code)?\\s*\\(no body\\)/i.test(message.errorMessage)) {\n\t\t\treturn true;\n\t\t}\n\t}\n\n\t// Some providers surface overflow as assistant text while putting a generic\n\t// classifier value in errorMessage (e.g. claude-code: errorMessage=\"success\",\n\t// text=\"Prompt is too long\"). Check rendered text as a fallback.\n\tif (message.stopReason === \"error\") {\n\t\tconst assistantText = message.content\n\t\t\t.filter((block) => block.type === \"text\")\n\t\t\t.map((block) => block.text)\n\t\t\t.join(\"\\n\");\n\t\tif (assistantText && OVERFLOW_PATTERNS.some((p) => p.test(assistantText))) {\n\t\t\treturn true;\n\t\t}\n\t}\n\n\t// Case 2: Silent overflow (z.ai style) - successful but usage exceeds context\n\tif (contextWindow && message.stopReason === \"stop\") {\n\t\tconst inputTokens = message.usage.input + message.usage.cacheRead;\n\t\tif (inputTokens > contextWindow) {\n\t\t\treturn true;\n\t\t}\n\t}\n\n\treturn false;\n}\n"]}

package/packages/pi-ai/dist/utils/tests/overflow.test.d.ts

@@ -0,0 +1,2 @@
+export {};
+//# sourceMappingURL=overflow.test.d.ts.map

package/packages/pi-ai/dist/utils/tests/overflow.test.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"overflow.test.d.ts","sourceRoot":"","sources":["../../../src/utils/tests/overflow.test.ts"],"names":[],"mappings":""}

package/packages/pi-ai/dist/utils/tests/overflow.test.js

@@ -0,0 +1,50 @@
+import { describe, test } from "node:test";
+import assert from "node:assert/strict";
+import { isContextOverflow } from "../overflow.js";
+function makeAssistantMessage(overrides = {}) {
+    return {
+        role: "assistant",
+        content: [],
+        api: "anthropic-messages",
+        provider: "anthropic",
+        model: "claude-sonnet-4-6",
+        usage: {
+            input: 0,
+            output: 0,
+            cacheRead: 0,
+            cacheWrite: 0,
+            totalTokens: 0,
+            cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },
+        },
+        stopReason: "error",
+        timestamp: Date.now(),
+        ...overrides,
+    };
+}
+describe("isContextOverflow", () => {
+    test("detects overflow from provider errorMessage", () => {
+        const message = makeAssistantMessage({
+            errorMessage: "prompt is too long: 213462 tokens > 200000 maximum",
+        });
+        assert.equal(isContextOverflow(message, 200000), true);
+    });
+    test("detects claude-code overflow when text contains the error but errorMessage is generic (#3925)", () => {
+        const message = makeAssistantMessage({
+            provider: "claude-code",
+            api: "anthropic-messages",
+            model: "claude-sonnet-4-6",
+            errorMessage: "success",
+            content: [{ type: "text", text: "Prompt is too long" }],
+        });
+        assert.equal(isContextOverflow(message, 200000), true);
+    });
+    test("does not treat normal non-error text as overflow", () => {
+        const message = makeAssistantMessage({
+            stopReason: "stop",
+            errorMessage: undefined,
+            content: [{ type: "text", text: "Prompt is too long" }],
+        });
+        assert.equal(isContextOverflow(message, 200000), false);
+    });
+});
+//# sourceMappingURL=overflow.test.js.map

package/packages/pi-ai/dist/utils/tests/overflow.test.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"overflow.test.js","sourceRoot":"","sources":["../../../src/utils/tests/overflow.test.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,QAAQ,EAAE,IAAI,EAAE,MAAM,WAAW,CAAC;AAC3C,OAAO,MAAM,MAAM,oBAAoB,CAAC;AAExC,OAAO,EAAE,iBAAiB,EAAE,MAAM,gBAAgB,CAAC;AAGnD,SAAS,oBAAoB,CAAC,YAAuC,EAAE;IACtE,OAAO;QACN,IAAI,EAAE,WAAW;QACjB,OAAO,EAAE,EAAE;QACX,GAAG,EAAE,oBAAoB;QACzB,QAAQ,EAAE,WAAW;QACrB,KAAK,EAAE,mBAAmB;QAC1B,KAAK,EAAE;YACN,KAAK,EAAE,CAAC;YACR,MAAM,EAAE,CAAC;YACT,SAAS,EAAE,CAAC;YACZ,UAAU,EAAE,CAAC;YACb,WAAW,EAAE,CAAC;YACd,IAAI,EAAE,EAAE,KAAK,EAAE,CAAC,EAAE,MAAM,EAAE,CAAC,EAAE,SAAS,EAAE,CAAC,EAAE,UAAU,EAAE,CAAC,EAAE,KAAK,EAAE,CAAC,EAAE;SACpE;QACD,UAAU,EAAE,OAAO;QACnB,SAAS,EAAE,IAAI,CAAC,GAAG,EAAE;QACrB,GAAG,SAAS;KACZ,CAAC;AACH,CAAC;AAED,QAAQ,CAAC,mBAAmB,EAAE,GAAG,EAAE;IAClC,IAAI,CAAC,6CAA6C,EAAE,GAAG,EAAE;QACxD,MAAM,OAAO,GAAG,oBAAoB,CAAC;YACpC,YAAY,EAAE,oDAAoD;SAClE,CAAC,CAAC;QAEH,MAAM,CAAC,KAAK,CAAC,iBAAiB,CAAC,OAAO,EAAE,MAAM,CAAC,EAAE,IAAI,CAAC,CAAC;IACxD,CAAC,CAAC,CAAC;IAEH,IAAI,CAAC,+FAA+F,EAAE,GAAG,EAAE;QAC1G,MAAM,OAAO,GAAG,oBAAoB,CAAC;YACpC,QAAQ,EAAE,aAAa;YACvB,GAAG,EAAE,oBAAoB;YACzB,KAAK,EAAE,mBAAmB;YAC1B,YAAY,EAAE,SAAS;YACvB,OAAO,EAAE,CAAC,EAAE,IAAI,EAAE,MAAM,EAAE,IAAI,EAAE,oBAAoB,EAAE,CAAC;SACvD,CAAC,CAAC;QAEH,MAAM,CAAC,KAAK,CAAC,iBAAiB,CAAC,OAAO,EAAE,MAAM,CAAC,EAAE,IAAI,CAAC,CAAC;IACxD,CAAC,CAAC,CAAC;IAEH,IAAI,CAAC,kDAAkD,EAAE,GAAG,EAAE;QAC7D,MAAM,OAAO,GAAG,oBAAoB,CAAC;YACpC,UAAU,EAAE,MAAM;YAClB,YAAY,EAAE,SAAS;YACvB,OAAO,EAAE,CAAC,EAAE,IAAI,EAAE,MAAM,EAAE,IAAI,EAAE,oBAAoB,EAAE,CAAC;SACvD,CAAC,CAAC;QAEH,MAAM,CAAC,KAAK,CAAC,iBAAiB,CAAC,OAAO,EAAE,MAAM,CAAC,EAAE,KAAK,CAAC,CAAC;IACzD,CAAC,CAAC,CAAC;AACJ,CAAC,CAAC,CAAC","sourcesContent":["import { describe, test } from \"node:test\";\nimport assert from \"node:assert/strict\";\n\nimport { isContextOverflow } from \"../overflow.js\";\nimport type { AssistantMessage } from \"../../types.js\";\n\nfunction makeAssistantMessage(overrides: Partial<AssistantMessage> = {}): AssistantMessage {\n\treturn {\n\t\trole: \"assistant\",\n\t\tcontent: [],\n\t\tapi: \"anthropic-messages\",\n\t\tprovider: \"anthropic\",\n\t\tmodel: \"claude-sonnet-4-6\",\n\t\tusage: {\n\t\t\tinput: 0,\n\t\t\toutput: 0,\n\t\t\tcacheRead: 0,\n\t\t\tcacheWrite: 0,\n\t\t\ttotalTokens: 0,\n\t\t\tcost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },\n\t\t},\n\t\tstopReason: \"error\",\n\t\ttimestamp: Date.now(),\n\t\t...overrides,\n\t};\n}\n\ndescribe(\"isContextOverflow\", () => {\n\ttest(\"detects overflow from provider errorMessage\", () => {\n\t\tconst message = makeAssistantMessage({\n\t\t\terrorMessage: \"prompt is too long: 213462 tokens > 200000 maximum\",\n\t\t});\n\n\t\tassert.equal(isContextOverflow(message, 200000), true);\n\t});\n\n\ttest(\"detects claude-code overflow when text contains the error but errorMessage is generic (#3925)\", () => {\n\t\tconst message = makeAssistantMessage({\n\t\t\tprovider: \"claude-code\",\n\t\t\tapi: \"anthropic-messages\",\n\t\t\tmodel: \"claude-sonnet-4-6\",\n\t\t\terrorMessage: \"success\",\n\t\t\tcontent: [{ type: \"text\", text: \"Prompt is too long\" }],\n\t\t});\n\n\t\tassert.equal(isContextOverflow(message, 200000), true);\n\t});\n\n\ttest(\"does not treat normal non-error text as overflow\", () => {\n\t\tconst message = makeAssistantMessage({\n\t\t\tstopReason: \"stop\",\n\t\t\terrorMessage: undefined,\n\t\t\tcontent: [{ type: \"text\", text: \"Prompt is too long\" }],\n\t\t});\n\n\t\tassert.equal(isContextOverflow(message, 200000), false);\n\t});\n});\n"]}

package/packages/pi-ai/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@gsd/pi-ai",
-  "version": "0.57.1",
+  "version": "2.74.0",
   "description": "Unified LLM API (vendored from pi-mono)",
   "type": "module",
   "main": "./dist/index.js",

package/packages/pi-ai/scripts/generate-models.ts

@@ -1,6 +1,6 @@
 #!/usr/bin/env tsx
 
-import { writeFileSync } from "fs";
+import { existsSync, mkdirSync, rmSync, writeFileSync } from "fs";
 import { join, dirname } from "path";
 import { fileURLToPath } from "url";
 import { Api, KnownProvider, Model } from "../src/types.js";
@@ -1470,61 +1470,95 @@ async function generateModels() {
 		}
 	}
 
-	// Generate TypeScript file
-	let output = `// This file is auto-generated by scripts/generate-models.ts
-// Do not edit manually - run 'npm run generate-models' to update
-
-import type { Model } from "./types.js";
-
-export const MODELS = {
-`;
-
-	// Generate provider sections (sorted for deterministic output)
-	const sortedProviderIds = Object.keys(providers).sort();
-	for (const providerId of sortedProviderIds) {
-		const models = providers[providerId];
-		output += `\t${JSON.stringify(providerId)}: {\n`;
+	const toConstName = (providerId: string): string => {
+		const normalized = providerId
+			.replace(/[^a-zA-Z0-9]+/g, "_")
+			.replace(/_+/g, "_")
+			.replace(/^_|_$/g, "")
+			.toUpperCase();
+		const safePrefix = /^[0-9]/.test(normalized) ? `_${normalized}` : normalized;
+		return `${safePrefix}_MODELS`;
+	};
 
+	const renderModelEntries = (models: Record<string, Model<any>>): string => {
+		let output = "{\n";
 		const sortedModelIds = Object.keys(models).sort();
 		for (const modelId of sortedModelIds) {
 			const model = models[modelId];
-			output += `\t\t"${model.id}": {\n`;
-			output += `\t\t\tid: "${model.id}",\n`;
-			output += `\t\t\tname: "${model.name}",\n`;
-			output += `\t\t\tapi: "${model.api}",\n`;
-			output += `\t\t\tprovider: "${model.provider}",\n`;
+			output += `\t"${model.id}": {\n`;
+			output += `\t\tid: "${model.id}",\n`;
+			output += `\t\tname: "${model.name}",\n`;
+			output += `\t\tapi: "${model.api}",\n`;
+			output += `\t\tprovider: "${model.provider}",\n`;
 			if (model.baseUrl !== undefined) {
-				output += `\t\t\tbaseUrl: "${model.baseUrl}",\n`;
+				output += `\t\tbaseUrl: "${model.baseUrl}",\n`;
 			}
 			if (model.headers) {
-				output += `\t\t\theaders: ${JSON.stringify(model.headers)},\n`;
+				output += `\t\theaders: ${JSON.stringify(model.headers)},\n`;
 			}
 			if (model.compat) {
-				output += `			compat: ${JSON.stringify(model.compat)},
-`;
+				output += `\t\tcompat: ${JSON.stringify(model.compat)},\n`;
 			}
-			output += `\t\t\treasoning: ${model.reasoning},\n`;
-			output += `\t\t\tinput: [${model.input.map(i => `"${i}"`).join(", ")}],\n`;
-			output += `\t\t\tcost: {\n`;
-			output += `\t\t\t\tinput: ${model.cost.input},\n`;
-			output += `\t\t\t\toutput: ${model.cost.output},\n`;
-			output += `\t\t\t\tcacheRead: ${model.cost.cacheRead},\n`;
-			output += `\t\t\t\tcacheWrite: ${model.cost.cacheWrite},\n`;
-			output += `\t\t\t},\n`;
-			output += `\t\t\tcontextWindow: ${model.contextWindow},\n`;
-			output += `\t\t\tmaxTokens: ${model.maxTokens},\n`;
-			output += `\t\t} satisfies Model<"${model.api}">,\n`;
+			output += `\t\treasoning: ${model.reasoning},\n`;
+			output += `\t\tinput: [${model.input.map(i => `"${i}"`).join(", ")}],\n`;
+			output += `\t\tcost: {\n`;
+			output += `\t\t\tinput: ${model.cost.input},\n`;
+			output += `\t\t\toutput: ${model.cost.output},\n`;
+			output += `\t\t\tcacheRead: ${model.cost.cacheRead},\n`;
+			output += `\t\t\tcacheWrite: ${model.cost.cacheWrite},\n`;
+			output += `\t\t},\n`;
+			output += `\t\tcontextWindow: ${model.contextWindow},\n`;
+			output += `\t\tmaxTokens: ${model.maxTokens},\n`;
+			output += `\t} satisfies Model<"${model.api}">,\n`;
 		}
+		output += "}";
+		return output;
+	};
 
-		output += `\t},\n`;
+	const sortedProviderIds = Object.keys(providers).sort();
+	const generatedDir = join(packageRoot, "src/models/generated");
+	rmSync(generatedDir, { recursive: true, force: true });
+	mkdirSync(generatedDir, { recursive: true });
+
+	for (const providerId of sortedProviderIds) {
+		const providerModels = providers[providerId];
+		const constName = toConstName(providerId);
+		const providerOutput = `// This file is auto-generated by scripts/generate-models.ts
+// Do not edit manually - run 'npm run generate-models' to update
+
+import type { Model } from "../../types.js";
+
+export const ${constName} = ${renderModelEntries(providerModels)} as const satisfies Record<string, Model<any>>;
+`;
+		writeFileSync(join(generatedDir, `${providerId}.ts`), providerOutput);
+		console.log(`Generated src/models/generated/${providerId}.ts`);
 	}
 
-	output += `} as const;
+	let indexOutput = `// This file is auto-generated by scripts/generate-models.ts
+// Do not edit manually - run 'npm run generate-models' to update
+
 `;
 
-	// Write file
-	writeFileSync(join(packageRoot, "src/models.generated.ts"), output);
-	console.log("Generated src/models.generated.ts");
+	for (const providerId of sortedProviderIds) {
+		const constName = toConstName(providerId);
+		indexOutput += `import { ${constName} } from "./${providerId}.js";\n`;
+	}
+
+	indexOutput += `\nexport const MODELS = {\n`;
+	for (const providerId of sortedProviderIds) {
+		const constName = toConstName(providerId);
+		indexOutput += `\t${JSON.stringify(providerId)}: ${constName},\n`;
+	}
+	indexOutput += `} as const;\n`;
+
+	writeFileSync(join(generatedDir, "index.ts"), indexOutput);
+	console.log("Generated src/models/generated/index.ts");
+
+	const legacyGeneratedPath = join(packageRoot, "src/models.generated.ts");
+	if (existsSync(legacyGeneratedPath)) {
+		rmSync(legacyGeneratedPath);
+		console.log("Removed legacy src/models.generated.ts");
+	}
 
 	// Print statistics
 	const totalModels = allModels.length;

package/packages/pi-ai/src/index.ts

@@ -3,15 +3,11 @@ export { Type } from "@sinclair/typebox";
 
 export * from "./api-registry.js";
 export * from "./env-api-keys.js";
-export * from "./models.js";
-export * from "./providers/anthropic.js";
-export * from "./providers/azure-openai-responses.js";
-export * from "./providers/google.js";
-export * from "./providers/google-gemini-cli.js";
-export * from "./providers/google-vertex.js";
-export * from "./providers/mistral.js";
-export * from "./providers/openai-completions.js";
-export * from "./providers/openai-responses.js";
+export * from "./models/index.js";
+export {
+	mapThinkingLevelToEffort,
+	supportsAdaptiveThinking,
+} from "./providers/anthropic-shared.js";
 export * from "./providers/provider-capabilities.js";
 export * from "./providers/register-builtins.js";
 export type { ProviderSwitchReport } from "./providers/transform-messages.js";

package/packages/pi-ai/src/models/capability-patches.ts

@@ -0,0 +1,40 @@
+import type { Api, Model, ModelCapabilities } from "../types.js";
+
+type CapabilityPatch = { match: (m: Model<Api>) => boolean; caps: ModelCapabilities };
+
+export const CAPABILITY_PATCHES: CapabilityPatch[] = [
+	// GPT-5.x supports xhigh thinking and OpenAI service tiers
+	{
+		match: (m) => m.id.includes("gpt-5.2") || m.id.includes("gpt-5.3") || m.id.includes("gpt-5.4"),
+		caps: { supportsXhigh: true, supportsServiceTier: true },
+	},
+	// Anthropic Opus 4.6 supports xhigh thinking
+	{
+		match: (m) => m.api === "anthropic-messages" && (m.id.includes("opus-4-6") || m.id.includes("opus-4.6")),
+		caps: { supportsXhigh: true },
+	},
+];
+
+/**
+ * Apply capability patches to a list of models.
+ *
+ * Models constructed outside the static pi-ai registry (custom models from
+ * models.json, extension-registered models, discovered models) do not pass
+ * through the module-init patch loop. Call this function after assembling
+ * any model list to ensure capabilities are set correctly.
+ *
+ * Explicit `capabilities` already set on a model take precedence over patches.
+ */
+export function applyCapabilityPatches(models: Model<Api>[]): Model<Api>[] {
+	return models.map((model) => {
+		for (const patch of CAPABILITY_PATCHES) {
+			if (patch.match(model)) {
+				return {
+					...model,
+					capabilities: { ...patch.caps, ...model.capabilities },
+				};
+			}
+		}
+		return model;
+	});
+}

package/packages/pi-ai/src/{models.custom.ts → models/custom.ts}

@@ -1,8 +1,8 @@
 // Manually-maintained model definitions for providers NOT tracked by models.dev.
 //
-// The auto-generated file (models.generated.ts) is rebuilt from the models.dev
-// third-party catalog. Providers that use proprietary endpoints and are not
-// listed on models.dev must be defined here so they survive regeneration.
+// The auto-generated files under models/generated/*.ts are rebuilt from the
+// models.dev third-party catalog. Providers that use proprietary endpoints and
+// are not listed on models.dev must be defined here so they survive regeneration.
 //
 // See: https://github.com/gsd-build/gsd-2/issues/2339
 //
@@ -11,7 +11,7 @@
 //   2. Add its API key mapping to env-api-keys.ts.
 //   3. Add its provider name to KnownProvider in types.ts (if not already there).
 
-import type { Model } from "./types.js";
+import type { Model } from "../types.js";
 
 export const CUSTOM_MODELS = {
 	// ─── Alibaba Coding Plan ─────────────────────────────────────────────