perf-lens/dist/ai/client.js

AI-powered frontend performance optimizer

import Anthropic from '@anthropic-ai/sdk';
import { zodOutputFormat } from '@anthropic-ai/sdk/helpers/zod';
import { FindingsSchema } from './schema.js';
import { recordUsage } from './cost.js';
/** Single source of truth for the default model. Override via config `ai.model`. */
export const DEFAULT_MODEL = 'claude-opus-4-8';
const DEFAULT_MAX_TOKENS = 16000;
// Adaptive thinking is only supported on Claude 4.6+ models; older/smaller models reject it.
export function thinkingFor(model) {
    return /opus-4-[678]|sonnet-4-6|sonnet-5|fable/.test(model)
        ? { thinking: { type: 'adaptive' } }
        : undefined;
}
/**
 * Thin wrapper around the Anthropic SDK. Structured outputs for code analysis,
 * streaming text for prose, prompt caching on the stable system prefix, and
 * usage accounting on every call. Retries are the SDK's built-in backoff.
 */
export class AIClient {
    constructor(config) {
        this.config = { ...config, model: config.model || DEFAULT_MODEL };
        this.client = new Anthropic({ apiKey: config.apiKey, maxRetries: 3 });
    }
    getConfig() {
        return this.config;
    }
    systemBlocks(systemPrompt, sharedContext) {
        const blocks = [];
        if (systemPrompt)
            blocks.push({ type: 'text', text: systemPrompt });
        if (sharedContext)
            blocks.push({ type: 'text', text: sharedContext });
        if (blocks.length > 0) {
            blocks[blocks.length - 1].cache_control = { type: 'ephemeral' };
        }
        return blocks;
    }
    /**
     * Analyzes code and returns schema-validated findings — no freeform text parsing.
     */
    async analyzeCode(prompt, options = {}) {
        const response = await this.client.messages.parse({
            model: this.config.model,
            max_tokens: this.config.maxTokens || DEFAULT_MAX_TOKENS,
            ...thinkingFor(this.config.model),
            system: this.systemBlocks(options.systemPrompt, options.sharedContext),
            messages: [{ role: 'user', content: prompt }],
            output_config: { format: zodOutputFormat(FindingsSchema) },
        });
        recordUsage(this.config.model, response.usage);
        return response.parsed_output?.findings ?? [];
    }
    /**
     * Streams a prose response (used for the Lighthouse report analysis).
     * Keeps the pre-revamp `generateSuggestions` surface so call sites stay small.
     */
    async generateSuggestions(prompt, options = {}) {
        const stream = this.client.messages.stream({
            model: this.config.model,
            max_tokens: this.config.maxTokens || DEFAULT_MAX_TOKENS,
            ...thinkingFor(this.config.model),
            system: this.systemBlocks(options.systemPrompt),
            messages: [{ role: 'user', content: prompt }],
        });
        if (options.onChunk) {
            let firstChunk = true;
            stream.on('text', text => {
                options.onChunk(text, firstChunk);
                firstChunk = false;
            });
        }
        const message = await stream.finalMessage();
        recordUsage(this.config.model, message.usage);
        return message.content
            .filter((block) => block.type === 'text')
            .map(block => block.text)
            .join('');
    }
    /** Raw SDK client — used by the agent loop, which drives its own tool-use turns. */
    get anthropic() {
        return this.client;
    }
}