@juspay/neurolink 9.15.0 → 9.17.0

package/dist/constants/enums.js

@@ -28,22 +28,36 @@ export var AIProviderName;
 export var OpenRouterModels;
 (function (OpenRouterModels) {
     // Anthropic Claude models
+    OpenRouterModels["CLAUDE_OPUS_4_6"] = "anthropic/claude-opus-4.6";
+    OpenRouterModels["CLAUDE_SONNET_4_6"] = "anthropic/claude-sonnet-4.6";
+    OpenRouterModels["CLAUDE_SONNET_4_5"] = "anthropic/claude-sonnet-4.5";
+    OpenRouterModels["CLAUDE_HAIKU_4_5"] = "anthropic/claude-haiku-4.5";
     OpenRouterModels["CLAUDE_3_5_SONNET"] = "anthropic/claude-3-5-sonnet";
     OpenRouterModels["CLAUDE_3_5_HAIKU"] = "anthropic/claude-3-5-haiku";
     OpenRouterModels["CLAUDE_3_OPUS"] = "anthropic/claude-3-opus";
     // OpenAI models
+    OpenRouterModels["GPT_5_2"] = "openai/gpt-5.2";
+    OpenRouterModels["GPT_5"] = "openai/gpt-5";
     OpenRouterModels["GPT_4O"] = "openai/gpt-4o";
     OpenRouterModels["GPT_4O_MINI"] = "openai/gpt-4o-mini";
     OpenRouterModels["GPT_4_TURBO"] = "openai/gpt-4-turbo";
     // Google models
+    OpenRouterModels["GEMINI_3_1_PRO_PREVIEW"] = "google/gemini-3.1-pro-preview";
+    OpenRouterModels["GEMINI_3_PRO_PREVIEW"] = "google/gemini-3-pro-preview";
+    OpenRouterModels["GEMINI_3_FLASH_PREVIEW"] = "google/gemini-3-flash-preview";
+    OpenRouterModels["GEMINI_2_5_FLASH"] = "google/gemini-2.5-flash";
+    OpenRouterModels["GEMINI_2_5_FLASH_LITE"] = "google/gemini-2.5-flash-lite";
     OpenRouterModels["GEMINI_2_0_FLASH"] = "google/gemini-2.0-flash";
-    OpenRouterModels["GEMINI_1_5_PRO"] = "google/gemini-1.5-pro";
     // Meta Llama models
     OpenRouterModels["LLAMA_3_1_70B"] = "meta-llama/llama-3.1-70b-instruct";
     OpenRouterModels["LLAMA_3_1_8B"] = "meta-llama/llama-3.1-8b-instruct";
     // Mistral models
     OpenRouterModels["MISTRAL_LARGE"] = "mistralai/mistral-large";
     OpenRouterModels["MIXTRAL_8X7B"] = "mistralai/mixtral-8x7b-instruct";
+    // DeepSeek models
+    OpenRouterModels["DEEPSEEK_R1"] = "deepseek/deepseek-r1";
+    // xAI models
+    OpenRouterModels["GROK_4_1_FAST"] = "xai/grok-4.1-fast";
 })(OpenRouterModels || (OpenRouterModels = {}));
 /**
  * Supported Models for Amazon Bedrock
@@ -53,7 +67,10 @@ export var BedrockModels;
     // ============================================================================
     // ANTHROPIC CLAUDE MODELS
     // ============================================================================
-    // Claude 4.5 Series (Latest - September-November 2025)
+    // Claude 4.6 Series (Latest - February 2026)
+    BedrockModels["CLAUDE_4_6_OPUS"] = "anthropic.claude-opus-4-6-v1:0";
+    BedrockModels["CLAUDE_4_6_SONNET"] = "anthropic.claude-sonnet-4-6";
+    // Claude 4.5 Series (September-November 2025)
     BedrockModels["CLAUDE_4_5_OPUS"] = "anthropic.claude-opus-4-5-20251124-v1:0";
     BedrockModels["CLAUDE_4_5_SONNET"] = "anthropic.claude-sonnet-4-5-20250929-v1:0";
     BedrockModels["CLAUDE_4_5_HAIKU"] = "anthropic.claude-haiku-4-5-20251001-v1:0";
@@ -160,20 +177,66 @@ export var BedrockModels;
     // AI21 Labs Models
     BedrockModels["JAMBA_1_5_LARGE"] = "ai21.jamba-1-5-large-v1:0";
     BedrockModels["JAMBA_1_5_MINI"] = "ai21.jamba-1-5-mini-v1:0";
+    // ============================================================================
+    // NEW PROVIDERS (February 2026)
+    // ============================================================================
+    // Writer Models
+    BedrockModels["WRITER_PALMYRA_X5"] = "writer.palmyra-x5-v1:0";
+    BedrockModels["WRITER_PALMYRA_X4"] = "writer.palmyra-x4-v1:0";
+    // MiniMax Models
+    BedrockModels["MINIMAX_M2_1"] = "minimax.minimax-m2.1";
+    BedrockModels["MINIMAX_M2"] = "minimax.minimax-m2";
+    // Moonshot AI (Kimi) Models
+    BedrockModels["KIMI_K2_THINKING"] = "moonshot.kimi-k2-thinking";
+    BedrockModels["KIMI_K2_5"] = "moonshotai.kimi-k2.5";
+    // NVIDIA Nemotron Models
+    BedrockModels["NVIDIA_NEMOTRON_NANO_3_30B"] = "nvidia.nemotron-nano-3-30b";
+    BedrockModels["NVIDIA_NEMOTRON_NANO_12B_V2"] = "nvidia.nemotron-nano-12b-v2";
+    BedrockModels["NVIDIA_NEMOTRON_NANO_9B_V2"] = "nvidia.nemotron-nano-9b-v2";
+    // OpenAI Open Source Models (Apache 2.0)
+    BedrockModels["OPENAI_GPT_OSS_120B"] = "openai.gpt-oss-120b-1:0";
+    BedrockModels["OPENAI_GPT_OSS_20B"] = "openai.gpt-oss-20b-1:0";
+    // Z.AI GLM Models
+    BedrockModels["GLM_4_7"] = "zai.glm-4.7";
+    BedrockModels["GLM_4_7_FLASH"] = "zai.glm-4.7-flash";
+    // Cohere Embedding & Reranking
+    BedrockModels["COHERE_EMBED_ENGLISH_V3"] = "cohere.embed-english-v3";
+    BedrockModels["COHERE_EMBED_MULTILINGUAL_V3"] = "cohere.embed-multilingual-v3";
+    BedrockModels["COHERE_EMBED_V4"] = "cohere.embed-v4:0";
+    BedrockModels["COHERE_RERANK_V3_5"] = "cohere.rerank-v3-5:0";
+    // Amazon Rerank
+    BedrockModels["AMAZON_RERANK_V1"] = "amazon.rerank-v1:0";
+    // Mistral Devstral 2
+    BedrockModels["DEVSTRAL_2_123B"] = "mistral.devstral-2-123b";
 })(BedrockModels || (BedrockModels = {}));
 /**
  * Supported Models for OpenAI
  */
 export var OpenAIModels;
 (function (OpenAIModels) {
-    // GPT-5.2 Series (Released December 11, 2025) - Latest flagship models
+    // GPT-5.3 Series (Released February 2026) - Latest coding models
+    OpenAIModels["GPT_5_3_CODEX"] = "gpt-5.3-codex";
+    // GPT-5.2 Series (Released December 11, 2025) - Flagship models
     OpenAIModels["GPT_5_2"] = "gpt-5.2";
     OpenAIModels["GPT_5_2_CHAT_LATEST"] = "gpt-5.2-chat-latest";
     OpenAIModels["GPT_5_2_PRO"] = "gpt-5.2-pro";
+    OpenAIModels["GPT_5_2_CODEX"] = "gpt-5.2-codex";
+    // GPT-5.1 Series (Released October 2025)
+    OpenAIModels["GPT_5_1"] = "gpt-5.1";
+    OpenAIModels["GPT_5_1_CHAT_LATEST"] = "gpt-5.1-chat-latest";
+    OpenAIModels["GPT_5_1_CODEX"] = "gpt-5.1-codex";
+    OpenAIModels["GPT_5_1_CODEX_MAX"] = "gpt-5.1-codex-max";
+    OpenAIModels["GPT_5_1_CODEX_MINI"] = "gpt-5.1-codex-mini";
     // GPT-5 Series (Released August 7, 2025)
     OpenAIModels["GPT_5"] = "gpt-5";
     OpenAIModels["GPT_5_MINI"] = "gpt-5-mini";
     OpenAIModels["GPT_5_NANO"] = "gpt-5-nano";
+    OpenAIModels["GPT_5_PRO"] = "gpt-5-pro";
+    OpenAIModels["GPT_5_CHAT_LATEST"] = "gpt-5-chat-latest";
+    OpenAIModels["GPT_5_CODEX"] = "gpt-5-codex";
+    // GPT Open Source (Apache 2.0 - January 2026, Responses API only)
+    OpenAIModels["GPT_OSS_120B"] = "gpt-oss-120b";
+    OpenAIModels["GPT_OSS_20B"] = "gpt-oss-20b";
     // GPT-4.1 Series (Released April 14, 2025)
     OpenAIModels["GPT_4_1"] = "gpt-4.1";
     OpenAIModels["GPT_4_1_MINI"] = "gpt-4.1-mini";
@@ -201,7 +264,12 @@ export var OpenAIModels;
  */
 export var AzureOpenAIModels;
 (function (AzureOpenAIModels) {
-    // GPT-5.1 Series (Latest - December 2025)
+    // GPT-5.2 Series (Latest - December 2025)
+    AzureOpenAIModels["GPT_5_2"] = "gpt-5.2";
+    AzureOpenAIModels["GPT_5_2_CHAT"] = "gpt-5.2-chat";
+    AzureOpenAIModels["GPT_5_2_PRO"] = "gpt-5.2-pro";
+    AzureOpenAIModels["GPT_5_2_CODEX"] = "gpt-5.2-codex";
+    // GPT-5.1 Series (October 2025)
     AzureOpenAIModels["GPT_5_1"] = "gpt-5.1";
     AzureOpenAIModels["GPT_5_1_CHAT"] = "gpt-5.1-chat";
     AzureOpenAIModels["GPT_5_1_CODEX"] = "gpt-5.1-codex";
@@ -244,7 +312,10 @@ export var AzureOpenAIModels;
  */
 export var VertexModels;
 (function (VertexModels) {
-    // Claude 4.5 Series (Latest - December 2025)
+    // Claude 4.6 Series (Latest - February 2026)
+    VertexModels["CLAUDE_4_6_OPUS"] = "claude-opus-4-6";
+    VertexModels["CLAUDE_4_6_SONNET"] = "claude-sonnet-4-6";
+    // Claude 4.5 Series (September-November 2025)
     VertexModels["CLAUDE_4_5_OPUS"] = "claude-opus-4-5@20251124";
     VertexModels["CLAUDE_4_5_SONNET"] = "claude-sonnet-4-5@20250929";
     VertexModels["CLAUDE_4_5_HAIKU"] = "claude-haiku-4-5@20251001";
@@ -260,33 +331,29 @@ export var VertexModels;
     VertexModels["CLAUDE_3_SONNET"] = "claude-3-sonnet-20240229";
     VertexModels["CLAUDE_3_OPUS"] = "claude-3-opus-20240229";
     VertexModels["CLAUDE_3_HAIKU"] = "claude-3-haiku-20240307";
+    // Gemini 3.1 Series (Released February 2026)
+    VertexModels["GEMINI_3_1_PRO_PREVIEW"] = "gemini-3.1-pro-preview";
     // Gemini 3 Series (Preview)
-    /** Gemini 3 Pro - Base model with adaptive thinking */
     VertexModels["GEMINI_3_PRO"] = "gemini-3-pro";
-    /** Gemini 3 Pro Preview - Versioned preview (November 2025) */
     VertexModels["GEMINI_3_PRO_PREVIEW_11_2025"] = "gemini-3-pro-preview-11-2025";
-    /** Gemini 3 Pro Latest - Auto-updated alias (always points to latest preview) */
     VertexModels["GEMINI_3_PRO_LATEST"] = "gemini-3-pro-latest";
-    /** Gemini 3 Pro Preview - Generic preview (legacy) */
     VertexModels["GEMINI_3_PRO_PREVIEW"] = "gemini-3-pro-preview";
-    /** Gemini 3 Flash - Base model with adaptive thinking */
     VertexModels["GEMINI_3_FLASH"] = "gemini-3-flash";
-    /** Gemini 3 Flash Preview - Versioned preview */
     VertexModels["GEMINI_3_FLASH_PREVIEW"] = "gemini-3-flash-preview";
-    /** Gemini 3 Flash Latest - Auto-updated alias (always points to latest preview) */
     VertexModels["GEMINI_3_FLASH_LATEST"] = "gemini-3-flash-latest";
-    // Gemini 2.5 Series (Latest - 2025)
+    // Gemini 2.5 Series (GA)
     VertexModels["GEMINI_2_5_PRO"] = "gemini-2.5-pro";
     VertexModels["GEMINI_2_5_FLASH"] = "gemini-2.5-flash";
     VertexModels["GEMINI_2_5_FLASH_LITE"] = "gemini-2.5-flash-lite";
     VertexModels["GEMINI_2_5_FLASH_IMAGE"] = "gemini-2.5-flash-image";
-    // Gemini 2.0 Series
+    // Gemini 2.0 Series (Deprecated - retiring Jun 2026)
     VertexModels["GEMINI_2_0_FLASH"] = "gemini-2.0-flash";
     VertexModels["GEMINI_2_0_FLASH_001"] = "gemini-2.0-flash-001";
-    /** Gemini 2.0 Flash Lite - GA, production-ready, cost-optimized */
     VertexModels["GEMINI_2_0_FLASH_LITE"] = "gemini-2.0-flash-lite";
-    // Gemini 1.5 Series (Legacy support)
+    // Gemini 1.5 Series (Retired - returns 404)
+    /** @deprecated Retired Sep 2025. Use Gemini 2.5+ */
     VertexModels["GEMINI_1_5_PRO"] = "gemini-1.5-pro-002";
+    /** @deprecated Retired Sep 2025. Use Gemini 2.5+ */
     VertexModels["GEMINI_1_5_FLASH"] = "gemini-1.5-flash-002";
 })(VertexModels || (VertexModels = {}));
 /**
@@ -294,27 +361,33 @@ export var VertexModels;
  */
 export var GoogleAIModels;
 (function (GoogleAIModels) {
-    // Gemini 3 Series
+    // Gemini 3.1 Series (Released February 2026)
+    GoogleAIModels["GEMINI_3_1_PRO_PREVIEW"] = "gemini-3.1-pro-preview";
+    // Gemini 3 Series (Preview)
     GoogleAIModels["GEMINI_3_PRO_PREVIEW"] = "gemini-3-pro-preview";
     GoogleAIModels["GEMINI_3_PRO_IMAGE_PREVIEW"] = "gemini-3-pro-image-preview";
     GoogleAIModels["GEMINI_3_FLASH"] = "gemini-3-flash";
     GoogleAIModels["GEMINI_3_FLASH_PREVIEW"] = "gemini-3-flash-preview";
-    // Gemini 2.5 Series
+    // Gemini 2.5 Series (GA)
     GoogleAIModels["GEMINI_2_5_PRO"] = "gemini-2.5-pro";
     GoogleAIModels["GEMINI_2_5_FLASH"] = "gemini-2.5-flash";
     GoogleAIModels["GEMINI_2_5_FLASH_LITE"] = "gemini-2.5-flash-lite";
     GoogleAIModels["GEMINI_2_5_FLASH_IMAGE"] = "gemini-2.5-flash-image";
-    GoogleAIModels["GEMINI_2_5_FLASH_LIVE"] = "gemini-2.5-flash-native-audio-preview-09-2025";
-    // Gemini 2.0 Series
+    GoogleAIModels["GEMINI_2_5_FLASH_PREVIEW_TTS"] = "gemini-2.5-flash-preview-tts";
+    GoogleAIModels["GEMINI_2_5_PRO_PREVIEW_TTS"] = "gemini-2.5-pro-preview-tts";
+    // Gemini 2.0 Series (Deprecated - retiring Jun 2026)
     GoogleAIModels["GEMINI_2_0_FLASH"] = "gemini-2.0-flash";
     GoogleAIModels["GEMINI_2_0_FLASH_001"] = "gemini-2.0-flash-001";
     GoogleAIModels["GEMINI_2_0_FLASH_LITE"] = "gemini-2.0-flash-lite";
     GoogleAIModels["GEMINI_2_0_FLASH_IMAGE"] = "gemini-2.0-flash-preview-image-generation";
-    // Gemini 1.5 Series (Legacy)
+    // Gemini 1.5 Series (Retired - returns 404)
+    /** @deprecated Retired Sep 2025. Use Gemini 2.5+ */
     GoogleAIModels["GEMINI_1_5_PRO"] = "gemini-1.5-pro";
+    /** @deprecated Retired Sep 2025. Use Gemini 2.5+ */
     GoogleAIModels["GEMINI_1_5_FLASH"] = "gemini-1.5-flash";
     // Embedding Models
     GoogleAIModels["GEMINI_EMBEDDING"] = "gemini-embedding-001";
+    /** @deprecated Shutdown Jan 2026. Use gemini-embedding-001 */
     GoogleAIModels["TEXT_EMBEDDING_004"] = "text-embedding-004";
 })(GoogleAIModels || (GoogleAIModels = {}));
 /**
@@ -322,7 +395,10 @@ export var GoogleAIModels;
  */
 export var AnthropicModels;
 (function (AnthropicModels) {
-    // Claude 4.5 Series (Latest - September-November 2025)
+    // Claude 4.6 Series (Latest - February 2026)
+    AnthropicModels["CLAUDE_OPUS_4_6"] = "claude-opus-4-6";
+    AnthropicModels["CLAUDE_SONNET_4_6"] = "claude-sonnet-4-6";
+    // Claude 4.5 Series (September-November 2025)
     AnthropicModels["CLAUDE_OPUS_4_5"] = "claude-opus-4-5-20251101";
     AnthropicModels["CLAUDE_SONNET_4_5"] = "claude-sonnet-4-5-20250929";
     AnthropicModels["CLAUDE_4_5_HAIKU"] = "claude-haiku-4-5-20251001";
@@ -375,6 +451,17 @@ export var MistralModels;
     // Voxtral (Audio)
     MistralModels["VOXTRAL_SMALL_LATEST"] = "voxtral-small-latest";
     MistralModels["VOXTRAL_MINI_LATEST"] = "voxtral-mini-latest";
+    // Devstral 2 Series (December 2025)
+    MistralModels["DEVSTRAL_2"] = "devstral-2512";
+    MistralModels["DEVSTRAL_SMALL_2"] = "devstral-small-2512";
+    // Magistral Versioned (September 2025)
+    MistralModels["MAGISTRAL_MEDIUM_2509"] = "magistral-medium-2509";
+    MistralModels["MAGISTRAL_SMALL_2509"] = "magistral-small-2509";
+    // Voxtral Transcribe 2 (February 2026)
+    MistralModels["VOXTRAL_MINI_TRANSCRIBE_2"] = "voxtral-mini-2602";
+    // OCR (December 2025)
+    MistralModels["MISTRAL_OCR_3"] = "mistral-ocr-2512";
+    MistralModels["MISTRAL_OCR_LATEST"] = "mistral-ocr-latest";
     // Specialized Models
     MistralModels["MISTRAL_NEMO"] = "mistral-nemo";
     MistralModels["MISTRAL_EMBED"] = "mistral-embed";
@@ -464,6 +551,24 @@ export var OllamaModels;
     OllamaModels["MIXTRAL_8X22B"] = "mixtral:8x22b";
     // Enterprise Models
     OllamaModels["COMMAND_R_PLUS"] = "command-r-plus:104b";
+    // Z.AI GLM-5 - Flagship reasoning model (February 2026)
+    OllamaModels["GLM_5_LATEST"] = "glm-5:latest";
+    // Kimi-K2.5 - Moonshot AI multimodal agentic model
+    OllamaModels["KIMI_K2_5_LATEST"] = "kimi-k2.5:latest";
+    // Qwen 3.5 - Multimodal native agents (February 2026)
+    OllamaModels["QWEN3_5_LATEST"] = "qwen3.5:latest";
+    // Qwen3-Coder - Coding-focused agentic model
+    OllamaModels["QWEN3_CODER_LATEST"] = "qwen3-coder:latest";
+    OllamaModels["QWEN3_CODER_30B"] = "qwen3-coder:30b";
+    // DeepSeek-V3.2 - Enhanced reasoning
+    OllamaModels["DEEPSEEK_V3_2_LATEST"] = "deepseek-v3.2:latest";
+    // NVIDIA Nemotron 3 Nano - Hybrid MoE, 1M context
+    OllamaModels["NEMOTRON_3_NANO_LATEST"] = "nemotron-3-nano:latest";
+    OllamaModels["NEMOTRON_3_NANO_30B"] = "nemotron-3-nano:30b";
+    // SmolLM3 - Compact dual-mode reasoning (HuggingFace)
+    OllamaModels["SMOLLM3_3B"] = "smollm3:3b";
+    // GPT-OSS - Open-source GPT (Apache 2.0)
+    OllamaModels["GPT_OSS_LATEST"] = "gpt-oss:latest";
 })(OllamaModels || (OllamaModels = {}));
 /**
  * Common Models for LiteLLM Proxy
@@ -513,6 +618,16 @@ export var LiteLLMModels;
     // AWS Bedrock via LiteLLM
     LiteLLMModels["BEDROCK_CLAUDE_3_5_SONNET"] = "bedrock/anthropic.claude-3-5-sonnet-20240620-v1:0";
     LiteLLMModels["BEDROCK_CLAUDE_3_HAIKU"] = "bedrock/anthropic.claude-3-haiku-20240307-v1:0";
+    // OpenAI GPT-5.2 via LiteLLM
+    LiteLLMModels["OPENAI_GPT_5_2"] = "openai/gpt-5.2";
+    LiteLLMModels["OPENAI_GPT_5_2_CODEX"] = "openai/gpt-5.2-codex";
+    // Anthropic Claude 4.6 via LiteLLM
+    LiteLLMModels["ANTHROPIC_CLAUDE_OPUS_4_6"] = "anthropic/claude-opus-4-6";
+    LiteLLMModels["ANTHROPIC_CLAUDE_SONNET_4_6"] = "anthropic/claude-sonnet-4-6";
+    // Google Gemini 3 via LiteLLM
+    LiteLLMModels["GEMINI_3_1_PRO"] = "gemini/gemini-3.1-pro-preview";
+    // xAI via LiteLLM
+    LiteLLMModels["XAI_GROK_4_1_FAST"] = "xai/grok-4.1-fast";
     // Perplexity AI via LiteLLM
     LiteLLMModels["PERPLEXITY_SONAR_PRO"] = "perplexity/sonar-pro";
     LiteLLMModels["PERPLEXITY_SONAR_REASONING_PRO"] = "perplexity/sonar-reasoning-pro";
@@ -588,6 +703,17 @@ export var HuggingFaceModels;
     // BLOOM
     HuggingFaceModels["BLOOM_7B1"] = "bigscience/bloom-7b1";
     HuggingFaceModels["BLOOM_1B3"] = "bigscience/bloom-1b3";
+    // Z.AI GLM-5 (February 2026)
+    HuggingFaceModels["GLM_5"] = "zai-org/GLM-5";
+    // Qwen 3.5 Multimodal (February 2026)
+    HuggingFaceModels["QWEN_3_5_397B_A17B"] = "Qwen/Qwen3.5-397B-A17B";
+    // NVIDIA Nemotron 3 Nano
+    HuggingFaceModels["NEMOTRON_3_NANO_30B"] = "nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16";
+    // HuggingFace SmolLM3
+    HuggingFaceModels["SMOLLM3_3B"] = "HuggingFaceTB/SmolLM3-3B";
+    // Falcon 3 Series
+    HuggingFaceModels["FALCON_3_7B_INSTRUCT"] = "tiiuae/Falcon3-7B-Instruct";
+    HuggingFaceModels["FALCON_3_10B_INSTRUCT"] = "tiiuae/Falcon3-10B-Instruct";
 })(HuggingFaceModels || (HuggingFaceModels = {}));
 /**
  * Supported Models for AWS SageMaker JumpStart
@@ -616,6 +742,10 @@ export var SageMakerModels;
     SageMakerModels["FALCON_3_10B"] = "tii-falcon-3-10b-instruct";
     SageMakerModels["FALCON_40B"] = "tii-falcon-40b-instruct";
     SageMakerModels["FALCON_180B"] = "tii-falcon-180b";
+    // NVIDIA Nemotron 3 Nano (February 2026)
+    SageMakerModels["NEMOTRON_3_NANO_30B"] = "nvidia-nemotron-3-nano-30b";
+    // Qwen3 VL - Vision-language
+    SageMakerModels["QWEN3_VL_8B"] = "qwen3-vl-8b-instruct";
 })(SageMakerModels || (SageMakerModels = {}));
 /**
  * API Versions for various providers

package/dist/context/budgetChecker.js

@@ -36,7 +36,7 @@ export function checkContextBudget(params) {
         ? toolDefinitions.reduce((sum, tool) => {
             try {
                 const serialized = JSON.stringify(tool);
-                return sum + Math.ceil(serialized.length / 4);
+                return sum + estimateTokens(serialized, provider);
             }
             catch {
                 return sum + TOKENS_PER_TOOL_DEFINITION;

package/dist/context/contextCompactor.js

@@ -8,6 +8,7 @@
  * Stage 3: LLM Summarization (expensive -- requires LLM call)
  * Stage 4: Sliding Window Truncation (fallback -- no LLM call)
  */
+import { trace, SpanStatusCode } from "@opentelemetry/api";
 import { estimateMessagesTokens } from "../utils/tokenEstimation.js";
 import { logger } from "../utils/logger.js";
 import { pruneToolOutputs } from "./stages/toolOutputPruner.js";
@@ -112,15 +113,36 @@ export class ContextCompactor {
                     saved: stageTokensBefore - stageTokensAfter,
                 });
             }
-            catch {
-                logger.info("[Compaction] Stage 3 (summarize)", {
+            catch (error) {
+                // Capture the actual error for debugging
+                const errorMessage = error instanceof Error ? error.message : String(error);
+                const errorName = error instanceof Error ? error.name : "UnknownError";
+                logger.warn("[Compaction] Stage 3 (summarize) FAILED", {
                     requestId,
-                    ran: false,
+                    error: errorMessage,
+                    errorName,
                     tokensBefore: stageTokensBefore,
                     tokensAfter: stageTokensBefore,
                     saved: 0,
                 });
-                // Summarization failed, fall through to truncation
+                // Record on OTel span for trace visibility
+                const activeSpan = trace.getActiveSpan();
+                if (activeSpan) {
+                    activeSpan.addEvent("compaction.stage3.failed", {
+                        "error.message": errorMessage,
+                        "error.name": errorName,
+                        "stage.tokens_before": stageTokensBefore,
+                    });
+                    if (error instanceof Error) {
+                        activeSpan.recordException(error);
+                    }
+                    // NLK-GAP-005 fix: set error status alongside recordException
+                    activeSpan.setStatus({
+                        code: SpanStatusCode.ERROR,
+                        message: `Compaction stage 3 (summarize) failed: ${errorMessage}`,
+                    });
+                }
+                // Fall through to Stage 4 truncation as before
             }
         }
         // Stage 4: Sliding Window Truncation (fallback)
@@ -129,6 +151,11 @@ export class ContextCompactor {
             const stageTokensBefore = estimateMessagesTokens(currentMessages, provider);
             const truncResult = truncateWithSlidingWindow(currentMessages, {
                 fraction: this.config.truncationFraction,
+                currentTokens: stageTokensBefore,
+                targetTokens: targetTokens,
+                provider: provider,
+                adaptiveBuffer: 0.15,
+                maxIterations: 3,
             });
             if (truncResult.truncated) {
                 currentMessages = truncResult.messages;

package/dist/context/emergencyTruncation.d.ts

@@ -0,0 +1,21 @@
+/**
+ * Emergency Content Truncation
+ *
+ * When message-level removal (sliding window) can't fit context into budget,
+ * this truncates the CONTENT of the longest messages as a last resort.
+ */
+import type { ChatMessage } from "../types/conversation.js";
+/**
+ * Emergency content truncation: truncate the content of the longest messages
+ * to fit within the available token budget.
+ *
+ * Strategy: Sort messages by content length (descending), truncate each
+ * to a proportional share of the available budget until total fits.
+ */
+export declare function emergencyContentTruncation(messages: ChatMessage[], availableTokensForHistory: number, breakdown: {
+    systemPrompt: number;
+    conversationHistory: number;
+    currentPrompt: number;
+    toolDefinitions: number;
+    fileAttachments: number;
+}, provider?: string): ChatMessage[];

package/dist/context/emergencyTruncation.js

@@ -0,0 +1,88 @@
+/**
+ * Emergency Content Truncation
+ *
+ * When message-level removal (sliding window) can't fit context into budget,
+ * this truncates the CONTENT of the longest messages as a last resort.
+ */
+import { estimateTokens, estimateMessagesTokens, truncateToTokenBudget, } from "../utils/tokenEstimation.js";
+import { logger } from "../utils/logger.js";
+/**
+ * Emergency content truncation: truncate the content of the longest messages
+ * to fit within the available token budget.
+ *
+ * Strategy: Sort messages by content length (descending), truncate each
+ * to a proportional share of the available budget until total fits.
+ */
+export function emergencyContentTruncation(messages, availableTokensForHistory, breakdown, provider) {
+    // Budget available for conversation history specifically
+    const historyBudget = availableTokensForHistory -
+        breakdown.systemPrompt -
+        breakdown.currentPrompt -
+        breakdown.toolDefinitions -
+        breakdown.fileAttachments;
+    if (historyBudget <= 0) {
+        // No room for history: return empty to guarantee budget safety
+        return [];
+    }
+    const currentHistoryTokens = estimateMessagesTokens(messages, provider);
+    if (currentHistoryTokens <= historyBudget) {
+        return messages; // Already fits
+    }
+    // Calculate per-message budgets proportional to original size,
+    // but cap large messages to free space for others
+    const result = [...messages];
+    const reductionNeeded = currentHistoryTokens - historyBudget;
+    const reductionRatio = reductionNeeded / currentHistoryTokens;
+    // Sort indices by content length descending (truncate biggest first)
+    const sortedIndices = result
+        .map((msg, idx) => ({ idx, len: msg.content.length }))
+        .sort((a, b) => b.len - a.len);
+    let tokensSaved = 0;
+    for (const { idx } of sortedIndices) {
+        if (tokensSaved >= reductionNeeded) {
+            break;
+        }
+        const msg = result[idx];
+        // Don't truncate system messages or very short messages
+        if (msg.role === "system" || msg.content.length < 200) {
+            continue;
+        }
+        const msgTokens = estimateTokens(msg.content, provider);
+        const targetTokens = Math.floor(msgTokens * (1 - reductionRatio - 0.05));
+        if (targetTokens < msgTokens && targetTokens > 50) {
+            const truncated = truncateToTokenBudget(msg.content, targetTokens, provider);
+            if (truncated.truncated) {
+                const savedThisMsg = msgTokens - estimateTokens(truncated.text, provider);
+                tokensSaved += savedThisMsg;
+                result[idx] = {
+                    ...msg,
+                    content: truncated.text,
+                    metadata: { ...msg.metadata, truncated: true },
+                };
+            }
+        }
+    }
+    logger.info("[EmergencyTruncation] Content truncation complete", {
+        tokensSaved,
+        reductionNeeded,
+        messagesModified: result.filter((m, i) => m !== messages[i]).length,
+    });
+    // Final safety check: guarantee returned history fits budget
+    if (estimateMessagesTokens(result, provider) <= historyBudget) {
+        return result;
+    }
+    // Hard fallback: keep newest non-system messages that fit
+    const fallback = [];
+    for (let i = result.length - 1; i >= 0; i--) {
+        const msg = result[i];
+        if (msg.role === "system") {
+            continue;
+        }
+        fallback.unshift(msg);
+        if (estimateMessagesTokens(fallback, provider) > historyBudget) {
+            fallback.shift();
+            break;
+        }
+    }
+    return fallback;
+}

package/dist/context/errorDetection.d.ts

@@ -12,3 +12,19 @@ export declare function isContextOverflowError(error: unknown): boolean;
  * Identify which provider produced the context overflow error.
  */
 export declare function getContextOverflowProvider(error: unknown): string | null;
+/**
+ * Extract actual token counts from provider overflow error messages.
+ *
+ * Many providers include the actual/max token counts in their error messages:
+ * - OpenAI: "This model's maximum context length is 128000 tokens. However, your messages resulted in 145000 tokens."
+ * - Anthropic: "prompt is too long: 180000 tokens > 200000 token limit"
+ * - Google: "exceeds the maximum number of tokens (180000 > 100000)"
+ */
+export declare function parseProviderOverflowDetails(error: unknown): {
+    actualTokens: number;
+    budgetTokens: number;
+} | null;
+/**
+ * Extract error message from various error formats.
+ */
+export declare function extractErrorMessage(error: unknown): string | null;

package/dist/context/errorDetection.js

@@ -86,10 +86,57 @@ export function getContextOverflowProvider(error) {
     }
     return null;
 }
+/**
+ * Extract actual token counts from provider overflow error messages.
+ *
+ * Many providers include the actual/max token counts in their error messages:
+ * - OpenAI: "This model's maximum context length is 128000 tokens. However, your messages resulted in 145000 tokens."
+ * - Anthropic: "prompt is too long: 180000 tokens > 200000 token limit"
+ * - Google: "exceeds the maximum number of tokens (180000 > 100000)"
+ */
+export function parseProviderOverflowDetails(error) {
+    const message = extractErrorMessage(error);
+    if (!message) {
+        return null;
+    }
+    // Guard against excessively long inputs that could slow regex matching
+    if (message.length > 2000) {
+        return null;
+    }
+    // OpenAI pattern: "resulted in X tokens" + "maximum context length is Y"
+    // Use single character-class number groups to prevent ReDoS (CodeQL: js/polynomial-redos)
+    const openaiActual = message.match(/resulted\s+in\s+(\d[\d,]{0,19})\s*tokens/i);
+    const openaiMax = message.match(/maximum\s+context\s+length\s+is\s+(\d[\d,]{0,19})/i);
+    if (openaiActual && openaiMax) {
+        return {
+            actualTokens: parseInt(openaiActual[1].replace(/,/g, ""), 10),
+            budgetTokens: parseInt(openaiMax[1].replace(/,/g, ""), 10),
+        };
+    }
+    // Anthropic pattern: "X tokens > Y token limit" or "X tokens, limit Y"
+    // Use single character-class number groups to prevent ReDoS (CodeQL: js/polynomial-redos)
+    const anthropicMatch = message.match(/(\d[\d,]{0,19})\s*tokens?\s*[>:]\s*(\d[\d,]{0,19})/i);
+    if (anthropicMatch) {
+        return {
+            actualTokens: parseInt(anthropicMatch[1].replace(/,/g, ""), 10),
+            budgetTokens: parseInt(anthropicMatch[2].replace(/,/g, ""), 10),
+        };
+    }
+    // Google pattern: "X > Y" or "X exceeds Y"
+    // Use single character-class number groups to prevent ReDoS (CodeQL: js/polynomial-redos)
+    const googleMatch = message.match(/(\d[\d,]{0,19})\s*(?:>|exceeds)\s*(\d[\d,]{0,19})/i);
+    if (googleMatch) {
+        return {
+            actualTokens: parseInt(googleMatch[1].replace(/,/g, ""), 10),
+            budgetTokens: parseInt(googleMatch[2].replace(/,/g, ""), 10),
+        };
+    }
+    return null;
+}
 /**
  * Extract error message from various error formats.
  */
-function extractErrorMessage(error) {
+export function extractErrorMessage(error) {
     if (!error) {
         return null;
     }

package/dist/context/errors.d.ts

@@ -0,0 +1,19 @@
+/**
+ * Context-specific error classes for budget and overflow scenarios.
+ */
+/**
+ * Thrown when context exceeds model budget after all compaction stages,
+ * preventing wasteful API calls to providers that will reject the request.
+ */
+export declare class ContextBudgetExceededError extends Error {
+    readonly estimatedTokens: number;
+    readonly availableTokens: number;
+    readonly stagesUsed: string[];
+    readonly breakdown: Record<string, number>;
+    constructor(message: string, details: {
+        estimatedTokens: number;
+        availableTokens: number;
+        stagesUsed: string[];
+        breakdown: Record<string, number>;
+    });
+}

package/dist/context/errors.js

@@ -0,0 +1,21 @@
+/**
+ * Context-specific error classes for budget and overflow scenarios.
+ */
+/**
+ * Thrown when context exceeds model budget after all compaction stages,
+ * preventing wasteful API calls to providers that will reject the request.
+ */
+export class ContextBudgetExceededError extends Error {
+    estimatedTokens;
+    availableTokens;
+    stagesUsed;
+    breakdown;
+    constructor(message, details) {
+        super(message);
+        this.name = "ContextBudgetExceededError";
+        this.estimatedTokens = details.estimatedTokens;
+        this.availableTokens = details.availableTokens;
+        this.stagesUsed = details.stagesUsed;
+        this.breakdown = details.breakdown;
+    }
+}

package/dist/context/stages/slidingWindowTruncator.d.ts

@@ -4,6 +4,12 @@
  * Non-destructive fallback: tags oldest messages as truncated
  * instead of deleting them. Always preserves first message pair.
  * Removes messages in pairs to maintain role alternation.
+ *
+ * Features:
+ * - Adaptive truncation (PERF-001): calculates fraction from actual overage
+ *   instead of fixed 50%, with iterative refinement up to 3 passes.
+ * - Small conversation handling (BUG-005): for <= 4 messages, truncates
+ *   message content proportionally instead of returning no-op.
  */
 import type { ChatMessage } from "../../types/conversation.js";
 import type { TruncationConfig, TruncationResult } from "../../types/contextTypes.js";