@juspay/neurolink 9.59.5 → 9.60.0

package/dist/cli/factories/commandFactory.js

@@ -51,6 +51,16 @@ export class CLICommandFactory {
                 "mistral",
                 "litellm",
                 "sagemaker",
+                "deepseek",
+                "ds",
+                "nvidia-nim",
+                "nim",
+                "nvidia",
+                "lm-studio",
+                "lmstudio",
+                "lms",
+                "llamacpp",
+                "llama.cpp",
             ],
             default: "auto",
             description: "AI provider to use (auto-selects best available). Use 'anthropic-subscription' for Claude subscription plans.",
@@ -1322,6 +1332,10 @@ export class CLICommandFactory {
                         "vertex",
                         "huggingface",
                         "mistral",
+                        "deepseek",
+                        "nvidia-nim",
+                        "lm-studio",
+                        "llamacpp",
                     ],
                 })
                     .option("list", {
@@ -2877,7 +2891,7 @@ export class CLICommandFactory {
                 "        generate|gen)\n" +
                 '            case "${prev}" in\n' +
                 "                --provider|-p)\n" +
-                '                    COMPREPLY=( $(compgen -W "auto openai bedrock vertex googleVertex anthropic azure google-ai huggingface ollama mistral litellm" -- ${cur}) )\n' +
+                '                    COMPREPLY=( $(compgen -W "auto openai openai-compatible openrouter or bedrock vertex googleVertex anthropic anthropic-subscription azure google-ai google-ai-studio huggingface ollama mistral litellm sagemaker deepseek ds nvidia-nim nim lm-studio lmstudio llamacpp llama.cpp" -- ${cur}) )\n' +
                 "                    return 0\n" +
                 "                    ;;\n" +
                 "                --format|-f|--output-format)\n" +

package/dist/cli/utils/interactiveSetup.js

@@ -204,6 +204,70 @@ export const PROVIDER_CONFIGS = [
             },
         ],
     },
+    {
+        id: AIProviderName.DEEPSEEK,
+        name: "DeepSeek",
+        description: "Cost-efficient frontier models (deepseek-chat V3, deepseek-reasoner R1)",
+        envVars: [
+            {
+                key: "DEEPSEEK_API_KEY",
+                prompt: "DeepSeek API Key (get one at https://platform.deepseek.com)",
+                secure: true,
+            },
+        ],
+    },
+    {
+        id: AIProviderName.NVIDIA_NIM,
+        name: "NVIDIA NIM",
+        description: "NVIDIA-hosted Llama, Nemotron, Mistral, and DeepSeek-R1 models",
+        envVars: [
+            {
+                key: "NVIDIA_NIM_API_KEY",
+                prompt: "NVIDIA NIM API Key (get one at https://build.nvidia.com/settings/api-keys)",
+                secure: true,
+            },
+        ],
+    },
+    {
+        id: AIProviderName.LM_STUDIO,
+        name: "LM Studio",
+        description: "Local inference via LM Studio desktop app (https://lmstudio.ai)",
+        envVars: [
+            {
+                key: "LM_STUDIO_BASE_URL",
+                prompt: "LM Studio server URL",
+                default: "http://localhost:1234/v1",
+                secure: false,
+                optional: true,
+            },
+            {
+                key: "LM_STUDIO_API_KEY",
+                prompt: "LM Studio API Key (leave blank — only needed behind an auth proxy)",
+                secure: false,
+                optional: true,
+            },
+        ],
+    },
+    {
+        id: AIProviderName.LLAMACPP,
+        name: "llama.cpp",
+        description: "Local inference via llama-server (https://github.com/ggerganov/llama.cpp). Start with: ./llama-server -m model.gguf --port 8080 --jinja",
+        envVars: [
+            {
+                key: "LLAMACPP_BASE_URL",
+                prompt: "llama-server URL",
+                default: "http://localhost:8080/v1",
+                secure: false,
+                optional: true,
+            },
+            {
+                key: "LLAMACPP_API_KEY",
+                prompt: "llama-server API Key (leave blank — only needed behind an auth proxy)",
+                secure: false,
+                optional: true,
+            },
+        ],
+    },
 ];
 /**
  * Run the interactive setup wizard

package/dist/constants/contextWindows.d.ts

@@ -36,9 +36,13 @@ export declare function getContextWindowSize(provider: string, model?: string):
 /**
  * Calculate output token reserve for a given context window.
  *
+ * Returns the *real* token count that will be reserved for output so callers
+ * (`getAvailableInputTokens`, `BudgetChecker`, conversation-memory pruning, file
+ * summarisation) compute input budget against the actual outgoing maxTokens.
+ *
  * @param contextWindow - Total context window size
  * @param maxTokens - Explicit maxTokens from user config (if set)
- * @returns Number of tokens reserved for output
+ * @returns Number of tokens reserved for output (matches what's sent upstream)
  */
 export declare function getOutputReserve(contextWindow: number, maxTokens?: number): number;
 /**

package/dist/constants/contextWindows.js

@@ -23,6 +23,34 @@ export const DEFAULT_OUTPUT_RESERVE_RATIO = 0.35;
  * The "_default" key is the fallback for unknown models within a provider.
  */
 export const MODEL_CONTEXT_WINDOWS = {
+    deepseek: {
+        _default: 64_000,
+        "deepseek-chat": 64_000,
+        "deepseek-reasoner": 64_000,
+    },
+    "nvidia-nim": {
+        _default: 128_000,
+        "meta/llama-3.3-70b-instruct": 128_000,
+        "meta/llama-3.1-405b-instruct": 128_000,
+        "meta/llama-3.1-70b-instruct": 128_000,
+        "meta/llama-3.2-90b-vision-instruct": 128_000,
+        "meta/llama-3.2-11b-vision-instruct": 128_000,
+        "nvidia/llama-3.3-nemotron-super-49b-v1": 128_000,
+        "nvidia/llama-3.1-nemotron-nano-8b-v1": 128_000,
+        "nvidia/llama-3.1-nemotron-70b-instruct": 128_000,
+        "deepseek-ai/deepseek-r1": 128_000,
+        "deepseek-ai/deepseek-r1-distill-llama-70b": 128_000,
+        "mistralai/mixtral-8x22b-instruct-v0.1": 65_536,
+        "mistralai/mixtral-8x7b-instruct-v0.1": 32_768,
+        "microsoft/phi-4": 16_384,
+        "google/gemma-3-27b-it": 8_192,
+    },
+    "lm-studio": {
+        _default: 8_192,
+    },
+    llamacpp: {
+        _default: 8_192,
+    },
     anthropic: {
         _default: 200_000,
         // Claude 4.6 (Feb 2026) — 1M context window
@@ -223,6 +251,36 @@ export const MODEL_CONTEXT_WINDOWS = {
         "qwen3-vl-8b-instruct": 32_768,
     },
 };
+/**
+ * Map of provider aliases to canonical MODEL_CONTEXT_WINDOWS keys.
+ *
+ * Callers reach `getContextWindowSize` via the unnormalized form on
+ * `options.provider` (e.g. CLI `--provider lmstudio`, alias `llama.cpp`),
+ * and `ProviderFactory.normalizeProviderName` runs only at instantiation —
+ * its output never reaches budget calculations. Without this normalization
+ * those alias forms miss the table and fall back to `DEFAULT_CONTEXT_WINDOW`,
+ * understating the budget for LM Studio / llama.cpp / NVIDIA NIM.
+ *
+ * The keys here are the result of stripping non-alpha characters, so
+ * `lm-studio` -> `lmstudio`, `nvidia-nim` -> `nvidianim`, `llama.cpp` -> `llamacpp`.
+ */
+const PROVIDER_ALIAS_MAP = {
+    googleaistudio: "google-ai-studio",
+    lmstudio: "lm-studio",
+    llamacpp: "llamacpp",
+    nvidianim: "nvidia-nim",
+    nim: "nvidia-nim",
+    nvidia: "nvidia-nim",
+    deepseek: "deepseek",
+};
+function normalizeProviderForLookup(provider) {
+    const stripped = provider.toLowerCase().replace(/[^a-z]/g, "");
+    // On alias miss, return the *stripped* key — not the raw input — so case /
+    // separator variants ("OpenAI", "open-ai", "Vertex AI") still find their
+    // table entry under the lowercase canonical key instead of falling through
+    // to DEFAULT_CONTEXT_WINDOW.
+    return PROVIDER_ALIAS_MAP[stripped] ?? stripped;
+}
 /**
  * Resolve context window size for a provider/model combination.
  *
@@ -252,8 +310,10 @@ export function getContextWindowSize(provider, model) {
             // Dynamic registry not initialized yet — fall through to static lookup
         }
     }
-    // Static fallback chain
-    const providerWindows = MODEL_CONTEXT_WINDOWS[provider];
+    // Static fallback chain — normalize aliases first so "lmstudio" / "llama.cpp" /
+    // "nvidianim" find their canonical entries instead of falling back to default.
+    const canonical = normalizeProviderForLookup(provider);
+    const providerWindows = MODEL_CONTEXT_WINDOWS[canonical] ?? MODEL_CONTEXT_WINDOWS[provider];
     if (!providerWindows) {
         return DEFAULT_CONTEXT_WINDOW;
     }
@@ -273,9 +333,13 @@ export function getContextWindowSize(provider, model) {
 /**
  * Calculate output token reserve for a given context window.
  *
+ * Returns the *real* token count that will be reserved for output so callers
+ * (`getAvailableInputTokens`, `BudgetChecker`, conversation-memory pruning, file
+ * summarisation) compute input budget against the actual outgoing maxTokens.
+ *
  * @param contextWindow - Total context window size
  * @param maxTokens - Explicit maxTokens from user config (if set)
- * @returns Number of tokens reserved for output
+ * @returns Number of tokens reserved for output (matches what's sent upstream)
  */
 export function getOutputReserve(contextWindow, maxTokens) {
     if (maxTokens !== undefined && maxTokens > 0) {

package/dist/constants/enums.d.ts

@@ -15,6 +15,10 @@ export declare enum AIProviderName {
     MISTRAL = "mistral",
     LITELLM = "litellm",
     SAGEMAKER = "sagemaker",
+    DEEPSEEK = "deepseek",
+    NVIDIA_NIM = "nvidia-nim",
+    LM_STUDIO = "lm-studio",
+    LLAMACPP = "llamacpp",
     AUTO = "auto"
 }
 /**
@@ -622,6 +626,54 @@ export declare enum AnthropicBetaFeature {
     INTERLEAVED_THINKING = "interleaved-thinking-2025-05-14",
     FINE_GRAINED_STREAMING = "fine-grained-tool-streaming-2025-05-14"
 }
+/**
+ * Supported Models for DeepSeek
+ * Docs: https://api-docs.deepseek.com/quick_start/pricing
+ */
+export declare enum DeepSeekModels {
+    /** General-purpose chat (DeepSeek-V3) */
+    DEEPSEEK_CHAT = "deepseek-chat",
+    /** Reasoning model (DeepSeek-R1) — exposes reasoning_content */
+    DEEPSEEK_REASONER = "deepseek-reasoner"
+}
+/**
+ * Selected NVIDIA NIM Models
+ * Full catalog: https://build.nvidia.com/models
+ * Note: NIM hosts hundreds of models; pass arbitrary IDs via --model.
+ */
+export declare enum NvidiaNimModels {
+    LLAMA_3_3_70B_INSTRUCT = "meta/llama-3.3-70b-instruct",
+    LLAMA_3_1_405B_INSTRUCT = "meta/llama-3.1-405b-instruct",
+    LLAMA_3_1_70B_INSTRUCT = "meta/llama-3.1-70b-instruct",
+    LLAMA_3_2_90B_VISION = "meta/llama-3.2-90b-vision-instruct",
+    LLAMA_3_2_11B_VISION = "meta/llama-3.2-11b-vision-instruct",
+    NEMOTRON_SUPER_49B = "nvidia/llama-3.3-nemotron-super-49b-v1",
+    NEMOTRON_NANO_8B = "nvidia/llama-3.1-nemotron-nano-8b-v1",
+    NEMOTRON_70B_INSTRUCT = "nvidia/llama-3.1-nemotron-70b-instruct",
+    DEEPSEEK_R1 = "deepseek-ai/deepseek-r1",
+    DEEPSEEK_R1_DISTILL_LLAMA_70B = "deepseek-ai/deepseek-r1-distill-llama-70b",
+    MIXTRAL_8X22B_INSTRUCT = "mistralai/mixtral-8x22b-instruct-v0.1",
+    MIXTRAL_8X7B_INSTRUCT = "mistralai/mixtral-8x7b-instruct-v0.1",
+    PHI_4 = "microsoft/phi-4",
+    GEMMA_3_27B_IT = "google/gemma-3-27b-it",
+    GLM_4_5 = "z-ai/glm4.5"
+}
+/**
+ * LM Studio loads any GGUF model the user has downloaded.
+ * Default: empty string → triggers /v1/models auto-discovery.
+ */
+export declare enum LMStudioModels {
+    /** Sentinel value — triggers auto-discovery from /v1/models */
+    AUTO_DISCOVER = ""
+}
+/**
+ * llama.cpp serves a single model loaded at server startup.
+ * Default: empty string → uses whatever is loaded.
+ */
+export declare enum LlamaCppModels {
+    /** Sentinel value — uses the model loaded by the llama-server process */
+    AUTO_DISCOVER = ""
+}
 /**
  * Buffer time in milliseconds before token expiry to trigger refresh
  *

package/dist/constants/enums.js

@@ -19,6 +19,10 @@ export var AIProviderName;
     AIProviderName["MISTRAL"] = "mistral";
     AIProviderName["LITELLM"] = "litellm";
     AIProviderName["SAGEMAKER"] = "sagemaker";
+    AIProviderName["DEEPSEEK"] = "deepseek";
+    AIProviderName["NVIDIA_NIM"] = "nvidia-nim";
+    AIProviderName["LM_STUDIO"] = "lm-studio";
+    AIProviderName["LLAMACPP"] = "llamacpp";
     AIProviderName["AUTO"] = "auto";
 })(AIProviderName || (AIProviderName = {}));
 /**
@@ -848,6 +852,65 @@ export var AnthropicBetaFeature;
     AnthropicBetaFeature["INTERLEAVED_THINKING"] = "interleaved-thinking-2025-05-14";
     AnthropicBetaFeature["FINE_GRAINED_STREAMING"] = "fine-grained-tool-streaming-2025-05-14";
 })(AnthropicBetaFeature || (AnthropicBetaFeature = {}));
+/**
+ * Supported Models for DeepSeek
+ * Docs: https://api-docs.deepseek.com/quick_start/pricing
+ */
+export var DeepSeekModels;
+(function (DeepSeekModels) {
+    /** General-purpose chat (DeepSeek-V3) */
+    DeepSeekModels["DEEPSEEK_CHAT"] = "deepseek-chat";
+    /** Reasoning model (DeepSeek-R1) — exposes reasoning_content */
+    DeepSeekModels["DEEPSEEK_REASONER"] = "deepseek-reasoner";
+})(DeepSeekModels || (DeepSeekModels = {}));
+/**
+ * Selected NVIDIA NIM Models
+ * Full catalog: https://build.nvidia.com/models
+ * Note: NIM hosts hundreds of models; pass arbitrary IDs via --model.
+ */
+export var NvidiaNimModels;
+(function (NvidiaNimModels) {
+    // Meta Llama
+    NvidiaNimModels["LLAMA_3_3_70B_INSTRUCT"] = "meta/llama-3.3-70b-instruct";
+    NvidiaNimModels["LLAMA_3_1_405B_INSTRUCT"] = "meta/llama-3.1-405b-instruct";
+    NvidiaNimModels["LLAMA_3_1_70B_INSTRUCT"] = "meta/llama-3.1-70b-instruct";
+    NvidiaNimModels["LLAMA_3_2_90B_VISION"] = "meta/llama-3.2-90b-vision-instruct";
+    NvidiaNimModels["LLAMA_3_2_11B_VISION"] = "meta/llama-3.2-11b-vision-instruct";
+    // NVIDIA Nemotron (reasoning)
+    NvidiaNimModels["NEMOTRON_SUPER_49B"] = "nvidia/llama-3.3-nemotron-super-49b-v1";
+    NvidiaNimModels["NEMOTRON_NANO_8B"] = "nvidia/llama-3.1-nemotron-nano-8b-v1";
+    NvidiaNimModels["NEMOTRON_70B_INSTRUCT"] = "nvidia/llama-3.1-nemotron-70b-instruct";
+    // DeepSeek hosted on NIM
+    NvidiaNimModels["DEEPSEEK_R1"] = "deepseek-ai/deepseek-r1";
+    NvidiaNimModels["DEEPSEEK_R1_DISTILL_LLAMA_70B"] = "deepseek-ai/deepseek-r1-distill-llama-70b";
+    // Mistral / Mixtral
+    NvidiaNimModels["MIXTRAL_8X22B_INSTRUCT"] = "mistralai/mixtral-8x22b-instruct-v0.1";
+    NvidiaNimModels["MIXTRAL_8X7B_INSTRUCT"] = "mistralai/mixtral-8x7b-instruct-v0.1";
+    // Microsoft Phi
+    NvidiaNimModels["PHI_4"] = "microsoft/phi-4";
+    // Google Gemma
+    NvidiaNimModels["GEMMA_3_27B_IT"] = "google/gemma-3-27b-it";
+    // Z.AI GLM
+    NvidiaNimModels["GLM_4_5"] = "z-ai/glm4.5";
+})(NvidiaNimModels || (NvidiaNimModels = {}));
+/**
+ * LM Studio loads any GGUF model the user has downloaded.
+ * Default: empty string → triggers /v1/models auto-discovery.
+ */
+export var LMStudioModels;
+(function (LMStudioModels) {
+    /** Sentinel value — triggers auto-discovery from /v1/models */
+    LMStudioModels["AUTO_DISCOVER"] = "";
+})(LMStudioModels || (LMStudioModels = {}));
+/**
+ * llama.cpp serves a single model loaded at server startup.
+ * Default: empty string → uses whatever is loaded.
+ */
+export var LlamaCppModels;
+(function (LlamaCppModels) {
+    /** Sentinel value — uses the model loaded by the llama-server process */
+    LlamaCppModels["AUTO_DISCOVER"] = "";
+})(LlamaCppModels || (LlamaCppModels = {}));
 // ============================================================================
 // ANTHROPIC OAUTH CONSTANTS
 // ============================================================================

package/dist/core/baseProvider.d.ts

@@ -9,7 +9,7 @@ import { TelemetryHandler } from "./modules/TelemetryHandler.js";
  * Tools are integrated as first-class citizens - always available by default
  */
 export declare abstract class BaseProvider implements AIProvider {
-    protected readonly modelName: string;
+    protected modelName: string;
     protected readonly providerName: AIProviderName;
     protected readonly defaultTimeout: number;
     protected middlewareOptions?: MiddlewareFactoryOptions;
@@ -29,13 +29,22 @@ export declare abstract class BaseProvider implements AIProvider {
         traceId: string;
         parentSpanId: string;
     } | null): void;
-    private readonly messageBuilder;
-    private readonly streamHandler;
-    private readonly generationHandler;
-    protected readonly telemetryHandler: TelemetryHandler;
-    private readonly utilities;
+    private messageBuilder;
+    private streamHandler;
+    private generationHandler;
+    protected telemetryHandler: TelemetryHandler;
+    private utilities;
     private readonly toolsManager;
     constructor(modelName?: string, providerName?: AIProviderName, neurolink?: NeuroLink, middleware?: MiddlewareFactoryOptions);
+    /**
+     * Update modelName and rebuild composition handlers with the new value.
+     *
+     * Auto-discovery providers (lm-studio, llamacpp) call this once they have
+     * resolved the loaded model from `/v1/models`. Without this, handlers
+     * (TelemetryHandler, MessageBuilder, ...) keep the pre-discovery name and
+     * pricing / span / log metadata reports the stale value.
+     */
+    protected refreshHandlersForModel(model: string): void;
     /**
      * Check if this provider supports tool/function calling
      * Override in subclasses to disable tools for specific providers or models

package/dist/core/baseProvider.js

@@ -23,6 +23,10 @@ import { Utilities } from "./modules/Utilities.js";
  * Tools are integrated as first-class citizens - always available by default
  */
 export class BaseProvider {
+    // Not `readonly` because providers that auto-discover the model from a
+    // /v1/models endpoint (lm-studio, llamacpp) need to update modelName after
+    // construction so handlers (TelemetryHandler, MessageBuilder) cache the
+    // resolved name. All other providers treat this as effectively readonly.
     modelName;
     providerName;
     defaultTimeout = 30000; // 30 seconds
@@ -43,6 +47,11 @@ export class BaseProvider {
         this._traceContext = ctx;
     }
     // Composition modules - Single Responsibility Principle
+    // Handlers below are not `readonly` so that providers which auto-discover
+    // their model after construction (lm-studio, llamacpp) can rebuild them
+    // via `refreshHandlersForModel(...)` and propagate the resolved name into
+    // pricing / telemetry / span attributes. All other providers leave these
+    // alone.
     messageBuilder;
     streamHandler;
     generationHandler;
@@ -67,6 +76,25 @@ export class BaseProvider {
             fixSchemaForOpenAIStrictMode: (schema) => this.fixSchemaForOpenAIStrictMode(schema),
         });
     }
+    /**
+     * Update modelName and rebuild composition handlers with the new value.
+     *
+     * Auto-discovery providers (lm-studio, llamacpp) call this once they have
+     * resolved the loaded model from `/v1/models`. Without this, handlers
+     * (TelemetryHandler, MessageBuilder, ...) keep the pre-discovery name and
+     * pricing / span / log metadata reports the stale value.
+     */
+    refreshHandlersForModel(model) {
+        this.modelName = model;
+        trace
+            .getSpan(context.active())
+            ?.setAttribute(ATTR.GEN_AI_MODEL, this.modelName);
+        this.messageBuilder = new MessageBuilder(this.providerName, this.modelName);
+        this.streamHandler = new StreamHandler(this.providerName, this.modelName);
+        this.telemetryHandler = new TelemetryHandler(this.providerName, this.modelName, this.neurolink);
+        this.generationHandler = new GenerationHandler(this.providerName, this.modelName, () => this.supportsTools(), (options, type) => this.telemetryHandler.getTelemetryConfig(options, type), (toolCalls, toolResults, options, timestamp) => this.handleToolExecutionStorage(toolCalls, toolResults, options, timestamp), () => this.neurolink?.getEventEmitter());
+        this.utilities = new Utilities(this.providerName, this.modelName, this.defaultTimeout, this.middlewareOptions);
+    }
     /**
      * Check if this provider supports tool/function calling
      * Override in subclasses to disable tools for specific providers or models

package/dist/factories/providerRegistry.js

@@ -1,6 +1,6 @@
 import { ProviderFactory } from "./providerFactory.js";
 import { logger } from "../utils/logger.js";
-import { AIProviderName, GoogleAIModels, OpenAIModels, AnthropicModels, VertexModels, MistralModels, OllamaModels, LiteLLMModels, HuggingFaceModels, } from "../constants/enums.js";
+import { AIProviderName, GoogleAIModels, OpenAIModels, AnthropicModels, VertexModels, MistralModels, OllamaModels, LiteLLMModels, HuggingFaceModels, DeepSeekModels, NvidiaNimModels, } from "../constants/enums.js";
 /**
  * Provider Registry - registers all providers with the factory
  * This is where we migrate providers one by one to the new pattern
@@ -128,6 +128,30 @@ export class ProviderRegistry {
                 const { AmazonSageMakerProvider } = await import("../providers/amazonSagemaker.js");
                 return new AmazonSageMakerProvider(modelName, undefined, region, undefined, sagemakerCreds);
             }, process.env.SAGEMAKER_MODEL || "sagemaker-model", ["sagemaker", "aws-sagemaker"]);
+            // Register DeepSeek provider
+            ProviderFactory.registerProvider(AIProviderName.DEEPSEEK, async (modelName, _providerName, sdk, _region, credentials) => {
+                const deepseekCreds = credentials;
+                const { DeepSeekProvider } = await import("../providers/deepseek.js");
+                return new DeepSeekProvider(modelName, sdk, undefined, deepseekCreds);
+            }, process.env.DEEPSEEK_MODEL || DeepSeekModels.DEEPSEEK_CHAT, ["deepseek", "ds"]);
+            // Register NVIDIA NIM provider
+            ProviderFactory.registerProvider(AIProviderName.NVIDIA_NIM, async (modelName, _providerName, sdk, _region, credentials) => {
+                const nimCreds = credentials;
+                const { NvidiaNimProvider } = await import("../providers/nvidiaNim.js");
+                return new NvidiaNimProvider(modelName, sdk, undefined, nimCreds);
+            }, process.env.NVIDIA_NIM_MODEL || NvidiaNimModels.LLAMA_3_3_70B_INSTRUCT, ["nvidia", "nim", "nvidia-nim"]);
+            // Register LM Studio provider (local)
+            ProviderFactory.registerProvider(AIProviderName.LM_STUDIO, async (modelName, _providerName, sdk, _region, credentials) => {
+                const lmStudioCreds = credentials;
+                const { LMStudioProvider } = await import("../providers/lmStudio.js");
+                return new LMStudioProvider(modelName, sdk, undefined, lmStudioCreds);
+            }, process.env.LM_STUDIO_MODEL || undefined, ["lmstudio", "lm-studio", "lms"]);
+            // Register llama.cpp provider (local)
+            ProviderFactory.registerProvider(AIProviderName.LLAMACPP, async (modelName, _providerName, sdk, _region, credentials) => {
+                const llamaCppCreds = credentials;
+                const { LlamaCppProvider } = await import("../providers/llamaCpp.js");
+                return new LlamaCppProvider(modelName, sdk, undefined, llamaCppCreds);
+            }, process.env.LLAMACPP_MODEL || undefined, ["llamacpp", "llama.cpp", "llama-cpp"]);
             logger.debug("All providers registered successfully");
             this.registered = true;
             // ===== TTS HANDLER REGISTRATION =====

package/dist/lib/adapters/providerImageAdapter.js

@@ -50,7 +50,19 @@ const PROXY_PROVIDERS = new Set(["litellm", "openrouter"]);
  */
 function normalizeVisionProvider(provider) {
     const lower = provider.toLowerCase();
-    switch (lower) {
+    // Strip non-alpha characters so alias forms (e.g. "lm-studio", "lm_studio",
+    // "llama.cpp", "nvidia_nim") all collapse onto a canonical key. Mirrors
+    // the alias-normalization pattern used in pricing.ts and contextWindows.ts.
+    const stripped = lower.replace(/[^a-z]/g, "");
+    switch (stripped) {
+        case "lmstudio":
+            return "lm-studio";
+        case "llamacpp":
+            return "llamacpp";
+        case "nvidianim":
+            return "nvidia-nim";
+        case "googleaistudio":
+            return "google-ai";
         case "or":
             return "openrouter";
         default:
@@ -436,6 +448,32 @@ const VISION_CAPABILITIES = {
         "meta-llama-4-maverick-17b-128e-instruct",
         "meta-llama-4-scout-17b-16e-instruct",
     ],
+    // DeepSeek has no vision support — empty list
+    deepseek: [],
+    "nvidia-nim": [
+        "meta/llama-3.2-90b-vision-instruct",
+        "meta/llama-3.2-11b-vision-instruct",
+    ],
+    // LM Studio + llama.cpp: vision depends on the loaded model.
+    // Substrings must point at known multimodal variants only — bare
+    // "llama-3.2" matches the text-only Llama-3.2-1B/3B chat models.
+    "lm-studio": [
+        "llava",
+        "llama-3.2-11b-vision",
+        "llama-3.2-90b-vision",
+        "vision-instruct",
+        "qwen2-vl",
+        "qwen2.5-vl",
+        "phi-3-vision",
+    ],
+    llamacpp: [
+        "llava",
+        "llama-3.2-11b-vision",
+        "llama-3.2-90b-vision",
+        "vision-instruct",
+        "qwen2-vl",
+        "phi-3-vision",
+    ],
 };
 /**
  * Provider Image Adapter - Smart routing and formatting
@@ -535,6 +573,13 @@ export class ProviderImageAdapter {
             if (!supportedModels) {
                 return false;
             }
+            // An empty list means the provider has NO vision support (e.g. deepseek).
+            // Without this guard, the no-model branch below would return `true` for
+            // every provider that has an entry in VISION_CAPABILITIES — even an empty
+            // one — letting vision requests through to a text-only API.
+            if (supportedModels.length === 0) {
+                return false;
+            }
             if (!model) {
                 return true; // Provider supports vision, but need to check specific model
             }
@@ -562,7 +607,12 @@ export class ProviderImageAdapter {
      * Get all vision-capable providers
      */
     static getVisionProviders() {
-        return Object.keys(VISION_CAPABILITIES);
+        // Filter out providers whose allowlist is empty (e.g. deepseek). They're
+        // listed in VISION_CAPABILITIES so supportsVision can return false for
+        // them, but they should not be advertised as vision-capable.
+        return Object.entries(VISION_CAPABILITIES)
+            .filter(([, models]) => models.length > 0)
+            .map(([provider]) => provider);
     }
     /**
      * Count total "images" in a message (actual images + PDF pages)

package/dist/lib/constants/contextWindows.d.ts

@@ -36,9 +36,13 @@ export declare function getContextWindowSize(provider: string, model?: string):
 /**
  * Calculate output token reserve for a given context window.
  *
+ * Returns the *real* token count that will be reserved for output so callers
+ * (`getAvailableInputTokens`, `BudgetChecker`, conversation-memory pruning, file
+ * summarisation) compute input budget against the actual outgoing maxTokens.
+ *
  * @param contextWindow - Total context window size
  * @param maxTokens - Explicit maxTokens from user config (if set)
- * @returns Number of tokens reserved for output
+ * @returns Number of tokens reserved for output (matches what's sent upstream)
  */
 export declare function getOutputReserve(contextWindow: number, maxTokens?: number): number;
 /**