@toolbaux/guardian 0.1.22 → 0.1.23

package/dist/db/sqlite-specs-store.js

@@ -0,0 +1,441 @@
+/**
+ * SqliteSpecsStore — SQLite implementation of SpecsStore.
+ *
+ * Stores everything that was previously scattered across .specs/machine/*.json
+ * and .specs/human/*.md into a single guardian.db file.
+ *
+ * Schema:
+ *   specs        — blob storage for all machine intelligence files
+ *   docs         — human-readable doc sections (markdown)
+ *   metrics_log  — append-only event log (replaces mcp-metrics.jsonl)
+ *   search_fts   — FTS5 virtual table built from specs content (extra index)
+ *
+ * Tier gating is stored per-row; the caller passes a tier filter when reading.
+ * This is the foundation for the pro/enterprise access control layer.
+ */
+import Database from "better-sqlite3";
+import path from "node:path";
+/**
+ * Normalise a file path to a canonical repo-relative form.
+ * Used consistently by the FTS builder, dep-graph builder, and search query.
+ * All paths stored in guardian.db go through this function.
+ *
+ *   "flask-full/src/flask/sessions.py"      → "src/flask/sessions.py"
+ *   "django/django/contrib/auth.py"         → "django/contrib/auth.py"
+ *   "sqlalchemy/lib/sqlalchemy/sql/base.py" → "lib/sqlalchemy/sql/base.py"
+ */
+export function normPath(p) {
+    // Strip leading reponame/src/ → src/
+    p = p.replace(/^[^/]+\/src\//, "src/");
+    // Strip double-prefix X/X/ → X/ (package namespace matches repo clone dir)
+    const dm = p.match(/^([^/]+)\/\1\//);
+    if (dm)
+        return p.slice(dm[1].length + 1);
+    // Strip leading repo segment when followed by a known source-directory name
+    if (/^[^/]+\/(?:lib|examples|pkg|packages|apps|internal|cmd|src)\//i.test(p)) {
+        p = p.slice(p.indexOf("/") + 1);
+    }
+    return p;
+}
+/**
+ * Split camelCase and snake_case identifiers into individual tokens so the
+ * porter stemmer can match partial terms.
+ *   getUserById  → "get user by id"
+ *   auth_service → "auth service"
+ */
+function splitIdentifiers(s) {
+    return s
+        .replace(/_/g, " ")
+        .replace(/([a-z])([A-Z])/g, "$1 $2")
+        .replace(/([A-Z]+)([A-Z][a-z])/g, "$1 $2")
+        .toLowerCase();
+}
+export const DB_FILENAME = "guardian.db";
+export class SqliteSpecsStore {
+    storeDir;
+    db;
+    constructor(storeDir) {
+        this.storeDir = storeDir;
+    }
+    async init() {
+        const dbPath = path.join(this.storeDir, DB_FILENAME);
+        this.db = new Database(dbPath);
+        this.db.pragma("journal_mode = WAL");
+        this.db.pragma("synchronous = NORMAL");
+        this._migrate();
+    }
+    async close() {
+        this.db?.close();
+    }
+    // ── Spec blobs ─────────────────────────────────────────────────────────────
+    async readSpec(name) {
+        const row = this.db
+            .prepare("SELECT * FROM specs WHERE name = ?")
+            .get(name);
+        return row ?? null;
+    }
+    async writeSpec(name, content, format, tier = "free") {
+        this.db
+            .prepare(`
+        INSERT INTO specs (name, format, content, tier, updated_at)
+        VALUES (?, ?, ?, ?, ?)
+        ON CONFLICT(name) DO UPDATE SET
+          content    = excluded.content,
+          format     = excluded.format,
+          tier       = excluded.tier,
+          updated_at = excluded.updated_at
+      `)
+            .run(name, format, content, tier, Date.now());
+    }
+    async listSpecs() {
+        const rows = this.db
+            .prepare("SELECT name FROM specs ORDER BY name")
+            .all();
+        return rows.map(r => r.name);
+    }
+    async hasSpec(name) {
+        const row = this.db
+            .prepare("SELECT 1 FROM specs WHERE name = ?")
+            .get(name);
+        return !!row;
+    }
+    // ── Human docs ─────────────────────────────────────────────────────────────
+    async readDoc(id) {
+        const row = this.db
+            .prepare("SELECT * FROM docs WHERE id = ?")
+            .get(id);
+        return row ?? null;
+    }
+    async writeDoc(entry) {
+        this.db
+            .prepare(`
+        INSERT INTO docs (id, section, title, body, tier, updated_at)
+        VALUES (?, ?, ?, ?, ?, ?)
+        ON CONFLICT(id) DO UPDATE SET
+          section    = excluded.section,
+          title      = excluded.title,
+          body       = excluded.body,
+          tier       = excluded.tier,
+          updated_at = excluded.updated_at
+      `)
+            .run(entry.id, entry.section, entry.title, entry.body, entry.tier, Date.now());
+    }
+    async listDocs(section) {
+        if (section) {
+            return this.db
+                .prepare("SELECT * FROM docs WHERE section = ? ORDER BY id")
+                .all(section);
+        }
+        return this.db
+            .prepare("SELECT * FROM docs ORDER BY section, id")
+            .all();
+    }
+    // ── Metrics log ────────────────────────────────────────────────────────────
+    async appendMetric(event, payload) {
+        this.db
+            .prepare("INSERT INTO metrics_log (ts, event, payload) VALUES (?, ?, ?)")
+            .run(Date.now(), event, JSON.stringify(payload));
+    }
+    async readMetrics(limit = 1000) {
+        return this.db
+            .prepare("SELECT * FROM metrics_log ORDER BY id DESC LIMIT ?")
+            .all(limit);
+    }
+    // ── FTS search (extra index, no equivalent in FileSpecsStore) ─────────────
+    /**
+     * Rebuild the FTS5 search index from extracted codebase data.
+     *
+     * Each row is one file. Symbol names and endpoints are pre-expanded with
+     * splitIdentifiers() so "getUserById" becomes "get user by id" before
+     * the porter stemmer runs — this gives sub-token recall without trigrams.
+     *
+     * Column BM25 weights (bm25 args, lower = more important):
+     *   file_path(1), symbol_name(2), endpoint(3), body(4), module(5)
+     *   weights:   1.0             0.5             0.7        1.0      0.6
+     */
+    rebuildSearchIndex(rows) {
+        this.db.prepare("DELETE FROM search_fts").run();
+        const insert = this.db.prepare("INSERT INTO search_fts (file_path, symbol_name, endpoint, body, module) VALUES (?, ?, ?, ?, ?)");
+        const insertAll = this.db.transaction((items) => {
+            for (const r of items) {
+                insert.run(r.file_path, splitIdentifiers(r.symbol_name), splitIdentifiers(r.endpoint), r.body, r.module ?? "");
+            }
+        });
+        insertAll(rows);
+    }
+    /** BM25-ranked full-text search over the indexed content. */
+    searchFTS(query, limit = 20) {
+        const tokens = this._buildTokens(query);
+        if (tokens.length === 0)
+            return [];
+        const ftsQuery = tokens.join(" OR ");
+        try {
+            return this.db
+                .prepare(`
+          SELECT file_path, symbol_name,
+                 bm25(search_fts, 1.0, 0.5, 0.7, 1.0, 0.6) AS rank
+          FROM search_fts
+          WHERE search_fts MATCH ?
+          ORDER BY rank
+          LIMIT ?
+        `)
+                .all(ftsQuery, limit);
+        }
+        catch {
+            return [];
+        }
+    }
+    /**
+     * Score how well a query maps to indexed codebase content.
+     *
+     * Returns a 0–1 confidence score and a short reason string.
+     * Useful for:
+     *   - Filtering low-quality benchmark tasks
+     *   - Returning confidence alongside guardian_search results
+     *   - Advising agents when a query needs reformulation
+     *
+     * Three signals (each 0–1, combined with weights):
+     *   token_coverage  0.4 — fraction of query tokens that hit anything in the index
+     *   top_bm25        0.4 — strength of the best match (normalised from BM25 score)
+     *   result_cluster  0.2 — do top results cluster in one module (high) or scatter (low)?
+     */
+    querySignal(query) {
+        const tokens = this._buildTokens(query);
+        if (tokens.length === 0) {
+            return { score: 0, confidence: "low", reason: "query produced no searchable tokens" };
+        }
+        // English stop words that appear everywhere — don't count as code signal.
+        const STOP = new Set(["the", "and", "for", "are", "but", "not", "you", "all", "can", "had", "her", "was", "one", "our", "out", "day", "get", "has", "him", "his", "how", "its", "let", "may", "new", "now", "old", "see", "two", "use", "way", "who", "did", "man", "use", "say", "she", "than", "then", "them", "these", "they", "this", "will", "with", "have", "from", "that", "been", "each", "into", "like", "make", "more", "other", "over", "same", "such", "take", "than", "them", "then", "they", "this", "when", "your", "also", "back", "came", "come", "does", "even", "find", "give", "good", "here", "just", "keep", "kind", "last", "left", "life", "long", "much", "must", "name", "need", "next", "only", "open", "own", "part", "plan", "play", "put", "read", "real", "said", "show", "side", "some", "tell", "time", "very", "well", "went", "what", "with", "work", "year", "change", "update", "remove", "add", "fix", "like", "file", "files", "other", "also"]);
+        // Domain-specific tokens: those NOT in the stop list.
+        const domainTokens = tokens.filter(t => !STOP.has(t.replace(/\*$/, "")));
+        // ── Signal 1: domain token coverage ──────────────────────────────────
+        // Only count tokens that are domain-specific AND match SOURCE files (not config).
+        const SOURCE_EXT_RE = /\.(py|ts|tsx|js|jsx|go|java|cs|rb|rs|cpp|c|php|swift|kt)$/;
+        let domainHits = 0;
+        for (const tok of domainTokens) {
+            try {
+                const row = this.db.prepare("SELECT file_path FROM search_fts WHERE search_fts MATCH ? LIMIT 5").all(tok);
+                // Token must hit at least one actual source file (not config/build)
+                if (row.some(r => SOURCE_EXT_RE.test(r.file_path)))
+                    domainHits++;
+            }
+            catch { /* skip */ }
+        }
+        const tokenCoverage = domainTokens.length > 0 ? domainHits / domainTokens.length : 0;
+        // ── Signal 2: joint match strength ───────────────────────────────────
+        // Use AND (not OR) to find files matching ALL domain tokens together.
+        // Joint co-occurrence in one file means the query is specific, not coincidental.
+        let jointStrength = 0;
+        if (domainTokens.length > 0) {
+            try {
+                const andQuery = domainTokens.join(" AND ");
+                const row = this.db.prepare(`
+          SELECT bm25(search_fts, 1.0, 0.5, 0.7, 1.0, 0.6) AS rank, file_path
+          FROM search_fts WHERE search_fts MATCH ? ORDER BY rank LIMIT 1
+        `).get(andQuery);
+                if (row && SOURCE_EXT_RE.test(row.file_path)) {
+                    // Clamp [-15, 0] → [1, 0]
+                    jointStrength = Math.min(1, Math.max(0, -row.rank / 8));
+                }
+            }
+            catch {
+                // AND query failed (no joint match) → fall back to OR top score
+                try {
+                    const orQuery = domainTokens.join(" OR ");
+                    const row = this.db.prepare(`
+            SELECT bm25(search_fts, 1.0, 0.5, 0.7, 1.0, 0.6) AS rank, file_path
+            FROM search_fts WHERE search_fts MATCH ? ORDER BY rank LIMIT 1
+          `).get(orQuery);
+                    if (row && SOURCE_EXT_RE.test(row.file_path)) {
+                        // OR match is weaker signal — scale down by 50%
+                        jointStrength = Math.min(0.5, Math.max(0, -row.rank / 16));
+                    }
+                }
+                catch { /* skip */ }
+            }
+        }
+        // ── Signal 3: result clustering ───────────────────────────────────────
+        let clustering = 0;
+        try {
+            const orQuery = domainTokens.length > 0 ? domainTokens.join(" OR ") : tokens.join(" OR ");
+            const rows = this.db.prepare(`
+        SELECT file_path FROM search_fts WHERE search_fts MATCH ? ORDER BY bm25(search_fts) LIMIT 5
+      `).all(orQuery);
+            const srcRows = rows.filter(r => SOURCE_EXT_RE.test(r.file_path));
+            if (srcRows.length > 1) {
+                const dirs = srcRows.map(r => r.file_path.split("/").slice(0, -1).join("/"));
+                const unique = new Set(dirs).size;
+                clustering = 1 - (unique - 1) / Math.max(srcRows.length - 1, 1);
+            }
+            else if (srcRows.length === 1) {
+                clustering = 1;
+            }
+        }
+        catch { /* skip */ }
+        const score = tokenCoverage * 0.35 + jointStrength * 0.45 + clustering * 0.2;
+        const confidence = score >= 0.55 ? "high" : score >= 0.25 ? "medium" : "low";
+        const noCodeTokens = domainTokens.length === 0;
+        const reason = noCodeTokens
+            ? "query contains only generic English words, no code-domain terms"
+            : tokenCoverage < 0.3
+                ? `only ${Math.round(tokenCoverage * 100)}% of domain tokens match indexed source files`
+                : jointStrength < 0.15
+                    ? "tokens don't co-occur in any single source file — query is too generic"
+                    : clustering < 0.3
+                        ? "matching files scatter across unrelated modules — query is ambiguous"
+                        : `${Math.round(tokenCoverage * 100)}% domain coverage, strong co-occurrence match`;
+        return { score: Math.round(score * 100) / 100, confidence, reason };
+    }
+    // ── Private ────────────────────────────────────────────────────────────────
+    /** Build FTS5 token list from a natural language query. */
+    _buildTokens(query) {
+        return splitIdentifiers(query)
+            .split(/\s+/)
+            .filter(t => t.length > 1)
+            .map(t => `${t.replace(/[^a-z0-9]/g, "")}*`)
+            .filter(Boolean);
+    }
+    _migrate() {
+        this.db.exec(`
+      CREATE TABLE IF NOT EXISTS specs (
+        name        TEXT PRIMARY KEY,
+        format      TEXT NOT NULL,
+        content     TEXT NOT NULL,
+        tier        TEXT NOT NULL DEFAULT 'free',
+        updated_at  INTEGER NOT NULL
+      );
+
+      CREATE TABLE IF NOT EXISTS docs (
+        id          TEXT PRIMARY KEY,
+        section     TEXT NOT NULL,
+        title       TEXT NOT NULL,
+        body        TEXT NOT NULL,
+        tier        TEXT NOT NULL DEFAULT 'free',
+        updated_at  INTEGER NOT NULL
+      );
+
+      CREATE INDEX IF NOT EXISTS docs_section ON docs(section);
+
+      CREATE TABLE IF NOT EXISTS metrics_log (
+        id          INTEGER PRIMARY KEY AUTOINCREMENT,
+        ts          INTEGER NOT NULL,
+        event       TEXT NOT NULL,
+        payload     TEXT NOT NULL
+      );
+
+      CREATE INDEX IF NOT EXISTS metrics_log_ts ON metrics_log(ts);
+
+      CREATE TABLE IF NOT EXISTS file_deps (
+        file     TEXT NOT NULL,
+        imports  TEXT NOT NULL,
+        PRIMARY KEY (file, imports)
+      );
+
+      CREATE INDEX IF NOT EXISTS file_deps_reverse ON file_deps(imports);
+    `);
+        // FTS5 table — recreate if module column is missing (no ALTER TABLE for virtual tables).
+        // search_fts is always rebuilt on extract, so drop+recreate is safe.
+        const existing = this.db
+            .prepare("SELECT sql FROM sqlite_master WHERE type='table' AND name='search_fts'")
+            .get();
+        if (!existing?.sql?.includes("module")) {
+            this.db.exec(`
+        DROP TABLE IF EXISTS search_fts;
+        CREATE VIRTUAL TABLE search_fts USING fts5(
+          file_path,
+          symbol_name,
+          endpoint,
+          body,
+          module,
+          tokenize='porter unicode61'
+        );
+      `);
+        }
+    }
+    // ── Dependency graph ────────────────────────────────────────────────────────
+    /** Replace all import edges (run once per guardian extract --backend sqlite). */
+    rebuildDeps(edges) {
+        const del = this.db.prepare("DELETE FROM file_deps");
+        const ins = this.db.prepare("INSERT OR IGNORE INTO file_deps (file, imports) VALUES (?, ?)");
+        this.db.transaction(() => {
+            del.run();
+            for (const e of edges)
+                ins.run(e.file, e.imports);
+        })();
+    }
+    /**
+     * BM25 search + dependency-graph quality reranking.
+     *
+     * Ranking model (inspired by HITS / PageRank applied to code):
+     *   - Source files are "authorities": many files import them (high used_by count)
+     *   - Test/example files are "hubs": they import source files but nothing imports them
+     *
+     * Quality score = authority_ratio = used_by / (used_by + imports)
+     * Combined score = bm25_rank / quality   (bm25 is negative; dividing dampens hubs)
+     *
+     * This naturally demotes test/example files without hardcoding path patterns.
+     * Files with no dependency data get a neutral quality (0.7) to avoid over-penalising
+     * isolated scripts or newly-added files not yet in the graph.
+     */
+    searchWithGraph(query, limit = 5) {
+        const tokens = this._buildTokens(query);
+        if (tokens.length === 0)
+            return [];
+        const ftsQuery = tokens.join(" OR ");
+        // Fetch a wider candidate pool so reranking has enough material.
+        const candidateLimit = Math.max(limit * 4, 60);
+        let rows;
+        try {
+            rows = this.db.prepare(`
+        WITH candidates AS (
+          SELECT file_path, symbol_name,
+                 bm25(search_fts, 1.0, 0.5, 0.7, 1.0, 0.6) AS rank
+          FROM search_fts
+          WHERE search_fts MATCH ?
+          ORDER BY rank
+          LIMIT ?
+        )
+        SELECT
+          c.file_path,
+          c.symbol_name,
+          c.rank,
+          GROUP_CONCAT(DISTINCT d.imports) AS imports_,
+          GROUP_CONCAT(DISTINCT r.file)    AS used_by_
+        FROM candidates c
+        LEFT JOIN file_deps d ON d.file    = c.file_path
+        LEFT JOIN file_deps r ON r.imports = c.file_path
+        GROUP BY c.file_path, c.symbol_name, c.rank
+        ORDER BY c.rank
+      `).all(ftsQuery, candidateLimit);
+        }
+        catch {
+            return [];
+        }
+        // Apply quality reranking using dependency-graph authority score.
+        const reranked = rows.map(r => {
+            const imports = r.imports_ ? r.imports_.split(",").filter(Boolean) : [];
+            const used_by = r.used_by_ ? r.used_by_.split(",").filter(Boolean) : [];
+            const usedByN = used_by.length;
+            const importsN = imports.length;
+            let quality;
+            if (usedByN === 0 && importsN === 0) {
+                // No dependency data — preserve BM25 rank entirely.
+                quality = 1.0;
+            }
+            else {
+                // authority_ratio ∈ [0, 1]: 1.0 = pure authority (many things import this file)
+                //                            0.0 = pure hub (imports many, nothing imports it)
+                const authority = usedByN / (usedByN + importsN);
+                // Gentle nudge: [0.7, 1.0] — hubs are demoted by at most 30%.
+                // BM25 relevance still dominates; this is a tiebreaker, not a hard filter.
+                quality = 0.7 + 0.3 * authority;
+            }
+            // bm25 is negative (more negative = better). Multiplying by quality < 1
+            // moves the score toward 0 — making low-quality files rank worse.
+            const combined = r.rank * quality;
+            return { file_path: r.file_path, symbol_name: r.symbol_name, rank: combined, imports, used_by };
+        });
+        reranked.sort((a, b) => a.rank - b.rank);
+        return reranked.slice(0, limit);
+    }
+}

package/dist/extract/codebase-intel.js

@@ -171,8 +171,10 @@ function buildEndpointPatternMap(architecture) {
     }
     return result;
 }
+// ── File-based IO (original implementation — unchanged) ────────────────────
 /**
- * Load snapshots and build CodebaseIntelligence, then write to disk.
+ * Load snapshots and write codebase-intelligence.json to disk.
+ * This is the original file-based implementation, kept intact.
  */
 export async function writeCodebaseIntelligence(specsDir, outputPath) {
     const machineDir = await resolveMachineInputDir(specsDir);
@@ -187,9 +189,36 @@ export async function writeCodebaseIntelligence(specsDir, outputPath) {
     await fs.writeFile(outputPath, JSON.stringify(intel, null, 2), "utf8");
 }
 /**
- * Load an existing codebase-intelligence.json from disk.
+ * Load an existing codebase-intelligence.json from a file path.
+ * Original file-based implementation, kept intact.
  */
 export async function loadCodebaseIntelligence(intelPath) {
     const raw = await fs.readFile(intelPath, "utf8");
     return JSON.parse(raw);
 }
+// ── Store-based IO (new — works with both FileSpecsStore and SqliteSpecsStore) ─
+/**
+ * Build CodebaseIntelligence and write it via a SpecsStore.
+ * Use this when operating on a guardian.db or when you already have a store open.
+ */
+export async function writeCodebaseIntelligenceViaStore(store) {
+    const archEntry = await store.readSpec("architecture.snapshot");
+    const uxEntry = await store.readSpec("ux.snapshot");
+    if (!archEntry || !uxEntry) {
+        throw new Error("architecture.snapshot or ux.snapshot not found in store. Run `guardian extract` first.");
+    }
+    const architecture = yaml.load(archEntry.content);
+    const ux = yaml.load(uxEntry.content);
+    const intel = buildCodebaseIntelligence(architecture, ux);
+    await store.writeSpec("codebase-intelligence", JSON.stringify(intel, null, 2), "json");
+}
+/**
+ * Load CodebaseIntelligence from a SpecsStore.
+ * Returns null if not yet built.
+ */
+export async function loadCodebaseIntelligenceViaStore(store) {
+    const entry = await store.readSpec("codebase-intelligence");
+    if (!entry)
+        return null;
+    return JSON.parse(entry.content);
+}

package/dist/extract/compress.js

@@ -319,6 +319,7 @@ function buildHeatmapFromGraph(level, nodes, edges, nodeLayers) {
         }
     }
     const cycleNodes = findCycleNodes(nodes, adjacency, reverse);
+    const pageRank = computePageRank(nodes, adjacency, reverse);
     const degreeValues = nodes.map((node) => (outbound.get(node) ?? 0) + (inbound.get(node) ?? 0));
     const maxDegree = Math.max(1, ...degreeValues);
     const maxCrossRatio = Math.max(1, ...nodes.map((node) => {
@@ -332,15 +333,22 @@ function buildHeatmapFromGraph(level, nodes, edges, nodeLayers) {
         const out = outbound.get(node) ?? 0;
         const crossRatio = out === 0 ? 0 : crossOut / out;
         const cycleFlag = cycleNodes.has(node) ? 1 : 0;
-        const score = 0.5 * (degree / maxDegree) +
-            0.3 * (crossRatio / maxCrossRatio) +
-            0.2 * cycleFlag;
+        const pr = pageRank.get(node) ?? 0;
+        // PageRank (40%) — importance by what depends on this node
+        // Degree   (30%) — raw connectivity (fallback signal)
+        // Cross-layer (20%) — architectural violation risk
+        // Cycle    (10%) — circular dependency penalty
+        const score = 0.4 * pr +
+            0.3 * (degree / maxDegree) +
+            0.2 * (crossRatio / maxCrossRatio) +
+            0.1 * cycleFlag;
         return {
             id: node,
             layer: nodeLayers.get(node) ?? "unknown",
             score: round(score, 4),
             components: {
                 degree,
+                pagerank: round(pr, 4),
                 cross_layer_ratio: round(crossRatio, 4),
                 cycle: cycleFlag
             }
@@ -368,6 +376,65 @@ function resolveDomainForModule(moduleId, domainMap) {
     }
     return null;
 }
+/**
+ * Iterative PageRank over a directed graph.
+ * Returns a map of node → normalized score in [0, 1].
+ *
+ * Semantics: a node is important if many important nodes import/depend on it.
+ * Damping factor α=0.85 (web-standard). Converges in ~20 iterations for
+ * codebases with <10K files.
+ *
+ * Edge direction follows dependency arrows (A imports B → edge A→B).
+ * Rank flows *backward*: B gains rank because A depends on it, meaning
+ * files that many other files rely on get high scores — exactly what we
+ * want to surface in AI context.
+ */
+function computePageRank(nodes, adjacency, // forward edges (importer → imported)
+reverse // backward edges (imported → importers)
+) {
+    const N = nodes.length;
+    if (N === 0)
+        return new Map();
+    const DAMPING = 0.85;
+    const ITERATIONS = 30;
+    const BASE = (1 - DAMPING) / N;
+    // Initialize uniform rank
+    const rank = new Map();
+    for (const node of nodes)
+        rank.set(node, 1 / N);
+    // Precompute out-degrees (how many nodes each node imports)
+    const outDeg = new Map();
+    for (const node of nodes)
+        outDeg.set(node, (adjacency.get(node) ?? []).length);
+    // Dangling nodes (no outgoing edges) distribute rank uniformly
+    for (let iter = 0; iter < ITERATIONS; iter++) {
+        const next = new Map();
+        // Dangling mass: sum of ranks of sink nodes spread across all nodes
+        let danglingMass = 0;
+        for (const node of nodes) {
+            if ((outDeg.get(node) ?? 0) === 0) {
+                danglingMass += (rank.get(node) ?? 0);
+            }
+        }
+        const danglingContrib = DAMPING * danglingMass / N;
+        for (const node of nodes) {
+            let incoming = 0;
+            for (const importer of (reverse.get(node) ?? [])) {
+                const d = outDeg.get(importer) ?? 1;
+                incoming += (rank.get(importer) ?? 0) / d;
+            }
+            next.set(node, BASE + danglingContrib + DAMPING * incoming);
+        }
+        for (const node of nodes)
+            rank.set(node, next.get(node) ?? 0);
+    }
+    // Normalize to [0, 1] relative to max
+    const max = Math.max(1e-10, ...Array.from(rank.values()));
+    const normalized = new Map();
+    for (const [node, r] of rank.entries())
+        normalized.set(node, r / max);
+    return normalized;
+}
 function findCycleNodes(nodes, adjacency, reverse) {
     const visited = new Set();
     const order = [];

package/dist/extract/context-block.js

@@ -29,8 +29,17 @@ export function renderContextBlock(architecture, ux, options) {
         }
         lines.push("");
     }
-    // Cross-module dependencies
-    const crossEdges = architecture.dependencies.module_graph.filter(e => e.from !== e.to);
+    // Cross-module dependencies (deduplicated)
+    const seenEdges = new Set();
+    const crossEdges = architecture.dependencies.module_graph.filter(e => {
+        if (e.from === e.to)
+            return false;
+        const key = `${e.from}→${e.to}`;
+        if (seenEdges.has(key))
+            return false;
+        seenEdges.add(key);
+        return true;
+    });
     if (crossEdges.length > 0) {
         lines.push("### Module Dependencies");
         for (const edge of crossEdges.slice(0, 10)) {

package/dist/extract/function-intel.js

@@ -160,8 +160,10 @@ async function listSourceFiles(dir, config, results = []) {
  * Scan one or more project roots, run adapters on every source file, and
  * return the aggregated FunctionIntelligence index.
  */
-export async function buildFunctionIntelligenceFromRoots(roots, config) {
+export async function buildFunctionIntelligenceFromRoots(roots, config, projectRoot) {
     const allFunctions = [];
+    // Relativize against project root if provided; otherwise fall back to the scan root
+    const baseDir = projectRoot ?? roots[0];
     for (const root of roots) {
         const files = await listSourceFiles(root, config);
         await Promise.all(files.map(async (filePath) => {
@@ -177,7 +179,8 @@ export async function buildFunctionIntelligenceFromRoots(roots, config) {
             }
             try {
                 const result = runAdapter(adapter, filePath, source);
-                allFunctions.push(...result.functions);
+                const relPath = path.relative(baseDir, filePath);
+                allFunctions.push(...result.functions.map(fn => ({ ...fn, file: relPath })));
             }
             catch {
                 // Skip files that fail to parse (malformed source, encoding issues)

package/dist/extract/index.js

@@ -191,8 +191,7 @@ export async function extractProject(options) {
     // Generate Function Intelligence — call graph, literal index across all languages.
     // Runs as an additive second pass; never modifies the architecture snapshot.
     try {
-        const allRoots = (architecture.project.roots ?? [projectRoot]).map((r) => path.isAbsolute(r) ? r : path.join(projectRoot, r));
-        const funcIntel = await buildFunctionIntelligenceFromRoots(allRoots, config);
+        const funcIntel = await buildFunctionIntelligenceFromRoots([projectRoot], config, projectRoot);
         await writeFunctionIntelligence(layout.machineDir, funcIntel);
     }
     catch (err) {
@@ -421,27 +420,6 @@ function mergeFrontendAnalyses(results, _roots, _workspaceRoot) {
         tests: results.flatMap(r => r.tests)
     };
 }
-function findCommonRoot(paths) {
-    if (paths.length === 0) {
-        return process.cwd();
-    }
-    const splitPaths = paths.map((entry) => path.resolve(entry).split(path.sep));
-    const minLength = Math.min(...splitPaths.map((parts) => parts.length));
-    const shared = [];
-    for (let i = 0; i < minLength; i += 1) {
-        const segment = splitPaths[0][i];
-        if (splitPaths.every((parts) => parts[i] === segment)) {
-            shared.push(segment);
-        }
-        else {
-            break;
-        }
-    }
-    if (shared.length === 0) {
-        return path.parse(paths[0]).root;
-    }
-    return shared.join(path.sep);
-}
 async function loadPreviousSnapshots(machineDir, rootDir) {
     const result = {};
     const candidates = [