@bastani/pi-ai 0.9.71-alpha.1 → 0.9.72-alpha.1
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/CHANGELOG.md +64 -1
- package/README.md +78 -8
- package/dist/api/anthropic-messages.d.ts.map +1 -1
- package/dist/api/anthropic-messages.js +50 -10
- package/dist/api/anthropic-messages.js.map +1 -1
- package/dist/api/azure-openai-responses.d.ts.map +1 -1
- package/dist/api/azure-openai-responses.js +7 -2
- package/dist/api/azure-openai-responses.js.map +1 -1
- package/dist/api/bedrock-converse-stream.d.ts.map +1 -1
- package/dist/api/bedrock-converse-stream.js +28 -1
- package/dist/api/bedrock-converse-stream.js.map +1 -1
- package/dist/api/bedrock-converse-stream.lazy.d.ts +7 -0
- package/dist/api/bedrock-converse-stream.lazy.d.ts.map +1 -1
- package/dist/api/bedrock-converse-stream.lazy.js +18 -5
- package/dist/api/bedrock-converse-stream.lazy.js.map +1 -1
- package/dist/api/google-generative-ai.d.ts +3 -1
- package/dist/api/google-generative-ai.d.ts.map +1 -1
- package/dist/api/google-generative-ai.js +4 -2
- package/dist/api/google-generative-ai.js.map +1 -1
- package/dist/api/google-shared.d.ts +10 -2
- package/dist/api/google-shared.d.ts.map +1 -1
- package/dist/api/google-shared.js +31 -9
- package/dist/api/google-shared.js.map +1 -1
- package/dist/api/google-vertex.d.ts +3 -1
- package/dist/api/google-vertex.d.ts.map +1 -1
- package/dist/api/google-vertex.js +4 -2
- package/dist/api/google-vertex.js.map +1 -1
- package/dist/api/lazy.d.ts +2 -2
- package/dist/api/lazy.d.ts.map +1 -1
- package/dist/api/lazy.js +10 -9
- package/dist/api/lazy.js.map +1 -1
- package/dist/api/mistral-conversations.d.ts.map +1 -1
- package/dist/api/mistral-conversations.js +12 -1
- package/dist/api/mistral-conversations.js.map +1 -1
- package/dist/api/openai-codex-responses.d.ts.map +1 -1
- package/dist/api/openai-codex-responses.js +6 -1
- package/dist/api/openai-codex-responses.js.map +1 -1
- package/dist/api/openai-completions.d.ts +2 -1
- package/dist/api/openai-completions.d.ts.map +1 -1
- package/dist/api/openai-completions.js +112 -15
- package/dist/api/openai-completions.js.map +1 -1
- package/dist/api/openai-images-options.d.ts +107 -0
- package/dist/api/openai-images-options.d.ts.map +1 -0
- package/dist/api/openai-images-options.js +106 -0
- package/dist/api/openai-images-options.js.map +1 -0
- package/dist/api/openai-images.d.ts +3 -11
- package/dist/api/openai-images.d.ts.map +1 -1
- package/dist/api/openai-images.js +117 -16
- package/dist/api/openai-images.js.map +1 -1
- package/dist/api/openai-responses-shared.js +6 -2
- package/dist/api/openai-responses-shared.js.map +1 -1
- package/dist/api/openai-responses.d.ts.map +1 -1
- package/dist/api/openai-responses.js +29 -7
- package/dist/api/openai-responses.js.map +1 -1
- package/dist/auth/oauth/load.d.ts.map +1 -1
- package/dist/auth/oauth/load.js +5 -2
- package/dist/auth/oauth/load.js.map +1 -1
- package/dist/env-api-keys.d.ts.map +1 -1
- package/dist/env-api-keys.js +1 -0
- package/dist/env-api-keys.js.map +1 -1
- package/dist/image-generation-error.d.ts +203 -0
- package/dist/image-generation-error.d.ts.map +1 -0
- package/dist/image-generation-error.js +441 -0
- package/dist/image-generation-error.js.map +1 -0
- package/dist/index.d.ts +4 -1
- package/dist/index.d.ts.map +1 -1
- package/dist/index.js +4 -0
- package/dist/index.js.map +1 -1
- package/dist/lazy-bedrock-provider.d.ts +10 -0
- package/dist/lazy-bedrock-provider.d.ts.map +1 -0
- package/dist/lazy-bedrock-provider.js +13 -0
- package/dist/lazy-bedrock-provider.js.map +1 -0
- package/dist/lazy-oauth.d.ts +10 -0
- package/dist/lazy-oauth.d.ts.map +1 -0
- package/dist/lazy-oauth.js +23 -0
- package/dist/lazy-oauth.js.map +1 -0
- package/dist/models.d.ts.map +1 -1
- package/dist/models.generated.d.ts +4 -0
- package/dist/models.generated.d.ts.map +1 -1
- package/dist/models.generated.js +4 -0
- package/dist/models.generated.js.map +1 -1
- package/dist/models.js +8 -8
- package/dist/models.js.map +1 -1
- package/dist/providers/all.d.ts.map +1 -1
- package/dist/providers/all.js +2 -0
- package/dist/providers/all.js.map +1 -1
- package/dist/providers/azure.d.ts.map +1 -1
- package/dist/providers/azure.js +2 -2
- package/dist/providers/azure.js.map +1 -1
- package/dist/providers/compat-schema.d.ts +15 -0
- package/dist/providers/compat-schema.d.ts.map +1 -1
- package/dist/providers/compat-schema.js +19 -1
- package/dist/providers/compat-schema.js.map +1 -1
- package/dist/providers/data/.manifest.json +1 -1
- package/dist/providers/data/anthropic.json +1 -1
- package/dist/providers/data/azure.json +1 -1
- package/dist/providers/data/baseten.json +1 -1
- package/dist/providers/data/cloudflare-ai-gateway.json +1 -1
- package/dist/providers/data/fireworks.json +1 -1
- package/dist/providers/data/github-copilot.json +1 -1
- package/dist/providers/data/google-vertex.json +1 -1
- package/dist/providers/data/google.json +1 -1
- package/dist/providers/data/nvidia.json +1 -1
- package/dist/providers/data/openai-images.json +1 -1
- package/dist/providers/data/opencode-go.json +1 -1
- package/dist/providers/data/opencode.json +1 -1
- package/dist/providers/data/openrouter.json +1 -1
- package/dist/providers/data/qwen-token-plan-individual-cn.json +1 -0
- package/dist/providers/data/qwen-token-plan-individual.json +1 -1
- package/dist/providers/data/together.json +1 -1
- package/dist/providers/data/vercel-ai-gateway.json +1 -1
- package/dist/providers/github-copilot-capabilities.d.ts +67 -0
- package/dist/providers/github-copilot-capabilities.d.ts.map +1 -0
- package/dist/providers/github-copilot-capabilities.js +81 -0
- package/dist/providers/github-copilot-capabilities.js.map +1 -0
- package/dist/providers/github-copilot-errors.d.ts +243 -0
- package/dist/providers/github-copilot-errors.d.ts.map +1 -0
- package/dist/providers/github-copilot-errors.js +720 -0
- package/dist/providers/github-copilot-errors.js.map +1 -0
- package/dist/providers/github-copilot.d.ts +1 -2
- package/dist/providers/github-copilot.d.ts.map +1 -1
- package/dist/providers/github-copilot.js +29 -4
- package/dist/providers/github-copilot.js.map +1 -1
- package/dist/providers/openai-codex.js +1 -1
- package/dist/providers/openai-codex.js.map +1 -1
- package/dist/providers/qwen-token-plan-individual-cn.d.ts +3 -0
- package/dist/providers/qwen-token-plan-individual-cn.d.ts.map +1 -0
- package/dist/providers/qwen-token-plan-individual-cn.js +15 -0
- package/dist/providers/qwen-token-plan-individual-cn.js.map +1 -0
- package/dist/providers/qwen-token-plan-individual-cn.models.d.ts +5 -0
- package/dist/providers/qwen-token-plan-individual-cn.models.d.ts.map +1 -0
- package/dist/providers/qwen-token-plan-individual-cn.models.js +8 -0
- package/dist/providers/qwen-token-plan-individual-cn.models.js.map +1 -0
- package/dist/types.d.ts +29 -3
- package/dist/types.d.ts.map +1 -1
- package/dist/types.js.map +1 -1
- package/dist/utils/error-body.d.ts +8 -0
- package/dist/utils/error-body.d.ts.map +1 -1
- package/dist/utils/error-body.js +28 -1
- package/dist/utils/error-body.js.map +1 -1
- package/dist/utils/model-operations.d.ts.map +1 -1
- package/dist/utils/model-operations.js +6 -0
- package/dist/utils/model-operations.js.map +1 -1
- package/dist/utils/overflow.d.ts +4 -0
- package/dist/utils/overflow.d.ts.map +1 -1
- package/dist/utils/overflow.js +16 -3
- package/dist/utils/overflow.js.map +1 -1
- package/dist/utils/provider-retry.d.ts +2 -0
- package/dist/utils/provider-retry.d.ts.map +1 -1
- package/dist/utils/provider-retry.js +2 -0
- package/dist/utils/provider-retry.js.map +1 -1
- package/dist/utils/retry.d.ts +3 -1
- package/dist/utils/retry.d.ts.map +1 -1
- package/dist/utils/retry.js +9 -0
- package/dist/utils/retry.js.map +1 -1
- package/dist/utils/thinking-off.d.ts +23 -0
- package/dist/utils/thinking-off.d.ts.map +1 -0
- package/dist/utils/thinking-off.js +51 -0
- package/dist/utils/thinking-off.js.map +1 -0
- package/dist/utils/version.d.ts +1 -1
- package/dist/utils/version.js +1 -1
- package/dist/utils/version.js.map +1 -1
- package/package.json +9 -1
|
@@ -1 +1 @@
|
|
|
1
|
-
{"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,iDAAiD;CACjD,CAAC;AAEF,SAAS,0BAA0B,CAAC,YAAoB;IACvD,MAAM,UAAU,GAAG,YAAY,CAAC,IAAI,EAAE,CAAC,WAAW,EAAE,CAAC;IACrD,OAAO,CACN,UAAU,KAAK,2BAA2B;QAC1C,UAAU,KAAK,2BAA2B;QAC1C,UAAU,KAAK,eAAe;QAC9B,UAAU,KAAK,eAAe,CAC9B,CAAC;AACH,CAAC;AAED;;;;;;;;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,0BAA0B,CAAC,OAAO,CAAC,YAAY,CAAC,EAAE,CAAC;gBACzF,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\nfunction isCerebrasBodylessOverflow(errorMessage: string): boolean {\n\tconst normalized = errorMessage.trim().toLowerCase();\n\treturn (\n\t\tnormalized === \"400 status code (no body)\" ||\n\t\tnormalized === \"413 status code (no body)\" ||\n\t\tnormalized === \"400 (no body)\" ||\n\t\tnormalized === \"413 (no body)\"\n\t);\n}\n\n/**\n * Patterns that indicate non-overflow errors (e.g. rate limiting, server errors).\n * Error messages matching any of these are excluded from overflow detection\n * even if they also match an OVERFLOW_PATTERN.\n *\n * Example: Bedrock formats throttling errors as \"ThrottlingException: Too many tokens,\n * please wait before trying again.\" which would match the /too many tokens/i overflow\n * pattern without this exclusion.\n */\nconst NON_OVERFLOW_PATTERNS = [\n\t/^(Throttling error|Service unavailable):/i, // AWS Bedrock non-overflow errors (human-readable prefixes from formatBedrockError)\n\t/rate limit/i, // Generic rate limiting\n\t/too many requests/i, // Generic HTTP 429 style\n];\n\n/**\n * Check if an assistant message represents a context overflow error.\n *\n * This handles three cases:\n * 1. Error-based overflow: Most providers return stopReason \"error\" with a\n * specific error message pattern.\n * 2. Silent overflow: Some providers accept overflow requests and return\n * successfully. For these, we check if usage.input exceeds the context window.\n * 3. Length-stop overflow: Xiaomi MiMo can return \"length\" with zero output when\n * the input fills the context window.\n *\n * ## Reliability by Provider\n *\n * **Reliable detection (returns error with detectable message):**\n * - Anthropic: \"prompt is too long: X tokens > Y maximum\" or \"request_too_large\"\n * - OpenAI (Completions & Responses): \"exceeds the context window\", \"exceeds the model's maximum context length of X tokens\", or \"exceeds model's maximum context length (X)\"\n * - Google Gemini: \"input token count exceeds the maximum\"\n * - xAI (Grok): \"maximum prompt length is X but request contains Y\"\n * - Groq: \"reduce the length of the messages\"\n * - Cerebras: 400/413 status code (no body)\n * - Mistral: \"Prompt contains X tokens ... too large for model with Y maximum context length\"\n * - OpenRouter (most backends): \"maximum context length is X tokens\"\n * - OpenRouter/Poolside: \"Input length X exceeds the maximum allowed input length of Y tokens.\"\n * - Together AI: \"The input (X tokens) is longer than the model's context length (Y tokens).\"\n * - llama.cpp: \"exceeds the available context size\"\n * - LM Studio: \"greater than the context length\"\n * - Kimi For Coding: \"exceeded model token limit: X (requested: Y)\"\n * - DS4: \"Prompt has X tokens, but the configured context size is Y tokens\"\n * - DashScope/Qwen: \"Range of input length should be [1, X]\"\n * - 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\" && isCerebrasBodylessOverflow(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"]}
|
|
1
|
+
{"version":3,"file":"overflow.js","sourceRoot":"","sources":["../../src/utils/overflow.ts"],"names":[],"mappings":"AAAA,OAAO,EAA+B,qBAAqB,EAAE,MAAM,uCAAuC,CAAC;AAG3G;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;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,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,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,iDAAiD;CACjD,CAAC;AAEF,gHAAgH;AAChH,MAAM,0BAA0B,GAAwC,IAAI,GAAG,CAAyB;IACvG,iBAAiB;IACjB,uBAAuB;CACvB,CAAC,CAAC;AAEH,SAAS,0BAA0B,CAAC,YAAoB;IACvD,MAAM,UAAU,GAAG,YAAY,CAAC,IAAI,EAAE,CAAC,WAAW,EAAE,CAAC;IACrD,OAAO,CACN,UAAU,KAAK,2BAA2B;QAC1C,UAAU,KAAK,2BAA2B;QAC1C,UAAU,KAAK,eAAe;QAC9B,UAAU,KAAK,eAAe,CAC9B,CAAC;AACH,CAAC;AAED;;;;;;;;GAQG;AACH,MAAM,qBAAqB,GAAG;IAC7B,2CAA2C,EAAE,oFAAoF;IACjI,aAAa,EAAE,wBAAwB;IACvC,oBAAoB,EAAE,yBAAyB;CAC/C,CAAC;AAEF;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA0DG;AACH,MAAM,UAAU,iBAAiB,CAAC,OAAyB,EAAE,aAAsB;IAClF,gHAAgH;IAChH,yEAAyE;IACzE,MAAM,YAAY,GAAG,OAAO,CAAC,UAAU,KAAK,OAAO,CAAC,CAAC,CAAC,qBAAqB,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,SAAS,CAAC;IACjG,IAAI,YAAY,KAAK,SAAS,IAAI,CAAC,0BAA0B,CAAC,GAAG,CAAC,YAAY,CAAC,IAAI,CAAC,EAAE,CAAC;QACtF,OAAO,YAAY,CAAC,IAAI,KAAK,kBAAkB,CAAC;IACjD,CAAC;IAED,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,0BAA0B,CAAC,OAAO,CAAC,YAAY,CAAC,EAAE,CAAC;gBACzF,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 GitHubCopilotErrorCode, getGitHubCopilotError } from \"../providers/github-copilot-errors.ts\";\nimport 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 * - 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 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\n/** Typed GitHub Copilot rejections too general to rule overflow out; their provider text is checked instead. */\nconst TEXT_CHECKED_COPILOT_CODES: ReadonlySet<GitHubCopilotErrorCode> = new Set<GitHubCopilotErrorCode>([\n\t\"invalid-request\",\n\t\"unrecognized-response\",\n]);\n\nfunction isCerebrasBodylessOverflow(errorMessage: string): boolean {\n\tconst normalized = errorMessage.trim().toLowerCase();\n\treturn (\n\t\tnormalized === \"400 status code (no body)\" ||\n\t\tnormalized === \"413 status code (no body)\" ||\n\t\tnormalized === \"400 (no body)\" ||\n\t\tnormalized === \"413 (no body)\"\n\t);\n}\n\n/**\n * Patterns that indicate non-overflow errors (e.g. rate limiting, server errors).\n * Error messages matching any of these are excluded from overflow detection\n * even if they also match an OVERFLOW_PATTERN.\n *\n * Example: Bedrock formats throttling errors as \"ThrottlingException: Too many tokens,\n * please wait before trying again.\" which would match the /too many tokens/i overflow\n * pattern without this exclusion.\n */\nconst NON_OVERFLOW_PATTERNS = [\n\t/^(Throttling error|Service unavailable):/i, // AWS Bedrock non-overflow errors (human-readable prefixes from formatBedrockError)\n\t/rate limit/i, // Generic rate limiting\n\t/too many requests/i, // Generic HTTP 429 style\n];\n\n/**\n * Check if an assistant message represents a context overflow error.\n *\n * This handles three cases:\n * 1. Error-based overflow: Most providers return stopReason \"error\" with a\n * specific error message pattern.\n * 2. Silent overflow: Some providers accept overflow requests and return\n * successfully. For these, we check if usage.input exceeds the context window.\n * 3. Length-stop overflow: Xiaomi MiMo can return \"length\" with zero output when\n * the input fills the context window.\n *\n * ## Reliability by Provider\n *\n * **Typed detection:** a GitHub Copilot failure is an overflow when its typed error (see\n * `getGitHubCopilotError`) has the code `prompt-too-large`, and never for another specific code. A typed\n * `invalid-request` or `unrecognized-response` falls through to the text patterns below.\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// A GitHub Copilot failure carries its typed error, whose code decides, except for a rejection Atomic could not\n\t// name more precisely: its provider text is checked like any provider's.\n\tconst copilotError = message.stopReason === \"error\" ? getGitHubCopilotError(message) : undefined;\n\tif (copilotError !== undefined && !TEXT_CHECKED_COPILOT_CODES.has(copilotError.code)) {\n\t\treturn copilotError.code === \"prompt-too-large\";\n\t}\n\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\" && isCerebrasBodylessOverflow(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"]}
|
|
@@ -4,6 +4,8 @@ interface ProviderRetryOptions {
|
|
|
4
4
|
signal?: AbortSignal;
|
|
5
5
|
/** HTTP statuses that fail at once although the default policy would retry them. */
|
|
6
6
|
noRetryStatuses?: readonly number[];
|
|
7
|
+
/** Failures that fail at once although the default policy would retry them, for example a 429 for exhausted quota. */
|
|
8
|
+
noRetry?: (error: Error) => boolean;
|
|
7
9
|
}
|
|
8
10
|
/**
|
|
9
11
|
* Reproduce the retry behavior used by the OpenAI and Anthropic SDKs while making
|
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"provider-retry.d.ts","sourceRoot":"","sources":["../../src/utils/provider-retry.ts"],"names":[],"mappings":"AAEA,UAAU,oBAAoB;IAC7B,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,eAAe,CAAC,EAAE,MAAM,CAAC;IACzB,MAAM,CAAC,EAAE,WAAW,CAAC;IACrB,oFAAoF;IACpF,eAAe,CAAC,EAAE,SAAS,MAAM,EAAE,CAAC;CACpC;AA+FD;;;;;;;GAOG;AACH,wBAAsB,oBAAoB,CAAC,CAAC,EAC3C,OAAO,EAAE,MAAM,OAAO,CAAC,CAAC,CAAC,EACzB,OAAO,GAAE,oBAAyB,GAChC,OAAO,CAAC,CAAC,CAAC,
|
|
1
|
+
{"version":3,"file":"provider-retry.d.ts","sourceRoot":"","sources":["../../src/utils/provider-retry.ts"],"names":[],"mappings":"AAEA,UAAU,oBAAoB;IAC7B,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,eAAe,CAAC,EAAE,MAAM,CAAC;IACzB,MAAM,CAAC,EAAE,WAAW,CAAC;IACrB,oFAAoF;IACpF,eAAe,CAAC,EAAE,SAAS,MAAM,EAAE,CAAC;IACpC,sHAAsH;IACtH,OAAO,CAAC,EAAE,CAAC,KAAK,EAAE,KAAK,KAAK,OAAO,CAAC;CACpC;AA+FD;;;;;;;GAOG;AACH,wBAAsB,oBAAoB,CAAC,CAAC,EAC3C,OAAO,EAAE,MAAM,OAAO,CAAC,CAAC,CAAC,EACzB,OAAO,GAAE,oBAAyB,GAChC,OAAO,CAAC,CAAC,CAAC,CAmBZ"}
|
|
@@ -89,6 +89,8 @@ export async function retryProviderRequest(request, options = {}) {
|
|
|
89
89
|
throw error;
|
|
90
90
|
if (error.status !== undefined && options.noRetryStatuses?.includes(error.status))
|
|
91
91
|
throw error;
|
|
92
|
+
if (options.noRetry?.(error))
|
|
93
|
+
throw error;
|
|
92
94
|
const retryIndex = maxRetries - retriesRemaining;
|
|
93
95
|
retriesRemaining--;
|
|
94
96
|
await abortableSleep(getRetryDelayMs(error, retryIndex, options.maxRetryDelayMs), options.signal);
|
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"provider-retry.js","sourceRoot":"","sources":["../../src/utils/provider-retry.ts"],"names":[],"mappings":"AAAA,MAAM,0BAA0B,GAAG,MAAM,CAAC;
|
|
1
|
+
{"version":3,"file":"provider-retry.js","sourceRoot":"","sources":["../../src/utils/provider-retry.ts"],"names":[],"mappings":"AAAA,MAAM,0BAA0B,GAAG,MAAM,CAAC;AAiB1C,SAAS,eAAe,CAAC,KAAc;IACtC,IAAI,CAAC,CAAC,KAAK,YAAY,KAAK,CAAC,IAAI,CAAC,CAAC,QAAQ,IAAI,KAAK,CAAC,IAAI,CAAC,CAAC,SAAS,IAAI,KAAK,CAAC;QAAE,OAAO,KAAK,CAAC;IAC7F,OAAO,CACN,CAAC,KAAK,CAAC,MAAM,KAAK,SAAS,IAAI,OAAO,KAAK,CAAC,MAAM,KAAK,QAAQ,CAAC;QAChE,CAAC,KAAK,CAAC,OAAO,KAAK,SAAS,IAAI,KAAK,CAAC,OAAO,YAAY,OAAO,CAAC,CACjE,CAAC;AACH,CAAC;AAED,gGAAgG;AAChG,SAAS,wBAAwB,CAAC,KAAoB;IACrD,MAAM,WAAW,GAAG,KAAK,CAAC,OAAO,EAAE,GAAG,CAAC,gBAAgB,CAAC,CAAC;IACzD,IAAI,WAAW,KAAK,MAAM;QAAE,OAAO,IAAI,CAAC;IACxC,IAAI,WAAW,KAAK,OAAO;QAAE,OAAO,KAAK,CAAC;IAE1C,IAAI,KAAK,CAAC,MAAM,KAAK,SAAS;QAAE,OAAO,IAAI,CAAC;IAC5C,OAAO,CACN,KAAK,CAAC,MAAM,KAAK,GAAG;QACpB,KAAK,CAAC,MAAM,KAAK,GAAG;QACpB,KAAK,CAAC,MAAM,KAAK,GAAG;QACpB,CAAC,OAAO,KAAK,CAAC,MAAM,KAAK,QAAQ,IAAI,KAAK,CAAC,MAAM,IAAI,GAAG,CAAC,CACzD,CAAC;AACH,CAAC;AAED,SAAS,0BAA0B,CAClC,OAAe,EACf,eAAmC,EACnC,KAAoB;IAEpB,MAAM,UAAU,GAAG,eAAe,IAAI,0BAA0B,CAAC;IACjE,IAAI,UAAU,GAAG,CAAC,IAAI,OAAO,GAAG,UAAU,EAAE,CAAC;QAC5C,kGAAkG;QAClG,MAAM,MAAM,CAAC,MAAM,CAClB,IAAI,KAAK,CACR,oBAAoB,IAAI,CAAC,IAAI,CAAC,OAAO,GAAG,IAAI,CAAC,uBAAuB,IAAI,CAAC,IAAI,CAAC,UAAU,GAAG,IAAI,CAAC,OAAO,KAAK,CAAC,OAAO,EAAE,EACtH,EAAE,KAAK,EAAE,KAAK,EAAE,CAChB,EACD,EAAE,MAAM,EAAE,KAAK,CAAC,MAAM,EAAE,YAAY,EAAE,OAAO,EAAE,CAC/C,CAAC;IACH,CAAC;IACD,OAAO,OAAO,CAAC;AAChB,CAAC;AAED,SAAS,eAAe,CAAC,KAAoB,EAAE,UAAkB,EAAE,eAAmC;IACrG,MAAM,YAAY,GAAG,KAAK,CAAC,OAAO,EAAE,GAAG,CAAC,gBAAgB,CAAC,CAAC;IAC1D,IAAI,YAAY,EAAE,CAAC;QAClB,MAAM,KAAK,GAAG,MAAM,CAAC,UAAU,CAAC,YAAY,CAAC,CAAC;QAC9C,IAAI,MAAM,CAAC,QAAQ,CAAC,KAAK,CAAC;YAAE,OAAO,0BAA0B,CAAC,KAAK,EAAE,eAAe,EAAE,KAAK,CAAC,CAAC;IAC9F,CAAC;IAED,MAAM,UAAU,GAAG,KAAK,CAAC,OAAO,EAAE,GAAG,CAAC,aAAa,CAAC,CAAC;IACrD,IAAI,UAAU,EAAE,CAAC;QAChB,MAAM,OAAO,GAAG,MAAM,CAAC,UAAU,CAAC,UAAU,CAAC,CAAC;QAC9C,MAAM,OAAO,GAAG,MAAM,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,KAAK,CAAC,UAAU,CAAC,GAAG,IAAI,CAAC,GAAG,EAAE,CAAC,CAAC,CAAC,OAAO,GAAG,IAAI,CAAC;QAC7F,IAAI,MAAM,CAAC,QAAQ,CAAC,OAAO,CAAC;YAAE,OAAO,0BAA0B,CAAC,OAAO,EAAE,eAAe,EAAE,KAAK,CAAC,CAAC;IAClG,CAAC;IAED,MAAM,gBAAgB,GAAG,IAAI,CAAC,GAAG,CAAC,GAAG,GAAG,CAAC,IAAI,UAAU,EAAE,CAAC,CAAC,GAAG,IAAI,CAAC;IACnE,OAAO,gBAAgB,GAAG,CAAC,CAAC,GAAG,IAAI,CAAC,MAAM,EAAE,GAAG,IAAI,CAAC,CAAC;AACtD,CAAC;AAED,SAAS,gBAAgB;IACxB,MAAM,KAAK,GAAG,IAAI,KAAK,CAAC,iBAAiB,CAAC,CAAC;IAC3C,KAAK,CAAC,IAAI,GAAG,YAAY,CAAC;IAC1B,OAAO,KAAK,CAAC;AACd,CAAC;AAED,SAAS,cAAc,CAAC,EAAU,EAAE,MAAoB;IACvD,OAAO,IAAI,OAAO,CAAC,CAAC,OAAO,EAAE,MAAM,EAAE,EAAE;QACtC,IAAI,MAAM,EAAE,OAAO,EAAE,CAAC;YACrB,MAAM,CAAC,gBAAgB,EAAE,CAAC,CAAC;YAC3B,OAAO;QACR,CAAC;QAED,MAAM,OAAO,GAAG,GAAG,EAAE;YACpB,YAAY,CAAC,OAAO,CAAC,CAAC;YACtB,MAAM,CAAC,gBAAgB,EAAE,CAAC,CAAC;QAC5B,CAAC,CAAC;QACF,MAAM,OAAO,GAAG,UAAU,CACzB,GAAG,EAAE;YACJ,MAAM,EAAE,mBAAmB,CAAC,OAAO,EAAE,OAAO,CAAC,CAAC;YAC9C,OAAO,EAAE,CAAC;QACX,CAAC,EACD,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,EAAE,CAAC,CACf,CAAC;QACF,MAAM,EAAE,gBAAgB,CAAC,OAAO,EAAE,OAAO,EAAE,EAAE,IAAI,EAAE,IAAI,EAAE,CAAC,CAAC;IAC5D,CAAC,CAAC,CAAC;AACJ,CAAC;AAED;;;;;;;GAOG;AACH,MAAM,CAAC,KAAK,UAAU,oBAAoB,CACzC,OAAyB,EACzB,OAAO,GAAyB,EAAE;IAElC,MAAM,UAAU,GAAG,OAAO,CAAC,UAAU,IAAI,CAAC,CAAC;IAC3C,IAAI,gBAAgB,GAAG,UAAU,CAAC;IAElC,SAAS,CAAC;QACT,IAAI,CAAC;YACJ,8EAA8E;YAC9E,OAAO,MAAM,OAAO,EAAE,CAAC;QACxB,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YAChB,IAAI,OAAO,CAAC,MAAM,EAAE,OAAO;gBAAE,MAAM,gBAAgB,EAAE,CAAC;YACtD,IAAI,gBAAgB,IAAI,CAAC,IAAI,CAAC,eAAe,CAAC,KAAK,CAAC,IAAI,CAAC,wBAAwB,CAAC,KAAK,CAAC;gBAAE,MAAM,KAAK,CAAC;YACtG,IAAI,KAAK,CAAC,MAAM,KAAK,SAAS,IAAI,OAAO,CAAC,eAAe,EAAE,QAAQ,CAAC,KAAK,CAAC,MAAM,CAAC;gBAAE,MAAM,KAAK,CAAC;YAC/F,IAAI,OAAO,CAAC,OAAO,EAAE,CAAC,KAAK,CAAC;gBAAE,MAAM,KAAK,CAAC;YAE1C,MAAM,UAAU,GAAG,UAAU,GAAG,gBAAgB,CAAC;YACjD,gBAAgB,EAAE,CAAC;YACnB,MAAM,cAAc,CAAC,eAAe,CAAC,KAAK,EAAE,UAAU,EAAE,OAAO,CAAC,eAAe,CAAC,EAAE,OAAO,CAAC,MAAM,CAAC,CAAC;QACnG,CAAC;IACF,CAAC;AACF,CAAC","sourcesContent":["const DEFAULT_MAX_RETRY_DELAY_MS = 60_000;\n\ninterface ProviderRetryOptions {\n\tmaxRetries?: number;\n\tmaxRetryDelayMs?: number;\n\tsignal?: AbortSignal;\n\t/** HTTP statuses that fail at once although the default policy would retry them. */\n\tnoRetryStatuses?: readonly number[];\n\t/** Failures that fail at once although the default policy would retry them, for example a 429 for exhausted quota. */\n\tnoRetry?: (error: Error) => boolean;\n}\n\ninterface ProviderError extends Error {\n\tstatus: number | undefined;\n\theaders: Headers | undefined;\n}\n\nfunction isProviderError(error: unknown): error is ProviderError {\n\tif (!(error instanceof Error) || !(\"status\" in error) || !(\"headers\" in error)) return false;\n\treturn (\n\t\t(error.status === undefined || typeof error.status === \"number\") &&\n\t\t(error.headers === undefined || error.headers instanceof Headers)\n\t);\n}\n\n/** Mirrors the pinned OpenAI/Anthropic SDK retry policy; review when either SDK is upgraded. */\nfunction isRetryableProviderError(error: ProviderError): boolean {\n\tconst shouldRetry = error.headers?.get(\"x-should-retry\");\n\tif (shouldRetry === \"true\") return true;\n\tif (shouldRetry === \"false\") return false;\n\n\tif (error.status === undefined) return true;\n\treturn (\n\t\terror.status === 408 ||\n\t\terror.status === 409 ||\n\t\terror.status === 429 ||\n\t\t(typeof error.status === \"number\" && error.status >= 500)\n\t);\n}\n\nfunction validateServerRetryDelayMs(\n\tdelayMs: number,\n\tmaxRetryDelayMs: number | undefined,\n\terror: ProviderError,\n): number {\n\tconst maxDelayMs = maxRetryDelayMs ?? DEFAULT_MAX_RETRY_DELAY_MS;\n\tif (maxDelayMs > 0 && delayMs > maxDelayMs) {\n\t\t// Keep the server's hint and the original failure so callers can still schedule around the reset.\n\t\tthrow Object.assign(\n\t\t\tnew Error(\n\t\t\t\t`Server requested ${Math.ceil(delayMs / 1000)}s retry delay (max: ${Math.ceil(maxDelayMs / 1000)}s). ${error.message}`,\n\t\t\t\t{ cause: error },\n\t\t\t),\n\t\t\t{ status: error.status, retryAfterMs: delayMs },\n\t\t);\n\t}\n\treturn delayMs;\n}\n\nfunction getRetryDelayMs(error: ProviderError, retryIndex: number, maxRetryDelayMs: number | undefined): number {\n\tconst retryAfterMs = error.headers?.get(\"retry-after-ms\");\n\tif (retryAfterMs) {\n\t\tconst value = Number.parseFloat(retryAfterMs);\n\t\tif (Number.isFinite(value)) return validateServerRetryDelayMs(value, maxRetryDelayMs, error);\n\t}\n\n\tconst retryAfter = error.headers?.get(\"retry-after\");\n\tif (retryAfter) {\n\t\tconst seconds = Number.parseFloat(retryAfter);\n\t\tconst delayMs = Number.isNaN(seconds) ? Date.parse(retryAfter) - Date.now() : seconds * 1000;\n\t\tif (Number.isFinite(delayMs)) return validateServerRetryDelayMs(delayMs, maxRetryDelayMs, error);\n\t}\n\n\tconst exponentialDelay = Math.min(0.5 * 2 ** retryIndex, 8) * 1000;\n\treturn exponentialDelay * (1 - Math.random() * 0.25);\n}\n\nfunction createAbortError(): Error {\n\tconst error = new Error(\"Request aborted\");\n\terror.name = \"AbortError\";\n\treturn error;\n}\n\nfunction abortableSleep(ms: number, signal?: AbortSignal): Promise<void> {\n\treturn new Promise((resolve, reject) => {\n\t\tif (signal?.aborted) {\n\t\t\treject(createAbortError());\n\t\t\treturn;\n\t\t}\n\n\t\tconst onAbort = () => {\n\t\t\tclearTimeout(timeout);\n\t\t\treject(createAbortError());\n\t\t};\n\t\tconst timeout = setTimeout(\n\t\t\t() => {\n\t\t\t\tsignal?.removeEventListener(\"abort\", onAbort);\n\t\t\t\tresolve();\n\t\t\t},\n\t\t\tMath.max(0, ms),\n\t\t);\n\t\tsignal?.addEventListener(\"abort\", onAbort, { once: true });\n\t});\n}\n\n/**\n * Reproduce the retry behavior used by the OpenAI and Anthropic SDKs while making\n * their backoff sleep interruptible. Their built-in retry timers ignore the\n * request AbortSignal, so callers must invoke the SDK with `maxRetries: 0` and\n * wrap the request with this helper. Provider-requested delays above\n * `maxRetryDelayMs` fail immediately (60 seconds by default); set it to zero to\n * disable the limit.\n */\nexport async function retryProviderRequest<T>(\n\trequest: () => Promise<T>,\n\toptions: ProviderRetryOptions = {},\n): Promise<T> {\n\tconst maxRetries = options.maxRetries ?? 0;\n\tlet retriesRemaining = maxRetries;\n\n\tfor (;;) {\n\t\ttry {\n\t\t\t// Each retry is a fresh SDK request, so X-Stainless-Retry-Count remains zero.\n\t\t\treturn await request();\n\t\t} catch (error) {\n\t\t\tif (options.signal?.aborted) throw createAbortError();\n\t\t\tif (retriesRemaining <= 0 || !isProviderError(error) || !isRetryableProviderError(error)) throw error;\n\t\t\tif (error.status !== undefined && options.noRetryStatuses?.includes(error.status)) throw error;\n\t\t\tif (options.noRetry?.(error)) throw error;\n\n\t\t\tconst retryIndex = maxRetries - retriesRemaining;\n\t\t\tretriesRemaining--;\n\t\t\tawait abortableSleep(getRetryDelayMs(error, retryIndex, options.maxRetryDelayMs), options.signal);\n\t\t}\n\t}\n}\n"]}
|
package/dist/utils/retry.d.ts
CHANGED
|
@@ -22,7 +22,7 @@ export interface RetryPolicy {
|
|
|
22
22
|
}
|
|
23
23
|
export declare const DEFAULT_MAX_AGENT_RETRY_DELAY_MS = 60000;
|
|
24
24
|
export declare function retryDelayMs(policy: Pick<RetryPolicy, "baseDelayMs" | "maxAgentDelayMs">, attempt: number): number;
|
|
25
|
-
type RetryFailure = Pick<AssistantMessage, "stopReason" | "errorMessage" | "retryAfterMs">;
|
|
25
|
+
type RetryFailure = Pick<AssistantMessage, "stopReason" | "errorMessage" | "retryAfterMs" | "diagnostics">;
|
|
26
26
|
/**
|
|
27
27
|
* Wait before retry `attempt` (1-indexed): the server retry hint when the failure
|
|
28
28
|
* carries one, capped by `maxRetryDelayMs`, otherwise the exponential backoff.
|
|
@@ -67,6 +67,8 @@ export declare function retryAssistantCall(produce: () => Promise<AssistantMessa
|
|
|
67
67
|
* or transport error, so callers can decide if the last assistant turn should be
|
|
68
68
|
* restarted.
|
|
69
69
|
*
|
|
70
|
+
* A GitHub Copilot failure carries its typed error, whose `retryable` decides.
|
|
71
|
+
*
|
|
70
72
|
* This does not implement retry policy. Callers should first handle context
|
|
71
73
|
* overflow separately, then apply their own retry budget, backoff, and reporting
|
|
72
74
|
* before restarting the assistant turn.
|
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"retry.d.ts","sourceRoot":"","sources":["../../src/utils/retry.ts"],"names":[],"mappings":"
|
|
1
|
+
{"version":3,"file":"retry.d.ts","sourceRoot":"","sources":["../../src/utils/retry.ts"],"names":[],"mappings":"AACA,OAAO,KAAK,EAAE,gBAAgB,EAAE,MAAM,aAAa,CAAC;AAsIpD;;;;;;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,sEAAsE;IACtE,eAAe,CAAC,EAAE,MAAM,CAAC;IACzB;;;OAGG;IACH,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,KAAK,YAAY,GAAG,IAAI,CAAC,gBAAgB,EAAE,YAAY,GAAG,cAAc,GAAG,cAAc,GAAG,aAAa,CAAC,CAAC;AAE3G;;;GAGG;AACH,wBAAgB,mBAAmB,CAClC,MAAM,EAAE,IAAI,CAAC,WAAW,EAAE,aAAa,GAAG,iBAAiB,GAAG,iBAAiB,CAAC,EAChF,OAAO,EAAE,MAAM,EACf,OAAO,EAAE,IAAI,CAAC,gBAAgB,EAAE,cAAc,CAAC,GAC7C,MAAM,CAKR;AAED;;;GAGG;AACH,wBAAgB,qBAAqB,CAAC,MAAM,EAAE,WAAW,EAAE,OAAO,EAAE,MAAM,EAAE,OAAO,EAAE,YAAY,GAAG,MAAM,GAAG,SAAS,CAGrH;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;;;;;;;;;;;;;;;;;;GAkBG;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,CA4C3B;AAED;;;;;;;;;;GAUG;AACH,wBAAgB,yBAAyB,CAAC,OAAO,EAAE,YAAY,GAAG,OAAO,CAOxE"}
|
package/dist/utils/retry.js
CHANGED
|
@@ -1,3 +1,4 @@
|
|
|
1
|
+
import { getGitHubCopilotError } from "../providers/github-copilot-errors.js";
|
|
1
2
|
function buildProviderErrorPattern(patterns) {
|
|
2
3
|
return new RegExp(patterns.join("|"), "i");
|
|
3
4
|
}
|
|
@@ -98,6 +99,9 @@ const RETRYABLE_PROVIDER_ERROR_PATTERN = buildProviderErrorPattern([
|
|
|
98
99
|
// Codex CLI mid-stream disconnect, relayed by proxies that wrap the Codex CLI,
|
|
99
100
|
// e.g. "stream disconnected before completion: stream closed before response.completed" (#11057).
|
|
100
101
|
"stream disconnected before completion",
|
|
102
|
+
// Magpie gateway relaying a dropped upstream Anthropic stream, e.g.
|
|
103
|
+
// "Claude Code's stream from Anthropic was cut short after the reply had begun" (#11879).
|
|
104
|
+
"cut short after the reply had begun",
|
|
101
105
|
"pending stream has been canceled",
|
|
102
106
|
// Provider-requested retry delay cap failures should flow through the outer
|
|
103
107
|
// retry policy so callers can surface/abort the backoff (#1123).
|
|
@@ -223,6 +227,8 @@ export async function retryAssistantCall(produce, policy, signal, callbacks) {
|
|
|
223
227
|
* or transport error, so callers can decide if the last assistant turn should be
|
|
224
228
|
* restarted.
|
|
225
229
|
*
|
|
230
|
+
* A GitHub Copilot failure carries its typed error, whose `retryable` decides.
|
|
231
|
+
*
|
|
226
232
|
* This does not implement retry policy. Callers should first handle context
|
|
227
233
|
* overflow separately, then apply their own retry budget, backoff, and reporting
|
|
228
234
|
* before restarting the assistant turn.
|
|
@@ -230,6 +236,9 @@ export async function retryAssistantCall(produce, policy, signal, callbacks) {
|
|
|
230
236
|
export function isRetryableAssistantError(message) {
|
|
231
237
|
if (message.stopReason !== "error" || !message.errorMessage)
|
|
232
238
|
return false;
|
|
239
|
+
const copilotError = getGitHubCopilotError(message);
|
|
240
|
+
if (copilotError !== undefined)
|
|
241
|
+
return copilotError.retryable;
|
|
233
242
|
const errorMessage = message.errorMessage;
|
|
234
243
|
if (NON_RETRYABLE_PROVIDER_LIMIT_ERROR_PATTERN.test(errorMessage))
|
|
235
244
|
return false;
|
package/dist/utils/retry.js.map
CHANGED
|
@@ -1 +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;IACT,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,6EAA6E;IAC7E,8DAA8D;IAC9D,kEAAkE;IAClE,iDAAiD;IACjD,eAAe;IACf,iBAAiB;IAEjB,8EAA8E;IAC9E,+CAA+C;IAC/C,oCAAoC;IACpC,uDAAuD;IACvD,4BAA4B;IAC5B,2BAA2B;IAC3B,4BAA4B;IAC5B,wBAAwB;IACxB,uBAAuB;IAEvB,2EAA2E;IAC3E,2EAA2E;IAC3E,2EAA2E;IAC3E,gBAAgB;IAChB,mBAAmB;IACnB,qBAAqB;IACrB,kBAAkB;IAClB,YAAY;IACZ,mBAAmB;IACnB,mBAAmB;IACnB,cAAc;IACd,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,2EAA2E;IAC3E,0EAA0E;IAC1E,wEAAwE;IACxE,6DAA6D;IAC7D,4EAA4E;IAC5E,6EAA6E;IAC7E,8EAA8E;IAC9E,MAAM;IACN,wBAAwB;IACxB,YAAY;IACZ,eAAe;IAEf,6EAA6E;IAC7E,yEAAyE;IACzE,0BAA0B;IAE1B,yEAAyE;IACzE,8EAA8E;IAC9E,yEAAyE;IACzE,eAAe;IACf,kCAAkC;IAClC,+CAA+C;IAC/C,sCAAsC;IACtC,+EAA+E;IAC/E,kGAAkG;IAClG,uCAAuC;IACvC,kCAAkC;IAElC,4EAA4E;IAC5E,iEAAiE;IACjE,aAAa;IAEb,6EAA6E;IAC7E,6BAA6B;IAC7B,4BAA4B;IAC5B,wBAAwB;IACxB,2BAA2B;IAE3B,yCAAyC;IACzC,mBAAmB;IACnB,wCAAwC;IACxC,uCAAuC;CACvC,CAAC,CAAC;AAwBH,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;AAID;;;GAGG;AACH,MAAM,UAAU,mBAAmB,CAClC,MAAgF,EAChF,OAAe,EACf,OAA+C;IAE/C,MAAM,MAAM,GAAG,OAAO,CAAC,YAAY,CAAC;IACpC,IAAI,MAAM,KAAK,SAAS,IAAI,CAAC,MAAM,CAAC,QAAQ,CAAC,MAAM,CAAC,IAAI,MAAM,IAAI,CAAC;QAAE,OAAO,YAAY,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC;IAC1G,MAAM,KAAK,GAAG,MAAM,CAAC,eAAe,IAAI,CAAC,MAAM,CAAC,eAAe,IAAI,gCAAgC,CAAC,CAAC;IACrG,OAAO,IAAI,CAAC,GAAG,CAAC,MAAM,EAAE,KAAK,CAAC,CAAC;AAChC,CAAC;AAED;;;GAGG;AACH,MAAM,UAAU,qBAAqB,CAAC,MAAmB,EAAE,OAAe,EAAE,OAAqB;IAChG,IAAI,CAAC,MAAM,CAAC,OAAO,IAAI,OAAO,GAAG,MAAM,CAAC,UAAU,IAAI,CAAC,yBAAyB,CAAC,OAAO,CAAC;QAAE,OAAO,SAAS,CAAC;IAC5G,OAAO,mBAAmB,CAAC,MAAM,EAAE,OAAO,EAAE,OAAO,CAAC,CAAC;AACtD,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;;;;;;;;;;;;;;;;;;GAkBG;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,MAAM,OAAO,GAAG,MAAM,CAAC,CAAC,CAAC,qBAAqB,CAAC,MAAM,EAAE,OAAO,GAAG,CAAC,EAAE,QAAQ,CAAC,CAAC,CAAC,CAAC,SAAS,CAAC;QAC1F,IAAI,OAAO,KAAK,SAAS,EAAE,CAAC;YAC3B,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,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,OAAO,EAAE,GAAG,QAAQ,EAAE,UAAU,EAAE,SAAS,EAAE,YAAY,EAAE,SAAS,EAAE,CAAC;YACxE,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,OAAqB;IAC9D,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\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// 521-523 are Cloudflare origin errors (#3623); 529 is an overload response.\n\t\"(?<![a-z0-9])(?:429|500|502|503|504|52[0-4]|529)(?![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).\n\t\"provider.?returned.?(?:an.?)?error\",\n\t\"exceeded request buffer limit while retrying upstream\",\n\t\"unable to forward upstream\",\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\"ECONNRESET\",\n\t\"connection.?reset\",\n\t\"other side closed\",\n\t\"fetch failed\",\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// Response-body decompression failures. Node/Bun zlib rejects a body whose\n\t// Content-Encoding lied or whose compressed bytes were corrupted, and the\n\t// fetch/WebSocket layer wraps it with a \"Library error:\" prefix. GitHub\n\t// Copilot on the default `transport: \"auto\"` reports this as\n\t// \"Library error: zlib error: incorrect header check\" (#2553). The transfer\n\t// is already unusable, and a fresh attempt normally succeeds, so treat every\n\t// variant as a transient transport failure rather than a deterministic error.\n\t\"zlib\",\n\t\"incorrect header check\",\n\t\"decompress\",\n\t\"library error\",\n\n\t// Idle-stream deadline enforced below the HTTP layer by `withStreamDeadline`\n\t// so a stream that never settles cannot hang an attempt forever (#2553).\n\t\"stream deadline exceeded\",\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\"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\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 * Each delay is capped by `maxAgentDelayMs` (60 seconds by default).\n * Matches `settings.retry` (`enabled`, `maxRetries`, `baseDelayMs`) 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/** Cap for agent-level retry delays in ms. Defaults to 60 seconds. */\n\tmaxAgentDelayMs?: number;\n\t/**\n\t * Cap for a server retry hint (`Retry-After-Ms`, `Retry-After`, a body `retry_after`).\n\t * Defaults to 60 seconds; 0 caps the hint by `maxAgentDelayMs` instead.\n\t */\n\tmaxRetryDelayMs?: 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\ntype RetryFailure = Pick<AssistantMessage, \"stopReason\" | \"errorMessage\" | \"retryAfterMs\">;\n\n/**\n * Wait before retry `attempt` (1-indexed): the server retry hint when the failure\n * carries one, capped by `maxRetryDelayMs`, otherwise the exponential backoff.\n */\nexport function failureRetryDelayMs(\n\tpolicy: Pick<RetryPolicy, \"baseDelayMs\" | \"maxAgentDelayMs\" | \"maxRetryDelayMs\">,\n\tattempt: number,\n\tfailure: Pick<AssistantMessage, \"retryAfterMs\">,\n): number {\n\tconst hintMs = failure.retryAfterMs;\n\tif (hintMs === undefined || !Number.isFinite(hintMs) || hintMs <= 0) return retryDelayMs(policy, attempt);\n\tconst capMs = policy.maxRetryDelayMs || (policy.maxAgentDelayMs ?? DEFAULT_MAX_AGENT_RETRY_DELAY_MS);\n\treturn Math.min(hintMs, capMs);\n}\n\n/**\n * Delay before retry `attempt` (1-indexed) of a failed assistant call, or `undefined`\n * when it is not retryable or `maxRetries` is spent.\n */\nexport function assistantRetryDelayMs(policy: RetryPolicy, attempt: number, message: RetryFailure): number | undefined {\n\tif (!policy.enabled || attempt > policy.maxRetries || !isRetryableAssistantError(message)) return undefined;\n\treturn failureRetryDelayMs(policy, attempt, message);\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, waiting for the server retry hint or\n * the exponential backoff (see {@link assistantRetryDelayMs}), 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\tconst delayMs = policy ? assistantRetryDelayMs(policy, attempt + 1, response) : undefined;\n\t\tif (delayMs === undefined) {\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\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\treturn { ...response, stopReason: \"aborted\", errorMessage: undefined };\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: RetryFailure): 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"]}
|
|
1
|
+
{"version":3,"file":"retry.js","sourceRoot":"","sources":["../../src/utils/retry.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,qBAAqB,EAAE,MAAM,uCAAuC,CAAC;AAG9E,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;IACT,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,6EAA6E;IAC7E,8DAA8D;IAC9D,kEAAkE;IAClE,iDAAiD;IACjD,eAAe;IACf,iBAAiB;IAEjB,8EAA8E;IAC9E,+CAA+C;IAC/C,oCAAoC;IACpC,uDAAuD;IACvD,4BAA4B;IAC5B,2BAA2B;IAC3B,4BAA4B;IAC5B,wBAAwB;IACxB,uBAAuB;IAEvB,2EAA2E;IAC3E,2EAA2E;IAC3E,2EAA2E;IAC3E,gBAAgB;IAChB,mBAAmB;IACnB,qBAAqB;IACrB,kBAAkB;IAClB,YAAY;IACZ,mBAAmB;IACnB,mBAAmB;IACnB,cAAc;IACd,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,2EAA2E;IAC3E,0EAA0E;IAC1E,wEAAwE;IACxE,6DAA6D;IAC7D,4EAA4E;IAC5E,6EAA6E;IAC7E,8EAA8E;IAC9E,MAAM;IACN,wBAAwB;IACxB,YAAY;IACZ,eAAe;IAEf,6EAA6E;IAC7E,yEAAyE;IACzE,0BAA0B;IAE1B,yEAAyE;IACzE,8EAA8E;IAC9E,yEAAyE;IACzE,eAAe;IACf,kCAAkC;IAClC,+CAA+C;IAC/C,sCAAsC;IACtC,+EAA+E;IAC/E,kGAAkG;IAClG,uCAAuC;IACvC,oEAAoE;IACpE,0FAA0F;IAC1F,qCAAqC;IACrC,kCAAkC;IAElC,4EAA4E;IAC5E,iEAAiE;IACjE,aAAa;IAEb,6EAA6E;IAC7E,6BAA6B;IAC7B,4BAA4B;IAC5B,wBAAwB;IACxB,2BAA2B;IAE3B,yCAAyC;IACzC,mBAAmB;IACnB,wCAAwC;IACxC,uCAAuC;CACvC,CAAC,CAAC;AAwBH,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;AAID;;;GAGG;AACH,MAAM,UAAU,mBAAmB,CAClC,MAAgF,EAChF,OAAe,EACf,OAA+C;IAE/C,MAAM,MAAM,GAAG,OAAO,CAAC,YAAY,CAAC;IACpC,IAAI,MAAM,KAAK,SAAS,IAAI,CAAC,MAAM,CAAC,QAAQ,CAAC,MAAM,CAAC,IAAI,MAAM,IAAI,CAAC;QAAE,OAAO,YAAY,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC;IAC1G,MAAM,KAAK,GAAG,MAAM,CAAC,eAAe,IAAI,CAAC,MAAM,CAAC,eAAe,IAAI,gCAAgC,CAAC,CAAC;IACrG,OAAO,IAAI,CAAC,GAAG,CAAC,MAAM,EAAE,KAAK,CAAC,CAAC;AAChC,CAAC;AAED;;;GAGG;AACH,MAAM,UAAU,qBAAqB,CAAC,MAAmB,EAAE,OAAe,EAAE,OAAqB;IAChG,IAAI,CAAC,MAAM,CAAC,OAAO,IAAI,OAAO,GAAG,MAAM,CAAC,UAAU,IAAI,CAAC,yBAAyB,CAAC,OAAO,CAAC;QAAE,OAAO,SAAS,CAAC;IAC5G,OAAO,mBAAmB,CAAC,MAAM,EAAE,OAAO,EAAE,OAAO,CAAC,CAAC;AACtD,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;;;;;;;;;;;;;;;;;;GAkBG;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,MAAM,OAAO,GAAG,MAAM,CAAC,CAAC,CAAC,qBAAqB,CAAC,MAAM,EAAE,OAAO,GAAG,CAAC,EAAE,QAAQ,CAAC,CAAC,CAAC,CAAC,SAAS,CAAC;QAC1F,IAAI,OAAO,KAAK,SAAS,EAAE,CAAC;YAC3B,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,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,OAAO,EAAE,GAAG,QAAQ,EAAE,UAAU,EAAE,SAAS,EAAE,YAAY,EAAE,SAAS,EAAE,CAAC;YACxE,CAAC;YACD,MAAM,KAAK,CAAC;QACb,CAAC;QACD,MAAM,SAAS,EAAE,mBAAmB,EAAE,EAAE,CAAC;IAC1C,CAAC;AACF,CAAC;AAED;;;;;;;;;;GAUG;AACH,MAAM,UAAU,yBAAyB,CAAC,OAAqB;IAC9D,IAAI,OAAO,CAAC,UAAU,KAAK,OAAO,IAAI,CAAC,OAAO,CAAC,YAAY;QAAE,OAAO,KAAK,CAAC;IAC1E,MAAM,YAAY,GAAG,qBAAqB,CAAC,OAAO,CAAC,CAAC;IACpD,IAAI,YAAY,KAAK,SAAS;QAAE,OAAO,YAAY,CAAC,SAAS,CAAC;IAC9D,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 { getGitHubCopilotError } from \"../providers/github-copilot-errors.ts\";\nimport 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\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// 521-523 are Cloudflare origin errors (#3623); 529 is an overload response.\n\t\"(?<![a-z0-9])(?:429|500|502|503|504|52[0-4]|529)(?![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).\n\t\"provider.?returned.?(?:an.?)?error\",\n\t\"exceeded request buffer limit while retrying upstream\",\n\t\"unable to forward upstream\",\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\"ECONNRESET\",\n\t\"connection.?reset\",\n\t\"other side closed\",\n\t\"fetch failed\",\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// Response-body decompression failures. Node/Bun zlib rejects a body whose\n\t// Content-Encoding lied or whose compressed bytes were corrupted, and the\n\t// fetch/WebSocket layer wraps it with a \"Library error:\" prefix. GitHub\n\t// Copilot on the default `transport: \"auto\"` reports this as\n\t// \"Library error: zlib error: incorrect header check\" (#2553). The transfer\n\t// is already unusable, and a fresh attempt normally succeeds, so treat every\n\t// variant as a transient transport failure rather than a deterministic error.\n\t\"zlib\",\n\t\"incorrect header check\",\n\t\"decompress\",\n\t\"library error\",\n\n\t// Idle-stream deadline enforced below the HTTP layer by `withStreamDeadline`\n\t// so a stream that never settles cannot hang an attempt forever (#2553).\n\t\"stream deadline exceeded\",\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// Magpie gateway relaying a dropped upstream Anthropic stream, e.g.\n\t// \"Claude Code's stream from Anthropic was cut short after the reply had begun\" (#11879).\n\t\"cut short after the reply had begun\",\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\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 * Each delay is capped by `maxAgentDelayMs` (60 seconds by default).\n * Matches `settings.retry` (`enabled`, `maxRetries`, `baseDelayMs`) 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/** Cap for agent-level retry delays in ms. Defaults to 60 seconds. */\n\tmaxAgentDelayMs?: number;\n\t/**\n\t * Cap for a server retry hint (`Retry-After-Ms`, `Retry-After`, a body `retry_after`).\n\t * Defaults to 60 seconds; 0 caps the hint by `maxAgentDelayMs` instead.\n\t */\n\tmaxRetryDelayMs?: 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\ntype RetryFailure = Pick<AssistantMessage, \"stopReason\" | \"errorMessage\" | \"retryAfterMs\" | \"diagnostics\">;\n\n/**\n * Wait before retry `attempt` (1-indexed): the server retry hint when the failure\n * carries one, capped by `maxRetryDelayMs`, otherwise the exponential backoff.\n */\nexport function failureRetryDelayMs(\n\tpolicy: Pick<RetryPolicy, \"baseDelayMs\" | \"maxAgentDelayMs\" | \"maxRetryDelayMs\">,\n\tattempt: number,\n\tfailure: Pick<AssistantMessage, \"retryAfterMs\">,\n): number {\n\tconst hintMs = failure.retryAfterMs;\n\tif (hintMs === undefined || !Number.isFinite(hintMs) || hintMs <= 0) return retryDelayMs(policy, attempt);\n\tconst capMs = policy.maxRetryDelayMs || (policy.maxAgentDelayMs ?? DEFAULT_MAX_AGENT_RETRY_DELAY_MS);\n\treturn Math.min(hintMs, capMs);\n}\n\n/**\n * Delay before retry `attempt` (1-indexed) of a failed assistant call, or `undefined`\n * when it is not retryable or `maxRetries` is spent.\n */\nexport function assistantRetryDelayMs(policy: RetryPolicy, attempt: number, message: RetryFailure): number | undefined {\n\tif (!policy.enabled || attempt > policy.maxRetries || !isRetryableAssistantError(message)) return undefined;\n\treturn failureRetryDelayMs(policy, attempt, message);\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, waiting for the server retry hint or\n * the exponential backoff (see {@link assistantRetryDelayMs}), 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\tconst delayMs = policy ? assistantRetryDelayMs(policy, attempt + 1, response) : undefined;\n\t\tif (delayMs === undefined) {\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\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\treturn { ...response, stopReason: \"aborted\", errorMessage: undefined };\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 * A GitHub Copilot failure carries its typed error, whose `retryable` decides.\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: RetryFailure): boolean {\n\tif (message.stopReason !== \"error\" || !message.errorMessage) return false;\n\tconst copilotError = getGitHubCopilotError(message);\n\tif (copilotError !== undefined) return copilotError.retryable;\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"]}
|
|
@@ -0,0 +1,23 @@
|
|
|
1
|
+
import type { Api, AssistantMessage, Model, ThinkingLevel } from "../types.ts";
|
|
2
|
+
/** Diagnostic type recorded when thinking `off` ran at a model's lowest supported level instead. */
|
|
3
|
+
export declare const THINKING_OFF_FALLBACK_DIAGNOSTIC = "thinking_off_fallback";
|
|
4
|
+
/**
|
|
5
|
+
* The level a request for thinking `off` runs at on a reasoning model whose `thinkingLevelMap.off` is
|
|
6
|
+
* `null`: its lowest supported level. `undefined` when the model can disable thinking or has no level.
|
|
7
|
+
*/
|
|
8
|
+
export declare function getThinkingOffFallbackLevel(model: Model<Api>): ThinkingLevel | undefined;
|
|
9
|
+
export declare function appendThinkingOffFallbackDiagnostic(output: AssistantMessage, thinkingLevel: ThinkingLevel, providerThinkingLevel: string): void;
|
|
10
|
+
/**
|
|
11
|
+
* Resolve thinking `off` for a request that sends a level's `thinkingLevelMap` value as the effort, such as
|
|
12
|
+
* the Responses, Chat Completions and Mistral APIs. Falls back to the lowest supported level the map lists
|
|
13
|
+
* an effort for, because these endpoints reject an effort the model does not support, and a level the map
|
|
14
|
+
* leaves unlisted does not say which effort is safe. Returns the fallback level, recorded on `output`.
|
|
15
|
+
*/
|
|
16
|
+
export declare function recordThinkingOffFallback(model: Model<Api>, output: AssistantMessage): ThinkingLevel | undefined;
|
|
17
|
+
/**
|
|
18
|
+
* Resolve thinking `off` to the model's lowest supported level, for an endpoint that maps any effort onto the
|
|
19
|
+
* ones the model supports. The level's `thinkingLevelMap` value, or the level itself when the map lists none,
|
|
20
|
+
* is the effort sent. Returns the fallback level, recorded on `output`.
|
|
21
|
+
*/
|
|
22
|
+
export declare function recordLowestThinkingOffFallback(model: Model<Api>, output: AssistantMessage): ThinkingLevel | undefined;
|
|
23
|
+
//# sourceMappingURL=thinking-off.d.ts.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"thinking-off.d.ts","sourceRoot":"","sources":["../../src/utils/thinking-off.ts"],"names":[],"mappings":"AACA,OAAO,KAAK,EAAE,GAAG,EAAE,gBAAgB,EAAE,KAAK,EAAE,aAAa,EAAE,MAAM,aAAa,CAAC;AAG/E,oGAAoG;AACpG,eAAO,MAAM,gCAAgC,0BAA0B,CAAC;AAExE;;;GAGG;AACH,wBAAgB,2BAA2B,CAAC,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,GAAG,aAAa,GAAG,SAAS,CAGxF;AAED,wBAAgB,mCAAmC,CAClD,MAAM,EAAE,gBAAgB,EACxB,aAAa,EAAE,aAAa,EAC5B,qBAAqB,EAAE,MAAM,GAC3B,IAAI,CAMN;AAED;;;;;GAKG;AACH,wBAAgB,yBAAyB,CAAC,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,EAAE,MAAM,EAAE,gBAAgB,GAAG,aAAa,GAAG,SAAS,CAShH;AAED;;;;GAIG;AACH,wBAAgB,+BAA+B,CAC9C,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,EACjB,MAAM,EAAE,gBAAgB,GACtB,aAAa,GAAG,SAAS,CAM3B"}
|
|
@@ -0,0 +1,51 @@
|
|
|
1
|
+
import { clampThinkingLevel, getSupportedThinkingLevels } from "../models.js";
|
|
2
|
+
import { appendAssistantMessageDiagnostic } from "./diagnostics.js";
|
|
3
|
+
/** Diagnostic type recorded when thinking `off` ran at a model's lowest supported level instead. */
|
|
4
|
+
export const THINKING_OFF_FALLBACK_DIAGNOSTIC = "thinking_off_fallback";
|
|
5
|
+
/**
|
|
6
|
+
* The level a request for thinking `off` runs at on a reasoning model whose `thinkingLevelMap.off` is
|
|
7
|
+
* `null`: its lowest supported level. `undefined` when the model can disable thinking or has no level.
|
|
8
|
+
*/
|
|
9
|
+
export function getThinkingOffFallbackLevel(model) {
|
|
10
|
+
const level = clampThinkingLevel(model, "off");
|
|
11
|
+
return level === "off" ? undefined : level;
|
|
12
|
+
}
|
|
13
|
+
export function appendThinkingOffFallbackDiagnostic(output, thinkingLevel, providerThinkingLevel) {
|
|
14
|
+
appendAssistantMessageDiagnostic(output, {
|
|
15
|
+
type: THINKING_OFF_FALLBACK_DIAGNOSTIC,
|
|
16
|
+
timestamp: Date.now(),
|
|
17
|
+
details: { requestedThinkingLevel: "off", thinkingLevel, providerThinkingLevel },
|
|
18
|
+
});
|
|
19
|
+
}
|
|
20
|
+
/**
|
|
21
|
+
* Resolve thinking `off` for a request that sends a level's `thinkingLevelMap` value as the effort, such as
|
|
22
|
+
* the Responses, Chat Completions and Mistral APIs. Falls back to the lowest supported level the map lists
|
|
23
|
+
* an effort for, because these endpoints reject an effort the model does not support, and a level the map
|
|
24
|
+
* leaves unlisted does not say which effort is safe. Returns the fallback level, recorded on `output`.
|
|
25
|
+
*/
|
|
26
|
+
export function recordThinkingOffFallback(model, output) {
|
|
27
|
+
for (const level of getSupportedThinkingLevels(model)) {
|
|
28
|
+
if (level === "off")
|
|
29
|
+
return undefined;
|
|
30
|
+
const effort = model.thinkingLevelMap?.[level];
|
|
31
|
+
if (typeof effort !== "string")
|
|
32
|
+
continue;
|
|
33
|
+
appendThinkingOffFallbackDiagnostic(output, level, effort);
|
|
34
|
+
return level;
|
|
35
|
+
}
|
|
36
|
+
return undefined;
|
|
37
|
+
}
|
|
38
|
+
/**
|
|
39
|
+
* Resolve thinking `off` to the model's lowest supported level, for an endpoint that maps any effort onto the
|
|
40
|
+
* ones the model supports. The level's `thinkingLevelMap` value, or the level itself when the map lists none,
|
|
41
|
+
* is the effort sent. Returns the fallback level, recorded on `output`.
|
|
42
|
+
*/
|
|
43
|
+
export function recordLowestThinkingOffFallback(model, output) {
|
|
44
|
+
const level = getThinkingOffFallbackLevel(model);
|
|
45
|
+
if (level === undefined)
|
|
46
|
+
return undefined;
|
|
47
|
+
const effort = model.thinkingLevelMap?.[level];
|
|
48
|
+
appendThinkingOffFallbackDiagnostic(output, level, typeof effort === "string" ? effort : level);
|
|
49
|
+
return level;
|
|
50
|
+
}
|
|
51
|
+
//# sourceMappingURL=thinking-off.js.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"thinking-off.js","sourceRoot":"","sources":["../../src/utils/thinking-off.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,kBAAkB,EAAE,0BAA0B,EAAE,MAAM,cAAc,CAAC;AAE9E,OAAO,EAAE,gCAAgC,EAAE,MAAM,kBAAkB,CAAC;AAEpE,oGAAoG;AACpG,MAAM,CAAC,MAAM,gCAAgC,GAAG,uBAAuB,CAAC;AAExE;;;GAGG;AACH,MAAM,UAAU,2BAA2B,CAAC,KAAiB;IAC5D,MAAM,KAAK,GAAG,kBAAkB,CAAC,KAAK,EAAE,KAAK,CAAC,CAAC;IAC/C,OAAO,KAAK,KAAK,KAAK,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,KAAK,CAAC;AAC5C,CAAC;AAED,MAAM,UAAU,mCAAmC,CAClD,MAAwB,EACxB,aAA4B,EAC5B,qBAA6B;IAE7B,gCAAgC,CAAC,MAAM,EAAE;QACxC,IAAI,EAAE,gCAAgC;QACtC,SAAS,EAAE,IAAI,CAAC,GAAG,EAAE;QACrB,OAAO,EAAE,EAAE,sBAAsB,EAAE,KAAK,EAAE,aAAa,EAAE,qBAAqB,EAAE;KAChF,CAAC,CAAC;AACJ,CAAC;AAED;;;;;GAKG;AACH,MAAM,UAAU,yBAAyB,CAAC,KAAiB,EAAE,MAAwB;IACpF,KAAK,MAAM,KAAK,IAAI,0BAA0B,CAAC,KAAK,CAAC,EAAE,CAAC;QACvD,IAAI,KAAK,KAAK,KAAK;YAAE,OAAO,SAAS,CAAC;QACtC,MAAM,MAAM,GAAG,KAAK,CAAC,gBAAgB,EAAE,CAAC,KAAK,CAAC,CAAC;QAC/C,IAAI,OAAO,MAAM,KAAK,QAAQ;YAAE,SAAS;QACzC,mCAAmC,CAAC,MAAM,EAAE,KAAK,EAAE,MAAM,CAAC,CAAC;QAC3D,OAAO,KAAK,CAAC;IACd,CAAC;IACD,OAAO,SAAS,CAAC;AAClB,CAAC;AAED;;;;GAIG;AACH,MAAM,UAAU,+BAA+B,CAC9C,KAAiB,EACjB,MAAwB;IAExB,MAAM,KAAK,GAAG,2BAA2B,CAAC,KAAK,CAAC,CAAC;IACjD,IAAI,KAAK,KAAK,SAAS;QAAE,OAAO,SAAS,CAAC;IAC1C,MAAM,MAAM,GAAG,KAAK,CAAC,gBAAgB,EAAE,CAAC,KAAK,CAAC,CAAC;IAC/C,mCAAmC,CAAC,MAAM,EAAE,KAAK,EAAE,OAAO,MAAM,KAAK,QAAQ,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC;IAChG,OAAO,KAAK,CAAC;AACd,CAAC","sourcesContent":["import { clampThinkingLevel, getSupportedThinkingLevels } from \"../models.ts\";\nimport type { Api, AssistantMessage, Model, ThinkingLevel } from \"../types.ts\";\nimport { appendAssistantMessageDiagnostic } from \"./diagnostics.ts\";\n\n/** Diagnostic type recorded when thinking `off` ran at a model's lowest supported level instead. */\nexport const THINKING_OFF_FALLBACK_DIAGNOSTIC = \"thinking_off_fallback\";\n\n/**\n * The level a request for thinking `off` runs at on a reasoning model whose `thinkingLevelMap.off` is\n * `null`: its lowest supported level. `undefined` when the model can disable thinking or has no level.\n */\nexport function getThinkingOffFallbackLevel(model: Model<Api>): ThinkingLevel | undefined {\n\tconst level = clampThinkingLevel(model, \"off\");\n\treturn level === \"off\" ? undefined : level;\n}\n\nexport function appendThinkingOffFallbackDiagnostic(\n\toutput: AssistantMessage,\n\tthinkingLevel: ThinkingLevel,\n\tproviderThinkingLevel: string,\n): void {\n\tappendAssistantMessageDiagnostic(output, {\n\t\ttype: THINKING_OFF_FALLBACK_DIAGNOSTIC,\n\t\ttimestamp: Date.now(),\n\t\tdetails: { requestedThinkingLevel: \"off\", thinkingLevel, providerThinkingLevel },\n\t});\n}\n\n/**\n * Resolve thinking `off` for a request that sends a level's `thinkingLevelMap` value as the effort, such as\n * the Responses, Chat Completions and Mistral APIs. Falls back to the lowest supported level the map lists\n * an effort for, because these endpoints reject an effort the model does not support, and a level the map\n * leaves unlisted does not say which effort is safe. Returns the fallback level, recorded on `output`.\n */\nexport function recordThinkingOffFallback(model: Model<Api>, output: AssistantMessage): ThinkingLevel | undefined {\n\tfor (const level of getSupportedThinkingLevels(model)) {\n\t\tif (level === \"off\") return undefined;\n\t\tconst effort = model.thinkingLevelMap?.[level];\n\t\tif (typeof effort !== \"string\") continue;\n\t\tappendThinkingOffFallbackDiagnostic(output, level, effort);\n\t\treturn level;\n\t}\n\treturn undefined;\n}\n\n/**\n * Resolve thinking `off` to the model's lowest supported level, for an endpoint that maps any effort onto the\n * ones the model supports. The level's `thinkingLevelMap` value, or the level itself when the map lists none,\n * is the effort sent. Returns the fallback level, recorded on `output`.\n */\nexport function recordLowestThinkingOffFallback(\n\tmodel: Model<Api>,\n\toutput: AssistantMessage,\n): ThinkingLevel | undefined {\n\tconst level = getThinkingOffFallbackLevel(model);\n\tif (level === undefined) return undefined;\n\tconst effort = model.thinkingLevelMap?.[level];\n\tappendThinkingOffFallbackDiagnostic(output, level, typeof effort === \"string\" ? effort : level);\n\treturn level;\n}\n"]}
|
package/dist/utils/version.d.ts
CHANGED
|
@@ -1,2 +1,2 @@
|
|
|
1
|
-
export declare const PI_AI_VERSION = "0.9.
|
|
1
|
+
export declare const PI_AI_VERSION = "0.9.72-alpha.1";
|
|
2
2
|
//# sourceMappingURL=version.d.ts.map
|
package/dist/utils/version.js
CHANGED
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"version.js","sourceRoot":"","sources":["../../src/utils/version.ts"],"names":[],"mappings":"AAAA,kGAAkG;AAClG,MAAM,CAAC,MAAM,aAAa,GAAG,gBAAgB,CAAC","sourcesContent":["// The @bastani/pi-ai package version. scripts/bump-version.ts keeps it in sync with package.json.\nexport const PI_AI_VERSION = \"0.9.
|
|
1
|
+
{"version":3,"file":"version.js","sourceRoot":"","sources":["../../src/utils/version.ts"],"names":[],"mappings":"AAAA,kGAAkG;AAClG,MAAM,CAAC,MAAM,aAAa,GAAG,gBAAgB,CAAC","sourcesContent":["// The @bastani/pi-ai package version. scripts/bump-version.ts keeps it in sync with package.json.\nexport const PI_AI_VERSION = \"0.9.72-alpha.1\";\n"]}
|
package/package.json
CHANGED
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
{
|
|
2
2
|
"name": "@bastani/pi-ai",
|
|
3
|
-
"version": "0.9.
|
|
3
|
+
"version": "0.9.72-alpha.1",
|
|
4
4
|
"description": "Bastani-branded unified LLM API with automatic model discovery and provider configuration",
|
|
5
5
|
"type": "module",
|
|
6
6
|
"main": "./dist/index.js",
|
|
@@ -46,6 +46,14 @@
|
|
|
46
46
|
"./bun-oauth": {
|
|
47
47
|
"types": "./dist/bun-oauth.d.ts",
|
|
48
48
|
"import": "./dist/bun-oauth.js"
|
|
49
|
+
},
|
|
50
|
+
"./lazy-bedrock-provider": {
|
|
51
|
+
"types": "./dist/lazy-bedrock-provider.d.ts",
|
|
52
|
+
"import": "./dist/lazy-bedrock-provider.js"
|
|
53
|
+
},
|
|
54
|
+
"./lazy-oauth": {
|
|
55
|
+
"types": "./dist/lazy-oauth.d.ts",
|
|
56
|
+
"import": "./dist/lazy-oauth.js"
|
|
49
57
|
}
|
|
50
58
|
},
|
|
51
59
|
"bin": {
|