hoomanjs 1.41.1 → 1.42.0

package/dist/core/models/llama-cpp/strands-llama-cpp.js

@@ -0,0 +1,422 @@
+import { ContextWindowOverflowError, Model, ModelError, } from "@strands-agents/sdk";
+import { ModelContentBlockDeltaEvent, ModelContentBlockStartEvent, ModelContentBlockStopEvent, ModelMessageStartEvent, ModelMessageStopEvent, ModelMetadataEvent, } from "@strands-agents/sdk";
+import { jsonSchemaToGbnf, pruneOptionalNulls } from "./gbnf-schema.js";
+import { resolveModelFile } from "./resolve-model.js";
+/**
+ * Loaded llama.cpp runtimes are expensive (weights in RAM/VRAM), so share them
+ * process-wide keyed by resolved GGUF path. Contexts stay per Strands model
+ * instance so concurrent agents (e.g. subagents) get independent sequences.
+ */
+const llamaRuntimePromises = new Map();
+const loadedModelPromises = new Map();
+function extractSystemText(system) {
+    if (system === undefined) {
+        return undefined;
+    }
+    if (typeof system === "string") {
+        return system;
+    }
+    const parts = [];
+    for (const block of system) {
+        if (block.type === "textBlock") {
+            parts.push(block.text);
+        }
+    }
+    const joined = parts.join("\n").trim();
+    return joined.length > 0 ? joined : undefined;
+}
+function contentBlockToText(block) {
+    if (block.type === "textBlock") {
+        return block.text;
+    }
+    if (block.type === "imageBlock") {
+        return "(llama.cpp: image content is not supported by this provider)";
+    }
+    if (block.type === "videoBlock" || block.type === "documentBlock") {
+        return `(llama.cpp: ${block.type} content is not supported by this provider)`;
+    }
+    return undefined;
+}
+function toolResultToText(block) {
+    const parts = [];
+    for (const c of block.content) {
+        if (c.type === "textBlock") {
+            parts.push(c.text);
+        }
+        else if (c.type === "jsonBlock") {
+            parts.push(JSON.stringify(c.json));
+        }
+        else if (c.type === "imageBlock") {
+            parts.push("(llama.cpp: image tool results are not supported)");
+        }
+    }
+    const joined = parts.join("\n");
+    if (joined.length > 0) {
+        return joined;
+    }
+    return block.status === "error" ? "(tool error)" : "";
+}
+/**
+ * Convert Strands conversation history to node-llama-cpp `ChatHistoryItem[]`.
+ * Tool calls are paired with their results (`ChatModelFunctionCall` requires
+ * both) by `toolUseId`: the toolUse blocks live on assistant messages and the
+ * matching toolResult blocks on the following user message, so results are
+ * folded back into the preceding model response.
+ */
+function strandsMessagesToHistory(messages, systemText) {
+    const history = [];
+    if (systemText) {
+        history.push({ type: "system", text: systemText });
+    }
+    const pendingToolUses = new Map();
+    const lastModelResponse = () => {
+        const last = history.at(-1);
+        if (last?.type === "model") {
+            return last;
+        }
+        const created = { type: "model", response: [] };
+        history.push(created);
+        return created;
+    };
+    const pushUserText = (text) => {
+        const last = history.at(-1);
+        if (last?.type === "user") {
+            last.text = last.text.length > 0 ? `${last.text}\n${text}` : text;
+        }
+        else {
+            history.push({ type: "user", text });
+        }
+    };
+    for (const msg of messages) {
+        for (const block of msg.content) {
+            if (msg.role === "assistant") {
+                if (block.type === "toolUseBlock") {
+                    const b = block;
+                    pendingToolUses.set(b.toolUseId, b);
+                    continue;
+                }
+                if (block.type === "reasoningBlock") {
+                    // Prior turns' chain of thought is not replayed into the context.
+                    continue;
+                }
+                const text = contentBlockToText(block);
+                if (text !== undefined && text.length > 0) {
+                    lastModelResponse().response.push(text);
+                }
+                continue;
+            }
+            // user role
+            if (block.type === "toolResultBlock") {
+                const b = block;
+                const use = pendingToolUses.get(b.toolUseId);
+                pendingToolUses.delete(b.toolUseId);
+                const call = {
+                    type: "functionCall",
+                    name: use?.name ?? b.toolUseId,
+                    params: use?.input ?? {},
+                    result: toolResultToText(b),
+                };
+                const model = lastModelResponse();
+                if (model.response.filter((item) => typeof item !== "string").length === 0) {
+                    call.startsNewChunk = true;
+                }
+                model.response.push(call);
+                continue;
+            }
+            const text = contentBlockToText(block);
+            if (text !== undefined && text.length > 0) {
+                pushUserText(text);
+            }
+        }
+    }
+    return history;
+}
+/**
+ * Convert Strands tool specs to node-llama-cpp function definitions. Tool
+ * inputSchemas are standard JSON Schema (Strands emits them via
+ * `z.toJSONSchema`), while node-llama-cpp only accepts its GBNF subset —
+ * passing e.g. `anyOf` through unconverted fails generation with
+ * `Unknown immutable type undefined`. See `gbnf-schema.ts`.
+ */
+function strandsToolsToFunctions(toolSpecs) {
+    if (!toolSpecs?.length) {
+        return undefined;
+    }
+    const functions = {};
+    for (const spec of toolSpecs) {
+        functions[spec.name] = {
+            ...(spec.description ? { description: spec.description } : {}),
+            ...(spec.inputSchema
+                ? { params: jsonSchemaToGbnf(spec.inputSchema) }
+                : {}),
+        };
+    }
+    return functions;
+}
+/** Strands {@link Model} backed by in-process llama.cpp via `node-llama-cpp`. */
+export class StrandsLlamaCppModel extends Model {
+    config;
+    chatPromise;
+    context;
+    constructor(config) {
+        super();
+        this.config = { ...config };
+    }
+    updateConfig(modelConfig) {
+        const chatKey = (c) => JSON.stringify([c.modelId, c.gpu, c.contextSize, c.reasoning ?? null]);
+        const before = chatKey(this.config);
+        this.config = { ...this.config, ...modelConfig };
+        if (chatKey(this.config) !== before) {
+            // The chat wrapper bakes in reasoning settings and the context bakes in
+            // its size, so drop the lazily-built chat/context and rebuild on the
+            // next stream call (loaded weights are cached process-wide).
+            this.chatPromise = undefined;
+            this.context = undefined;
+        }
+    }
+    getConfig() {
+        return { ...this.config };
+    }
+    getChat() {
+        this.chatPromise ??= this.initChat();
+        return this.chatPromise;
+    }
+    async initChat() {
+        const modelId = this.config.modelId;
+        if (!modelId) {
+            throw new ModelError("llama.cpp model is not configured");
+        }
+        const { getLlama, LlamaChat, LlamaLogLevel, resolveChatWrapper } = await import("node-llama-cpp");
+        const modelPath = await resolveModelFile(modelId, this.config.hfToken);
+        const gpu = this.config.gpu ?? "auto";
+        const runtimeKey = JSON.stringify(gpu);
+        let llamaPromise = llamaRuntimePromises.get(runtimeKey);
+        if (!llamaPromise) {
+            // Keep llama.cpp's native logging out of the TUI transcript.
+            llamaPromise = getLlama({ gpu, logLevel: LlamaLogLevel.error });
+            llamaRuntimePromises.set(runtimeKey, llamaPromise);
+        }
+        const llama = await llamaPromise;
+        let modelPromise = loadedModelPromises.get(modelPath);
+        if (!modelPromise) {
+            modelPromise = llama.loadModel({ modelPath });
+            loadedModelPromises.set(modelPath, modelPromise);
+        }
+        const model = await modelPromise;
+        this.context = await model.createContext({
+            ...(this.config.contextSize !== undefined
+                ? { contextSize: this.config.contextSize }
+                : {}),
+        });
+        // Reasoning on/off lives in the chat template, so resolve the wrapper
+        // explicitly (same auto-detection LlamaChat would do) with per-wrapper
+        // thinking settings derived from the shared `reasoning` option.
+        const reasoningEnabled = this.config.reasoning !== undefined;
+        const effort = this.config.reasoning?.effort;
+        const harmonyEffort = effort === "high" || effort === "medium" ? effort : "low";
+        const chatWrapper = resolveChatWrapper(model, {
+            customWrapperSettings: {
+                qwen: { thoughts: reasoningEnabled ? "auto" : "discourage" },
+                gemma4: { reasoning: reasoningEnabled },
+                harmony: {
+                    reasoningEffort: reasoningEnabled ? harmonyEffort : "low",
+                },
+            },
+        });
+        return new LlamaChat({
+            contextSequence: this.context.getSequence(),
+            chatWrapper,
+        });
+    }
+    async *stream(messages, options) {
+        let chat;
+        try {
+            chat = await this.getChat();
+        }
+        catch (e) {
+            // Let a failed init (bad model spec, download failure) be retried.
+            this.chatPromise = undefined;
+            if (e instanceof ModelError) {
+                throw e;
+            }
+            const msg = e instanceof Error ? e.message : String(e);
+            throw new ModelError(`llama.cpp initialization failed: ${msg}`, {
+                cause: e,
+            });
+        }
+        const systemText = extractSystemText(options?.systemPrompt);
+        const history = strandsMessagesToHistory(messages, systemText);
+        if (history.at(-1)?.type === "user") {
+            history.push({ type: "model", response: [] });
+        }
+        const functions = strandsToolsToFunctions(options?.toolSpecs);
+        const originalSchemas = new Map((options?.toolSpecs ?? []).map((spec) => [spec.name, spec.inputSchema]));
+        yield new ModelMessageStartEvent({
+            type: "modelMessageStartEvent",
+            role: "assistant",
+        });
+        const queue = [];
+        let wake;
+        const push = (event) => {
+            queue.push(event);
+            wake?.();
+        };
+        let textBlockOpen = false;
+        let reasoningBlockOpen = false;
+        const closeOpenBlocks = () => {
+            if (reasoningBlockOpen) {
+                push(new ModelContentBlockStopEvent({
+                    type: "modelContentBlockStopEvent",
+                }));
+                reasoningBlockOpen = false;
+            }
+            if (textBlockOpen) {
+                push(new ModelContentBlockStopEvent({
+                    type: "modelContentBlockStopEvent",
+                }));
+                textBlockOpen = false;
+            }
+        };
+        const meterBefore = {
+            input: chat.sequence.tokenMeter.usedInputTokens,
+            output: chat.sequence.tokenMeter.usedOutputTokens,
+        };
+        // Budget 0 closes any thought segment as soon as it opens, covering models
+        // whose template ignores the wrapper's discourage setting (e.g. DeepSeek
+        // R1 distills that always think).
+        const thoughtBudget = this.config.reasoning === undefined ? 0 : this.config.thoughtBudgetTokens;
+        let done = false;
+        let failure;
+        const generation = chat
+            .generateResponse(history, {
+            ...(functions ? { functions } : {}),
+            ...(this.config.maxTokens !== undefined
+                ? { maxTokens: this.config.maxTokens }
+                : {}),
+            ...(this.config.temperature !== undefined
+                ? { temperature: this.config.temperature }
+                : {}),
+            ...(this.config.topP !== undefined ? { topP: this.config.topP } : {}),
+            ...(thoughtBudget !== undefined
+                ? { budgets: { thoughtTokens: thoughtBudget } }
+                : {}),
+            onResponseChunk: (chunk) => {
+                if (chunk.type === "segment") {
+                    if (chunk.segmentType !== "thought" || chunk.text.length === 0) {
+                        return;
+                    }
+                    if (textBlockOpen) {
+                        push(new ModelContentBlockStopEvent({
+                            type: "modelContentBlockStopEvent",
+                        }));
+                        textBlockOpen = false;
+                    }
+                    if (!reasoningBlockOpen) {
+                        push(new ModelContentBlockStartEvent({
+                            type: "modelContentBlockStartEvent",
+                        }));
+                        reasoningBlockOpen = true;
+                    }
+                    push(new ModelContentBlockDeltaEvent({
+                        type: "modelContentBlockDeltaEvent",
+                        delta: { type: "reasoningContentDelta", text: chunk.text },
+                    }));
+                    return;
+                }
+                if (chunk.text.length === 0) {
+                    return;
+                }
+                if (reasoningBlockOpen) {
+                    push(new ModelContentBlockStopEvent({
+                        type: "modelContentBlockStopEvent",
+                    }));
+                    reasoningBlockOpen = false;
+                }
+                if (!textBlockOpen) {
+                    push(new ModelContentBlockStartEvent({
+                        type: "modelContentBlockStartEvent",
+                    }));
+                    textBlockOpen = true;
+                }
+                push(new ModelContentBlockDeltaEvent({
+                    type: "modelContentBlockDeltaEvent",
+                    delta: { type: "textDelta", text: chunk.text },
+                }));
+            },
+        })
+            .then((res) => {
+            closeOpenBlocks();
+            const functionCalls = res.functionCalls ?? [];
+            let toolIndex = 0;
+            for (const call of functionCalls) {
+                const name = call.functionName;
+                const toolUseId = functionCalls.length > 1 ? `${name}_${toolIndex}` : name;
+                toolIndex += 1;
+                // Drop the `null` markers the GBNF grammar uses for optional keys
+                // (see gbnf-schema.ts) so tools receive clean params.
+                const params = pruneOptionalNulls(call.params ?? {}, originalSchemas.get(name));
+                push(new ModelContentBlockStartEvent({
+                    type: "modelContentBlockStartEvent",
+                    start: { type: "toolUseStart", name, toolUseId },
+                }));
+                push(new ModelContentBlockDeltaEvent({
+                    type: "modelContentBlockDeltaEvent",
+                    delta: {
+                        type: "toolUseInputDelta",
+                        input: JSON.stringify(params ?? {}),
+                    },
+                }));
+                push(new ModelContentBlockStopEvent({
+                    type: "modelContentBlockStopEvent",
+                }));
+            }
+            const stopReason = functionCalls.length > 0
+                ? "toolUse"
+                : res.metadata.stopReason === "maxTokens"
+                    ? "maxTokens"
+                    : "endTurn";
+            push(new ModelMessageStopEvent({
+                type: "modelMessageStopEvent",
+                stopReason,
+            }));
+            const inputTokens = chat.sequence.tokenMeter.usedInputTokens - meterBefore.input;
+            const outputTokens = chat.sequence.tokenMeter.usedOutputTokens - meterBefore.output;
+            push(new ModelMetadataEvent({
+                type: "modelMetadataEvent",
+                usage: {
+                    inputTokens,
+                    outputTokens,
+                    totalTokens: inputTokens + outputTokens,
+                },
+            }));
+        })
+            .catch((e) => {
+            failure = e;
+        })
+            .finally(() => {
+            done = true;
+            wake?.();
+        });
+        while (!done || queue.length > 0) {
+            if (queue.length > 0) {
+                yield queue.shift();
+                continue;
+            }
+            await new Promise((resolve) => {
+                wake = resolve;
+            });
+            wake = undefined;
+        }
+        await generation;
+        if (failure !== undefined) {
+            const msg = failure instanceof Error ? failure.message : String(failure);
+            if (msg.toLowerCase().includes("context size")) {
+                throw new ContextWindowOverflowError(msg);
+            }
+            throw new ModelError(`llama.cpp generation error: ${msg}`, {
+                cause: failure,
+            });
+        }
+    }
+}
+//# sourceMappingURL=strands-llama-cpp.js.map

package/dist/core/models/llama-cpp/strands-llama-cpp.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"strands-llama-cpp.js","sourceRoot":"","sources":["../../../../src/core/models/llama-cpp/strands-llama-cpp.ts"],"names":[],"mappings":"AAAA,OAAO,EACL,0BAA0B,EAC1B,KAAK,EACL,UAAU,GACX,MAAM,qBAAqB,CAAC;AAS7B,OAAO,EACL,2BAA2B,EAC3B,2BAA2B,EAC3B,0BAA0B,EAC1B,sBAAsB,EACtB,qBAAqB,EACrB,kBAAkB,GACnB,MAAM,qBAAqB,CAAC;AAa7B,OAAO,EAAE,gBAAgB,EAAE,kBAAkB,EAAE,MAAM,kBAAkB,CAAC;AACxE,OAAO,EAAE,gBAAgB,EAAE,MAAM,oBAAoB,CAAC;AA0BtD;;;;GAIG;AACH,MAAM,oBAAoB,GAAG,IAAI,GAAG,EAA0B,CAAC;AAC/D,MAAM,mBAAmB,GAAG,IAAI,GAAG,EAA+B,CAAC;AAEnE,SAAS,iBAAiB,CAAC,MAAqB;IAC9C,IAAI,MAAM,KAAK,SAAS,EAAE,CAAC;QACzB,OAAO,SAAS,CAAC;IACnB,CAAC;IACD,IAAI,OAAO,MAAM,KAAK,QAAQ,EAAE,CAAC;QAC/B,OAAO,MAAM,CAAC;IAChB,CAAC;IACD,MAAM,KAAK,GAAa,EAAE,CAAC;IAC3B,KAAK,MAAM,KAAK,IAAI,MAAM,EAAE,CAAC;QAC3B,IAAI,KAAK,CAAC,IAAI,KAAK,WAAW,EAAE,CAAC;YAC/B,KAAK,CAAC,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC;QACzB,CAAC;IACH,CAAC;IACD,MAAM,MAAM,GAAG,KAAK,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC,IAAI,EAAE,CAAC;IACvC,OAAO,MAAM,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,SAAS,CAAC;AAChD,CAAC;AAED,SAAS,kBAAkB,CAAC,KAAmB;IAC7C,IAAI,KAAK,CAAC,IAAI,KAAK,WAAW,EAAE,CAAC;QAC/B,OAAO,KAAK,CAAC,IAAI,CAAC;IACpB,CAAC;IACD,IAAI,KAAK,CAAC,IAAI,KAAK,YAAY,EAAE,CAAC;QAChC,OAAO,8DAA8D,CAAC;IACxE,CAAC;IACD,IAAI,KAAK,CAAC,IAAI,KAAK,YAAY,IAAI,KAAK,CAAC,IAAI,KAAK,eAAe,EAAE,CAAC;QAClE,OAAO,eAAe,KAAK,CAAC,IAAI,6CAA6C,CAAC;IAChF,CAAC;IACD,OAAO,SAAS,CAAC;AACnB,CAAC;AAED,SAAS,gBAAgB,CAAC,KAAsB;IAC9C,MAAM,KAAK,GAAa,EAAE,CAAC;IAC3B,KAAK,MAAM,CAAC,IAAI,KAAK,CAAC,OAAO,EAAE,CAAC;QAC9B,IAAI,CAAC,CAAC,IAAI,KAAK,WAAW,EAAE,CAAC;YAC3B,KAAK,CAAC,IAAI,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC;QACrB,CAAC;aAAM,IAAI,CAAC,CAAC,IAAI,KAAK,WAAW,EAAE,CAAC;YAClC,KAAK,CAAC,IAAI,CAAC,IAAI,CAAC,SAAS,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC;QACrC,CAAC;aAAM,IAAI,CAAC,CAAC,IAAI,KAAK,YAAY,EAAE,CAAC;YACnC,KAAK,CAAC,IAAI,CAAC,mDAAmD,CAAC,CAAC;QAClE,CAAC;IACH,CAAC;IACD,MAAM,MAAM,GAAG,KAAK,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;IAChC,IAAI,MAAM,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;QACtB,OAAO,MAAM,CAAC;IAChB,CAAC;IACD,OAAO,KAAK,CAAC,MAAM,KAAK,OAAO,CAAC,CAAC,CAAC,cAAc,CAAC,CAAC,CAAC,EAAE,CAAC;AACxD,CAAC;AAED;;;;;;GAMG;AACH,SAAS,wBAAwB,CAC/B,QAAmB,EACnB,UAA8B;IAE9B,MAAM,OAAO,GAAsB,EAAE,CAAC;IACtC,IAAI,UAAU,EAAE,CAAC;QACf,OAAO,CAAC,IAAI,CAAC,EAAE,IAAI,EAAE,QAAQ,EAAE,IAAI,EAAE,UAAU,EAAE,CAAC,CAAC;IACrD,CAAC;IACD,MAAM,eAAe,GAAG,IAAI,GAAG,EAAwB,CAAC;IAExD,MAAM,iBAAiB,GAAG,GAAsB,EAAE;QAChD,MAAM,IAAI,GAAG,OAAO,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC,CAAC;QAC5B,IAAI,IAAI,EAAE,IAAI,KAAK,OAAO,EAAE,CAAC;YAC3B,OAAO,IAAI,CAAC;QACd,CAAC;QACD,MAAM,OAAO,GAAsB,EAAE,IAAI,EAAE,OAAO,EAAE,QAAQ,EAAE,EAAE,EAAE,CAAC;QACnE,OAAO,CAAC,IAAI,CAAC,OAAO,CAAC,CAAC;QACtB,OAAO,OAAO,CAAC;IACjB,CAAC,CAAC;IAEF,MAAM,YAAY,GAAG,CAAC,IAAY,EAAE,EAAE;QACpC,MAAM,IAAI,GAAG,OAAO,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC,CAAC;QAC5B,IAAI,IAAI,EAAE,IAAI,KAAK,MAAM,EAAE,CAAC;YAC1B,IAAI,CAAC,IAAI,GAAG,IAAI,CAAC,IAAI,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,CAAC,GAAG,IAAI,CAAC,IAAI,KAAK,IAAI,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC;QACpE,CAAC;aAAM,CAAC;YACN,OAAO,CAAC,IAAI,CAAC,EAAE,IAAI,EAAE,MAAM,EAAE,IAAI,EAAE,CAAC,CAAC;QACvC,CAAC;IACH,CAAC,CAAC;IAEF,KAAK,MAAM,GAAG,IAAI,QAAQ,EAAE,CAAC;QAC3B,KAAK,MAAM,KAAK,IAAI,GAAG,CAAC,OAAO,EAAE,CAAC;YAChC,IAAI,GAAG,CAAC,IAAI,KAAK,WAAW,EAAE,CAAC;gBAC7B,IAAI,KAAK,CAAC,IAAI,KAAK,cAAc,EAAE,CAAC;oBAClC,MAAM,CAAC,GAAG,KAAqB,CAAC;oBAChC,eAAe,CAAC,GAAG,CAAC,CAAC,CAAC,SAAS,EAAE,CAAC,CAAC,CAAC;oBACpC,SAAS;gBACX,CAAC;gBACD,IAAI,KAAK,CAAC,IAAI,KAAK,gBAAgB,EAAE,CAAC;oBACpC,kEAAkE;oBAClE,SAAS;gBACX,CAAC;gBACD,MAAM,IAAI,GAAG,kBAAkB,CAAC,KAAK,CAAC,CAAC;gBACvC,IAAI,IAAI,KAAK,SAAS,IAAI,IAAI,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;oBAC1C,iBAAiB,EAAE,CAAC,QAAQ,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;gBAC1C,CAAC;gBACD,SAAS;YACX,CAAC;YACD,YAAY;YACZ,IAAI,KAAK,CAAC,IAAI,KAAK,iBAAiB,EAAE,CAAC;gBACrC,MAAM,CAAC,GAAG,KAAwB,CAAC;gBACnC,MAAM,GAAG,GAAG,eAAe,CAAC,GAAG,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC;gBAC7C,eAAe,CAAC,MAAM,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC;gBACpC,MAAM,IAAI,GAA0B;oBAClC,IAAI,EAAE,cAAc;oBACpB,IAAI,EAAE,GAAG,EAAE,IAAI,IAAI,CAAC,CAAC,SAAS;oBAC9B,MAAM,EAAE,GAAG,EAAE,KAAK,IAAI,EAAE;oBACxB,MAAM,EAAE,gBAAgB,CAAC,CAAC,CAAC;iBAC5B,CAAC;gBACF,MAAM,KAAK,GAAG,iBAAiB,EAAE,CAAC;gBAClC,IACE,KAAK,CAAC,QAAQ,CAAC,MAAM,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,OAAO,IAAI,KAAK,QAAQ,CAAC,CAAC,MAAM,KAAK,CAAC,EACtE,CAAC;oBACD,IAAI,CAAC,cAAc,GAAG,IAAI,CAAC;gBAC7B,CAAC;gBACD,KAAK,CAAC,QAAQ,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;gBAC1B,SAAS;YACX,CAAC;YACD,MAAM,IAAI,GAAG,kBAAkB,CAAC,KAAK,CAAC,CAAC;YACvC,IAAI,IAAI,KAAK,SAAS,IAAI,IAAI,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;gBAC1C,YAAY,CAAC,IAAI,CAAC,CAAC;YACrB,CAAC;QACH,CAAC;IACH,CAAC;IACD,OAAO,OAAO,CAAC;AACjB,CAAC;AAED;;;;;;GAMG;AACH,SAAS,uBAAuB,CAC9B,SAAiC;IAEjC,IAAI,CAAC,SAAS,EAAE,MAAM,EAAE,CAAC;QACvB,OAAO,SAAS,CAAC;IACnB,CAAC;IACD,MAAM,SAAS,GACb,EAAE,CAAC;IACL,KAAK,MAAM,IAAI,IAAI,SAAS,EAAE,CAAC;QAC7B,SAAS,CAAC,IAAI,CAAC,IAAI,CAAC,GAAG;YACrB,GAAG,CAAC,IAAI,CAAC,WAAW,CAAC,CAAC,CAAC,EAAE,WAAW,EAAE,IAAI,CAAC,WAAW,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;YAC9D,GAAG,CAAC,IAAI,CAAC,WAAW;gBAClB,CAAC,CAAC,EAAE,MAAM,EAAE,gBAAgB,CAAC,IAAI,CAAC,WAAW,CAAC,EAAE;gBAChD,CAAC,CAAC,EAAE,CAAC;SACR,CAAC;IACJ,CAAC;IACD,OAAO,SAA+B,CAAC;AACzC,CAAC;AAED,iFAAiF;AACjF,MAAM,OAAO,oBAAqB,SAAQ,KAA0B;IAC1D,MAAM,CAAsB;IAC5B,WAAW,CAAiC;IAC5C,OAAO,CAA2B;IAE1C,YAAY,MAA2B;QACrC,KAAK,EAAE,CAAC;QACR,IAAI,CAAC,MAAM,GAAG,EAAE,GAAG,MAAM,EAAE,CAAC;IAC9B,CAAC;IAED,YAAY,CAAC,WAAgC;QAC3C,MAAM,OAAO,GAAG,CAAC,CAAsB,EAAE,EAAE,CACzC,IAAI,CAAC,SAAS,CAAC,CAAC,CAAC,CAAC,OAAO,EAAE,CAAC,CAAC,GAAG,EAAE,CAAC,CAAC,WAAW,EAAE,CAAC,CAAC,SAAS,IAAI,IAAI,CAAC,CAAC,CAAC;QACzE,MAAM,MAAM,GAAG,OAAO,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC;QACpC,IAAI,CAAC,MAAM,GAAG,EAAE,GAAG,IAAI,CAAC,MAAM,EAAE,GAAG,WAAW,EAAE,CAAC;QACjD,IAAI,OAAO,CAAC,IAAI,CAAC,MAAM,CAAC,KAAK,MAAM,EAAE,CAAC;YACpC,wEAAwE;YACxE,qEAAqE;YACrE,6DAA6D;YAC7D,IAAI,CAAC,WAAW,GAAG,SAAS,CAAC;YAC7B,IAAI,CAAC,OAAO,GAAG,SAAS,CAAC;QAC3B,CAAC;IACH,CAAC;IAED,SAAS;QACP,OAAO,EAAE,GAAG,IAAI,CAAC,MAAM,EAAE,CAAC;IAC5B,CAAC;IAEO,OAAO;QACb,IAAI,CAAC,WAAW,KAAK,IAAI,CAAC,QAAQ,EAAE,CAAC;QACrC,OAAO,IAAI,CAAC,WAAW,CAAC;IAC1B,CAAC;IAEO,KAAK,CAAC,QAAQ;QACpB,MAAM,OAAO,GAAG,IAAI,CAAC,MAAM,CAAC,OAAO,CAAC;QACpC,IAAI,CAAC,OAAO,EAAE,CAAC;YACb,MAAM,IAAI,UAAU,CAAC,mCAAmC,CAAC,CAAC;QAC5D,CAAC;QACD,MAAM,EAAE,QAAQ,EAAE,SAAS,EAAE,aAAa,EAAE,kBAAkB,EAAE,GAC9D,MAAM,MAAM,CAAC,gBAAgB,CAAC,CAAC;QACjC,MAAM,SAAS,GAAG,MAAM,gBAAgB,CAAC,OAAO,EAAE,IAAI,CAAC,MAAM,CAAC,OAAO,CAAC,CAAC;QACvE,MAAM,GAAG,GAAG,IAAI,CAAC,MAAM,CAAC,GAAG,IAAI,MAAM,CAAC;QACtC,MAAM,UAAU,GAAG,IAAI,CAAC,SAAS,CAAC,GAAG,CAAC,CAAC;QACvC,IAAI,YAAY,GAAG,oBAAoB,CAAC,GAAG,CAAC,UAAU,CAAC,CAAC;QACxD,IAAI,CAAC,YAAY,EAAE,CAAC;YAClB,6DAA6D;YAC7D,YAAY,GAAG,QAAQ,CAAC,EAAE,GAAG,EAAE,QAAQ,EAAE,aAAa,CAAC,KAAK,EAAE,CAAC,CAAC;YAChE,oBAAoB,CAAC,GAAG,CAAC,UAAU,EAAE,YAAY,CAAC,CAAC;QACrD,CAAC;QACD,MAAM,KAAK,GAAG,MAAM,YAAY,CAAC;QACjC,IAAI,YAAY,GAAG,mBAAmB,CAAC,GAAG,CAAC,SAAS,CAAC,CAAC;QACtD,IAAI,CAAC,YAAY,EAAE,CAAC;YAClB,YAAY,GAAG,KAAK,CAAC,SAAS,CAAC,EAAE,SAAS,EAAE,CAAC,CAAC;YAC9C,mBAAmB,CAAC,GAAG,CAAC,SAAS,EAAE,YAAY,CAAC,CAAC;QACnD,CAAC;QACD,MAAM,KAAK,GAAG,MAAM,YAAY,CAAC;QACjC,IAAI,CAAC,OAAO,GAAG,MAAM,KAAK,CAAC,aAAa,CAAC;YACvC,GAAG,CAAC,IAAI,CAAC,MAAM,CAAC,WAAW,KAAK,SAAS;gBACvC,CAAC,CAAC,EAAE,WAAW,EAAE,IAAI,CAAC,MAAM,CAAC,WAAW,EAAE;gBAC1C,CAAC,CAAC,EAAE,CAAC;SACR,CAAC,CAAC;QACH,sEAAsE;QACtE,uEAAuE;QACvE,gEAAgE;QAChE,MAAM,gBAAgB,GAAG,IAAI,CAAC,MAAM,CAAC,SAAS,KAAK,SAAS,CAAC;QAC7D,MAAM,MAAM,GAAG,IAAI,CAAC,MAAM,CAAC,SAAS,EAAE,MAAM,CAAC;QAC7C,MAAM,aAAa,GACjB,MAAM,KAAK,MAAM,IAAI,MAAM,KAAK,QAAQ,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,KAAK,CAAC;QAC5D,MAAM,WAAW,GAAG,kBAAkB,CAAC,KAAK,EAAE;YAC5C,qBAAqB,EAAE;gBACrB,IAAI,EAAE,EAAE,QAAQ,EAAE,gBAAgB,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,YAAY,EAAE;gBAC5D,MAAM,EAAE,EAAE,SAAS,EAAE,gBAAgB,EAAE;gBACvC,OAAO,EAAE;oBACP,eAAe,EAAE,gBAAgB,CAAC,CAAC,CAAC,aAAa,CAAC,CAAC,CAAC,KAAK;iBAC1D;aACF;SACF,CAAC,CAAC;QACH,OAAO,IAAI,SAAS,CAAC;YACnB,eAAe,EAAE,IAAI,CAAC,OAAO,CAAC,WAAW,EAAE;YAC3C,WAAW;SACZ,CAAC,CAAC;IACL,CAAC;IAED,KAAK,CAAC,CAAC,MAAM,CACX,QAAmB,EACnB,OAAuB;QAEvB,IAAI,IAAe,CAAC;QACpB,IAAI,CAAC;YACH,IAAI,GAAG,MAAM,IAAI,CAAC,OAAO,EAAE,CAAC;QAC9B,CAAC;QAAC,OAAO,CAAC,EAAE,CAAC;YACX,mEAAmE;YACnE,IAAI,CAAC,WAAW,GAAG,SAAS,CAAC;YAC7B,IAAI,CAAC,YAAY,UAAU,EAAE,CAAC;gBAC5B,MAAM,CAAC,CAAC;YACV,CAAC;YACD,MAAM,GAAG,GAAG,CAAC,YAAY,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC;YACvD,MAAM,IAAI,UAAU,CAAC,oCAAoC,GAAG,EAAE,EAAE;gBAC9D,KAAK,EAAE,CAAC;aACT,CAAC,CAAC;QACL,CAAC;QAED,MAAM,UAAU,GAAG,iBAAiB,CAAC,OAAO,EAAE,YAAY,CAAC,CAAC;QAC5D,MAAM,OAAO,GAAG,wBAAwB,CAAC,QAAQ,EAAE,UAAU,CAAC,CAAC;QAC/D,IAAI,OAAO,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC,EAAE,IAAI,KAAK,MAAM,EAAE,CAAC;YACpC,OAAO,CAAC,IAAI,CAAC,EAAE,IAAI,EAAE,OAAO,EAAE,QAAQ,EAAE,EAAE,EAAE,CAAC,CAAC;QAChD,CAAC;QACD,MAAM,SAAS,GAAG,uBAAuB,CAAC,OAAO,EAAE,SAAS,CAAC,CAAC;QAC9D,MAAM,eAAe,GAAG,IAAI,GAAG,CAC7B,CAAC,OAAO,EAAE,SAAS,IAAI,EAAE,CAAC,CAAC,GAAG,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,CAAC,IAAI,CAAC,IAAI,EAAE,IAAI,CAAC,WAAW,CAAC,CAAC,CACxE,CAAC;QAEF,MAAM,IAAI,sBAAsB,CAAC;YAC/B,IAAI,EAAE,wBAAwB;YAC9B,IAAI,EAAE,WAAW;SAClB,CAAC,CAAC;QAEH,MAAM,KAAK,GAAuB,EAAE,CAAC;QACrC,IAAI,IAA8B,CAAC;QACnC,MAAM,IAAI,GAAG,CAAC,KAAuB,EAAE,EAAE;YACvC,KAAK,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC;YAClB,IAAI,EAAE,EAAE,CAAC;QACX,CAAC,CAAC;QAEF,IAAI,aAAa,GAAG,KAAK,CAAC;QAC1B,IAAI,kBAAkB,GAAG,KAAK,CAAC;QAC/B,MAAM,eAAe,GAAG,GAAG,EAAE;YAC3B,IAAI,kBAAkB,EAAE,CAAC;gBACvB,IAAI,CACF,IAAI,0BAA0B,CAAC;oBAC7B,IAAI,EAAE,4BAA4B;iBACnC,CAAC,CACH,CAAC;gBACF,kBAAkB,GAAG,KAAK,CAAC;YAC7B,CAAC;YACD,IAAI,aAAa,EAAE,CAAC;gBAClB,IAAI,CACF,IAAI,0BAA0B,CAAC;oBAC7B,IAAI,EAAE,4BAA4B;iBACnC,CAAC,CACH,CAAC;gBACF,aAAa,GAAG,KAAK,CAAC;YACxB,CAAC;QACH,CAAC,CAAC;QAEF,MAAM,WAAW,GAAG;YAClB,KAAK,EAAE,IAAI,CAAC,QAAQ,CAAC,UAAU,CAAC,eAAe;YAC/C,MAAM,EAAE,IAAI,CAAC,QAAQ,CAAC,UAAU,CAAC,gBAAgB;SAClD,CAAC;QAEF,2EAA2E;QAC3E,yEAAyE;QACzE,kCAAkC;QAClC,MAAM,aAAa,GACjB,IAAI,CAAC,MAAM,CAAC,SAAS,KAAK,SAAS,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,MAAM,CAAC,mBAAmB,CAAC;QAE5E,IAAI,IAAI,GAAG,KAAK,CAAC;QACjB,IAAI,OAAgB,CAAC;QACrB,MAAM,UAAU,GAAG,IAAI;aACpB,gBAAgB,CAAC,OAAO,EAAE;YACzB,GAAG,CAAC,SAAS,CAAC,CAAC,CAAC,EAAE,SAAS,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;YACnC,GAAG,CAAC,IAAI,CAAC,MAAM,CAAC,SAAS,KAAK,SAAS;gBACrC,CAAC,CAAC,EAAE,SAAS,EAAE,IAAI,CAAC,MAAM,CAAC,SAAS,EAAE;gBACtC,CAAC,CAAC,EAAE,CAAC;YACP,GAAG,CAAC,IAAI,CAAC,MAAM,CAAC,WAAW,KAAK,SAAS;gBACvC,CAAC,CAAC,EAAE,WAAW,EAAE,IAAI,CAAC,MAAM,CAAC,WAAW,EAAE;gBAC1C,CAAC,CAAC,EAAE,CAAC;YACP,GAAG,CAAC,IAAI,CAAC,MAAM,CAAC,IAAI,KAAK,SAAS,CAAC,CAAC,CAAC,EAAE,IAAI,EAAE,IAAI,CAAC,MAAM,CAAC,IAAI,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;YACrE,GAAG,CAAC,aAAa,KAAK,SAAS;gBAC7B,CAAC,CAAC,EAAE,OAAO,EAAE,EAAE,aAAa,EAAE,aAAa,EAAE,EAAE;gBAC/C,CAAC,CAAC,EAAE,CAAC;YACP,eAAe,EAAE,CAAC,KAAK,EAAE,EAAE;gBACzB,IAAI,KAAK,CAAC,IAAI,KAAK,SAAS,EAAE,CAAC;oBAC7B,IAAI,KAAK,CAAC,WAAW,KAAK,SAAS,IAAI,KAAK,CAAC,IAAI,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;wBAC/D,OAAO;oBACT,CAAC;oBACD,IAAI,aAAa,EAAE,CAAC;wBAClB,IAAI,CACF,IAAI,0BAA0B,CAAC;4BAC7B,IAAI,EAAE,4BAA4B;yBACnC,CAAC,CACH,CAAC;wBACF,aAAa,GAAG,KAAK,CAAC;oBACxB,CAAC;oBACD,IAAI,CAAC,kBAAkB,EAAE,CAAC;wBACxB,IAAI,CACF,IAAI,2BAA2B,CAAC;4BAC9B,IAAI,EAAE,6BAA6B;yBACpC,CAAC,CACH,CAAC;wBACF,kBAAkB,GAAG,IAAI,CAAC;oBAC5B,CAAC;oBACD,IAAI,CACF,IAAI,2BAA2B,CAAC;wBAC9B,IAAI,EAAE,6BAA6B;wBACnC,KAAK,EAAE,EAAE,IAAI,EAAE,uBAAuB,EAAE,IAAI,EAAE,KAAK,CAAC,IAAI,EAAE;qBAC3D,CAAC,CACH,CAAC;oBACF,OAAO;gBACT,CAAC;gBACD,IAAI,KAAK,CAAC,IAAI,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;oBAC5B,OAAO;gBACT,CAAC;gBACD,IAAI,kBAAkB,EAAE,CAAC;oBACvB,IAAI,CACF,IAAI,0BAA0B,CAAC;wBAC7B,IAAI,EAAE,4BAA4B;qBACnC,CAAC,CACH,CAAC;oBACF,kBAAkB,GAAG,KAAK,CAAC;gBAC7B,CAAC;gBACD,IAAI,CAAC,aAAa,EAAE,CAAC;oBACnB,IAAI,CACF,IAAI,2BAA2B,CAAC;wBAC9B,IAAI,EAAE,6BAA6B;qBACpC,CAAC,CACH,CAAC;oBACF,aAAa,GAAG,IAAI,CAAC;gBACvB,CAAC;gBACD,IAAI,CACF,IAAI,2BAA2B,CAAC;oBAC9B,IAAI,EAAE,6BAA6B;oBACnC,KAAK,EAAE,EAAE,IAAI,EAAE,WAAW,EAAE,IAAI,EAAE,KAAK,CAAC,IAAI,EAAE;iBAC/C,CAAC,CACH,CAAC;YACJ,CAAC;SACF,CAAC;aACD,IAAI,CAAC,CAAC,GAAG,EAAE,EAAE;YACZ,eAAe,EAAE,CAAC;YAElB,MAAM,aAAa,GAAG,GAAG,CAAC,aAAa,IAAI,EAAE,CAAC;YAC9C,IAAI,SAAS,GAAG,CAAC,CAAC;YAClB,KAAK,MAAM,IAAI,IAAI,aAAa,EAAE,CAAC;gBACjC,MAAM,IAAI,GAAG,IAAI,CAAC,YAAY,CAAC;gBAC/B,MAAM,SAAS,GACb,aAAa,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,CAAC,GAAG,IAAI,IAAI,SAAS,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC;gBAC3D,SAAS,IAAI,CAAC,CAAC;gBACf,kEAAkE;gBAClE,sDAAsD;gBACtD,MAAM,MAAM,GAAG,kBAAkB,CAC/B,IAAI,CAAC,MAAM,IAAI,EAAE,EACjB,eAAe,CAAC,GAAG,CAAC,IAAI,CAAC,CAC1B,CAAC;gBACF,IAAI,CACF,IAAI,2BAA2B,CAAC;oBAC9B,IAAI,EAAE,6BAA6B;oBACnC,KAAK,EAAE,EAAE,IAAI,EAAE,cAAc,EAAE,IAAI,EAAE,SAAS,EAAE;iBACjD,CAAC,CACH,CAAC;gBACF,IAAI,CACF,IAAI,2BAA2B,CAAC;oBAC9B,IAAI,EAAE,6BAA6B;oBACnC,KAAK,EAAE;wBACL,IAAI,EAAE,mBAAmB;wBACzB,KAAK,EAAE,IAAI,CAAC,SAAS,CAAC,MAAM,IAAI,EAAE,CAAC;qBACpC;iBACF,CAAC,CACH,CAAC;gBACF,IAAI,CACF,IAAI,0BAA0B,CAAC;oBAC7B,IAAI,EAAE,4BAA4B;iBACnC,CAAC,CACH,CAAC;YACJ,CAAC;YAED,MAAM,UAAU,GACd,aAAa,CAAC,MAAM,GAAG,CAAC;gBACtB,CAAC,CAAE,SAAmB;gBACtB,CAAC,CAAC,GAAG,CAAC,QAAQ,CAAC,UAAU,KAAK,WAAW;oBACvC,CAAC,CAAE,WAAqB;oBACxB,CAAC,CAAE,SAAmB,CAAC;YAC7B,IAAI,CACF,IAAI,qBAAqB,CAAC;gBACxB,IAAI,EAAE,uBAAuB;gBAC7B,UAAU;aACX,CAAC,CACH,CAAC;YAEF,MAAM,WAAW,GACf,IAAI,CAAC,QAAQ,CAAC,UAAU,CAAC,eAAe,GAAG,WAAW,CAAC,KAAK,CAAC;YAC/D,MAAM,YAAY,GAChB,IAAI,CAAC,QAAQ,CAAC,UAAU,CAAC,gBAAgB,GAAG,WAAW,CAAC,MAAM,CAAC;YACjE,IAAI,CACF,IAAI,kBAAkB,CAAC;gBACrB,IAAI,EAAE,oBAAoB;gBAC1B,KAAK,EAAE;oBACL,WAAW;oBACX,YAAY;oBACZ,WAAW,EAAE,WAAW,GAAG,YAAY;iBACxC;aACF,CAAC,CACH,CAAC;QACJ,CAAC,CAAC;aACD,KAAK,CAAC,CAAC,CAAU,EAAE,EAAE;YACpB,OAAO,GAAG,CAAC,CAAC;QACd,CAAC,CAAC;aACD,OAAO,CAAC,GAAG,EAAE;YACZ,IAAI,GAAG,IAAI,CAAC;YACZ,IAAI,EAAE,EAAE,CAAC;QACX,CAAC,CAAC,CAAC;QAEL,OAAO,CAAC,IAAI,IAAI,KAAK,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;YACjC,IAAI,KAAK,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;gBACrB,MAAM,KAAK,CAAC,KAAK,EAAG,CAAC;gBACrB,SAAS;YACX,CAAC;YACD,MAAM,IAAI,OAAO,CAAO,CAAC,OAAO,EAAE,EAAE;gBAClC,IAAI,GAAG,OAAO,CAAC;YACjB,CAAC,CAAC,CAAC;YACH,IAAI,GAAG,SAAS,CAAC;QACnB,CAAC;QACD,MAAM,UAAU,CAAC;QAEjB,IAAI,OAAO,KAAK,SAAS,EAAE,CAAC;YAC1B,MAAM,GAAG,GAAG,OAAO,YAAY,KAAK,CAAC,CAAC,CAAC,OAAO,CAAC,OAAO,CAAC,CAAC,CAAC,MAAM,CAAC,OAAO,CAAC,CAAC;YACzE,IAAI,GAAG,CAAC,WAAW,EAAE,CAAC,QAAQ,CAAC,cAAc,CAAC,EAAE,CAAC;gBAC/C,MAAM,IAAI,0BAA0B,CAAC,GAAG,CAAC,CAAC;YAC5C,CAAC;YACD,MAAM,IAAI,UAAU,CAAC,+BAA+B,GAAG,EAAE,EAAE;gBACzD,KAAK,EAAE,OAAO;aACf,CAAC,CAAC;QACL,CAAC;IACH,CAAC;CACF"}

package/dist/core/models/types.d.ts

@@ -5,6 +5,7 @@ export declare enum LlmProvider {
     Bedrock = "bedrock",
     Google = "google",
     Groq = "groq",
+    LlamaCpp = "llama-cpp",
     Minimax = "minimax",
     Moonshot = "moonshot",
     Ollama = "ollama",
@@ -130,6 +131,32 @@ export type GroqProviderOptions = {
      */
     reasoning?: ReasoningOptions;
 };
+export type LlamaCppProviderOptions = {
+    /**
+     * Hugging Face access token used when downloading GGUF weights from the Hub
+     * (gated/private repos). Falls back to the `HF_TOKEN` env var when unset.
+     */
+    hfToken?: string;
+    /**
+     * GPU backend forwarded to node-llama-cpp's `getLlama` (default `"auto"`).
+     * Set `false` to force CPU-only inference.
+     */
+    gpu?: "auto" | "metal" | "cuda" | "vulkan" | false;
+    /**
+     * Context size in tokens for the llama.cpp context. When omitted,
+     * node-llama-cpp adapts it to the model's training context and free memory.
+     */
+    contextSize?: number;
+    /**
+     * Reasoning controls. Providing `reasoning` enables thinking: the chat
+     * template is configured to allow thought segments (Qwen `thoughts: "auto"`,
+     * Gemma 4 `reasoning: true`, gpt-oss/Harmony native reasoning-effort) and
+     * `effort` caps thought tokens via node-llama-cpp's thought budget
+     * (1024/2048/4096/8192, default `medium`). Omit `reasoning` to disable
+     * thinking (templates discourage thoughts, thought budget forced to 0).
+     */
+    reasoning?: ReasoningOptions;
+};
 export type MinimaxProviderOptions = {
     apiKey?: string;
     baseURL?: string;
@@ -211,7 +238,7 @@ export type XaiProviderOptions = {
      */
     reasoning?: ReasoningOptions;
 };
-export type ProviderOptions = AnthropicProviderOptions | AzureProviderOptions | BedrockProviderOptions | GoogleProviderOptions | GroqProviderOptions | MinimaxProviderOptions | MoonshotProviderOptions | OllamaProviderOptions | OpenAIProviderOptions | OpenRouterProviderOptions | XaiProviderOptions;
+export type ProviderOptions = AnthropicProviderOptions | AzureProviderOptions | BedrockProviderOptions | GoogleProviderOptions | GroqProviderOptions | LlamaCppProviderOptions | MinimaxProviderOptions | MoonshotProviderOptions | OllamaProviderOptions | OpenAIProviderOptions | OpenRouterProviderOptions | XaiProviderOptions;
 export declare const OpenAIApiSchema: z.ZodEnum<{
     chat: "chat";
     responses: "responses";
@@ -387,6 +414,34 @@ export declare const GroqProviderOptionsSchema: z.ZodObject<{
         }>>;
     }, z.core.$strict>>;
 }, z.core.$strict>;
+export declare const LlamaCppProviderOptionsSchema: z.ZodObject<{
+    hfToken: z.ZodOptional<z.ZodString>;
+    gpu: z.ZodOptional<z.ZodUnion<readonly [z.ZodEnum<{
+        auto: "auto";
+        metal: "metal";
+        cuda: "cuda";
+        vulkan: "vulkan";
+    }>, z.ZodLiteral<false>]>>;
+    contextSize: z.ZodOptional<z.ZodNumber>;
+    reasoning: z.ZodOptional<z.ZodObject<{
+        effort: z.ZodOptional<z.ZodEnum<{
+            minimal: "minimal";
+            low: "low";
+            medium: "medium";
+            high: "high";
+        }>>;
+        summary: z.ZodOptional<z.ZodEnum<{
+            auto: "auto";
+            concise: "concise";
+            detailed: "detailed";
+            none: "none";
+        }>>;
+        display: z.ZodOptional<z.ZodEnum<{
+            summarized: "summarized";
+            omitted: "omitted";
+        }>>;
+    }, z.core.$strict>>;
+}, z.core.$strict>;
 export declare const MinimaxProviderOptionsSchema: z.ZodObject<{
     apiKey: z.ZodOptional<z.ZodString>;
     baseURL: z.ZodOptional<z.ZodString>;
@@ -646,6 +701,34 @@ export declare const ProviderOptionsSchemas: {
             }>>;
         }, z.core.$strict>>;
     }, z.core.$strict>;
+    readonly "llama-cpp": z.ZodObject<{
+        hfToken: z.ZodOptional<z.ZodString>;
+        gpu: z.ZodOptional<z.ZodUnion<readonly [z.ZodEnum<{
+            auto: "auto";
+            metal: "metal";
+            cuda: "cuda";
+            vulkan: "vulkan";
+        }>, z.ZodLiteral<false>]>>;
+        contextSize: z.ZodOptional<z.ZodNumber>;
+        reasoning: z.ZodOptional<z.ZodObject<{
+            effort: z.ZodOptional<z.ZodEnum<{
+                minimal: "minimal";
+                low: "low";
+                medium: "medium";
+                high: "high";
+            }>>;
+            summary: z.ZodOptional<z.ZodEnum<{
+                auto: "auto";
+                concise: "concise";
+                detailed: "detailed";
+                none: "none";
+            }>>;
+            display: z.ZodOptional<z.ZodEnum<{
+                summarized: "summarized";
+                omitted: "omitted";
+            }>>;
+        }, z.core.$strict>>;
+    }, z.core.$strict>;
     readonly minimax: z.ZodObject<{
         apiKey: z.ZodOptional<z.ZodString>;
         baseURL: z.ZodOptional<z.ZodString>;
@@ -920,6 +1003,37 @@ export declare const NamedProviderConfigSchema: z.ZodDiscriminatedUnion<[z.ZodOb
             }>>;
         }, z.core.$strict>>;
     }, z.core.$strict>;
+}, z.core.$strict>, z.ZodObject<{
+    name: z.ZodString;
+    provider: z.ZodLiteral<LlmProvider.LlamaCpp>;
+    options: z.ZodObject<{
+        hfToken: z.ZodOptional<z.ZodString>;
+        gpu: z.ZodOptional<z.ZodUnion<readonly [z.ZodEnum<{
+            auto: "auto";
+            metal: "metal";
+            cuda: "cuda";
+            vulkan: "vulkan";
+        }>, z.ZodLiteral<false>]>>;
+        contextSize: z.ZodOptional<z.ZodNumber>;
+        reasoning: z.ZodOptional<z.ZodObject<{
+            effort: z.ZodOptional<z.ZodEnum<{
+                minimal: "minimal";
+                low: "low";
+                medium: "medium";
+                high: "high";
+            }>>;
+            summary: z.ZodOptional<z.ZodEnum<{
+                auto: "auto";
+                concise: "concise";
+                detailed: "detailed";
+                none: "none";
+            }>>;
+            display: z.ZodOptional<z.ZodEnum<{
+                summarized: "summarized";
+                omitted: "omitted";
+            }>>;
+        }, z.core.$strict>>;
+    }, z.core.$strict>;
 }, z.core.$strict>, z.ZodObject<{
     name: z.ZodString;
     provider: z.ZodLiteral<LlmProvider.Minimax>;

package/dist/core/models/types.js

@@ -6,6 +6,7 @@ export var LlmProvider;
     LlmProvider["Bedrock"] = "bedrock";
     LlmProvider["Google"] = "google";
     LlmProvider["Groq"] = "groq";
+    LlmProvider["LlamaCpp"] = "llama-cpp";
     LlmProvider["Minimax"] = "minimax";
     LlmProvider["Moonshot"] = "moonshot";
     LlmProvider["Ollama"] = "ollama";
@@ -121,6 +122,16 @@ export const GroqProviderOptionsSchema = z
     reasoning: ReasoningOptionsSchema.optional(),
 })
     .strict();
+export const LlamaCppProviderOptionsSchema = z
+    .object({
+    hfToken: NonEmptyStringSchema.optional(),
+    gpu: z
+        .union([z.enum(["auto", "metal", "cuda", "vulkan"]), z.literal(false)])
+        .optional(),
+    contextSize: z.number().int().positive().optional(),
+    reasoning: ReasoningOptionsSchema.optional(),
+})
+    .strict();
 export const MinimaxProviderOptionsSchema = z
     .object({
     apiKey: NonEmptyStringSchema.optional(),
@@ -174,6 +185,7 @@ export const ProviderOptionsSchemas = {
     [LlmProvider.Bedrock]: BedrockProviderOptionsSchema,
     [LlmProvider.Google]: GoogleProviderOptionsSchema,
     [LlmProvider.Groq]: GroqProviderOptionsSchema,
+    [LlmProvider.LlamaCpp]: LlamaCppProviderOptionsSchema,
     [LlmProvider.Minimax]: MinimaxProviderOptionsSchema,
     [LlmProvider.Moonshot]: MoonshotProviderOptionsSchema,
     [LlmProvider.Ollama]: OllamaProviderOptionsSchema,
@@ -217,6 +229,13 @@ export const NamedProviderConfigSchema = z.discriminatedUnion("provider", [
         options: GroqProviderOptionsSchema,
     })
         .strict(),
+    z
+        .object({
+        name: NonEmptyStringSchema,
+        provider: z.literal(LlmProvider.LlamaCpp),
+        options: LlamaCppProviderOptionsSchema,
+    })
+        .strict(),
     z
         .object({
         name: NonEmptyStringSchema,