scai/dist/agents/selectRelevantSourcesStep.js

> **AI-powered CLI for local code analysis, commit message suggestions, and natural-language queries.** 100% local, private, GDPR-friendly, made in Denmark/EU with ❤️.

// File: src/modules/selectRelevantSourcesStep.ts
import fs from "fs";
import chalk from "chalk";
import { generate } from "../lib/generate.js";
import { cleanupModule } from "../pipeline/modules/cleanupModule.js";
import { logInputOutput } from "../utils/promptLogHelper.js";
export const selectRelevantSourcesStep = {
    name: "selectRelevantSources",
    description: "Selects the most relevant files for the query from relatedFiles + workingFiles, loads their code, and updates context.workingFiles.",
    groups: ["analysis"],
    run: async (input) => {
        const query = input.query ?? "";
        const context = input.context;
        if (!context) {
            throw new Error("[selectRelevantSources] StructuredContext is required.");
        }
        console.log(chalk.blueBright(`📁 [selectRelevantSources] Selecting relevant files for query "${query}"...`));
        // Merge candidate paths (relatedFiles + existing workingFiles)
        // At this point in-depth context build or searchfiles may have placed files in 
        // working files. 
        const candidatePaths = [
            ...(context.analysis?.focus?.relevantFiles ?? []),
            ...(context.initContext?.relatedFiles ?? []),
            ...(context.workingFiles?.map(f => f.path) ?? [])
        ];
        const uniquePaths = Array.from(new Set(candidatePaths));
        // -----------------------------
        // Prompt the LLM to rank files
        // -----------------------------
        const prompt = `
You are given:

Query:
"${query}"

Candidate file paths:
${JSON.stringify(uniquePaths, null, 2)}

Task:
- Identify ONLY the files directly you find to be directly relevant to the query. 
- Return ONLY a JSON array of objects.
- Each object must have at least a "path" field.
- Optional fields: "summary".
- Do NOT include explanations.
    `.trim();
        let topFiles = [];
        try {
            const response = await generate({ content: prompt, query: "" });
            const cleaned = await cleanupModule.run({
                query,
                content: response.data,
            });
            const rawFiles = Array.isArray(cleaned.data) ? cleaned.data : [];
            // Load code for selected files
            topFiles = rawFiles
                .map((f) => {
                if (typeof f?.path !== "string")
                    return null;
                let code;
                try {
                    code = fs.readFileSync(f.path, "utf-8");
                }
                catch (err) {
                    console.warn(`⚠️ Failed to read file ${f.path}:`, err.message);
                }
                return { path: f.path, code };
            })
                .filter(Boolean);
        }
        catch (err) {
            console.error("❌ [selectRelevantSources] Model selection failed:", err);
            topFiles = [];
        }
        // -----------------------------
        // Persist authoritative context (append, do not overwrite)
        // -----------------------------
        context.workingFiles = [
            ...(context.workingFiles ?? []), // existing working files
            ...topFiles, // new top files
        ];
        // Deduplicate by path
        const seenPaths = new Set();
        context.workingFiles = context.workingFiles.filter(f => {
            if (!f.path)
                return false; // safety check
            if (seenPaths.has(f.path))
                return false;
            seenPaths.add(f.path);
            return true;
        });
        const output = {
            query,
            data: {
                workingFiles: topFiles.map(f => ({ path: f.path })),
            },
        };
        logInputOutput("selectRelevantSources", "output", output.data);
        return output;
    },
};