vibn-agent-runner/dist/llm.js

"use strict";
Object.defineProperty(exports, "__esModule", { value: true });
exports.GeminiClient = exports.VertexOpenAIClient = void 0;
exports.createLLM = createLLM;
exports.toOAITools = toOAITools;
const child_process_1 = require("child_process");
const genai_1 = require("@google/genai");
const uuid_1 = require("uuid");
// ---------------------------------------------------------------------------
// Vertex AI OpenAI-compatible client
// Used for: zai-org/glm-5-maas, anthropic/claude-sonnet-4-6, etc.
// ---------------------------------------------------------------------------
let _cachedToken = '';
let _tokenExpiry = 0;
function getVertexToken() {
    const now = Date.now();
    if (_cachedToken && now < _tokenExpiry)
        return _cachedToken;
    _cachedToken = (0, child_process_1.execSync)('gcloud auth print-access-token', { encoding: 'utf8' }).trim();
    _tokenExpiry = now + 55 * 60 * 1000; // tokens last 1hr, refresh at 55min
    return _cachedToken;
}
class VertexOpenAIClient {
    constructor(modelId, opts) {
        this.modelId = modelId;
        this.projectId = opts?.projectId ?? process.env.GCP_PROJECT_ID ?? 'master-ai-484822';
        this.region = opts?.region ?? 'global';
        this.temperature = opts?.temperature ?? 0.3;
    }
    async chat(messages, tools, maxTokens = 4096) {
        const token = getVertexToken();
        const base = this.region === 'global'
            ? 'https://aiplatform.googleapis.com'
            : `https://${this.region}-aiplatform.googleapis.com`;
        const url = `${base}/v1/projects/${this.projectId}/locations/${this.region}/endpoints/openapi/chat/completions`;
        const body = {
            model: this.modelId,
            messages,
            max_tokens: maxTokens,
            temperature: this.temperature,
            stream: false
        };
        if (tools && tools.length > 0) {
            body.tools = tools;
            body.tool_choice = 'auto';
        }
        const res = await fetch(url, {
            method: 'POST',
            headers: {
                'Authorization': `Bearer ${token}`,
                'Content-Type': 'application/json'
            },
            body: JSON.stringify(body)
        });
        if (!res.ok) {
            const errText = await res.text();
            // Force token refresh on 401
            if (res.status === 401)
                _tokenExpiry = 0;
            throw new Error(`Vertex API ${res.status}: ${errText.slice(0, 400)}`);
        }
        const data = await res.json();
        const choice = data.choices?.[0];
        const message = choice?.message ?? {};
        return {
            content: message.content ?? null,
            reasoning: message.reasoning_content ?? null,
            tool_calls: message.tool_calls ?? [],
            finish_reason: choice?.finish_reason ?? 'stop',
            usage: data.usage
        };
    }
}
exports.VertexOpenAIClient = VertexOpenAIClient;
// ---------------------------------------------------------------------------
// Gemini client via @google/genai SDK
// Used for: Tier A (fast/cheap routing, summaries, log parsing)
// Converts to/from OpenAI message format internally.
// ---------------------------------------------------------------------------
class GeminiClient {
    constructor(modelId = 'gemini-2.5-flash', opts) {
        this.modelId = modelId;
        this.temperature = opts?.temperature ?? 0.2;
    }
    async chat(messages, tools, maxTokens = 8192) {
        const apiKey = process.env.GOOGLE_API_KEY;
        if (!apiKey)
            throw new Error('GOOGLE_API_KEY not set');
        const genai = new genai_1.GoogleGenAI({ apiKey });
        const systemMsg = messages.find(m => m.role === 'system');
        const nonSystem = messages.filter(m => m.role !== 'system');
        const functionDeclarations = (tools ?? []).map(t => ({
            name: t.function.name,
            description: t.function.description,
            parameters: t.function.parameters
        }));
        const response = await genai.models.generateContent({
            model: this.modelId,
            contents: toGeminiContents(nonSystem),
            config: {
                systemInstruction: systemMsg?.content ?? undefined,
                tools: functionDeclarations.length > 0 ? [{ functionDeclarations }] : undefined,
                temperature: this.temperature,
                maxOutputTokens: maxTokens
            }
        });
        const candidate = response.candidates?.[0];
        if (!candidate)
            throw new Error('No response from Gemini');
        const parts = candidate.content?.parts ?? [];
        const textContent = parts.filter(p => p.text).map(p => p.text).join('') || null;
        const fnCalls = parts.filter(p => p.functionCall);
        const tool_calls = fnCalls.map(p => ({
            id: `call_${(0, uuid_1.v4)().replace(/-/g, '').slice(0, 12)}`,
            type: 'function',
            function: {
                name: p.functionCall.name ?? '',
                arguments: JSON.stringify(p.functionCall.args ?? {})
            }
        }));
        return {
            content: textContent,
            reasoning: null,
            tool_calls,
            finish_reason: fnCalls.length > 0 ? 'tool_calls' : 'stop'
        };
    }
}
exports.GeminiClient = GeminiClient;
/** Convert OpenAI message format → Gemini Content[] format */
function toGeminiContents(messages) {
    const contents = [];
    for (const msg of messages) {
        if (msg.role === 'assistant') {
            const parts = [];
            if (msg.content)
                parts.push({ text: msg.content });
            for (const tc of msg.tool_calls ?? []) {
                parts.push({
                    functionCall: {
                        name: tc.function.name,
                        args: JSON.parse(tc.function.arguments || '{}')
                    }
                });
            }
            contents.push({ role: 'model', parts });
        }
        else if (msg.role === 'tool') {
            // Parse content back — could be JSON or plain text
            let resultValue = msg.content;
            try {
                resultValue = JSON.parse(msg.content ?? 'null');
            }
            catch { /* keep as string */ }
            contents.push({
                role: 'user',
                parts: [{
                        functionResponse: {
                            name: msg.name ?? 'tool',
                            response: { result: resultValue }
                        }
                    }]
            });
        }
        else {
            contents.push({ role: 'user', parts: [{ text: msg.content ?? '' }] });
        }
    }
    return contents;
}
const TIER_MODELS = {
    A: process.env.TIER_A_MODEL ?? 'gemini-2.5-flash',
    B: process.env.TIER_B_MODEL ?? 'zai-org/glm-5-maas',
    C: process.env.TIER_C_MODEL ?? 'zai-org/glm-5-maas'
};
function createLLM(modelOrTier, opts) {
    const modelId = (modelOrTier === 'A' || modelOrTier === 'B' || modelOrTier === 'C')
        ? TIER_MODELS[modelOrTier]
        : modelOrTier;
    if (modelId.startsWith('gemini-')) {
        return new GeminiClient(modelId, opts);
    }
    return new VertexOpenAIClient(modelId, { temperature: opts?.temperature });
}
// ---------------------------------------------------------------------------
// Helper — convert our ToolDefinition[] → LLMTool[] (OpenAI format)
// ---------------------------------------------------------------------------
function toOAITools(tools) {
    return tools.map(t => ({
        type: 'function',
        function: {
            name: t.name,
            description: t.description,
            parameters: t.parameters
        }
    }));
}