devsmind-mcp 2.4.0 → 3.0.0

package/dist/db/database.js

@@ -36,30 +36,95 @@ var __importDefault = (this && this.__importDefault) || function (mod) {
     return (mod && mod.__esModule) ? mod : { "default": mod };
 };
 Object.defineProperty(exports, "__esModule", { value: true });
-exports.DevMindDatabase = void 0;
+exports.DevMindDatabase = exports.NO_STATIC_CALLERS_NOTE = exports.WORKFLOW_SIDECAR_FILE = exports.WORKFLOW_SCHEMA_VERSION = void 0;
+exports.toCompactSearchResult = toCompactSearchResult;
 exports.formatReasoning = formatReasoning;
 exports.parseReasoningBlocks = parseReasoningBlocks;
 const better_sqlite3_1 = __importDefault(require("better-sqlite3"));
 const crypto = __importStar(require("crypto"));
 const fs = __importStar(require("fs"));
 const path = __importStar(require("path"));
-const zlib = __importStar(require("zlib"));
 const schema_1 = require("./schema");
 const config_1 = require("../utils/config");
 const ast_1 = require("../utils/ast");
-function compressText(text) {
-    return zlib.deflateSync(Buffer.from(text, 'utf-8'));
-}
-function decompressText(val) {
-    if (val instanceof Buffer || Buffer.isBuffer(val)) {
-        try {
-            return zlib.inflateSync(val).toString('utf-8');
-        }
-        catch {
-            return val.toString('utf-8');
-        }
-    }
-    return String(val);
+const tokenize_1 = require("../utils/tokenize");
+const search_index_1 = require("./search-index");
+const embedder_1 = require("./embedder");
+const grep_1 = require("./grep");
+/** A confident "0 callers" is a false negative on dynamic-dispatch/generated-binding code (RTK
+ * hooks, DI containers) — the AST resolver can't see those calls at all. Shared between
+ * `search_nodes`' drill-in hooks and `get_node_code`'s `used_by_note` so the two tools can never
+ * say this in two slightly different ways. */
+/** How many budget-dropped node ids `getGraph` will name before it stops listing them — enough to
+ * act on, not so many that the omission list becomes its own oversized payload. */
+const OMITTED_NODE_ID_CAP = 20;
+/** Bumped when the shape written to `.devmind/workflows/<id>/workflow.json` changes. v2 replaced
+ * per-step `history_ids`/`pending_tasks` with `reasoning`/`node_ids`/`doc_paths`, and workflow
+ * `status`/`is_active` with `archived`. */
+exports.WORKFLOW_SCHEMA_VERSION = 2;
+/** The half of a workflow a v1 client has no field for, and therefore cannot overwrite when it
+ * rewrites `workflow.json` from its own columns. See `writeWorkflowToDisk`. */
+exports.WORKFLOW_SIDECAR_FILE = 'v2.json';
+exports.NO_STATIC_CALLERS_NOTE = 'no static callers found — may be used via dynamic dispatch or a generated binding; verify before assuming unused';
+/** Tier 1 keeps a couple of sample lines; anything past this is bulk, not evidence. */
+const COMPACT_SAMPLE_CAP = 2;
+/** Tier 1 line truncation. 200 chars is enough to read a matching line in context; the full
+ * 400-char lines are the single biggest contributor to an oversized files bucket. */
+const COMPACT_LINE_CAP = 200;
+/**
+ * Shrink a search result to fit, in two tiers.
+ *
+ * Why two rather than an on/off switch: the sample lines and `code_matches` are genuinely the
+ * most useful part of a result — real agent feedback credits them with catching a live bug — so
+ * throwing all of them away at the first byte over a threshold overcorrects. Tier 1 drops what is
+ * bulk-without-signal (`match_counts`, `matched_terms`, `aliases`, `created_at`, `deprecated`, and
+ * the repeated `used_by_note` boilerplate) and thins the rest; only tier 2 gives up the evidence
+ * lines entirely and becomes a pure triage list.
+ *
+ * `confidence`/`relevance`/`found_by` and the `uses`/`used_by`/`history_count` drill-in hooks
+ * survive BOTH tiers on purpose. They are a handful of bytes each and they are precisely what a
+ * caller uses to decide which result to open next — dropping them would make a compact response
+ * smaller and useless at the same time.
+ *
+ * Pure: no DB access, no I/O. Kept here rather than in the MCP handler so it is unit-testable
+ * directly, and so it sits inside the coverage gate.
+ */
+function toCompactSearchResult(result, tier) {
+    const keepEvidence = tier === 1;
+    const trimLines = (lines) => lines.slice(0, COMPACT_SAMPLE_CAP).map(l => ({ ...l, line_content: l.line_content.slice(0, COMPACT_LINE_CAP) }));
+    return {
+        nodes: result.nodes.map(n => {
+            // Cast rather than narrow the RankedNode union: `code_matches` exists only on the fuzzy
+            // variant, and the identifier variant simply has no evidence lines to trim.
+            const src = n;
+            return {
+                id: src.id,
+                name: src.name,
+                type: src.type,
+                confidence: src.confidence,
+                relevance: src.relevance,
+                found_by: src.found_by,
+                file_path: src.file_path,
+                signature: src.signature,
+                description: src.description,
+                uses: src.uses,
+                used_by: src.used_by,
+                history_count: src.history_count,
+                code_matches: keepEvidence && src.code_matches ? trimLines(src.code_matches) : undefined
+            };
+        }),
+        files: result.files.map(f => ({
+            file_path: f.file_path,
+            total_matches: f.total_matches,
+            sample_lines: keepEvidence ? trimLines(f.sample_lines) : undefined
+        })),
+        files_total: result.files_total,
+        files_offset: result.files_offset,
+        nodes_total: result.nodes_total,
+        hint: result.hint,
+        truncated: result.truncated,
+        scope_note: result.scope_note
+    };
 }
 function formatReasoning(r) {
     if (typeof r === 'string') {
@@ -77,6 +142,42 @@ function formatReasoning(r) {
     ];
     return lines.join('\n');
 }
+/** Matches the `── Update @ … ──` separator `updateHistory` appends each same-session update under. */
+const REASONING_SEPARATOR = /(\n*── Update @ [^\n]*──\n)/g;
+/**
+ * Removes one block from an accumulated reasoning log, matched by its exact text and searched
+ * from the newest end.
+ *
+ * Not simply "drop the last block": an update that carries no code change (a `stage_change`, a
+ * bare `update_history`) appends reasoning without recording an edit, so blocks and edits are not
+ * one-to-one and positional removal would take the wrong one. Matching on content is exact when
+ * the block is there, and when it isn't the log is returned untouched — leaving a stale line is
+ * recoverable, mangling someone else's reasoning is not.
+ *
+ * Splitting on a capturing group keeps the separators in the result, so every block that stays
+ * keeps its original timestamp rather than being re-stamped on the way out.
+ */
+function dropReasoningBlock(raw, block) {
+    if (!raw || !block)
+        return raw;
+    const target = block.trim();
+    if (!target)
+        return raw;
+    // [block, sep, block, sep, block, …] — blocks at even indices, separators at odd.
+    const parts = raw.split(REASONING_SEPARATOR);
+    for (let i = parts.length - 1; i >= 0; i -= 2) {
+        if (parts[i].trim() !== target)
+            continue;
+        // Drop the block with the separator that introduced it. The first block has none, so it
+        // takes the separator that follows instead — whatever came after now leads the log.
+        if (i > 0)
+            parts.splice(i - 1, 2);
+        else
+            parts.splice(0, 2);
+        return parts.join('').replace(/^\n+/, '');
+    }
+    return raw;
+}
 /**
  * Inverse of `formatReasoning`. A single history row accumulates every later update appended
  * under a `── Update @ … ──` separator, so one stored blob can hold several changes — this
@@ -122,7 +223,15 @@ class DevMindDatabase {
     db;
     dbPath;
     context = null;
-    constructor(dbPath) {
+    /**
+     * `onSyncProgress`: optional, fires during the constructor's initial `syncFromDisk()` pass —
+     * the one silent stretch every CLI command pays on `new DevMindDatabase(...)` before it can
+     * print anything else. Large `.devmind` folders (mainly `history/`, which grows one file per
+     * edit — much faster than node count) can make that pass take minutes; without this, a caller
+     * has no way to tell "still working" from "hung". Omit it for silent construction (the MCP
+     * server's normal path, where per-open console spam would be noise, not signal).
+     */
+    constructor(dbPath, opts) {
         this.dbPath = dbPath;
         // Open SQLite database
         this.db = new better_sqlite3_1.default(dbPath);
@@ -138,7 +247,14 @@ class DevMindDatabase {
             // Ignore context errors (e.g. running from scratch scripts)
         }
         // Auto-sync history and graph from disk JSONs
-        this.syncFromDisk();
+        this.syncFromDisk(opts?.onSyncProgress);
+    }
+    /** Throttles progress callbacks to ~100 updates across `total` items, regardless of scale —
+     * so a 500-file sync and a 500,000-file sync both report about as often, and the callback
+     * itself (a stdout write) never becomes the bottleneck it was being added to diagnose. */
+    static shouldReport(done, total) {
+        const every = Math.max(1, Math.floor(total / 100));
+        return done === total || done % every === 0;
     }
     initSchema() {
         this.db.exec(schema_1.INIT_SCHEMA_SQL);
@@ -148,12 +264,45 @@ class DevMindDatabase {
         catch {
             // Column already exists, ignore
         }
-        this.db.exec(`
-      CREATE TABLE IF NOT EXISTS system_meta (
-        key TEXT PRIMARY KEY,
-        value TEXT NOT NULL,
-        updated_at DATETIME DEFAULT CURRENT_TIMESTAMP
-      )
+        try {
+            this.db.exec('ALTER TABLE nodes ADD COLUMN description TEXT');
+        }
+        catch {
+            // Column already exists, ignore
+        }
+        try {
+            this.db.exec("ALTER TABLE nodes ADD COLUMN aliases TEXT DEFAULT '[]'");
+        }
+        catch {
+            // Column already exists, ignore
+        }
+        // Workflow v2 columns. Same additive, idempotent shape as the three above: on a fresh brain
+        // INIT_SCHEMA_SQL already created them so the ALTER throws and the catch absorbs it; on a brain
+        // created before v2 the ALTER is what actually adds them. Deliberately NOT paired with a
+        // DROP COLUMN for the fields they replace — `status`/`pending_tasks`/`history_ids` stay as
+        // vestigial columns, because dropping them buys nothing (they are nullable or defaulted, so
+        // nothing has to write them) and would break an older globally-installed CLI opening the same
+        // brain.db with "no such column".
+        for (const ddl of [
+            'ALTER TABLE workflow_steps ADD COLUMN reasoning TEXT',
+            'ALTER TABLE workflow_steps ADD COLUMN node_ids TEXT',
+            'ALTER TABLE workflow_steps ADD COLUMN doc_paths TEXT',
+            'ALTER TABLE workflows ADD COLUMN archived INTEGER NOT NULL DEFAULT 0'
+        ]) {
+            try {
+                this.db.exec(ddl);
+            }
+            catch {
+                // Column already exists, ignore
+            }
+        }
+        this.backfillWorkflowStepNodeIds();
+        this.db.exec(`
+      CREATE TABLE IF NOT EXISTS system_meta (
+        key TEXT PRIMARY KEY,
+        value TEXT NOT NULL,
+        updated_at DATETIME DEFAULT CURRENT_TIMESTAMP
+      )
     `);
     }
     getContext() {
@@ -170,10 +319,10 @@ class DevMindDatabase {
         }
     }
     setSystemMeta(key, value) {
-        const stmt = this.db.prepare(`
-      INSERT INTO system_meta (key, value, updated_at)
-      VALUES (?, ?, CURRENT_TIMESTAMP)
-      ON CONFLICT(key) DO UPDATE SET value = ?, updated_at = CURRENT_TIMESTAMP
+        const stmt = this.db.prepare(`
+      INSERT INTO system_meta (key, value, updated_at)
+      VALUES (?, ?, CURRENT_TIMESTAMP)
+      ON CONFLICT(key) DO UPDATE SET value = ?, updated_at = CURRENT_TIMESTAMP
     `);
         stmt.run(key, value, value);
     }
@@ -186,7 +335,7 @@ class DevMindDatabase {
      */
     getNodesByFilePath(filePath) {
         const stmt = this.db.prepare(`SELECT * FROM nodes WHERE deprecated = 0 AND REPLACE(LOWER(file_path), '\\', '/') = ?`);
-        return stmt.all((0, ast_1.normalizeFsPath)(filePath));
+        return DevMindDatabase.parseNodeRows(stmt.all((0, ast_1.normalizeFsPath)(filePath)));
     }
     close() {
         this.db.close();
@@ -196,6 +345,10 @@ class DevMindDatabase {
         const one = (sql) => {
             try {
                 const row = this.db.prepare(sql).get();
+                /* istanbul ignore next -- every call site here is a `SELECT COUNT(*) AS c FROM ...`,
+                   which always returns exactly one row; `row` can only be undefined if this helper is
+                   ever repurposed for a query that can return zero rows. Kept as a real guard, not
+                   because today's three call sites can hit it. */
                 return row ? row.c : 0;
             }
             catch {
@@ -226,9 +379,10 @@ class DevMindDatabase {
         this.db.exec('DELETE FROM node_connections');
         this.db.exec('DELETE FROM history');
         this.db.exec('DELETE FROM nodes');
+        this.db.exec('DELETE FROM node_vectors');
         this.db.exec('DELETE FROM system_meta');
         const workspaceRoot = path.dirname(this.dbPath);
-        for (const dir of ['graph', 'history']) {
+        for (const dir of ['graph', 'history', 'vectors']) {
             const p = path.join(workspaceRoot, dir);
             if (fs.existsSync(p)) {
                 fs.rmSync(p, { recursive: true, force: true });
@@ -289,6 +443,11 @@ class DevMindDatabase {
     upsertNode(node) {
         const canonicalFp = (0, config_1.canonicalizePath)(node.file_path);
         const existing = this.getNode(node.id);
+        // aliases JSON is only computed when the caller actually passed some — an ordinary
+        // edit/re-index of a node that ISN'T alias-bearing must not blank out aliases a prior
+        // detector pass (or Phase E's record_alias) already attached, same COALESCE idiom as
+        // description/signature below.
+        const aliasesJson = node.aliases ? JSON.stringify(Array.from(new Set(node.aliases))) : null;
         if (existing) {
             let finalPath = existing.file_path;
             const paths = existing.file_path.split(',').map(p => p.trim()).filter(Boolean);
@@ -297,36 +456,143 @@ class DevMindDatabase {
                 paths.push(incoming);
                 finalPath = paths.join(', ');
             }
-            const stmt = this.db.prepare(`
-        UPDATE nodes
-        SET type = ?,
-            name = ?,
-            file_path = ?,
-            signature = COALESCE(?, signature),
-            deprecated = 0
-        WHERE id = ?
+            // description follows the same COALESCE idiom as signature: an unspecified description
+            // on an ordinary edit must never blank out one already written — only an explicit new
+            // value (from add_description, or a description passed alongside this edit) overwrites it.
+            const stmt = this.db.prepare(`
+        UPDATE nodes
+        SET type = ?,
+            name = ?,
+            file_path = ?,
+            signature = COALESCE(?, signature),
+            description = COALESCE(?, description),
+            aliases = COALESCE(?, aliases),
+            deprecated = 0
+        WHERE id = ?
       `);
-            stmt.run(node.type, node.name, finalPath, node.signature || null, node.id);
+            stmt.run(node.type, node.name, finalPath, node.signature || null, node.description || null, aliasesJson, node.id);
         }
         else {
-            const stmt = this.db.prepare(`
-        INSERT INTO nodes (id, type, name, file_path, signature)
-        VALUES (?, ?, ?, ?, ?)
+            const stmt = this.db.prepare(`
+        INSERT INTO nodes (id, type, name, file_path, signature, description, aliases)
+        VALUES (?, ?, ?, ?, ?, ?, ?)
       `);
-            stmt.run(node.id, node.type, node.name, canonicalFp, node.signature || null);
+            stmt.run(node.id, node.type, node.name, canonicalFp, node.signature || null, node.description || null, aliasesJson ?? '[]');
         }
         this.writeGraphToDisk(canonicalFp);
     }
+    /**
+     * Adds one alias to a node WITHOUT touching any it already has — the merge-safe counterpart to
+     * `upsertNode`'s replace-if-given aliases. This is what the batch graph-fix session's
+     * `record_alias` correction tool (Phase E) uses: it should never be able to accidentally drop an
+     * alias a deterministic detector pass already attached.
+     */
+    addAlias(nodeId, alias) {
+        const node = this.getNode(nodeId);
+        if (!node)
+            return;
+        if (node.aliases.includes(alias))
+            return;
+        const next = [...node.aliases, alias];
+        this.db.prepare('UPDATE nodes SET aliases = ? WHERE id = ?').run(JSON.stringify(next), node.id);
+        this.writeGraphToDisk(node.file_path);
+    }
+    /**
+     * Stores a node's semantic vector and writes it to the committed `vectors/*.json` tree.
+     * `vector` must already be int8-quantized (`embedTextInt8`/`embedTextsInt8` in embedder.ts) and
+     * `descriptionHash` must be `hashDescription()` of the exact description it was computed from —
+     * this is the staleness key `getNodesNeedingEmbedding` checks against.
+     */
+    upsertNodeVector(nodeId, vector, descriptionHash) {
+        const node = this.getNode(nodeId);
+        if (!node)
+            return;
+        const stmt = this.db.prepare(`
+      INSERT INTO node_vectors (node_id, model_id, dim, description_hash, vector)
+      VALUES (?, ?, ?, ?, ?)
+      ON CONFLICT(node_id) DO UPDATE SET
+        model_id = excluded.model_id,
+        dim = excluded.dim,
+        description_hash = excluded.description_hash,
+        vector = excluded.vector
+    `);
+        stmt.run(node.id, embedder_1.EMBEDDING_MODEL_ID, embedder_1.EMBEDDING_DIM, descriptionHash, Buffer.from(vector.buffer, vector.byteOffset, vector.byteLength));
+        this.writeVectorsToDisk(node.file_path);
+    }
+    getNodeVector(nodeId) {
+        const stmt = this.db.prepare('SELECT model_id, dim, description_hash, vector FROM node_vectors WHERE node_id = ?');
+        const row = stmt.get(nodeId);
+        if (!row)
+            return null;
+        return {
+            modelId: row.model_id,
+            dim: row.dim,
+            descriptionHash: row.description_hash,
+            vector: new Int8Array(row.vector.buffer, row.vector.byteOffset, row.vector.byteLength)
+        };
+    }
+    /**
+     * Every non-deprecated, described node whose vector is missing, from a different model
+     * (`model_id` mismatch — e.g. a mismatched vector ignored during `syncFromDisk`), or stale
+     * (its description changed since the vector was computed). This is the work queue for both
+     * `devsmind embed` and the auto-embed hooks in `describe`/`add_description` — resumable and
+     * idempotent by construction, same shape as `describe.ts`'s own `WHERE description IS NULL`.
+     * `force: true` returns every described node regardless of vector state (model upgrades).
+     */
+    getNodesNeedingEmbedding(force = false) {
+        const described = this.getAllNodes().filter(n => !n.deprecated && n.description);
+        if (force)
+            return described;
+        const vecRows = this.db.prepare('SELECT node_id, model_id, description_hash FROM node_vectors').all();
+        const vecMap = new Map(vecRows.map(r => [r.node_id, r]));
+        return described.filter(n => {
+            const v = vecMap.get(n.id);
+            if (!v)
+                return true;
+            if (v.model_id !== embedder_1.EMBEDDING_MODEL_ID)
+                return true;
+            if (v.description_hash !== (0, embedder_1.hashDescription)(n.description))
+                return true;
+            return false;
+        });
+    }
+    /**
+     * The `nodes.aliases` column is a JSON-array-in-TEXT blob — better-sqlite3 hands it back as a
+     * raw string, not a parsed array, on every `stmt.all()`/`stmt.get()`. A JSON string masquerading
+     * as `string[]` is a silent-wrong-answer hazard (both have `.length`, so a bug here would not
+     * throw, just quietly misbehave — e.g. counting characters instead of aliases). EVERY raw SQL
+     * read of the `nodes` table must route its rows through {@link parseNodeRow}/{@link
+     * parseNodeRows}, never cast `as DbNode`/`as DbNode[]` directly.
+     */
+    static parseNodeAliases(raw) {
+        if (Array.isArray(raw))
+            return raw.filter(a => typeof a === 'string');
+        if (typeof raw !== 'string' || raw.length === 0)
+            return [];
+        try {
+            const parsed = JSON.parse(raw);
+            return Array.isArray(parsed) ? parsed.filter(a => typeof a === 'string') : [];
+        }
+        catch {
+            return [];
+        }
+    }
+    static parseNodeRow(row) {
+        return { ...row, aliases: DevMindDatabase.parseNodeAliases(row.aliases) };
+    }
+    static parseNodeRows(rows) {
+        return rows.map(DevMindDatabase.parseNodeRow);
+    }
     getNode(id) {
         const stmt = this.db.prepare('SELECT * FROM nodes WHERE id = ?');
         const direct = stmt.get(id);
         if (direct)
-            return direct;
+            return DevMindDatabase.parseNodeRow(direct);
         if (!id.includes('#')) {
             const suffixStmt = this.db.prepare("SELECT * FROM nodes WHERE id LIKE ? ESCAPE '\\' AND deprecated = 0");
             const matches = suffixStmt.all(`%#${this.likeEscape(id)}`);
             if (matches.length === 1) {
-                return matches[0];
+                return DevMindDatabase.parseNodeRow(matches[0]);
             }
         }
         return null;
@@ -341,8 +607,13 @@ class DevMindDatabase {
         this.deleteHistoryFilesForNode(resolvedId);
         const stmt = this.db.prepare('DELETE FROM nodes WHERE id = ?');
         stmt.run(resolvedId);
+        // node_vectors has no FK to nodes (see schema.ts) so this doesn't cascade — delete it
+        // explicitly rather than waiting for the next syncFromDisk() orphan sweep, or a vector
+        // search could keep surfacing a node that no longer exists until the next server restart.
+        this.db.prepare('DELETE FROM node_vectors WHERE node_id = ?').run(resolvedId);
         if (node && node.file_path) {
             this.writeGraphToDisk(node.file_path);
+            this.writeVectorsToDisk(node.file_path);
         }
         for (const p of inboundSourceFiles) {
             this.writeGraphToDisk(p);
@@ -356,14 +627,19 @@ class DevMindDatabase {
         const inboundSourceFiles = this.collectInboundSourceFiles(resolvedId);
         const updateStmt = this.db.prepare('UPDATE nodes SET deprecated = 1 WHERE id = ?');
         const deleteConnStmt = this.db.prepare('DELETE FROM node_connections WHERE source_node_id = ? OR target_node_id = ?');
+        // A deprecated node's vector is pure dead weight (writeVectorsToDisk already excludes
+        // deprecated nodes from the JSON) — drop it now rather than leaving an unused row behind.
+        const deleteVectorStmt = this.db.prepare('DELETE FROM node_vectors WHERE node_id = ?');
         const tx = this.db.transaction(() => {
             updateStmt.run(resolvedId);
             deleteConnStmt.run(resolvedId, resolvedId);
+            deleteVectorStmt.run(resolvedId);
         });
         tx();
         // Rewrite the node's own file (now carrying deprecated:1) and every caller file (so their
         // stale inbound edges don't resurrect the connection on the next syncFromDisk()).
         if (node && node.file_path) {
+            this.writeVectorsToDisk(node.file_path);
             this.writeGraphToDisk(node.file_path);
         }
         for (const p of inboundSourceFiles) {
@@ -388,23 +664,29 @@ class DevMindDatabase {
         this.db.pragma('foreign_keys = OFF');
         try {
             const runTx = this.db.transaction(() => {
-                const insertStmt = this.db.prepare(`
-          INSERT INTO nodes (id, type, name, file_path, signature, created_at, deprecated)
-          VALUES (?, ?, ?, ?, ?, ?, ?)
+                const insertStmt = this.db.prepare(`
+          INSERT INTO nodes (id, type, name, file_path, signature, description, aliases, created_at, deprecated)
+          VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?)
         `);
-                insertStmt.run(newId, node.type, name, filePath, node.signature, node.created_at, node.deprecated ? 1 : 0);
-                const updateSourceStmt = this.db.prepare(`
-          UPDATE node_connections SET source_node_id = ? WHERE source_node_id = ?
+                insertStmt.run(newId, node.type, name, filePath, node.signature, node.description, JSON.stringify(node.aliases), node.created_at, node.deprecated ? 1 : 0);
+                const updateSourceStmt = this.db.prepare(`
+          UPDATE node_connections SET source_node_id = ? WHERE source_node_id = ?
         `);
                 updateSourceStmt.run(newId, resolvedOldId);
-                const updateTargetStmt = this.db.prepare(`
-          UPDATE node_connections SET target_node_id = ? WHERE target_node_id = ?
+                const updateTargetStmt = this.db.prepare(`
+          UPDATE node_connections SET target_node_id = ? WHERE target_node_id = ?
         `);
                 updateTargetStmt.run(newId, resolvedOldId);
-                const updateHistoryStmt = this.db.prepare(`
-          UPDATE history SET node_id = ? WHERE node_id = ?
+                const updateHistoryStmt = this.db.prepare(`
+          UPDATE history SET node_id = ? WHERE node_id = ?
         `);
                 updateHistoryStmt.run(newId, resolvedOldId);
+                // Carry any existing vector row to the new id — same reasoning as description above:
+                // a rename shouldn't force a re-embed. No-op if the node had no vector yet.
+                const updateVectorStmt = this.db.prepare(`
+          UPDATE node_vectors SET node_id = ? WHERE node_id = ?
+        `);
+                updateVectorStmt.run(newId, resolvedOldId);
                 const deleteOldStmt = this.db.prepare('DELETE FROM nodes WHERE id = ?');
                 deleteOldStmt.run(resolvedOldId);
             });
@@ -413,9 +695,11 @@ class DevMindDatabase {
                 // Rewrite the OLD file's graph JSON too when the file itself moved, so the
                 // stale node entry doesn't linger under the old path's JSON on disk.
                 this.writeGraphToDisk(node.file_path);
+                this.writeVectorsToDisk(node.file_path);
             }
             if (filePath && filePath !== node.file_path) {
                 this.writeGraphToDisk(filePath);
+                this.writeVectorsToDisk(filePath);
             }
             // Edges pointing INTO the renamed node live in the SOURCE nodes' files' graph JSONs
             // (which still reference oldId on disk). The DB was already repointed to newId above,
@@ -434,6 +718,69 @@ class DevMindDatabase {
             this.db.pragma('foreign_keys = ON');
         }
     }
+    /**
+     * Merges `fromId` into `intoId` — the batch graph-fix session's `merge_nodes` correction, for
+     * when curation (or a human reviewing feedback) decides two node candidates were never really
+     * distinct entities. Unlike `renameNode` (which moves everything to a FRESH id), `intoId`
+     * already exists with its own rows here: `fromId`'s connections (both directions) and history
+     * are reassigned onto it, `fromId`'s aliases (plus its own name, so old references by that name
+     * still resolve) are folded into `intoId`'s alias set, and `fromId` is deprecated — not hard
+     * deleted, so its history stays reachable and the merge itself stays a reversible correction,
+     * not a destructive one.
+     */
+    mergeNodes(fromId, intoId) {
+        const fromNode = this.getNode(fromId);
+        const intoNode = this.getNode(intoId);
+        if (!fromNode)
+            throw new Error(`mergeNodes: source node not found: ${fromId}`);
+        if (!intoNode)
+            throw new Error(`mergeNodes: target node not found: ${intoId}`);
+        const resolvedFrom = fromNode.id;
+        const resolvedInto = intoNode.id;
+        if (resolvedFrom === resolvedInto)
+            return; // already the same node — no-op, not an error
+        this.db.pragma('foreign_keys = OFF');
+        try {
+            const runTx = this.db.transaction(() => {
+                // Reassign fromId's connections onto intoId in both directions. INSERT OR IGNORE
+                // absorbs a duplicate-PK conflict when both nodes already shared a target/source; the
+                // `!== resolvedInto` guard drops what would otherwise become a self-referencing edge.
+                const outRows = this.db.prepare('SELECT target_node_id FROM node_connections WHERE source_node_id = ?').all(resolvedFrom);
+                for (const r of outRows) {
+                    if (r.target_node_id === resolvedInto)
+                        continue;
+                    this.db.prepare('INSERT OR IGNORE INTO node_connections (source_node_id, target_node_id) VALUES (?, ?)').run(resolvedInto, r.target_node_id);
+                }
+                const inRows = this.db.prepare('SELECT source_node_id FROM node_connections WHERE target_node_id = ?').all(resolvedFrom);
+                for (const r of inRows) {
+                    if (r.source_node_id === resolvedInto)
+                        continue;
+                    this.db.prepare('INSERT OR IGNORE INTO node_connections (source_node_id, target_node_id) VALUES (?, ?)').run(r.source_node_id, resolvedInto);
+                }
+                this.db.prepare('DELETE FROM node_connections WHERE source_node_id = ? OR target_node_id = ?').run(resolvedFrom, resolvedFrom);
+                this.db.prepare('UPDATE history SET node_id = ? WHERE node_id = ?').run(resolvedInto, resolvedFrom);
+                const mergedAliases = Array.from(new Set([...intoNode.aliases, ...fromNode.aliases, fromNode.name]));
+                this.db.prepare('UPDATE nodes SET aliases = ? WHERE id = ?').run(JSON.stringify(mergedAliases), resolvedInto);
+                this.db.prepare('UPDATE nodes SET deprecated = 1 WHERE id = ?').run(resolvedFrom);
+            });
+            runTx();
+            if (fromNode.file_path)
+                this.writeGraphToDisk(fromNode.file_path);
+            if (intoNode.file_path && intoNode.file_path !== fromNode.file_path)
+                this.writeGraphToDisk(intoNode.file_path);
+            // Edges that pointed INTO fromId now point at intoId in the DB, but the SOURCE files'
+            // on-disk JSON still says fromId until rewritten — same reasoning as renameNode's own use
+            // of this helper. Queried AFTER the transaction, so it reflects the already-repointed state.
+            this.rewriteInboundSourceFiles(resolvedInto);
+            const historyIds = this.db.prepare('SELECT id FROM history WHERE node_id = ?').all(resolvedInto);
+            for (const row of historyIds) {
+                this.patchHistoryDiskIdentity(row.id, resolvedInto, intoNode.name, intoNode.type, intoNode.file_path, intoNode.signature);
+            }
+        }
+        finally {
+            this.db.pragma('foreign_keys = ON');
+        }
+    }
     /**
      * Rewrites a history/[id].json file's identifying fields (node_id, node_metadata) in
      * place, leaving code_snapshot/reasoning/timestamps untouched. Used after a rename so
@@ -468,10 +815,10 @@ class DevMindDatabase {
      * callers that merely repoint them (rename) can rewrite after the fact.
      */
     collectInboundSourceFiles(nodeId) {
-        const rows = this.db.prepare(`
-      SELECT DISTINCT n.file_path AS file_path
-      FROM node_connections c JOIN nodes n ON n.id = c.source_node_id
-      WHERE c.target_node_id = ?
+        const rows = this.db.prepare(`
+      SELECT DISTINCT n.file_path AS file_path
+      FROM node_connections c JOIN nodes n ON n.id = c.source_node_id
+      WHERE c.target_node_id = ?
     `).all(nodeId);
         const files = new Set();
         for (const row of rows) {
@@ -531,9 +878,9 @@ class DevMindDatabase {
         }
         this.db.pragma('foreign_keys = OFF');
         try {
-            const stmt = this.db.prepare(`
-        INSERT OR IGNORE INTO node_connections (source_node_id, target_node_id)
-        VALUES (?, ?)
+            const stmt = this.db.prepare(`
+        INSERT OR IGNORE INTO node_connections (source_node_id, target_node_id)
+        VALUES (?, ?)
       `);
             stmt.run(resolvedSrc, resolvedTgt);
             if (srcNode.file_path) {
@@ -549,42 +896,119 @@ class DevMindDatabase {
         const tgtNode = this.getNode(targetNodeId);
         const resolvedSrc = srcNode ? srcNode.id : sourceNodeId;
         const resolvedTgt = tgtNode ? tgtNode.id : targetNodeId;
-        const stmt = this.db.prepare(`
-      DELETE FROM node_connections
-      WHERE source_node_id = ? AND target_node_id = ?
+        const stmt = this.db.prepare(`
+      DELETE FROM node_connections
+      WHERE source_node_id = ? AND target_node_id = ?
     `);
         stmt.run(resolvedSrc, resolvedTgt);
         if (srcNode && srcNode.file_path) {
             this.writeGraphToDisk(srcNode.file_path);
         }
     }
-    getConnections(nodeId) {
+    /**
+     * `opts.limit`/`opts.offset` page a hub node's caller/callee list deterministically —
+     * `ORDER BY file_path, name` so a repeated call with the same offset returns the same slice,
+     * and so the local/nearby callers a reader actually wants tend to sort ahead of a scattered
+     * cross-repo tail (same file_path groups together). Omitting `opts` returns every row, exactly
+     * as before — every pre-existing call site keeps working unchanged.
+     */
+    getConnections(nodeId, opts = {}) {
         const node = this.getNode(nodeId);
         const resolvedId = node ? node.id : nodeId;
-        const usesStmt = this.db.prepare(`
-      SELECT n.* FROM nodes n
-      JOIN node_connections c ON n.id = c.target_node_id
-      WHERE c.source_node_id = ?
+        const paging = opts.limit !== undefined ? ' ORDER BY n.file_path, n.name LIMIT ? OFFSET ?' : '';
+        const usesStmt = this.db.prepare(`
+      SELECT n.* FROM nodes n
+      JOIN node_connections c ON n.id = c.target_node_id
+      WHERE c.source_node_id = ?${paging}
     `);
-        const usedByStmt = this.db.prepare(`
-      SELECT n.* FROM nodes n
-      JOIN node_connections c ON n.id = c.source_node_id
-      WHERE c.target_node_id = ?
+        const usedByStmt = this.db.prepare(`
+      SELECT n.* FROM nodes n
+      JOIN node_connections c ON n.id = c.source_node_id
+      WHERE c.target_node_id = ?${paging}
     `);
+        const args = opts.limit !== undefined ? [resolvedId, opts.limit, opts.offset ?? 0] : [resolvedId];
         return {
-            uses: usesStmt.all(resolvedId),
-            usedBy: usedByStmt.all(resolvedId)
+            uses: DevMindDatabase.parseNodeRows(usesStmt.all(...args)),
+            usedBy: DevMindDatabase.parseNodeRows(usedByStmt.all(...args))
         };
     }
+    /**
+     * Batched, COUNT-only connection degree for many nodes at once — the search-result drill-in
+     * hooks need this for ~20 nodes per call, and `getConnections` per-node would mean 20 pairs of
+     * full-row-fetching queries. One grouped COUNT each way instead (same shape as the degree
+     * subquery in {@link getGodEntities}). IDs not present in `node_connections` still get a
+     * `{uses:0, usedBy:0}` entry so callers never need an existence check.
+     */
+    getConnectionCounts(ids) {
+        const result = new Map();
+        if (ids.length === 0)
+            return result;
+        for (const id of ids)
+            result.set(id, { uses: 0, usedBy: 0 });
+        const placeholders = ids.map(() => '?').join(',');
+        const usesStmt = this.db.prepare(`
+      SELECT source_node_id AS id, COUNT(*) AS n FROM node_connections
+      WHERE source_node_id IN (${placeholders}) GROUP BY source_node_id
+    `);
+        const usedByStmt = this.db.prepare(`
+      SELECT target_node_id AS id, COUNT(*) AS n FROM node_connections
+      WHERE target_node_id IN (${placeholders}) GROUP BY target_node_id
+    `);
+        for (const row of usesStmt.all(...ids)) {
+            const e = result.get(row.id);
+            if (e)
+                e.uses = row.n;
+        }
+        for (const row of usedByStmt.all(...ids)) {
+            const e = result.get(row.id);
+            if (e)
+                e.usedBy = row.n;
+        }
+        return result;
+    }
+    /** Batched history-entry count for many nodes at once — metadata only, no disk reads. */
+    getHistoryCounts(ids) {
+        const result = new Map();
+        if (ids.length === 0)
+            return result;
+        const placeholders = ids.map(() => '?').join(',');
+        const stmt = this.db.prepare(`
+      SELECT node_id AS id, COUNT(*) AS n FROM history
+      WHERE node_id IN (${placeholders}) GROUP BY node_id
+    `);
+        for (const row of stmt.all(...ids)) {
+            result.set(row.id, row.n);
+        }
+        return result;
+    }
+    /**
+     * Batched most-recent history timestamp for many nodes at once. Deliberately SQL-only (`MAX`
+     * over the indexed `updated_at` column) — unlike {@link getLatestHistory}, this never touches
+     * `populateHistoryFromDisk`, so it costs nothing beyond the query itself.
+     */
+    getLastUpdatedMap(ids) {
+        const result = new Map();
+        if (ids.length === 0)
+            return result;
+        const placeholders = ids.map(() => '?').join(',');
+        const stmt = this.db.prepare(`
+      SELECT node_id AS id, MAX(updated_at) AS last FROM history
+      WHERE node_id IN (${placeholders}) GROUP BY node_id
+    `);
+        for (const row of stmt.all(...ids)) {
+            result.set(row.id, row.last);
+        }
+        return result;
+    }
     // --- History Operations ---
     getLatestHistory(nodeId) {
         const node = this.getNode(nodeId);
         const resolvedId = node ? node.id : nodeId;
-        const stmt = this.db.prepare(`
-      SELECT id, node_id, session_id, created_at, updated_at FROM history
-      WHERE node_id = ?
-      ORDER BY updated_at DESC
-      LIMIT 1
+        const stmt = this.db.prepare(`
+      SELECT id, node_id, session_id, created_at, updated_at FROM history
+      WHERE node_id = ?
+      ORDER BY updated_at DESC
+      LIMIT 1
     `);
         const row = stmt.get(resolvedId);
         if (!row)
@@ -594,11 +1018,11 @@ class DevMindDatabase {
     listHistory(nodeId) {
         const node = this.getNode(nodeId);
         const resolvedId = node ? node.id : nodeId;
-        const stmt = this.db.prepare(`
-      SELECT id, node_id, session_id, created_at, updated_at
-      FROM history
-      WHERE node_id = ?
-      ORDER BY updated_at DESC
+        const stmt = this.db.prepare(`
+      SELECT id, node_id, session_id, created_at, updated_at
+      FROM history
+      WHERE node_id = ?
+      ORDER BY updated_at DESC
     `);
         return stmt.all(resolvedId);
     }
@@ -612,15 +1036,62 @@ class DevMindDatabase {
     getFullHistory(nodeId) {
         const node = this.getNode(nodeId);
         const resolvedId = node ? node.id : nodeId;
-        const stmt = this.db.prepare(`
-      SELECT id, node_id, session_id, created_at, updated_at
-      FROM history
-      WHERE node_id = ?
-      ORDER BY updated_at DESC
+        const stmt = this.db.prepare(`
+      SELECT id, node_id, session_id, created_at, updated_at
+      FROM history
+      WHERE node_id = ?
+      ORDER BY updated_at DESC
     `);
         const rows = stmt.all(resolvedId);
         return rows.map(row => this.populateHistoryFromDisk(row));
     }
+    /**
+     * The last `limit` history entries' reasoning + timestamps only — no `code_snapshot`/`edits`.
+     * Built for `get_node_code`'s default `history:"recent"` mode, which already returns the
+     * CURRENT code: repeating past snapshots inline would just duplicate what's already in the
+     * response. The full trail (snapshots + diffable edits) is `history:"full"`, served by
+     * {@link getHistoryPage}. This answers "why does this look the way it does" cheaply enough to
+     * attach to every get_node_code call by default, instead of leaving that as a round trip an AI
+     * has to remember to make (or skip, and re-break a decision it never saw).
+     */
+    getRecentHistorySummaries(nodeId, limit) {
+        const node = this.getNode(nodeId);
+        const resolvedId = node ? node.id : nodeId;
+        const stmt = this.db.prepare(`
+      SELECT id, node_id, session_id, created_at, updated_at
+      FROM history
+      WHERE node_id = ?
+      ORDER BY updated_at DESC
+      LIMIT ?
+    `);
+        const rows = stmt.all(resolvedId, limit);
+        return rows.map(row => {
+            const full = this.populateHistoryFromDisk(row);
+            return { id: full.id, session_id: full.session_id, created_at: full.created_at, updated_at: full.updated_at, reasoning: full.reasoning };
+        });
+    }
+    /**
+     * `history:"full"`'s backing query — the same full-fidelity payload as {@link getFullHistory}
+     * (code_snapshot + diffable edits per entry), but LIMIT/OFFSET applied IN SQL before any disk
+     * read happens, not by slicing an already-fully-loaded array. `getFullHistory` reads every
+     * revision's JSON off disk unconditionally; for a node with dozens of revisions that is dozens
+     * of synchronous file reads to serve a request for the newest 5. `total` is the true count
+     * before paging, same honesty contract as `nodes_total`/`files_total` elsewhere.
+     */
+    getHistoryPage(nodeId, limit, offset) {
+        const node = this.getNode(nodeId);
+        const resolvedId = node ? node.id : nodeId;
+        const totalRow = this.db.prepare('SELECT COUNT(*) AS c FROM history WHERE node_id = ?').get(resolvedId);
+        const stmt = this.db.prepare(`
+      SELECT id, node_id, session_id, created_at, updated_at
+      FROM history
+      WHERE node_id = ?
+      ORDER BY updated_at DESC
+      LIMIT ? OFFSET ?
+    `);
+        const rows = stmt.all(resolvedId, limit, offset);
+        return { entries: rows.map(row => this.populateHistoryFromDisk(row)), total: totalRow.c };
+    }
     /** Distinct source node ids of edges pointing INTO this node (its "used-by" callers). */
     getInboundSources(nodeId) {
         const rows = this.db
@@ -671,15 +1142,22 @@ class DevMindDatabase {
         if (node) {
             const live = this.extractLiveCode(node);
             if (live !== null) {
+                const outdated = snapshot ? snapshot.code_snapshot !== live : undefined;
                 return {
                     exists: true,
                     node_id: node.id,
                     file_path: node.file_path,
                     code: live,
                     source: 'live',
-                    // Snapshot exists but disagrees with disk → the graph has drifted.
-                    snapshot_outdated: snapshot ? snapshot.code_snapshot !== live : undefined,
-                    updated_at: snapshot?.updated_at
+                    // Snapshot exists but disagrees with disk → the RECORDED HISTORY has drifted, not the
+                    // `code` above — that was just read fresh from disk this call. Worth a message, not
+                    // just a bare flag: without it, this reads as "don't trust what you were just handed,"
+                    // which is backwards — it's the opposite thing (the graph's history) that's behind.
+                    snapshot_outdated: outdated,
+                    updated_at: snapshot?.updated_at,
+                    message: outdated
+                        ? 'The code above is current — read live from disk this call. Only the recorded history snapshot is stale (it predates this edit); no need to re-read the file to double-check.'
+                        : undefined
                 };
             }
         }
@@ -705,7 +1183,7 @@ class DevMindDatabase {
     getGraph(nodeId, maxDepth = 6, opts = {}) {
         const direction = opts.direction ?? 'both';
         const codeCharBudget = opts.codeCharBudget ?? 60_000;
-        const maxNodesLimit = 500;
+        const maxNodesLimit = opts.maxNodes ?? 500;
         const visited = new Set();
         const nodes = [];
         const connections = [];
@@ -719,12 +1197,19 @@ class DevMindDatabase {
         // argument would find zero edges and return a lone root.
         const queue = [{ id: rootNode.id, depth: 0 }];
         visited.add(rootNode.id);
+        rootNode.depth = 0;
         nodes.push(rootNode);
-        const usesStmt = this.db.prepare(`
-      SELECT target_node_id FROM node_connections WHERE source_node_id = ?
+        // ORDER BY on both directions so the walk is REPRODUCIBLE: without it, sibling order within a
+        // depth is whatever physical row order SQLite happens to return, which shifts after inserts,
+        // deletes, or a VACUUM. That made two identical getGraph calls able to disagree about which
+        // nodes got code when the budget ran out — the same query returning a different answer for no
+        // visible reason. Both are index-ordered already (the PK autoindex outbound,
+        // idx_node_connections_target inbound), so the planner elides the sort and this costs nothing.
+        const usesStmt = this.db.prepare(`
+      SELECT target_node_id FROM node_connections WHERE source_node_id = ? ORDER BY target_node_id
     `);
-        const usedByStmt = this.db.prepare(`
-      SELECT source_node_id FROM node_connections WHERE target_node_id = ?
+        const usedByStmt = this.db.prepare(`
+      SELECT source_node_id FROM node_connections WHERE target_node_id = ? ORDER BY source_node_id
     `);
         while (queue.length > 0 && nodes.length < maxNodesLimit) {
             const current = queue.shift();
@@ -745,10 +1230,16 @@ class DevMindDatabase {
                         visited.add(targetId);
                         const targetNode = this.getNode(targetId);
                         if (targetNode) {
+                            targetNode.depth = current.depth + 1;
                             nodes.push(targetNode);
                             if (nodes.length >= maxNodesLimit)
                                 break;
                         }
+                        // targetNode === null means `target_node_id` no longer resolves to a real node (the
+                        // node was deleted/renamed but the connections row survived). It still goes on the
+                        // queue/visited so a stale id is never re-processed, but — deliberately — no entry
+                        // is added to `nodes` for it. The edge pushed above therefore references an id that
+                        // will never appear in `nodes`; the dedup pass after the loop (see below) strips it.
                         queue.push({ id: targetId, depth: current.depth + 1 });
                     }
                 }
@@ -769,6 +1260,7 @@ class DevMindDatabase {
                         visited.add(sourceId);
                         const sourceNode = this.getNode(sourceId);
                         if (sourceNode) {
+                            sourceNode.depth = current.depth + 1;
                             nodes.push(sourceNode);
                             if (nodes.length >= maxNodesLimit)
                                 break;
@@ -778,24 +1270,50 @@ class DevMindDatabase {
                 }
             }
         }
-        const result = { nodes, connections };
+        // Two kinds of "more exists than was returned", reported honestly instead of silently:
+        //  1. The node cap cut the walk short with the queue still non-empty.
+        //  2. A connections row survives referencing a node this walk never added (deleted/renamed
+        //     node, or — same shape — a node dropped for being past the cap on a LATER queue entry
+        //     than the one whose edge pointed at it). Filtering here, once, after the walk, is
+        //     simpler and more certainly correct than trying to prevent every path that could create
+        //     one during the BFS itself.
+        const nodesTruncated = nodes.length >= maxNodesLimit && queue.length > 0;
+        const nodeIds = new Set(nodes.map(n => n.id));
+        const cleanConnections = connections.filter(c => nodeIds.has(c.source_node_id) && nodeIds.has(c.target_node_id));
+        const connectionsTruncated = cleanConnections.length < connections.length;
+        const result = { nodes, connections: cleanConnections };
+        if (nodesTruncated)
+            result.nodes_truncated = true;
+        if (connectionsTruncated)
+            result.connections_truncated = true;
         if (opts.includeCode) {
             let spent = 0;
-            let withoutCode = 0;
+            let noCodeAvailable = 0;
+            const omittedForBudget = [];
             // `nodes` is in BFS order (nearest the root first), so the budget is spent on the most
             // relevant code before anything is dropped.
             for (const [i, n] of nodes.entries()) {
                 const live = this.extractLiveCode(n);
                 const code = live ?? this.getLatestCode(n.id)?.code_snapshot ?? null;
                 if (!code) {
-                    withoutCode++;
+                    // Nothing to attach — a node whose symbol no longer resolves on disk and has no cached
+                    // snapshot. Counted SEPARATELY from a budget drop: these two used to share one
+                    // `nodes_without_code` counter, which meant a graph full of unresolvable nodes reported
+                    // `code_truncated: true` on a completely unspent budget. Raising the budget would then
+                    // do nothing, and the caller had no way to tell that from the response.
+                    noCodeAvailable++;
                     continue;
                 }
                 // The root always gets its code — it is what was asked for, and dropping it would make
                 // the response useless. Every other node must fit in the REMAINING budget, so a single
                 // large node can't blow past the cap (it is skipped and counted, not truncated).
                 if (i > 0 && spent + code.length > codeCharBudget) {
-                    withoutCode++;
+                    // Recorded BY ID, not merely counted. An id is a valid argument to get_node_code, so a
+                    // caller who needs the rest can fetch exactly those nodes — where a bare count, or an
+                    // array index into a graph that gets re-derived on every call, tells them nothing they
+                    // can act on. Capped so a wide graph can't turn the omission list into its own payload.
+                    if (omittedForBudget.length < OMITTED_NODE_ID_CAP)
+                        omittedForBudget.push(n.id);
                     continue;
                 }
                 n.code = code;
@@ -803,10 +1321,14 @@ class DevMindDatabase {
                 spent += code.length;
             }
             result.code_chars = spent;
-            if (withoutCode > 0) {
+            if (noCodeAvailable > 0 || omittedForBudget.length > 0) {
                 result.code_truncated = true;
-                result.nodes_without_code = withoutCode;
+                result.nodes_without_code = noCodeAvailable + omittedForBudget.length;
             }
+            if (noCodeAvailable > 0)
+                result.nodes_no_code_available = noCodeAvailable;
+            if (omittedForBudget.length > 0)
+                result.code_omitted_node_ids = omittedForBudget;
         }
         return result;
     }
@@ -826,7 +1348,9 @@ class DevMindDatabase {
         const resolvedId = node ? node.id : node_id;
         const formattedReasoning = formatReasoning(reasoning);
         const nowStr = new Date().toISOString();
-        const compressedCode = compressText(code_snapshot);
+        const newEdit = params.code_before === undefined
+            ? null
+            : { at: nowStr, before: params.code_before ?? '', after: code_snapshot, reasoning: formattedReasoning };
         // 1-hour session boundary rule check
         const latest = this.getLatestHistory(resolvedId);
         if (latest) {
@@ -838,25 +1362,34 @@ class DevMindDatabase {
             // session). code_snapshot is always the latest state (git already owns version history for
             // code). reasoning is APPENDED, not overwritten — an earlier commit's "why" in this same
             // session is still real and still worth keeping; losing it silently is worse than a few
-            // extra lines in one file. This also keeps any workflow step whose history_ids point at
-            // this row valid: it never loses what it originally linked to, only gains more below it.
+            // extra lines in one file.
             if (diffMs < 3600000) {
+                /* istanbul ignore next -- `latest` came from `getLatestHistory()` -> `populateHistoryFromDisk()`,
+                   which always returns a STRING `reasoning` (either the disk JSON's string field, or the
+                   result of `formatReasoning(...)`, itself always a string) — so the `: ''` fallback here
+                   can never actually run. Kept as a real guard against a future change to
+                   `populateHistoryFromDisk`'s return shape, not because today's flow can reach it. */
                 const previousReasoning = typeof latest.reasoning === 'string' ? latest.reasoning : '';
                 const mergedReasoning = previousReasoning.trim().length > 0
                     ? `${previousReasoning}\n\n── Update @ ${nowStr} ──\n${formattedReasoning}`
                     : formattedReasoning;
-                const updateStmt = this.db.prepare(`
-          UPDATE history
-          SET code_snapshot = '', reasoning = ?, updated_at = ?
-          WHERE id = ?
+                const updateStmt = this.db.prepare(`
+          UPDATE history
+          SET code_snapshot = '', reasoning = ?, updated_at = ?
+          WHERE id = ?
         `);
                 updateStmt.run(mergedReasoning, nowStr, latest.id);
+                // The edit trail APPENDS for the same reason reasoning does: this row now covers several
+                // edits, and only a per-edit before/after lets a revert undo the last one rather than the
+                // whole session's work — the window slides off updated_at, so one row can span hours.
+                const mergedEdits = newEdit ? [...latest.edits, newEdit] : latest.edits;
                 // Write/Update on disk
-                this.writeHistoryToDisk(latest.id, resolvedId, latest.session_id, latest.created_at, nowStr, code_snapshot, mergedReasoning);
+                this.writeHistoryToDisk(latest.id, resolvedId, latest.session_id, latest.created_at, nowStr, code_snapshot, mergedReasoning, mergedEdits);
                 return {
                     ...latest,
                     code_snapshot,
                     reasoning: mergedReasoning,
+                    edits: mergedEdits,
                     updated_at: nowStr
                 };
             }
@@ -864,13 +1397,14 @@ class DevMindDatabase {
         // Otherwise (or if no record exists), insert new history block
         const newId = crypto.randomUUID();
         const sessionId = params.session_id || crypto.randomUUID();
-        const insertStmt = this.db.prepare(`
-      INSERT INTO history (id, node_id, session_id, created_at, updated_at, code_snapshot, reasoning)
-      VALUES (?, ?, ?, ?, ?, '', ?)
+        const insertStmt = this.db.prepare(`
+      INSERT INTO history (id, node_id, session_id, created_at, updated_at, code_snapshot, reasoning)
+      VALUES (?, ?, ?, ?, ?, '', ?)
     `);
         insertStmt.run(newId, resolvedId, sessionId, nowStr, nowStr, formattedReasoning);
+        const newEdits = newEdit ? [newEdit] : [];
         // Write to disk
-        this.writeHistoryToDisk(newId, resolvedId, sessionId, nowStr, nowStr, code_snapshot, formattedReasoning);
+        this.writeHistoryToDisk(newId, resolvedId, sessionId, nowStr, nowStr, code_snapshot, formattedReasoning, newEdits);
         return {
             id: newId,
             node_id: resolvedId,
@@ -878,195 +1412,688 @@ class DevMindDatabase {
             created_at: nowStr,
             updated_at: nowStr,
             code_snapshot,
-            reasoning: formattedReasoning
+            reasoning: formattedReasoning,
+            edits: newEdits
         };
     }
+    /**
+     * Removes the newest recorded edit from a history entry, leaving no trace of it.
+     *
+     * There used to be a citation guard here: a history row cited by a workflow step was emptied
+     * rather than deleted, so the step was not left pointing at nothing. Workflow steps record
+     * `node_ids` now, not history ids, so nothing cites a history row any more and the guard had
+     * nothing left to check. Keeping it would have meant keeping the `history_ids` column alive
+     * purely to protect a reference nothing makes.
+     *
+     * Callers are expected to have restored the file already; this only unwinds what was written
+     * about it.
+     */
+    eraseLastEdit(historyId) {
+        const entry = this.getHistoryEntry(historyId);
+        if (!entry)
+            return { erased: false, entry_deleted: false, reason: 'history entry not found' };
+        if (!entry.edits.length)
+            return { erased: false, entry_deleted: false, reason: 'entry has no recorded edits' };
+        const remaining = entry.edits.slice(0, -1);
+        const dropped = entry.edits[entry.edits.length - 1];
+        if (!remaining.length) {
+            // Nothing left to keep the row for — no workflow step references a history id any more.
+            this.db.prepare('DELETE FROM history WHERE id = ?').run(historyId);
+            try {
+                const f = path.join(path.dirname(this.dbPath), 'history', `${historyId}.json`);
+                if (fs.existsSync(f))
+                    fs.unlinkSync(f);
+            }
+            catch { /* ignore */ }
+            return { erased: true, entry_deleted: true };
+        }
+        const nowStr = new Date().toISOString();
+        const newSnapshot = remaining[remaining.length - 1].after;
+        const newReasoning = dropReasoningBlock(entry.reasoning, dropped.reasoning);
+        this.db.prepare('UPDATE history SET reasoning = ?, updated_at = ? WHERE id = ?')
+            .run(newReasoning, nowStr, historyId);
+        this.writeHistoryToDisk(historyId, entry.node_id, entry.session_id, entry.created_at, nowStr, newSnapshot, newReasoning, remaining);
+        return { erased: true, entry_deleted: false };
+    }
     // --- Search Operations ---
     /**
-     * Search for nodes by name/id/reasoning first (cheap, SQL-only). If that finds
-     * nothing, transparently fall back to a code-content search (same engine as
-     * {@link searchCode}) so a query like "alipay" still succeeds even when no
-     * node's name/id/reasoning mentions it but the code itself does. Every result
-     * is tagged `matched_via` so the caller knows which path found it.
+     * The one search tool, covering both worlds in a single call:
+     *  - a primary `nodes` bucket — the indexed graph, found by exact identifier, then by three
+     *    fused rankers (BM25 over metadata, vector over descriptions, and code-body match), and
+     *  - a last-resort `files` bucket — a real filesystem grep of the configured repos, so files
+     *    the graph never models (CSS, JSON, config, markup, un-indexed code) are finally covered
+     *    in the same call instead of sending the caller off to an external grep.
+     *
+     * Inputs play to each layer's strength: the natural-language `query` drives the semantic vector
+     * layer (and BM25); `opts.keywords` (literal, OR) drive grep and the code-body match, and also
+     * feed BM25. If no keywords are given they're derived from the query's significant tokens, so a
+     * natural-only call still gets code + file coverage.
+     *
+     * Speed is the point — the two slow layers (vector, grep) run concurrently, and the old ~9k
+     * per-node snapshot read (the tool's former ~10-30s cost) is gone: code-body matching now rides
+     * the single grep walk. See `grep.ts` and Phase 4 of the plan.
      */
-    searchNodes(query, opts = {}) {
-        const stmt = this.db.prepare(`
-      SELECT DISTINCT n.* FROM nodes n
-      LEFT JOIN history h ON n.id = h.node_id
-      WHERE n.name LIKE ? ESCAPE '\\' OR n.id LIKE ? ESCAPE '\\' OR h.reasoning LIKE ? ESCAPE '\\'
-      LIMIT 50
-    `);
-        // Escaped so a query containing '%'/'_' (a real identifier fragment like "CartService_addItem"
-        // matches the literal underscore, not "any single character") searches for those characters
-        // rather than acting as SQL LIKE wildcards — sibling methods (getDeveloperActivity,
-        // searchDecisions) already do this; this one didn't.
-        const wildcard = `%${this.likeEscape(query)}%`;
-        const identifierMatches = stmt.all(wildcard, wildcard, wildcard);
-        if (identifierMatches.length > 0) {
-            return identifierMatches.map(n => ({ ...n, matched_via: 'identifier' }));
-        }
-        const codeMatches = this.searchCode({
-            query,
-            is_regex: opts.is_regex,
-            case_insensitive: opts.case_insensitive
-        });
-        if (codeMatches.length > 0) {
-            return codeMatches.map(m => ({ ...m, matched_via: 'code' }));
+    /**
+     * Attaches the drill-in hooks (`uses`/`used_by`/`history_count`/`last_updated`) to a batch of
+     * search results in place, via ONE grouped query per hook instead of per-node fetches — see
+     * {@link getConnectionCounts}/{@link getHistoryCounts}/{@link getLastUpdatedMap}. This is the
+     * signal that turns a search result from a dead end into something worth drilling into with
+     * `get_node_graph`/`get_node_history` — without it, nothing hints there's more to find.
+     */
+    attachDrillInHooks(nodes) {
+        if (nodes.length === 0)
+            return nodes;
+        const ids = nodes.map(n => n.id);
+        const connCounts = this.getConnectionCounts(ids);
+        const historyCounts = this.getHistoryCounts(ids);
+        const lastUpdated = this.getLastUpdatedMap(ids);
+        for (const n of nodes) {
+            /* istanbul ignore next -- `connCounts` is built from `getConnectionCounts(ids)` on this
+               SAME `ids` array a few lines up, and that helper pre-seeds a {uses:0,usedBy:0} entry for
+               every id it's given before querying — so `connCounts.get(n.id)` can never miss here.
+               Kept as a real guard against a future refactor decoupling the two, not because today's
+               flow can reach the fallback. */
+            const conn = connCounts.get(n.id) ?? { uses: 0, usedBy: 0 };
+            n.uses = conn.uses;
+            n.used_by = conn.usedBy;
+            n.history_count = historyCounts.get(n.id) ?? 0;
+            const lu = lastUpdated.get(n.id);
+            if (lu)
+                n.last_updated = lu;
+            // Say so instead of asserting "unused", or this hook actively misleads instead of helping.
+            if (conn.usedBy === 0) {
+                n.used_by_note = exports.NO_STATIC_CALLERS_NOTE;
+            }
         }
-        const tokens = this.tokenizeQuery(query);
-        if (tokens.length === 0) {
+        return nodes;
+    }
+    async searchNodes(query, opts = {}) {
+        const trimmedQuery = query?.trim();
+        const hasQuery = !!trimmedQuery;
+        const trimmedPattern = opts.pattern?.trim();
+        const hasPattern = !!trimmedPattern;
+        if (!hasQuery && !hasPattern) {
+            throw new Error('searchNodes requires at least one of `query` or `pattern`.');
+        }
+        const caseInsensitive = opts.case_insensitive !== false;
+        const scopePath = this.resolveSearchScopePath(opts.path);
+        // Scoping `path` straight AT a lockfile or build artifact is honored, not overridden — the
+        // same rule that has always applied to scoping at an ignored directory. What makes that
+        // confusing is the SILENCE, not the exclusion: an empty result is indistinguishable from
+        // "the pattern isn't in that file". Say so instead, so the agent stops rather than retrying
+        // variations of a search that can never return anything.
+        const scopeNote = scopePath && (0, grep_1.isDefaultIgnoredFile)(scopePath)
+            ? `path "${scopePath}" is a lockfile or generated artifact, excluded from search by default — no file content was scanned. Read the file directly if you genuinely need it.`
+            : undefined;
+        const filesOffset = opts.offset ?? 0;
+        const filesLimit = opts.limit ?? 25;
+        // `compact` is not just a projection flag — it SKIPS work. `annotateSampleLinesWithSymbol` is
+        // the AST path (the same per-node span resolution that was 8.2s of a 9.6s query before its
+        // file cap), and its only product is the `symbol` field ON the sample lines. A compact result
+        // drops those lines entirely, so resolving them first would be pure waste. This is exactly
+        // why compaction can't live wholly in the MCP handler: a projection applied after the fact
+        // can trim the response, but it cannot un-spend the time that produced it.
+        const compact = opts.compact === true;
+        const grepOpts = { ignoredPaths: this.context?.config.ignored_paths, caseInsensitive, scopePath };
+        // The pattern actually handed to the filesystem grep: the caller's own regex when given, else
+        // a literal (escaped) OR of the query's own significant tokens — the same "derive from query"
+        // fallback as before, just built from a real regex now instead of a keyword array. This is the
+        // ONLY layer that ever runs — everything meaning-driven below is gated on `hasQuery`.
+        const grepPattern = hasPattern
+            ? trimmedPattern
+            : Array.from(new Set((0, tokenize_1.tokenizeText)(trimmedQuery))).map(grep_1.escapeRegExp).join('|');
+        // BM25/vector/the identifier short-circuit are all driven by natural-language MEANING — a bare
+        // regex has no meaning for them to tokenize or embed, so none of them run without a `query`.
+        // Grep (and the code-match nodes it feeds via mapGrepHitsToNodes) is unconditional.
+        const bm25Tokens = hasQuery ? Array.from(new Set((0, tokenize_1.tokenizeText)(trimmedQuery))) : [];
+        if (hasQuery) {
+            // Deliberately name/id ONLY — NOT description/reasoning. Those are free-text natural-language
+            // fields (a description is a whole sentence); matching the query as a raw substring against
+            // them turns "any short natural-language query that happens to appear inside some node's
+            // description" into a false "exact identifier hit," which then skipped vector search AND the
+            // grep-derived code layer entirely (see the comment below) for what was never actually an
+            // identifier lookup. description/reasoning are already covered properly — with real ranking,
+            // not a blind substring guess — by BM25 (`tokenSearchNodes`/`search-index.ts`) on the path
+            // below, so nothing is lost by dropping them here; queries that used to wrongly short-circuit
+            // now correctly fall through to BM25 + vector + code instead.
+            //
+            // Gated on `hasQuery` for a second, sharper reason too: this LIKE builds `%<query>%`, and an
+            // ABSENT query would make that `%%` — matching every node in the database. A regex-only
+            // search on a small repo (≤10 nodes) would then wrongly "short-circuit" and hand back the
+            // ENTIRE graph as high-confidence identifier hits. Never reachable when `query` is empty.
+            const stmt = this.db.prepare(`
+        SELECT DISTINCT n.* FROM nodes n
+        WHERE n.name LIKE ? ESCAPE '\\' OR n.id LIKE ? ESCAPE '\\'
+        LIMIT 50
+      `);
+            const wildcard = `%${this.likeEscape(trimmedQuery)}%`;
+            const identifierMatches = DevMindDatabase.parseNodeRows(stmt.all(wildcard, wildcard));
+            // Exact-identifier short-circuit: a small, unambiguous set of literal name/id hits is trusted
+            // outright and skips the rankers (semantic blur only hurts an exact symbol lookup). Grep
+            // always runs here too (using `grepPattern`, the same derive-from-query-if-absent value the
+            // main path below uses), so an agent that only ever sends `query` still gets a populated
+            // `files` bucket instead of an empty one for no reason.
+            if (identifierMatches.length > 0 && identifierMatches.length <= 10) {
+                // An exact identifier hit is the most trustworthy result there is → high confidence, top
+                // relevance, found by name.
+                // Trust signals FIRST, before `description`. JSON.stringify preserves insertion order and
+                // an agent reads a result top-down; with the plain `{...n, confidence}` spread these
+                // landed after a full sentence of description, and real session feedback was that they got
+                // skipped in favour of eyeballing node names. Ordering is the whole fix — the fields were
+                // always there. Applied at BOTH construction sites (see the fused path below); doing only
+                // one would leave the two paths disagreeing, and this identifier path is the more common
+                // one for a symbol lookup.
+                const nodes = identifierMatches.map(({ id, name, type, ...rest }) => ({
+                    id, name, type,
+                    confidence: 'high', relevance: 100, found_by: ['name'], matched_via: 'identifier',
+                    ...rest,
+                    uses: 0, used_by: 0, history_count: 0
+                }));
+                this.attachDrillInHooks(nodes);
+                const grep = await (0, grep_1.grepRepos)(this.repoRoots(), grepPattern, grepOpts);
+                const filesResult = (0, grep_1.rankGrepHits)(grep.hits, { offset: filesOffset, maxFiles: filesLimit });
+                if (!compact)
+                    this.annotateSampleLinesWithSymbol(filesResult.files);
+                return {
+                    nodes,
+                    files: filesResult.files,
+                    files_total: filesResult.total,
+                    files_offset: filesOffset,
+                    nodes_total: nodes.length,
+                    truncated: grep.truncated || undefined,
+                    scope_note: scopeNote
+                };
+            }
+        }
+        // The two slow layers run CONCURRENTLY — vector inference and the filesystem grep walk. BM25
+        // is synchronous indexed SQL, so it runs inline for free. Wall-clock ≈ max(vector, grep).
+        //
+        // Timed at every stage, opt-in via DEVSMIND_PERF_DEBUG — kept permanently (not stripped after
+        // this pass) because it's how the mapGrepHitsToNodes bottleneck below was actually found: on a
+        // real 8-repo query it was 8.2s of a 9.6s total, invisible from the outside since the tool
+        // just looked uniformly slow. Zero cost when unset (`process.hrtime.bigint()` calls only;
+        // the env check gates the one string-building/console.error).
+        const perfDebug = !!process.env.DEVSMIND_PERF_DEBUG;
+        const ms = (a, b) => (Number(b - a) / 1e6).toFixed(0);
+        const perfStart = process.hrtime.bigint();
+        const bm25Ranked = this.tokenSearchNodes(bm25Tokens);
+        const perfAfterBm25 = process.hrtime.bigint();
+        // Timed individually (not just the combined Promise.all) so a slow run can be attributed to
+        // ONE of the two instead of leaving both under suspicion.
+        let vectorMs = '?', grepMs = '?';
+        const [vectorIds, grep] = await Promise.all([
+            (hasQuery ? this.vectorSearchNodes(trimmedQuery) : Promise.resolve([])).then(r => { if (perfDebug)
+                vectorMs = hasQuery ? ms(perfAfterBm25, process.hrtime.bigint()) : 'skipped(no query)'; return r; }),
+            (0, grep_1.grepRepos)(this.repoRoots(), grepPattern, grepOpts).then(r => { if (perfDebug)
+                grepMs = ms(perfAfterBm25, process.hrtime.bigint()); return r; })
+        ]);
+        const perfAfterVectorGrep = process.hrtime.bigint();
+        // The single grep walk feeds BOTH buckets: raw hits → the files bucket, and hits landing in
+        // indexed source files → code-match nodes (this is the "code search that returns nodes").
+        const filesResult = (0, grep_1.rankGrepHits)(grep.hits, { offset: filesOffset, maxFiles: filesLimit });
+        const perfAfterRank = process.hrtime.bigint();
+        const codeMatches = this.mapGrepHitsToNodes(grep.hits);
+        const perfAfterCodeMatch = process.hrtime.bigint();
+        const codeLinesById = new Map(codeMatches.map(c => [c.nodeId, c.lines]));
+        // Bounded annotation (only the page actually returned — see the doc comment on the helper for
+        // why this must NOT run over every raw hit). Timed and logged BEFORE the [perf] line below,
+        // not after — it shares `locateNodeInFile`, the exact primitive mapGrepHitsToNodes' own doc
+        // comment identifies as the historical bottleneck, so leaving it unmeasured would silently
+        // exempt a real cost from the one line this file's instrumentation exists to catch it with.
+        if (!compact)
+            this.annotateSampleLinesWithSymbol(filesResult.files);
+        const perfAfterAnnotate = process.hrtime.bigint();
+        if (perfDebug) {
+            console.error(`[perf] bm25=${ms(perfStart, perfAfterBm25)}ms vector=${vectorMs}ms grep=${grepMs}ms (combined=${ms(perfAfterBm25, perfAfterVectorGrep)}ms) rankGrepHits=${ms(perfAfterVectorGrep, perfAfterRank)}ms mapGrepHitsToNodes=${ms(perfAfterRank, perfAfterCodeMatch)}ms annotateSampleLinesWithSymbol=${ms(perfAfterCodeMatch, perfAfterAnnotate)}ms grepHits=${grep.hits.length} truncated=${grep.truncated}`);
+        }
+        if (bm25Ranked.length === 0 && vectorIds.length === 0 && codeMatches.length === 0) {
+            // Neither the graph (metadata, meaning, code body) nor grep found anything meaningful.
+            const base = {
+                nodes: [],
+                files: filesResult.files,
+                files_total: filesResult.total,
+                files_offset: filesOffset,
+                nodes_total: 0,
+                truncated: grep.truncated || undefined,
+                scope_note: scopeNote
+            };
+            if (filesResult.files.length === 0) {
+                base.hint = 'No meaningful match anywhere — not in any node\'s name, description, reasoning, or code body, and no file on disk contains this pattern. If you expected this to exist, retry with a broader pattern or different query terms; if it genuinely isn\'t in this codebase, that\'s a real answer — don\'t keep re-querying variations.';
+            }
+            return base;
+        }
+        // Fuse the THREE node rankings by rank (RRF) — BM25 (metadata), vector (meaning), code-match
+        // (body). Each is strong at something the others miss; rank-fusion needs no score calibration.
+        // But the RAW RRF float is a terrible thing to hand back (a #1-ranked hit tops out near 0.03,
+        // which reads like "3% confident"), so it drives ORDER only — every node is then re-described
+        // with human-meaningful signals: which layers found it, a high/medium/low confidence, and a
+        // 0-100 relevance relative to the top hit.
+        const bm25ById = new Map(bm25Ranked.map(n => [n.id, n]));
+        const simById = new Map(vectorIds.map(v => [v.id, v.sim]));
+        const fused = (0, search_index_1.reciprocalRankFusion)([bm25Ranked.map(n => n.id), vectorIds.map(v => v.id), codeMatches.map(c => c.nodeId)]);
+        /* istanbul ignore next -- unreachable here: reaching this line already required at least one
+           of bm25Ranked/vectorIds/codeMatches to be non-empty (the `if (... .length === 0 && ...)`
+           guard above returns early otherwise), and reciprocalRankFusion's output is the union of its
+           input rankings — so `fused` is always non-empty by the time this runs. Kept as a real guard
+           against a future change decoupling that invariant, not because today's flow can reach it. */
+        const topScore = fused.length ? fused[0].score : 1;
+        const nodes = [];
+        for (const { id, score } of fused.slice(0, 20)) {
+            const bm25Hit = bm25ById.get(id);
+            const sim = simById.get(id);
+            const codeLines = codeLinesById.get(id);
+            // Reconstruct a clean DbNode — never spread bm25Hit wholesale, or its internal fields leak.
+            const src = bm25Hit ?? this.getNode(id);
+            /* istanbul ignore next -- `fused`'s ids are drawn only from tokenSearchNodes/vectorSearchNodes/
+               mapGrepHitsToNodes, all three of which query `nodes` directly (deprecated = 0) at call time
+               a few lines above; nothing mutates the DB between those queries and this synchronous loop,
+               so `this.getNode(id)` (the fallback when `bm25Hit` is absent) always finds a row here today.
+               Kept as a real guard against a future async gap or query decoupling, not because today's
+               flow can reach it. */
+            if (!src)
+                continue; // orphaned vector row surviving between sweeps — skip, don't crash
+            /* istanbul ignore next -- `src` is always either a `tokenSearchNodes` row or a
+               `getNode()` result, both of which route through `parseNodeRow`/`parseNodeAliases` and
+               so always carry a real `string[]` aliases — the `?? []` fallback is unreachable in
+               practice, kept only as a type-level safety net. */
+            const srcAliases = src.aliases ?? [];
+            const node = {
+                id: src.id, type: src.type, name: src.name, file_path: src.file_path,
+                signature: src.signature, description: src.description,
+                aliases: srcAliases,
+                deprecated: src.deprecated, created_at: src.created_at
+            };
+            const found_by = [];
+            if (bm25Hit)
+                found_by.push('keyword');
+            if (sim !== undefined)
+                found_by.push('meaning');
+            if (codeLines)
+                found_by.push('code');
+            // Confidence from real evidence, not the fused float. Corroboration across ≥2 independent
+            // layers is the strongest signal there is → high. A lone semantic match is graded by its
+            // actual cosine (0.6+ is genuinely close; 0.35-0.45 is borderline). A lone keyword or code
+            // hit already cleared its own floor, so it's real-but-uncorroborated → medium.
+            let confidence;
+            if (found_by.length >= 2) {
+                confidence = 'high';
+            }
+            else if (found_by.length === 1 && found_by[0] === 'meaning' && sim !== undefined) {
+                confidence = sim >= 0.6 ? 'high' : sim >= 0.45 ? 'medium' : 'low';
+            }
+            else {
+                confidence = 'medium';
+            }
+            const matched_via = bm25Hit ? 'fuzzy' : (codeLines ? 'code' : 'semantic');
+            const relevance = Math.max(1, Math.round((score / topScore) * 100));
+            // Trust signals ahead of `description` — see the identifier short-circuit above for why.
+            const { id: nodeId, name: nodeName, type: nodeType, ...nodeRest } = node;
+            nodes.push({
+                id: nodeId, name: nodeName, type: nodeType,
+                confidence,
+                relevance,
+                found_by,
+                matched_via,
+                ...nodeRest,
+                matched_terms: bm25Hit ? bm25Hit.matched_terms : [],
+                code_matches: codeLines,
+                uses: 0, used_by: 0, history_count: 0
+            });
+        }
+        this.attachDrillInHooks(nodes);
+        return {
+            nodes,
+            files: filesResult.files,
+            files_total: filesResult.total,
+            files_offset: filesOffset,
+            nodes_total: fused.length,
+            truncated: grep.truncated || undefined,
+            scope_note: scopeNote
+        };
+    }
+    /**
+     * The token-ranked (BM25) half of {@link searchNodes}. Looks up every query/keyword token in
+     * the local `node_tokens` index (rebuilding it first if stale — see {@link ensureSearchIndexFresh}),
+     * scores each candidate with {@link scoreCandidate} (IDF-weighted, saturating TF, per-field
+     * weights favoring `description`), and applies the noise floor. Metadata only — identifier / id
+     * / path / description / reasoning. Code-body matching is NO LONGER folded in here: it used to
+     * read one history-JSON per node (~9k serial reads = the tool's ~10-30s cost), and is now
+     * served far faster by the real filesystem grep in `searchNodes` (see `mapGrepHitsToNodes`).
+     * Returns nodes tagged `matched_via:'fuzzy'`, ranked; `searchNodes` re-fuses them by RRF.
+     */
+    tokenSearchNodes(tokens) {
+        const uniq = Array.from(new Set(tokens));
+        if (uniq.length === 0)
+            return [];
+        this.ensureSearchIndexFresh();
+        const totalNodesRow = this.db.prepare('SELECT COUNT(*) as c FROM nodes WHERE deprecated = 0').get();
+        const totalNodes = Math.max(totalNodesRow.c, 1);
+        const placeholders = uniq.map(() => '?').join(',');
+        // Document frequency per (token, field) computed ONCE via a grouped aggregate, not as a
+        // per-ROW correlated subquery — that used to re-scan node_tokens once for EVERY matching row
+        // (thousands, for a common word across an 8k-node/468k-token real corpus), which alone
+        // measured at 9-10 SECONDS on a real production graph. One aggregate query + an in-memory
+        // lookup does the identical computation in a few milliseconds.
+        const docFreqRows = this.db.prepare(`
+      SELECT token, field, COUNT(DISTINCT node_id) AS doc_freq
+      FROM node_tokens
+      WHERE token IN (${placeholders})
+      GROUP BY token, field
+    `).all(...uniq);
+        const docFreqByKey = new Map();
+        for (const r of docFreqRows)
+            docFreqByKey.set(`${r.token}${r.field}`, r.doc_freq);
+        const rows = this.db.prepare(`
+      SELECT node_id, token, field, tf
+      FROM node_tokens
+      WHERE token IN (${placeholders})
+    `).all(...uniq);
+        const byNode = new Map();
+        for (const row of rows) {
+            const entry = byNode.get(row.node_id) || { matches: [], terms: new Set() };
+            /* istanbul ignore next -- `rows` and `docFreqRows` are both filtered from `node_tokens` by
+               the exact same `WHERE token IN (...)`, and `docFreqRows` is a `GROUP BY token, field` over
+               that identical row set — so every (token, field) pair appearing in `rows` necessarily has
+               a matching aggregate entry already. The `?? 1` fallback is unreachable in practice, kept
+               only as a defensive default if the two queries are ever edited out of lockstep. */
+            const docFreq = docFreqByKey.get(`${row.token}${row.field}`) ?? 1;
+            entry.matches.push({ field: row.field, tf: row.tf, docFreq, totalNodes });
+            entry.terms.add(row.token);
+            byNode.set(row.node_id, entry);
+        }
+        if (byNode.size === 0)
             return [];
+        // Reject the thinnest possible "match" instead of letting anything with one shared token
+        // through: a query like "pending-orders/process-order" sharing only the generic word "order"
+        // with an unrelated node is weak evidence, not a real hit. Two tunable gates:
+        //  (1) coverage — a multi-token query must match more than a single one of its distinct tokens
+        //      (a single-token query has nothing more to require, so this is a no-op for it);
+        //  (2) a minimum absolute score, since even one match CAN be strong (a rare word in the
+        //      description field) and shouldn't be discarded just for being one token.
+        const minCoverage = Math.min(2, uniq.length);
+        const MIN_BM25_SCORE = 0.75;
+        const out = [];
+        const nodeIds = Array.from(byNode.keys());
+        const nodePlaceholders = nodeIds.map(() => '?').join(',');
+        const nodeRows = DevMindDatabase.parseNodeRows(this.db.prepare(`SELECT * FROM nodes WHERE id IN (${nodePlaceholders}) AND deprecated = 0`).all(...nodeIds));
+        for (const node of nodeRows) {
+            const entry = byNode.get(node.id);
+            if (entry.terms.size < minCoverage)
+                continue;
+            const score = (0, search_index_1.scoreCandidate)(entry.matches, search_index_1.DEFAULT_FIELD_WEIGHTS);
+            if (score < MIN_BM25_SCORE)
+                continue;
+            out.push({ ...node, matched_via: 'fuzzy', matched_terms: Array.from(entry.terms), score, low_confidence: true });
+        }
+        return out.sort((a, b) => b.score - a.score).slice(0, 20);
+    }
+    /** Absolute filesystem roots for every configured repo — the search space for `grepRepos`. */
+    repoRoots() {
+        if (!this.context)
+            return [];
+        const roots = [];
+        for (const repo of this.context.config.repos) {
+            const root = (0, config_1.resolveRepoPath)(this.context, repo.name);
+            if (root)
+                roots.push(root);
         }
-        return this.fuzzySearchNodes(tokens);
+        return roots;
     }
     /**
-     * Splits a query string into lowercase word tokens for the fuzzy fallback
-     * stage of {@link searchNodes}. This is request-scoped tokenization only —
-     * nothing is persisted or indexed; the result is discarded after the call.
+     * Resolves and validates `search_nodes`' optional `path` scope — a single folder or file the
+     * grep walk restricts to, instead of always walking every configured repo. A path outside every
+     * repo root is REJECTED (thrown, not silently widened to "search everything") — a scope that
+     * can't be honored should fail loudly, since silently ignoring it would search far more than
+     * the caller asked for. `canonicalizePath` (already used throughout this file for path
+     * comparisons) lowercases the Windows drive letter so `C:\...` and `c:\...` compare equal.
      */
-    tokenizeQuery(query) {
-        const seen = new Set();
-        for (const raw of query.toLowerCase().split(/[^a-z0-9]+/i)) {
-            if (raw.length >= 2)
-                seen.add(raw);
+    resolveSearchScopePath(rawPath) {
+        if (!rawPath || !rawPath.trim())
+            return undefined;
+        const resolved = (0, config_1.canonicalizePath)(rawPath.trim());
+        const roots = this.repoRoots().map(config_1.canonicalizePath);
+        const contained = roots.some(root => resolved === root || resolved.startsWith(root + path.sep));
+        if (!contained) {
+            throw new Error(`search_nodes: path "${rawPath}" is outside every configured repo (${roots.join(', ') || 'none configured'}).`);
         }
-        return Array.from(seen);
+        return resolved;
     }
     /**
-     * Word-split relevance-ranked fallback for {@link searchNodes}. Runs only
-     * when the exact identifier and code stages both return nothing. Scores
-     * every non-deprecated node by how many distinct query tokens appear as a
-     * substring of its file_path/name/id (highest signal), latest reasoning,
-     * or code content (lowest signal, one point per matching line). No new
-     * data is written or synced — this is a plain in-memory scan reusing the
-     * same node/history sources searchCode already reads.
+     * Maps raw grep hits back to graph nodes — this is "code search that returns nodes", rebuilt
+     * on top of the single filesystem walk instead of the old ~9k per-node snapshot reads. For each
+     * file that had hits AND contains indexed nodes: an AST-parseable file is resolved precisely
+     * (each hit line → the one node whose line-range contains it, via {@link locateNodeInFile}), so
+     * a keyword hitting one method in a 10-method file surfaces THAT node, not all ten. A non-AST
+     * indexed file (a staged `.py`/`.go` node) falls back to coarse file→node. Returns node ids in
+     * match-strength order (files with more hits first) plus the matching lines, for RRF + display.
+     *
+     * Only the top {@link CODE_MATCH_FILE_CAP} files by hit count are AST-resolved — measured on a
+     * real 8-repo, 15k-grep-hit query, this step alone was 8.2s of a 9.6s total before the cap, one
+     * `locateNodeInFile` AST walk per node per matched file with no bound on how many files that
+     * could be. It doesn't cost result quality: the fused ranking downstream keeps only the top ~20
+     * nodes anyway, and files with more hits are exactly the ones most likely to place there — files
+     * past the cap would almost never have survived to the final result even fully resolved.
      */
-    fuzzySearchNodes(tokens) {
-        const historyDir = path.join(path.dirname(this.dbPath), 'history');
-        const stmt = this.db.prepare(`
-      SELECT n.*, h.id AS latest_history_id, h.reasoning AS reasoning
-      FROM nodes n
-      LEFT JOIN history h ON h.id = (
-        SELECT id FROM history WHERE node_id = n.id ORDER BY updated_at DESC LIMIT 1
-      )
-      WHERE n.deprecated = 0
-    `);
-        const rows = stmt.all();
-        const FIELD_WEIGHT = { path: 3, identifier: 3, reasoning: 2, code: 1 };
-        const scored = [];
-        for (const row of rows) {
-            const { latest_history_id, reasoning, ...node } = row;
-            const matchedTerms = new Set();
-            let score = 0;
-            const filePathLower = (node.file_path || '').toLowerCase();
-            const nameLower = (node.name || '').toLowerCase();
-            const idLower = (node.id || '').toLowerCase();
-            const reasoningLower = (reasoning || '').toLowerCase();
-            let code = '';
-            if (latest_history_id) {
-                const historyFile = path.join(historyDir, `${latest_history_id}.json`);
-                if (fs.existsSync(historyFile)) {
-                    try {
-                        const data = JSON.parse(fs.readFileSync(historyFile, 'utf-8'));
-                        code = (data.code_snapshot || '').toLowerCase();
-                    }
-                    catch {
-                        // Skip corrupted or unreadable history files
+    mapGrepHitsToNodes(hits) {
+        if (hits.length === 0)
+            return [];
+        // Group hit line numbers (+ content) by file, and rank files by hit count so the strongest
+        // code matches come first in the returned order (which becomes their RRF rank).
+        const byFile = new Map();
+        for (const h of hits) {
+            let arr = byFile.get(h.file_path);
+            if (!arr) {
+                arr = [];
+                byFile.set(h.file_path, arr);
+            }
+            arr.push({ line_number: h.line_number, line_content: h.line_content });
+        }
+        const CODE_MATCH_FILE_CAP = 30;
+        const filesByStrength = Array.from(byFile.entries())
+            .sort((a, b) => b[1].length - a[1].length)
+            .slice(0, CODE_MATCH_FILE_CAP);
+        const perNode = new Map();
+        for (const [filePath, lines] of filesByStrength) {
+            const nodes = this.getNodesByFilePath(filePath);
+            if (nodes.length === 0)
+                continue; // an un-indexed file — lives in the files bucket only
+            if ((0, ast_1.isAstParseable)(filePath)) {
+                // Precise: resolve each node's line span once, assign every hit line to its container.
+                const spans = this.computeSymbolSpans(filePath, nodes);
+                for (const line of lines) {
+                    const containing = spans.find(s => line.line_number >= s.startLine && line.line_number <= s.endLine);
+                    if (!containing)
+                        continue; // hit was between symbols (an import, a top-level const) — file bucket has it
+                    let acc = perNode.get(containing.id);
+                    if (!acc) {
+                        acc = [];
+                        perNode.set(containing.id, acc);
                     }
+                    if (acc.length < 5)
+                        acc.push(line);
                 }
             }
-            for (const token of tokens) {
-                let tokenMatched = false;
-                if (filePathLower.includes(token)) {
-                    score += FIELD_WEIGHT.path;
-                    tokenMatched = true;
-                }
-                if (nameLower.includes(token) || idLower.includes(token)) {
-                    score += FIELD_WEIGHT.identifier;
-                    tokenMatched = true;
-                }
-                if (reasoningLower.includes(token)) {
-                    score += FIELD_WEIGHT.reasoning;
-                    tokenMatched = true;
-                }
-                if (code && code.includes(token)) {
-                    score += FIELD_WEIGHT.code;
-                    tokenMatched = true;
+            else {
+                // Coarse fallback for indexed but non-AST files (staged .py/.go/etc): every node in the file.
+                for (const n of nodes) {
+                    if (!perNode.has(n.id))
+                        perNode.set(n.id, lines.slice(0, 5));
                 }
-                if (tokenMatched)
-                    matchedTerms.add(token);
             }
-            if (score > 0) {
-                scored.push({ ...node, matched_via: 'fuzzy', matched_terms: Array.from(matchedTerms), score });
+        }
+        // perNode preserves insertion order = file-strength order, which is the ranking we want.
+        return Array.from(perNode.entries()).map(([nodeId, lines]) => ({ nodeId, lines }));
+    }
+    /**
+     * Resolves every indexed node's line span within one AST-parseable file — extracted out of
+     * `mapGrepHitsToNodes` so `annotateSampleLinesWithSymbol` can reuse the identical span logic
+     * instead of re-deriving it. Computed ONCE per file (not per line/hit): a file with several
+     * sample lines would otherwise repeat the same `locateNodeInFile` AST walk once per line.
+     * `nodes` is accepted rather than re-fetched so an existing `getNodesByFilePath` result (as
+     * `mapGrepHitsToNodes` already has) isn't queried twice.
+     */
+    computeSymbolSpans(filePath, nodes) {
+        return nodes.map(n => {
+            const parsed = (0, ast_1.parseNodeId)(n.id);
+            const symbol = parsed ? parsed.symbolName : (n.id.split('#').pop() || n.name);
+            const loc = symbol ? (0, ast_1.locateNodeInFile)(filePath, symbol) : null;
+            return loc ? { id: n.id, name: n.name, startLine: loc.startLine, endLine: loc.endLine } : null;
+        }).filter((s) => s !== null);
+    }
+    /**
+     * Mutates each `RankedFile`'s `sample_lines` in place, tagging every line with the function/class
+     * that contains it — the insight a plain filesystem grep can never give: not just "line 87
+     * matched" but "line 87, inside `onLikeTap`". This is the annotation `search_nodes` offers that
+     * makes it worth more than grep for a hit that lands inside real source.
+     *
+     * Deliberately bounded to the ALREADY-CAPPED page (`files`, ≤`maxFiles` entries × ≤5 sample
+     * lines each — at most ~125 lookups) rather than every raw grep hit. `mapGrepHitsToNodes` is
+     * measured at 8.2s of a 9.6s query on a real 8-repo search (see its doc comment) precisely from
+     * unbounded per-hit AST resolution; annotating only what's actually returned keeps this at a
+     * small, constant added cost regardless of how broad the pattern was or how many total hits it
+     * produced. Silently no-ops for a file that isn't indexed or isn't AST-parseable — a `symbol` on
+     * a sample line is a bonus, never a requirement.
+     */
+    annotateSampleLinesWithSymbol(files) {
+        for (const file of files) {
+            if (!(0, ast_1.isAstParseable)(file.file_path))
+                continue;
+            const nodes = this.getNodesByFilePath(file.file_path);
+            if (nodes.length === 0)
+                continue;
+            const spans = this.computeSymbolSpans(file.file_path, nodes);
+            if (spans.length === 0)
+                continue;
+            for (const line of file.sample_lines) {
+                const containing = spans.find(s => line.line_number >= s.startLine && line.line_number <= s.endLine);
+                if (containing)
+                    line.symbol = containing.name;
             }
         }
-        return scored.sort((a, b) => b.score - a.score).slice(0, 20);
     }
-    getRecentChanges(hours = 24, analyzeImpact = true) {
-        const stmt = this.db.prepare(`
-      SELECT h.id, h.node_id, n.name as node_name, n.file_path, h.updated_at, h.reasoning
-      FROM history h
-      JOIN nodes n ON h.node_id = n.id
-      WHERE h.updated_at >= datetime('now', ?)
-      ORDER BY h.updated_at DESC
-    `);
-        const recentChanges = stmt.all(`-${hours} hours`);
-        for (const change of recentChanges) {
-            const populated = this.populateHistoryFromDisk({ id: change.id, reasoning: change.reasoning });
-            change.reasoning = populated.reasoning;
-            delete change.id;
-        }
-        if (!analyzeImpact) {
-            return recentChanges;
-        }
-        const modifiedSet = new Set(recentChanges.map(c => c.node_id));
-        const callersStmt = this.db.prepare(`
-      SELECT n.id as node_id, n.name as node_name, n.file_path
-      FROM nodes n
-      JOIN node_connections c ON n.id = c.source_node_id
-      WHERE c.target_node_id = ?
-    `);
-        for (const change of recentChanges) {
-            const callers = callersStmt.all(change.node_id);
-            change.downstream_impact = callers.map(caller => ({
-                node_id: caller.node_id,
-                node_name: caller.node_name,
-                file_path: caller.file_path,
-                status: modifiedSet.has(caller.node_id) ? 'already_updated' : 'stale_warning'
-            }));
-        }
-        return recentChanges;
-    }
-    getDeveloperActivity(developer, limit = 50) {
-        const stmt = this.db.prepare(`
-      SELECT h.node_id, n.name as node_name, h.updated_at, h.reasoning
-      FROM history h
-      JOIN nodes n ON h.node_id = n.id
-      WHERE h.reasoning LIKE ? ESCAPE '\\'
-      ORDER BY h.updated_at DESC
-      LIMIT ?
-    `);
-        const query = `%Developer: %${this.likeEscape(developer)}%`;
-        return stmt.all(query, limit);
-    }
-    getChangesByRequirement(requirementId) {
-        const stmt = this.db.prepare(`
-      SELECT h.node_id, n.name as node_name, h.updated_at, h.reasoning
-      FROM history h
-      JOIN nodes n ON h.node_id = n.id
-      WHERE h.reasoning LIKE ? ESCAPE '\\'
-      ORDER BY h.updated_at DESC
-    `);
-        const query = `%Requirement: %${this.likeEscape(requirementId)}%`;
-        return stmt.all(query);
+    /**
+     * The semantic half of hybrid search: embeds the query with the same vendored ONNX model used
+     * to embed every node's description (see embedder.ts), then linear-scans `node_vectors` for
+     * cosine similarity. Linear is fine at this scale — ~9k nodes × 384 int8 dims is a few million
+     * integer multiplications, well under a millisecond, far below the model's own inference time
+     * for the query itself; revisit only past ~100k nodes. Returns each surviving node's raw cosine
+     * alongside its id — the ORDER drives RRF, but the cosine value is kept so `searchNodes` can turn
+     * it into a human-meaningful confidence (a 0.7 cosine is a strong match; a 0.36 is a weak one —
+     * the fused RRF float can't express that). Returns [] — never throws — if the embedder is
+     * unavailable, so `searchNodes` degrades to BM25+grep exactly as before.
+     */
+    async vectorSearchNodes(query) {
+        const queryVector = await (0, embedder_1.embedTextInt8)(query);
+        if (!queryVector)
+            return [];
+        const rows = this.db.prepare(`
+      SELECT nv.node_id AS node_id, nv.vector AS vector
+      FROM node_vectors nv
+      JOIN nodes n ON n.id = nv.node_id
+      WHERE n.deprecated = 0 AND nv.model_id = ?
+    `).all(embedder_1.EMBEDDING_MODEL_ID);
+        if (rows.length === 0)
+            return [];
+        const scored = rows.map(r => ({
+            id: r.node_id,
+            sim: (0, embedder_1.cosineInt8)(queryVector, new Int8Array(r.vector.buffer, r.vector.byteOffset, r.vector.byteLength))
+        }));
+        scored.sort((a, b) => b.sim - a.sim);
+        // A floor, not just a top-N cut: padding the ranking with genuinely unrelated nodes just to
+        // fill a quota would inject noise into the RRF fusion. A low cosine here means "nothing is
+        // semantically close" — that's a real signal that vector search found nothing, not a reason
+        // to return its least-bad guesses. 0.35, not 0.2: MiniLM-class sentence embeddings commonly
+        // put two UNRELATED short texts around 0.1-0.25 just from shared sentence structure/English
+        // baseline, not real similarity — 0.2 was letting that noise floor through as if it meant
+        // something. Tunable; revisit against real query/result pairs if this starts rejecting
+        // matches that should have passed.
+        const MIN_COSINE_SIMILARITY = 0.35;
+        return scored.filter(s => s.sim > MIN_COSINE_SIMILARITY).slice(0, 50);
+    }
+    /**
+     * Cheap, single-pass signal for "has the graph changed since node_tokens was last built":
+     * how many non-deprecated nodes exist, the total length of every description (catches a
+     * description being added OR edited — length almost never stays identical), and how many
+     * history rows exist (catches new reasoning). Deliberately NOT based on an `updated_at`
+     * column on `nodes` — there isn't one, and `syncFromDisk`'s destructive graph-rebuild pass
+     * writes nodes via raw SQL, not through a single method that could easily be hooked, so a
+     * write-path-by-write-path invalidation scheme would silently miss that path (and did, in an
+     * earlier draft of this). A fingerprint checked lazily at query time can't be missed the same
+     * way — it doesn't matter HOW the data changed, only THAT it did.
+     */
+    searchIndexFingerprint() {
+        const row = this.db.prepare(`
+      SELECT COUNT(*) as node_count, COALESCE(SUM(LENGTH(description)), 0) as desc_len_sum
+      FROM nodes WHERE deprecated = 0
+    `).get();
+        const historyRow = this.db.prepare('SELECT COUNT(*) as c FROM history').get();
+        return `${row.node_count}:${row.desc_len_sum}:${historyRow.c}`;
+    }
+    /** Rebuilds `node_tokens` from scratch for every non-deprecated node's current
+     * identifier/id/path/description and EVERY revision's reasoning (not just the latest — a
+     * decision recorded three revisions ago must stay findable by `search_nodes`, since there is
+     * no separate decisions-only search tool anymore; `GROUP_CONCAT` folds every history row's
+     * reasoning into one field before tokenizing, so an older "Decision: …" is searchable exactly
+     * like the newest one). Cheap at realistic node counts (hundreds to low-thousands) — a single
+     * scan plus one batched transaction, not something that needs to be avoided; simplicity here
+     * is worth more than incremental upkeep that a bypassed write path could silently defeat. */
+    rebuildSearchIndex() {
+        const nodes = this.db.prepare(`
+      SELECT n.*, (
+        SELECT GROUP_CONCAT(h.reasoning, ' ') FROM history h WHERE h.node_id = n.id
+      ) AS all_reasoning
+      FROM nodes n
+      WHERE n.deprecated = 0
+    `).all();
+        const del = this.db.prepare('DELETE FROM node_tokens');
+        const ins = this.db.prepare('INSERT OR REPLACE INTO node_tokens (node_id, token, field, tf) VALUES (?, ?, ?, ?)');
+        const tx = this.db.transaction(() => {
+            del.run();
+            for (const node of nodes) {
+                const rows = [
+                    ...(0, search_index_1.tokenizeNodeField)(node.name, 'identifier'),
+                    ...(0, search_index_1.tokenizeNodeField)(node.id, 'identifier'),
+                    ...(0, search_index_1.tokenizeNodeField)(node.file_path, 'path'),
+                    ...(0, search_index_1.tokenizeNodeField)(node.description, 'description'),
+                    ...(0, search_index_1.tokenizeNodeField)(node.all_reasoning, 'reasoning')
+                ];
+                // Merge duplicate (field,token) pairs from the multiple sources above (e.g. name and id
+                // both contributing the same word) into one summed tf, rather than letting the LATER
+                // INSERT OR REPLACE silently discard the earlier one's count.
+                const merged = new Map();
+                for (const r of rows) {
+                    const key = `${r.field}:${r.token}`;
+                    const existing = merged.get(key);
+                    if (existing)
+                        existing.tf += r.tf;
+                    else
+                        merged.set(key, { ...r });
+                }
+                for (const r of merged.values())
+                    ins.run(node.id, r.token, r.field, r.tf);
+            }
+        });
+        tx();
+    }
+    /** Compares the current fingerprint against what `node_tokens` was last built from, stored in
+     * `system_meta`; rebuilds and updates the stored fingerprint only on a mismatch. A momentarily
+     * stale index (between a real change and the next search call) only costs ranking quality on
+     * that one call, never correctness — the next call rebuilds it. */
+    ensureSearchIndexFresh() {
+        const current = this.searchIndexFingerprint();
+        const row = this.db.prepare(`SELECT value FROM system_meta WHERE key = 'search_index_fingerprint'`).get();
+        if (row && row.value === current)
+            return;
+        this.rebuildSearchIndex();
+        this.db.prepare(`INSERT OR REPLACE INTO system_meta (key, value, updated_at) VALUES ('search_index_fingerprint', ?, CURRENT_TIMESTAMP)`).run(current);
     }
     searchDecisions(query) {
-        const stmt = this.db.prepare(`
-      SELECT h.node_id, n.name as node_name, h.updated_at, h.reasoning
-      FROM history h
-      JOIN nodes n ON h.node_id = n.id
-      WHERE h.reasoning LIKE ? ESCAPE '\\'
-      ORDER BY h.updated_at DESC
+        const stmt = this.db.prepare(`
+      SELECT h.node_id, n.name as node_name, h.updated_at, h.reasoning
+      FROM history h
+      JOIN nodes n ON h.node_id = n.id
+      WHERE h.reasoning LIKE ? ESCAPE '\\'
+      ORDER BY h.updated_at DESC
     `);
         const wildcard = `%Decision: %${this.likeEscape(query)}%`;
         return stmt.all(wildcard);
@@ -1074,17 +2101,17 @@ class DevMindDatabase {
     searchCode(params) {
         const { query, is_regex = false, case_insensitive = true } = params;
         const historyDir = path.join(path.dirname(this.dbPath), 'history');
-        const stmt = this.db.prepare(`
-      SELECT h.id, n.id AS node_id, n.name AS node_name, n.file_path
-      FROM nodes n
-      JOIN history h ON h.node_id = n.id
-      WHERE n.deprecated = 0
-        AND h.id = (
-          SELECT id FROM history
-          WHERE node_id = n.id
-          ORDER BY updated_at DESC
-          LIMIT 1
-        )
+        const stmt = this.db.prepare(`
+      SELECT h.id, n.id AS node_id, n.name AS node_name, n.file_path
+      FROM nodes n
+      JOIN history h ON h.node_id = n.id
+      WHERE n.deprecated = 0
+        AND h.id = (
+          SELECT id FROM history
+          WHERE node_id = n.id
+          ORDER BY updated_at DESC
+          LIMIT 1
+        )
     `);
         const rows = stmt.all();
         let matcher;
@@ -1139,20 +2166,25 @@ class DevMindDatabase {
         return results.sort((a, b) => b.match_count - a.match_count);
     }
     getOrphanedNodes() {
-        const stmt = this.db.prepare(`
-      SELECT * FROM nodes
-      WHERE deprecated = 0
-        AND id NOT IN (SELECT DISTINCT source_node_id FROM node_connections)
-        AND id NOT IN (SELECT DISTINCT target_node_id FROM node_connections)
+        const stmt = this.db.prepare(`
+      SELECT * FROM nodes
+      WHERE deprecated = 0
+        AND id NOT IN (SELECT DISTINCT source_node_id FROM node_connections)
+        AND id NOT IN (SELECT DISTINCT target_node_id FROM node_connections)
     `);
-        return stmt.all();
+        return DevMindDatabase.parseNodeRows(stmt.all());
     }
     getAllNodes() {
         const stmt = this.db.prepare('SELECT * FROM nodes');
-        return stmt.all();
+        return DevMindDatabase.parseNodeRows(stmt.all());
     }
-    listNodes(filter) {
-        let sql = 'SELECT * FROM nodes WHERE 1=1';
+    /**
+     * The shared WHERE clause behind {@link listNodes} and {@link countNodes}. Extracted so the page
+     * and its total can never drift apart — a `total` computed from even slightly different criteria
+     * than the rows it describes is worse than no total at all, since it reads as authoritative.
+     */
+    buildNodeFilterSql(filter) {
+        let sql = ' WHERE 1=1';
         const params = [];
         if (filter?.type) {
             sql += ' AND type = ?';
@@ -1172,8 +2204,34 @@ class DevMindDatabase {
         if (!filter?.include_deprecated) {
             sql += ' AND deprecated = 0';
         }
-        const stmt = this.db.prepare(sql);
-        return stmt.all(...params);
+        return { where: sql, params };
+    }
+    /**
+     * Total nodes matching a filter, independent of any page. Exists so `list_nodes` can tell
+     * "that's everything" from "there is more" — the same honesty contract `nodes_total` and
+     * `files_total` already keep for search.
+     */
+    countNodes(filter) {
+        const { where, params } = this.buildNodeFilterSql(filter);
+        const row = this.db.prepare(`SELECT COUNT(*) AS c FROM nodes${where}`).get(...params);
+        return row.c;
+    }
+    /**
+     * `limit`/`offset` are OPTIONAL and, when omitted, this returns every matching row exactly as
+     * before — the internal analysis callers (`analyze.ts`, `edges.ts`) legitimately need the whole
+     * graph, so paging is opt-in rather than a default that would silently truncate them. Ordering
+     * is applied only when paging, since an unordered LIMIT is a lottery: without it, "page 2" is
+     * not guaranteed to exclude what "page 1" already returned.
+     */
+    listNodes(filter) {
+        const { where, params } = this.buildNodeFilterSql(filter);
+        let sql = `SELECT * FROM nodes${where}`;
+        const args = [...params];
+        if (filter?.limit !== undefined) {
+            sql += ' ORDER BY file_path, name LIMIT ? OFFSET ?';
+            args.push(filter.limit, filter.offset ?? 0);
+        }
+        return DevMindDatabase.parseNodeRows(this.db.prepare(sql).all(...args));
     }
     getAllConnections() {
         const stmt = this.db.prepare('SELECT * FROM node_connections');
@@ -1188,17 +2246,17 @@ class DevMindDatabase {
     // All pure queries/graph traversal, no mutation, no LLM calls.
     /** Nodes whose total (in + out) connection degree meets/exceeds `threshold` — architectural bottleneck candidates. */
     getGodEntities(threshold = 15) {
-        const stmt = this.db.prepare(`
-      SELECT * FROM (
-        SELECT n.id, n.name, n.file_path, (
-          (SELECT COUNT(*) FROM node_connections c WHERE c.source_node_id = n.id) +
-          (SELECT COUNT(*) FROM node_connections c WHERE c.target_node_id = n.id)
-        ) AS degree
-        FROM nodes n
-        WHERE n.deprecated = 0
-      )
-      WHERE degree >= ?
-      ORDER BY degree DESC
+        const stmt = this.db.prepare(`
+      SELECT * FROM (
+        SELECT n.id, n.name, n.file_path, (
+          (SELECT COUNT(*) FROM node_connections c WHERE c.source_node_id = n.id) +
+          (SELECT COUNT(*) FROM node_connections c WHERE c.target_node_id = n.id)
+        ) AS degree
+        FROM nodes n
+        WHERE n.deprecated = 0
+      )
+      WHERE degree >= ?
+      ORDER BY degree DESC
     `);
         return stmt.all(threshold);
     }
@@ -1216,6 +2274,12 @@ class DevMindDatabase {
         const stack = [];
         const onStack = new Set();
         const dfs = (node) => {
+            /* istanbul ignore if -- both of dfs's call sites (the outer `for` loop below, and the
+               `for (const next of ...)` loop a few lines down) already check `cycles.length >=
+               maxCycles` immediately before every single call to `dfs(...)`, so this repeats a guard
+               that has always already passed by the time control reaches here. Kept as a real guard
+               against a future call site that skips that pre-check, not because today's two call
+               sites can reach it. */
             if (cycles.length >= maxCycles)
                 return;
             if (onStack.has(node)) {
@@ -1246,10 +2310,10 @@ class DevMindDatabase {
     }
     /** node_connections rows whose source or target no longer exists in `nodes` (broken by a non-transactional delete, or a sync race). */
     getDanglingEdges() {
-        const stmt = this.db.prepare(`
-      SELECT * FROM node_connections
-      WHERE source_node_id NOT IN (SELECT id FROM nodes)
-         OR target_node_id NOT IN (SELECT id FROM nodes)
+        const stmt = this.db.prepare(`
+      SELECT * FROM node_connections
+      WHERE source_node_id NOT IN (SELECT id FROM nodes)
+         OR target_node_id NOT IN (SELECT id FROM nodes)
     `);
         return stmt.all();
     }
@@ -1259,12 +2323,12 @@ class DevMindDatabase {
     }
     /** Node ids that differ only by case — a real collision risk on Windows's case-insensitive filesystem. */
     getDuplicateNodeIds() {
-        const stmt = this.db.prepare(`
-      SELECT LOWER(id) AS lower_id, GROUP_CONCAT(id, '|') AS ids
-      FROM nodes
-      WHERE deprecated = 0
-      GROUP BY lower_id
-      HAVING COUNT(*) > 1
+        const stmt = this.db.prepare(`
+      SELECT LOWER(id) AS lower_id, GROUP_CONCAT(id, '|') AS ids
+      FROM nodes
+      WHERE deprecated = 0
+      GROUP BY lower_id
+      HAVING COUNT(*) > 1
     `);
         const rows = stmt.all();
         return rows.map(r => ({ lowerId: r.lower_id, ids: r.ids.split('|') }));
@@ -1301,7 +2365,76 @@ class DevMindDatabase {
     workflowsDir() {
         return path.join(path.dirname(this.dbPath), 'workflows');
     }
-    /** Serializes the workflow + its steps + artifact index to disk so teammates can sync it via git. */
+    /**
+     * One-time backfill of `node_ids`/`reasoning` for steps written before v2, resolving each old
+     * `history_ids` entry to the node it belongs to. Runs at open, right after the ALTERs, and is
+     * self-limiting: it only touches rows where `node_ids IS NULL AND history_ids IS NOT NULL`, so
+     * the second open finds nothing and the query costs one indexed scan.
+     *
+     * Best-effort by nature. Because of the 1-hour history merge, an old step's `history_ids` can
+     * include rows an ADJACENT commit created, so a backfilled node list can be broader than what
+     * that step actually touched. That is acceptable for pre-v2 rows — they were already imprecise,
+     * which is exactly why the format changed — but it is why nothing presents backfilled data as
+     * exact. A step whose history rows have since been pruned keeps its `summary` and gets an empty
+     * list rather than being skipped, so it still appears on the timeline.
+     */
+    backfillWorkflowStepNodeIds() {
+        try {
+            const stale = this.db
+                .prepare(`SELECT id, history_ids FROM workflow_steps WHERE node_ids IS NULL AND history_ids IS NOT NULL`)
+                .all();
+            if (stale.length === 0)
+                return;
+            const lookup = this.db.prepare('SELECT node_id, reasoning FROM history WHERE id = ?');
+            const update = this.db.prepare('UPDATE workflow_steps SET node_ids = ?, reasoning = COALESCE(reasoning, ?) WHERE id = ?');
+            const run = this.db.transaction((rows) => {
+                for (const row of rows) {
+                    let ids;
+                    try {
+                        ids = JSON.parse(row.history_ids);
+                    }
+                    catch {
+                        ids = [];
+                    }
+                    const nodeIds = [];
+                    let reasoning = null;
+                    if (Array.isArray(ids)) {
+                        for (const historyId of ids) {
+                            const hit = lookup.get(String(historyId));
+                            if (!hit)
+                                continue;
+                            if (!nodeIds.includes(hit.node_id))
+                                nodeIds.push(hit.node_id);
+                            if (reasoning === null && hit.reasoning)
+                                reasoning = hit.reasoning;
+                        }
+                    }
+                    update.run(JSON.stringify(nodeIds), reasoning, row.id);
+                }
+            });
+            run(stale);
+        }
+        catch {
+            // A brain mid-migration (or a workflow_steps table that predates history_ids entirely)
+            // must not block opening the DB — the steps simply stay on their old shape.
+        }
+    }
+    /**
+     * Serializes the workflow + its steps + artifact index to disk so teammates can sync it via git.
+     *
+     * Written as TWO files, and the split is the whole point:
+     *
+     * - `workflow.json` keeps the shape a v1 client understands, so an older build reading it loses
+     *   nothing it ever had.
+     * - `v2.json` holds everything v1 has no field for (`archived`, and per-step `reasoning` /
+     *   `node_ids` / `doc_paths`).
+     *
+     * A single file could not be made safe. `devsmind sync` calls `syncToDisk`, which re-serializes
+     * every workflow.json from whatever columns the local build knows about — so a teammate who
+     * pulls on an older version and syncs would rewrite every workflow and silently strip the new
+     * fields, then commit that loss for everyone. An older build has no idea `v2.json` exists, so it
+     * cannot rewrite it; the data survives the round trip and is merged back on the next read.
+     */
     writeWorkflowToDisk(workflowId) {
         try {
             const workflow = this.db.prepare('SELECT * FROM workflows WHERE id = ?').get(workflowId);
@@ -1309,21 +2442,21 @@ class DevMindDatabase {
                 return;
             const steps = this.db.prepare('SELECT * FROM workflow_steps WHERE workflow_id = ? ORDER BY step_index ASC').all(workflowId);
             const artifacts = this.db.prepare('SELECT * FROM workflow_artifacts WHERE workflow_id = ? ORDER BY created_at ASC').all(workflowId);
-            const activeId = this.getSystemMeta('active_workflow_id');
             const data = {
+                schema_version: exports.WORKFLOW_SCHEMA_VERSION,
                 id: workflow.id,
                 name: workflow.name,
                 description: workflow.description,
-                status: workflow.status,
+                archived: workflow.archived ? 1 : 0,
                 created_at: workflow.created_at,
                 updated_at: workflow.updated_at,
-                is_active: activeId === workflowId,
                 steps: steps.map(s => ({
                     id: s.id,
                     step_index: s.step_index,
                     summary: s.summary,
-                    pending_tasks: s.pending_tasks,
-                    history_ids: s.history_ids,
+                    reasoning: s.reasoning,
+                    node_ids: s.node_ids,
+                    doc_paths: s.doc_paths,
                     session_id: s.session_id,
                     created_at: s.created_at
                 })),
@@ -1339,86 +2472,123 @@ class DevMindDatabase {
             const dir = path.join(this.workflowsDir(), workflowId);
             fs.mkdirSync(dir, { recursive: true });
             fs.writeFileSync(path.join(dir, 'workflow.json'), JSON.stringify(data, null, 2), 'utf-8');
+            // The v1-invisible half. Keyed by step id rather than positional, so it still merges
+            // correctly onto a workflow.json an older build reordered or rewrote.
+            const sidecar = {
+                schema_version: exports.WORKFLOW_SCHEMA_VERSION,
+                archived: workflow.archived ? 1 : 0,
+                steps: Object.fromEntries(steps
+                    .filter(s => s.reasoning || s.node_ids || s.doc_paths)
+                    .map(s => [s.id, { reasoning: s.reasoning, node_ids: s.node_ids, doc_paths: s.doc_paths }]))
+            };
+            fs.writeFileSync(path.join(dir, exports.WORKFLOW_SIDECAR_FILE), JSON.stringify(sidecar, null, 2), 'utf-8');
         }
         catch (err) {
             console.warn('⚠️ DevsMind: Failed to write workflow JSON to disk:', err);
         }
     }
+    /**
+     * Creates a workflow. Note what it deliberately does NOT do any more: set a global "active"
+     * pointer. Which workflow you are working on is a property of YOUR session, held locally — a
+     * workflow is a shared record, and one shared pointer meant two sessions (or two teammates,
+     * since the pointer synced through git) silently stole it from each other mid-work.
+     */
     createWorkflow(name, description) {
         const id = `wf_${crypto.randomUUID()}`;
         const now = new Date().toISOString();
-        this.db.prepare(`
-      INSERT INTO workflows (id, name, description, status, created_at, updated_at)
-      VALUES (?, ?, ?, 'active', ?, ?)
+        this.db.prepare(`
+      INSERT INTO workflows (id, name, description, archived, created_at, updated_at)
+      VALUES (?, ?, ?, 0, ?, ?)
     `).run(id, name, description, now, now);
-        this.setSystemMeta('active_workflow_id', id);
         this.writeWorkflowToDisk(id);
-        return { id, name, description, status: 'active', created_at: now, updated_at: now };
+        return { id, name, description, archived: 0, created_at: now, updated_at: now };
     }
     getWorkflow(id) {
         const row = this.db.prepare('SELECT * FROM workflows WHERE id = ?').get(id);
         return row || null;
     }
-    getActiveWorkflow() {
-        const id = this.getSystemMeta('active_workflow_id');
-        return id ? this.getWorkflow(id) : null;
-    }
-    listWorkflows(status) {
-        if (status) {
-            return this.db.prepare('SELECT * FROM workflows WHERE status = ? ORDER BY updated_at DESC').all(status);
+    /**
+     * Workflows newest-touched first — which is the ordering that replaces the old `status` field.
+     * Live work floats up and abandoned threads sink on their own, so nothing has to be marked
+     * "completed" by hand (nobody ever did, and a lifecycle field nobody maintains just lies).
+     *
+     * `query` matches name AND description, the search `searchWorkflows` never actually did: it
+     * scanned step summaries and artifact names only, so looking a workflow up by its own name
+     * returned nothing. Paging mirrors `listNodes` — `total` is the true count before the page.
+     */
+    listWorkflows(opts) {
+        const { where, params } = this.buildWorkflowFilterSql(opts);
+        let sql = `SELECT * FROM workflows${where} ORDER BY updated_at DESC`;
+        const args = [...params];
+        if (opts?.limit !== undefined) {
+            sql += ' LIMIT ? OFFSET ?';
+            args.push(opts.limit, opts.offset ?? 0);
+        }
+        return this.db.prepare(sql).all(...args);
+    }
+    countWorkflows(opts) {
+        const { where, params } = this.buildWorkflowFilterSql(opts);
+        const row = this.db.prepare(`SELECT COUNT(*) AS c FROM workflows${where}`).get(...params);
+        return row.c;
+    }
+    /** Shared WHERE builder, so a page and its `total` can never describe different criteria. */
+    buildWorkflowFilterSql(opts) {
+        let where = ' WHERE 1=1';
+        const params = [];
+        if (!opts?.includeArchived)
+            where += ' AND archived = 0';
+        if (opts?.query && opts.query.trim()) {
+            where += " AND (LOWER(name) LIKE ? ESCAPE '\\' OR LOWER(description) LIKE ? ESCAPE '\\')";
+            const like = `%${this.likeEscape(opts.query.trim().toLowerCase())}%`;
+            params.push(like, like);
         }
-        return this.db.prepare('SELECT * FROM workflows ORDER BY updated_at DESC').all();
-    }
-    /** Pauses the currently active workflow (if any) and clears the active pointer. */
-    pauseWorkflow() {
-        const active = this.getActiveWorkflow();
-        if (!active)
-            return null;
-        const now = new Date().toISOString();
-        this.db.prepare(`UPDATE workflows SET status = 'paused', updated_at = ? WHERE id = ?`).run(now, active.id);
-        this.setSystemMeta('active_workflow_id', '');
-        this.writeWorkflowToDisk(active.id);
-        return { ...active, status: 'paused', updated_at: now };
-    }
-    /** Resumes `id`, auto-pausing whatever was previously active (only one workflow is active at a time). */
-    resumeWorkflow(id) {
-        const workflow = this.getWorkflow(id);
-        if (!workflow)
-            throw new Error(`Workflow not found: ${id}`);
-        const currentActive = this.getActiveWorkflow();
-        if (currentActive && currentActive.id !== id)
-            this.pauseWorkflow();
-        const now = new Date().toISOString();
-        this.db.prepare(`UPDATE workflows SET status = 'active', updated_at = ? WHERE id = ?`).run(now, id);
-        this.setSystemMeta('active_workflow_id', id);
-        this.writeWorkflowToDisk(id);
-        return { ...workflow, status: 'active', updated_at: now };
+        return { where, params };
     }
-    completeWorkflow(id) {
+    /**
+     * Hides a workflow from the default listing. Deliberately NOT called "complete": a feature is
+     * never finished, it just stops being worked on, and the old `completed` status promised a
+     * lifecycle nobody maintained. Archiving claims only what it delivers, and is reversible.
+     */
+    setWorkflowArchived(id, archived) {
         const workflow = this.getWorkflow(id);
         if (!workflow)
             throw new Error(`Workflow not found: ${id}`);
         const now = new Date().toISOString();
-        this.db.prepare(`UPDATE workflows SET status = 'completed', updated_at = ? WHERE id = ?`).run(now, id);
-        if (this.getSystemMeta('active_workflow_id') === id)
-            this.setSystemMeta('active_workflow_id', '');
+        this.db.prepare('UPDATE workflows SET archived = ?, updated_at = ? WHERE id = ?').run(archived ? 1 : 0, now, id);
         this.writeWorkflowToDisk(id);
-        return { ...workflow, status: 'completed', updated_at: now };
+        return { ...workflow, archived: archived ? 1 : 0, updated_at: now };
     }
+    /**
+     * Appends one step. A step is either a COMMIT (summary + reasoning + the node ids it touched) or
+     * a RESEARCH finding (summary + reasoning + the docs behind it, no nodes) — the second is the
+     * only record of work that produced a decision but no code, which nothing else in DevsMind
+     * captures: git has the diff and history has the per-node reasoning, but neither can tell you
+     * what was evaluated and rejected.
+     *
+     * `reasoning` is stored, not joined from `history`, deliberately — see DbWorkflowStep.
+     */
     addWorkflowStep(workflowId, opts) {
         if (!this.getWorkflow(workflowId))
             throw new Error(`Workflow not found: ${workflowId}`);
         const id = crypto.randomUUID();
         const now = new Date().toISOString();
         const nextIndex = (this.db.prepare('SELECT MAX(step_index) AS m FROM workflow_steps WHERE workflow_id = ?').get(workflowId).m ?? 0) + 1;
-        const historyIdsJson = opts.historyIds && opts.historyIds.length ? JSON.stringify(opts.historyIds) : null;
-        this.db.prepare(`
-      INSERT INTO workflow_steps (id, workflow_id, step_index, summary, pending_tasks, history_ids, session_id, created_at)
-      VALUES (?, ?, ?, ?, ?, ?, ?, ?)
-    `).run(id, workflowId, nextIndex, opts.summary, opts.pendingTasks || null, historyIdsJson, opts.sessionId || null, now);
+        // Empty arrays store as NULL rather than "[]" so "this step touched nothing" and "this step
+        // predates the column" read the same downstream — neither is a list worth rendering.
+        const nodeIdsJson = opts.nodeIds && opts.nodeIds.length ? JSON.stringify(opts.nodeIds) : null;
+        const docPathsJson = opts.docPaths && opts.docPaths.length ? JSON.stringify(opts.docPaths) : null;
+        const reasoning = opts.reasoning || null;
+        this.db.prepare(`
+      INSERT INTO workflow_steps (id, workflow_id, step_index, summary, reasoning, node_ids, doc_paths, session_id, created_at)
+      VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?)
+    `).run(id, workflowId, nextIndex, opts.summary, reasoning, nodeIdsJson, docPathsJson, opts.sessionId || null, now);
         this.db.prepare(`UPDATE workflows SET updated_at = ? WHERE id = ?`).run(now, workflowId);
         this.writeWorkflowToDisk(workflowId);
-        return { id, workflow_id: workflowId, step_index: nextIndex, summary: opts.summary, pending_tasks: opts.pendingTasks || null, history_ids: historyIdsJson, session_id: opts.sessionId || null, created_at: now };
+        return {
+            id, workflow_id: workflowId, step_index: nextIndex, summary: opts.summary,
+            reasoning, node_ids: nodeIdsJson, doc_paths: docPathsJson,
+            session_id: opts.sessionId || null, created_at: now
+        };
     }
     /** Writes `content` to `.devmind/workflows/<workflowId>/<artifactId>_<sourceName>` and records the DB row. */
     addWorkflowArtifact(workflowId, opts) {
@@ -1431,32 +2601,38 @@ class DevMindDatabase {
         fs.mkdirSync(dir, { recursive: true });
         const filePath = path.join(dir, `${id}_${safeName}`);
         fs.writeFileSync(filePath, opts.content, 'utf-8');
-        this.db.prepare(`
-      INSERT INTO workflow_artifacts (id, workflow_id, step_id, type, source_name, file_path, created_at)
-      VALUES (?, ?, ?, ?, ?, ?, ?)
+        this.db.prepare(`
+      INSERT INTO workflow_artifacts (id, workflow_id, step_id, type, source_name, file_path, created_at)
+      VALUES (?, ?, ?, ?, ?, ?, ?)
     `).run(id, workflowId, opts.stepId || null, opts.type, opts.sourceName, filePath, now);
         this.db.prepare(`UPDATE workflows SET updated_at = ? WHERE id = ?`).run(now, workflowId);
         this.writeWorkflowToDisk(workflowId);
         return { id, workflow_id: workflowId, step_id: opts.stepId || null, type: opts.type, source_name: opts.sourceName, file_path: filePath, created_at: now };
     }
+    /**
+     * The workflow's story: its steps in order, plus the docs attached to it.
+     *
+     * Paged, because this is now the ONLY read (it absorbed the old `workflow_get_steps`) and steps
+     * carry their own reasoning, so an unbounded version of it would be the largest response the
+     * server can produce. `steps_total` is exact regardless of the page — a short page must never
+     * read as "that is the whole story".
+     *
+     * Artifact CONTENT is deliberately not returned. It used to be inlined whole, which on an
+     * imported architecture doc is trivially tens of KB; the file path is enough, since the file is
+     * on disk and the caller can read exactly the part it needs.
+     */
     getWorkflowContext(id, opts) {
         const workflow = this.getWorkflow(id);
         if (!workflow)
             throw new Error(`Workflow not found: ${id}`);
-        const steps = this.db.prepare('SELECT * FROM workflow_steps WHERE workflow_id = ? ORDER BY step_index ASC').all(id);
-        const artifactRows = this.db.prepare('SELECT * FROM workflow_artifacts WHERE workflow_id = ? ORDER BY created_at ASC').all(id);
-        const artifacts = artifactRows.map(a => {
-            if (!opts?.includeArtifactContent)
-                return a;
-            try {
-                const content = fs.existsSync(a.file_path) ? fs.readFileSync(a.file_path, 'utf-8') : undefined;
-                return { ...a, content };
-            }
-            catch {
-                return a;
-            }
-        });
-        return { workflow, steps, artifacts };
+        const steps_total = this.db.prepare('SELECT COUNT(*) AS c FROM workflow_steps WHERE workflow_id = ?').get(id).c;
+        const steps = this.getWorkflowSteps(id, opts);
+        // `last_n` walks backwards from the end, so its offset is wherever that tail begins.
+        const steps_offset = opts?.last_n && opts.last_n > 0
+            ? Math.max(0, steps_total - steps.length)
+            : (opts?.offset ?? 0);
+        const artifacts = this.db.prepare('SELECT * FROM workflow_artifacts WHERE workflow_id = ? ORDER BY created_at ASC').all(id);
+        return { workflow, steps, steps_total, steps_offset, artifacts };
     }
     /**
      * Returns steps for a workflow with optional pagination.
@@ -1476,99 +2652,13 @@ class DevMindDatabase {
         }
         return this.db.prepare('SELECT * FROM workflow_steps WHERE workflow_id = ? ORDER BY step_index ASC').all(workflowId);
     }
-    /**
-     * Reads a single workflow artifact's file content from disk.
-     * Accepts either an artifact_id or a source_name (first match used).
-     */
-    readWorkflowArtifact(workflowId, artifactId) {
-        if (!this.getWorkflow(workflowId))
-            throw new Error(`Workflow not found: ${workflowId}`);
-        const row = this.db.prepare('SELECT * FROM workflow_artifacts WHERE workflow_id = ? AND id = ?').get(workflowId, artifactId);
-        if (!row)
-            throw new Error(`Artifact not found: ${artifactId} in workflow ${workflowId}`);
-        if (!fs.existsSync(row.file_path))
-            throw new Error(`Artifact file missing on disk: ${row.file_path}`);
-        const content = fs.readFileSync(row.file_path, 'utf-8');
-        return { artifact: row, content };
-    }
-    /**
-     * Full-text keyword search across all workflows' step summaries, pending_tasks,
-     * and artifact source names. Optionally also searches artifact file content.
-     * Returns a list of matches grouped by workflow.
-     */
-    searchWorkflows(query, opts) {
-        // Escaped so a query containing '%' or '_' matches those characters literally instead of
-        // acting as SQL LIKE wildcards — otherwise `query: "%"` matches every row in the project.
-        const lq = `%${this.likeEscape(query.toLowerCase())}%`;
-        // Find matching steps
-        const matchedStepRows = this.db.prepare(`
-      SELECT ws.* FROM workflow_steps ws
-      JOIN workflows w ON w.id = ws.workflow_id
-      WHERE (LOWER(ws.summary) LIKE ? ESCAPE '\\' OR LOWER(IFNULL(ws.pending_tasks,'')) LIKE ? ESCAPE '\\')
-      ${opts?.status ? 'AND w.status = ?' : ''}
-      ORDER BY ws.workflow_id, ws.step_index ASC
-    `).all(...(opts?.status ? [lq, lq, opts.status] : [lq, lq]));
-        // Find matching artifacts by source_name
-        const matchedArtifactRows = this.db.prepare(`
-      SELECT wa.* FROM workflow_artifacts wa
-      JOIN workflows w ON w.id = wa.workflow_id
-      WHERE LOWER(wa.source_name) LIKE ? ESCAPE '\\'
-      ${opts?.status ? 'AND w.status = ?' : ''}
-      ORDER BY wa.workflow_id, wa.created_at ASC
-    `).all(...(opts?.status ? [lq, opts.status] : [lq]));
-        // If content search requested, also scan artifact files
-        const contentMatchedArtifactIds = new Set();
-        const artifactContentSnippets = new Map();
-        if (opts?.include_artifact_content) {
-            const allArtifacts = this.db.prepare(`SELECT wa.* FROM workflow_artifacts wa JOIN workflows w ON w.id = wa.workflow_id${opts.status ? ' WHERE w.status = ?' : ''}`).all(...(opts.status ? [opts.status] : []));
-            const lqPlain = query.toLowerCase();
-            for (const a of allArtifacts) {
-                if (contentMatchedArtifactIds.has(a.id))
-                    continue;
-                try {
-                    if (fs.existsSync(a.file_path)) {
-                        const text = fs.readFileSync(a.file_path, 'utf-8');
-                        const idx = text.toLowerCase().indexOf(lqPlain);
-                        if (idx !== -1) {
-                            contentMatchedArtifactIds.add(a.id);
-                            const start = Math.max(0, idx - 80);
-                            const end = Math.min(text.length, idx + query.length + 80);
-                            artifactContentSnippets.set(a.id, (start > 0 ? '…' : '') + text.slice(start, end) + (end < text.length ? '…' : ''));
-                        }
-                    }
-                }
-                catch { /* skip unreadable */ }
-            }
-        }
-        // Collect all relevant workflow IDs
-        const workflowIdSet = new Set([
-            ...matchedStepRows.map(s => s.workflow_id),
-            ...matchedArtifactRows.map(a => a.workflow_id),
-            ...Array.from(contentMatchedArtifactIds).map(id => {
-                const r = this.db.prepare('SELECT workflow_id FROM workflow_artifacts WHERE id = ?').get(id);
-                return r?.workflow_id || '';
-            }).filter(Boolean)
-        ]);
-        const results = [];
-        for (const wid of workflowIdSet) {
-            const workflow = this.getWorkflow(wid);
-            if (!workflow)
-                continue;
-            const steps = matchedStepRows.filter(s => s.workflow_id === wid);
-            const artByName = matchedArtifactRows.filter(a => a.workflow_id === wid);
-            const artByContent = opts?.include_artifact_content
-                ? this.db.prepare('SELECT * FROM workflow_artifacts WHERE workflow_id = ?').all(wid).filter(a => contentMatchedArtifactIds.has(a.id) && !artByName.find(x => x.id === a.id))
-                : [];
-            const allArtifacts = [
-                ...artByName.map(a => ({ ...a, content_snippet: artifactContentSnippets.get(a.id) })),
-                ...artByContent.map(a => ({ ...a, content_snippet: artifactContentSnippets.get(a.id) }))
-            ];
-            results.push({ workflow, matched_steps: steps, matched_artifacts: allArtifacts });
-        }
-        // Sort by most recently updated workflow first
-        results.sort((a, b) => b.workflow.updated_at.localeCompare(a.workflow.updated_at));
-        return results;
-    }
+    // NOTE: `readWorkflowArtifact` and `searchWorkflows` were removed here.
+    // Artifacts are referenced by PATH now (workflow_add_step's doc_paths, plus the file paths
+    // getWorkflowContext already returns), so nothing needs the DB to read a file back for it —
+    // and inlining whole imported docs was the single largest thing a workflow response could emit.
+    // searchWorkflows was replaced by listWorkflows({ query }): it scanned step summaries and
+    // artifact names but NOT workflow name/description, so looking a workflow up by its own name
+    // returned nothing — the one search anybody actually tries.
     /**
      * Imports an existing flow/architecture doc as a paused workflow (not active — importing
      * a doc isn't the same as declaring active work). Idempotent on `name`: re-importing the
@@ -1591,14 +2681,14 @@ class DevMindDatabase {
             return { workflow: { ...existing, description, updated_at: now }, created: false };
         }
         const id = `wf_${crypto.randomUUID()}`;
-        this.db.prepare(`
-      INSERT INTO workflows (id, name, description, status, created_at, updated_at)
-      VALUES (?, ?, ?, 'paused', ?, ?)
+        this.db.prepare(`
+      INSERT INTO workflows (id, name, description, archived, created_at, updated_at)
+      VALUES (?, ?, ?, 0, ?, ?)
     `).run(id, name, description, now, now);
         this.addWorkflowStep(id, { summary: `Imported existing flow documentation: ${sourceFileName}` });
         this.addWorkflowArtifact(id, { type: 'imported_doc', sourceName: sourceFileName, content });
         // writeWorkflowToDisk is already called inside addWorkflowArtifact/addWorkflowStep above
-        return { workflow: { id, name, description, status: 'paused', created_at: now, updated_at: now }, created: true };
+        return { workflow: { id, name, description, archived: 0, created_at: now, updated_at: now }, created: true };
     }
     static SPURIOUS_NODE_NAMES = new Set([
         'promise', 'map', 'set', 'json', 'console', 'error', 'object', 'function', 'array', 'string', 'number', 'boolean', 'regexp', 'date', 'math',
@@ -1610,9 +2700,9 @@ class DevMindDatabase {
      * analyze`'s dry-run report (which just lists it). Never mutates the DB.
      */
     findSpuriousAndMissingFileNodes(workspaceRoot) {
-        const stmt = this.db.prepare(`
-      SELECT id, name, file_path FROM nodes
-      WHERE deprecated = 0
+        const stmt = this.db.prepare(`
+      SELECT id, name, file_path FROM nodes
+      WHERE deprecated = 0
     `);
         const candidates = stmt.all();
         const spurious = [];
@@ -1695,7 +2785,10 @@ class DevMindDatabase {
                 return {
                     ...row,
                     code_snapshot: data.code_snapshot || '',
-                    reasoning: typeof data.reasoning === 'string' ? data.reasoning : formatReasoning(data.reasoning || '')
+                    reasoning: typeof data.reasoning === 'string' ? data.reasoning : formatReasoning(data.reasoning || ''),
+                    // Absent in every entry written before the edit trail existed — an empty trail is the
+                    // honest answer there: nothing to diff, nothing to revert.
+                    edits: Array.isArray(data.edits) ? data.edits : []
                 };
             }
         }
@@ -1705,10 +2798,11 @@ class DevMindDatabase {
         return {
             ...row,
             code_snapshot: '',
-            reasoning: ''
+            reasoning: '',
+            edits: []
         };
     }
-    writeHistoryToDisk(id, nodeId, sessionId, createdAt, updatedAt, codeSnapshot, reasoning) {
+    writeHistoryToDisk(id, nodeId, sessionId, createdAt, updatedAt, codeSnapshot, reasoning, edits = []) {
         try {
             const historyDir = path.join(path.dirname(this.dbPath), 'history');
             if (!fs.existsSync(historyDir)) {
@@ -1729,7 +2823,8 @@ class DevMindDatabase {
                 created_at: createdAt,
                 updated_at: updatedAt,
                 code_snapshot: codeSnapshot,
-                reasoning
+                reasoning,
+                edits
             };
             const filePath = path.join(historyDir, `${id}.json`);
             fs.writeFileSync(filePath, JSON.stringify(data, null, 2), 'utf-8');
@@ -1738,6 +2833,10 @@ class DevMindDatabase {
             console.warn('⚠️ SQLite warning: Failed to write history JSON to disk:', err);
         }
     }
+    /** The configured developer identity (`.devmind/.env`'s DEVELOPER_NAME), or null if unset. */
+    getDeveloperName() {
+        return this.context?.developer?.name || null;
+    }
     toRepoRelativePath(absolutePath) {
         if (!absolutePath || !this.context)
             return absolutePath;
@@ -1841,8 +2940,16 @@ class DevMindDatabase {
         // Fallback: resolve relative to workspace root
         return (0, config_1.canonicalizePath)(this.clampToRoot(workspaceRoot, path.resolve(workspaceRoot, repoRelativePath)));
     }
-    syncFromDisk() {
+    syncFromDisk(onProgress) {
         this.db.pragma('foreign_keys = OFF');
+        // Captured BEFORE the walk starts (not after it finishes) so a file touched WHILE this sync
+        // is running is still mtime >= this timestamp and gets correctly picked up on the NEXT sync,
+        // rather than silently missed by a checkpoint that raced ahead of it.
+        const syncStartedAtMs = Date.now();
+        // 0 (never synced before, or system_meta was wiped by resetAll) disables every mtime skip
+        // below unconditionally — every real file's mtimeMs is a large positive epoch value, so
+        // `mtimeMs < 0` is never true and the first pass always processes everything, as before.
+        const lastSyncedAtMs = Number(this.getSystemMeta('last_sync_checkpoint_ms')) || 0;
         try {
             const workspaceRoot = path.dirname(this.dbPath);
             // 0. Auto-heal any legacy relative path records in SQLite.
@@ -1879,21 +2986,37 @@ class DevMindDatabase {
                 if (files.length > 0) {
                     const checkHistoryStmt = this.db.prepare('SELECT id FROM history WHERE id = ?');
                     const checkNodeStmt = this.db.prepare('SELECT id FROM nodes WHERE id = ?');
-                    const insertNodeStmt = this.db.prepare(`
-            INSERT INTO nodes (id, type, name, file_path, signature, deprecated)
-            VALUES (?, ?, ?, ?, ?, 0)
+                    const insertNodeStmt = this.db.prepare(`
+            INSERT INTO nodes (id, type, name, file_path, signature, deprecated)
+            VALUES (?, ?, ?, ?, ?, 0)
           `);
-                    const insertHistoryStmt = this.db.prepare(`
-            INSERT INTO history (id, node_id, session_id, created_at, updated_at, code_snapshot, reasoning)
-            VALUES (?, ?, ?, ?, ?, '', ?)
+                    const insertHistoryStmt = this.db.prepare(`
+            INSERT INTO history (id, node_id, session_id, created_at, updated_at, code_snapshot, reasoning)
+            VALUES (?, ?, ?, ?, ?, '', ?)
           `);
                     const syncHistoryTx = this.db.transaction(() => {
+                        let done = 0;
                         for (const file of files) {
+                            done++;
+                            if (onProgress && DevMindDatabase.shouldReport(done, files.length))
+                                onProgress('history', done, files.length);
                             try {
+                                // History files are immutable and always named `${id}.json` (every writer uses
+                                // this convention — see updateHistory/eraseLastEdit/etc.), so an already-synced
+                                // file can be identified from its NAME alone, before ever reading it. This is the
+                                // single biggest cost in syncFromDisk: history/ grows one file per edit forever,
+                                // so on a mature repo the overwhelming majority of files here are already synced
+                                // and this check-before-read turns them from a readFileSync+JSON.parse into a
+                                // single indexed SQLite lookup.
+                                if (checkHistoryStmt.get(file.slice(0, -'.json'.length)))
+                                    continue;
                                 const filePath = path.join(historyDir, file);
                                 const data = JSON.parse(fs.readFileSync(filePath, 'utf-8'));
                                 if (!data.id || !data.node_id)
                                     continue;
+                                // Kept as a safety net in case a hand-written/legacy file's internal `id` ever
+                                // differs from its filename — the fast path above is an optimization, not a
+                                // replacement for this correctness check.
                                 if (checkHistoryStmt.get(data.id))
                                     continue;
                                 if (!checkNodeStmt.get(data.node_id) && data.node_metadata) {
@@ -1933,18 +3056,31 @@ class DevMindDatabase {
                 if (jsonFiles.length > 0) {
                     const deleteNodesForFileStmt = this.db.prepare('DELETE FROM nodes WHERE file_path = ?');
                     const deleteConnsForNodesStmt = this.db.prepare('DELETE FROM node_connections WHERE source_node_id = ?');
-                    const insertNodeStmt = this.db.prepare(`
-            INSERT OR REPLACE INTO nodes (id, type, name, file_path, signature, deprecated)
-            VALUES (?, ?, ?, ?, ?, ?)
+                    const insertNodeStmt = this.db.prepare(`
+            INSERT OR REPLACE INTO nodes (id, type, name, file_path, signature, description, aliases, deprecated)
+            VALUES (?, ?, ?, ?, ?, ?, ?, ?)
           `);
-                    const insertConnStmt = this.db.prepare(`
-            INSERT OR IGNORE INTO node_connections (source_node_id, target_node_id)
-            VALUES (?, ?)
+                    const insertConnStmt = this.db.prepare(`
+            INSERT OR IGNORE INTO node_connections (source_node_id, target_node_id)
+            VALUES (?, ?)
           `);
                     // Transaction for fast batch syncing
                     const syncGraphTx = this.db.transaction(() => {
+                        let done = 0;
                         for (const file of jsonFiles) {
+                            done++;
+                            if (onProgress && DevMindDatabase.shouldReport(done, jsonFiles.length))
+                                onProgress('graph', done, jsonFiles.length);
                             try {
+                                // Unlike history/, a graph JSON gets REWRITTEN in place on every edit to its
+                                // source file (same path, new content) — so identity alone can't tell us "already
+                                // synced." mtime can: if this file hasn't changed since the last successful sync,
+                                // the DB is already current for it (a `git pull`/checkout always bumps the mtime
+                                // of every file it actually changed, so this correctly still processes exactly
+                                // what came in on a pull). Skips the readFileSync+JSON.parse+delete+reinsert
+                                // entirely for the — typically overwhelming — majority of untouched files.
+                                if (fs.statSync(file).mtimeMs < lastSyncedAtMs)
+                                    continue;
                                 const data = JSON.parse(fs.readFileSync(file, 'utf-8'));
                                 if (!data.file_path)
                                     continue;
@@ -1960,7 +3096,8 @@ class DevMindDatabase {
                                 const nodes = data.nodes || [];
                                 for (const n of nodes) {
                                     deleteConnsForNodesStmt.run(n.id);
-                                    insertNodeStmt.run(n.id, n.type, n.name, fileAbsPath, n.signature || null, n.deprecated ? 1 : 0);
+                                    const aliasesJson = Array.isArray(n.aliases) ? JSON.stringify(n.aliases) : '[]';
+                                    insertNodeStmt.run(n.id, n.type, n.name, fileAbsPath, n.signature || null, n.description || null, aliasesJson, n.deprecated ? 1 : 0);
                                 }
                                 // Insert connections
                                 const connections = data.connections || [];
@@ -1976,29 +3113,99 @@ class DevMindDatabase {
                     syncGraphTx();
                 }
             }
+            // 2.5. Sync Vector JSONs — MUST run after the graph pass (2), since that pass just
+            // deleted and re-inserted the current node set; vectors are reconciled against nodes as
+            // they now stand, not as they stood before this sync. Model-mismatched vectors (a
+            // teammate on a different devsmind version) are ignored on import, never partially
+            // trusted — those nodes simply fall back into the local `devsmind embed` queue. Then an
+            // explicit orphan sweep, since node_vectors has no FK (this whole method runs with
+            // foreign_keys=OFF, so a cascade would silently never fire during the graph pass anyway).
+            const vectorsDir = path.join(workspaceRoot, 'vectors');
+            if (fs.existsSync(vectorsDir)) {
+                const walkSyncVec = (dir, fileList = []) => {
+                    const files = fs.readdirSync(dir);
+                    for (const file of files) {
+                        const filePath = path.join(dir, file);
+                        if (fs.statSync(filePath).isDirectory()) {
+                            walkSyncVec(filePath, fileList);
+                        }
+                        else if (file.endsWith('.json')) {
+                            fileList.push(filePath);
+                        }
+                    }
+                    return fileList;
+                };
+                const vectorJsonFiles = walkSyncVec(vectorsDir);
+                if (vectorJsonFiles.length > 0) {
+                    const insertVectorStmt = this.db.prepare(`
+            INSERT OR REPLACE INTO node_vectors (node_id, model_id, dim, description_hash, vector)
+            VALUES (?, ?, ?, ?, ?)
+          `);
+                    const syncVectorsTx = this.db.transaction(() => {
+                        let done = 0;
+                        for (const file of vectorJsonFiles) {
+                            done++;
+                            if (onProgress && DevMindDatabase.shouldReport(done, vectorJsonFiles.length))
+                                onProgress('vectors', done, vectorJsonFiles.length);
+                            try {
+                                // Same reasoning as the graph pass above: a vectors/*.json is rewritten in place
+                                // whenever its node's vector changes, so mtime — not identity — is what tells us
+                                // whether the DB is still current for this file.
+                                if (fs.statSync(file).mtimeMs < lastSyncedAtMs)
+                                    continue;
+                                const data = JSON.parse(fs.readFileSync(file, 'utf-8'));
+                                if (!data.model_id || data.model_id !== embedder_1.EMBEDDING_MODEL_ID)
+                                    continue;
+                                const vectors = data.vectors || {};
+                                for (const nodeId of Object.keys(vectors)) {
+                                    const entry = vectors[nodeId];
+                                    if (!entry || !entry.v || !entry.h)
+                                        continue;
+                                    const buf = Buffer.from(entry.v, 'base64');
+                                    insertVectorStmt.run(nodeId, embedder_1.EMBEDDING_MODEL_ID, data.dim || embedder_1.EMBEDDING_DIM, entry.h, buf);
+                                }
+                            }
+                            catch (err) {
+                                // ignore malformed vectors JSON
+                            }
+                        }
+                    });
+                    syncVectorsTx();
+                }
+            }
+            // Orphan sweep — always runs, even with no vectors/ dir, to catch nodes deleted or
+            // renamed by the graph pass above that still had a (now-dangling) vector row.
+            this.db.exec('DELETE FROM node_vectors WHERE node_id NOT IN (SELECT id FROM nodes)');
             // 3. Sync Workflow JSONs
             const workflowsDir = this.workflowsDir();
             if (fs.existsSync(workflowsDir)) {
-                const upsertWorkflow = this.db.prepare(`
-          INSERT INTO workflows (id, name, description, status, created_at, updated_at)
-          VALUES (?, ?, ?, ?, ?, ?)
-          ON CONFLICT(id) DO UPDATE SET
-            name = excluded.name,
-            description = excluded.description,
-            status = excluded.status,
-            updated_at = excluded.updated_at
+                const upsertWorkflow = this.db.prepare(`
+          INSERT INTO workflows (id, name, description, archived, created_at, updated_at)
+          VALUES (?, ?, ?, ?, ?, ?)
+          ON CONFLICT(id) DO UPDATE SET
+            name = excluded.name,
+            description = excluded.description,
+            archived = excluded.archived,
+            updated_at = excluded.updated_at
         `);
-                const upsertStep = this.db.prepare(`
-          INSERT OR IGNORE INTO workflow_steps (id, workflow_id, step_index, summary, pending_tasks, history_ids, session_id, created_at)
-          VALUES (?, ?, ?, ?, ?, ?, ?, ?)
+                // DO UPDATE, not INSERT OR IGNORE. A teammate who already has a step row would otherwise
+                // never pick up `reasoning`/`node_ids`/`doc_paths` from a newer workflow.json — the row
+                // exists, so the insert is ignored, and their brain stays permanently half-migrated with
+                // no sign anything went wrong. Safe to overwrite because steps are append-only: nothing
+                // edits one locally after it is written, so incoming disk state is always authoritative.
+                const upsertStep = this.db.prepare(`
+          INSERT INTO workflow_steps (id, workflow_id, step_index, summary, reasoning, node_ids, doc_paths, session_id, created_at)
+          VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?)
+          ON CONFLICT(id) DO UPDATE SET
+            summary = excluded.summary,
+            reasoning = COALESCE(excluded.reasoning, workflow_steps.reasoning),
+            node_ids = COALESCE(excluded.node_ids, workflow_steps.node_ids),
+            doc_paths = COALESCE(excluded.doc_paths, workflow_steps.doc_paths)
         `);
-                const upsertArtifact = this.db.prepare(`
-          INSERT OR IGNORE INTO workflow_artifacts (id, workflow_id, step_id, type, source_name, file_path, created_at)
-          VALUES (?, ?, ?, ?, ?, ?, ?)
+                const upsertArtifact = this.db.prepare(`
+          INSERT OR IGNORE INTO workflow_artifacts (id, workflow_id, step_id, type, source_name, file_path, created_at)
+          VALUES (?, ?, ?, ?, ?, ?, ?)
         `);
-                // Track which workflow.json has is_active:true with the latest updated_at
-                let bestActiveId = null;
-                let bestActiveUpdatedAt = '';
                 const syncWorkflowsTx = this.db.transaction(() => {
                     const subdirs = fs.readdirSync(workflowsDir);
                     for (const subdir of subdirs) {
@@ -2009,32 +3216,48 @@ class DevMindDatabase {
                             const data = JSON.parse(fs.readFileSync(jsonPath, 'utf-8'));
                             if (!data.id || !data.name)
                                 continue;
-                            upsertWorkflow.run(data.id, data.name, data.description || '', data.status || 'paused', data.created_at || new Date().toISOString(), data.updated_at || new Date().toISOString());
+                            // The sidecar is authoritative for the fields v1 has no place for. It matters most
+                            // in exactly the case that looks fine: a teammate on an older build rewrote
+                            // workflow.json from their own columns, so it came back v1-shaped — but they could
+                            // not touch v2.json, so `archived` and every step's reasoning/node_ids/doc_paths are
+                            // still here to merge back on top.
+                            let sidecar = {};
+                            try {
+                                const sidecarPath = path.join(workflowsDir, subdir, exports.WORKFLOW_SIDECAR_FILE);
+                                if (fs.existsSync(sidecarPath))
+                                    sidecar = JSON.parse(fs.readFileSync(sidecarPath, 'utf-8')) || {};
+                            }
+                            catch { /* a corrupt sidecar degrades to the v1 shape rather than losing the workflow */ }
+                            upsertWorkflow.run(data.id, data.name, data.description || '', (sidecar.archived ?? data.archived) ? 1 : 0, data.created_at || new Date().toISOString(), data.updated_at || new Date().toISOString());
                             for (const s of (data.steps || [])) {
                                 if (!s.id)
                                     continue;
-                                upsertStep.run(s.id, data.id, s.step_index, s.summary || '', s.pending_tasks || null, s.history_ids || null, s.session_id || null, s.created_at || new Date().toISOString());
+                                // A genuinely v1 workflow (never written by this version) has no sidecar entry
+                                // either — those fields land as NULL and the step shows its summary alone, which
+                                // is all v1 ever stored. Nothing is lost that the old format ever held.
+                                const extra = sidecar.steps?.[s.id] || {};
+                                upsertStep.run(s.id, data.id, s.step_index, s.summary || '', extra.reasoning ?? s.reasoning ?? null, extra.node_ids ?? s.node_ids ?? null, extra.doc_paths ?? s.doc_paths ?? null, s.session_id || null, s.created_at || new Date().toISOString());
                             }
                             for (const a of (data.artifact_index || [])) {
                                 if (!a.id)
                                     continue;
                                 upsertArtifact.run(a.id, data.id, a.step_id || null, a.type || 'unknown', a.source_name || '', a.file_path || '', a.created_at || new Date().toISOString());
                             }
-                            // Track which workflow declared itself active most recently
-                            if (data.is_active && data.updated_at > bestActiveUpdatedAt) {
-                                bestActiveId = data.id;
-                                bestActiveUpdatedAt = data.updated_at;
-                            }
+                            // A v1 JSON's `is_active` is deliberately ignored. That flag is exactly how one
+                            // developer's "currently working on" state used to travel through git and take over
+                            // everyone else's — which workflow you are on is now local to your session and never
+                            // synced.
                         }
                         catch { /* skip malformed */ }
                     }
                 });
                 syncWorkflowsTx();
-                // Restore active_workflow_id if not already set and a JSON claims active status
-                if (bestActiveId && !this.getSystemMeta('active_workflow_id')) {
-                    this.setSystemMeta('active_workflow_id', bestActiveId);
-                }
             }
+            // Only advance the checkpoint after every step above completed without throwing — an
+            // exception anywhere earlier jumps straight to `catch` below, so this line is never
+            // reached, and the NEXT sync correctly retries a full pass from the old (or absent)
+            // checkpoint instead of wrongly believing a failed run succeeded.
+            this.setSystemMeta('last_sync_checkpoint_ms', String(syncStartedAtMs));
         }
         catch (err) {
             console.warn('⚠️ SQLite warning: Failed to sync from disk:', err);
@@ -2069,16 +3292,16 @@ class DevMindDatabase {
             const absEsc = this.likeEscape(absPath);
             const absLower = absPath.toLowerCase();
             const absEscLower = absEsc.toLowerCase();
-            const stmtNodes = this.db.prepare(`
-        SELECT * FROM nodes
-        WHERE (
-          LOWER(file_path) = ? OR
-          LOWER(file_path) LIKE ? ESCAPE '\\' OR
-          LOWER(file_path) LIKE ? ESCAPE '\\' OR
-          LOWER(file_path) LIKE ? ESCAPE '\\'
-        )
+            const stmtNodes = this.db.prepare(`
+        SELECT * FROM nodes
+        WHERE (
+          LOWER(file_path) = ? OR
+          LOWER(file_path) LIKE ? ESCAPE '\\' OR
+          LOWER(file_path) LIKE ? ESCAPE '\\' OR
+          LOWER(file_path) LIKE ? ESCAPE '\\'
+        )
       `);
-            const nodes = stmtNodes.all(absLower, `${absEscLower}, %`, `%, ${absEscLower}`, `%, ${absEscLower}, %`);
+            const nodes = DevMindDatabase.parseNodeRows(stmtNodes.all(absLower, `${absEscLower}, %`, `%, ${absEscLower}`, `%, ${absEscLower}, %`));
             if (nodes.length === 0) {
                 // If no nodes left, delete the JSON file if it exists
                 if (fs.existsSync(graphJsonPath)) {
@@ -2090,9 +3313,9 @@ class DevMindDatabase {
             const nodeIds = nodes.map(n => n.id);
             const connections = [];
             if (nodeIds.length > 0) {
-                const stmtConn = this.db.prepare(`
-          SELECT * FROM node_connections
-          WHERE source_node_id = ?
+                const stmtConn = this.db.prepare(`
+          SELECT * FROM node_connections
+          WHERE source_node_id = ?
         `);
                 for (const id of nodeIds) {
                     const conns = stmtConn.all(id);
@@ -2107,6 +3330,8 @@ class DevMindDatabase {
                     name: n.name,
                     type: n.type,
                     signature: n.signature,
+                    description: n.description || undefined,
+                    aliases: n.aliases.length > 0 ? n.aliases : undefined,
                     deprecated: n.deprecated ? 1 : 0
                 })),
                 connections: connections.map(c => ({
@@ -2121,6 +3346,53 @@ class DevMindDatabase {
             console.warn('⚠️ SQLite warning: Failed to write graph JSON to disk:', err);
         }
     }
+    /**
+     * Mirrors `writeGraphToDisk` exactly (same file-matching logic, same directory shape) but into
+     * a separate `vectors/` tree rather than inside `graph/*.json` — deliberately, so opaque base64
+     * blobs never pollute the human-readable, merge-friendly graph JSON. Deprecated nodes are
+     * skipped here (unlike the graph, which keeps them): `searchNodes` never queries a deprecated
+     * node's vector, so writing one is pure dead weight.
+     */
+    writeVectorsToDisk(filePath) {
+        try {
+            if (!filePath)
+                return;
+            const workspaceRoot = (0, config_1.canonicalizePath)(path.dirname(this.dbPath));
+            const absPath = (0, config_1.canonicalizePath)(filePath);
+            const repoRelPath = this.toRepoRelativePath(absPath);
+            const diskRelPath = repoRelPath.replace(/^\{([^}]+)\}/, '$1').replace(/\.[^/.]+$/, '.json');
+            const vectorsJsonPath = path.join(workspaceRoot, 'vectors', diskRelPath);
+            const absLower = absPath.toLowerCase();
+            const absEscLower = this.likeEscape(absPath).toLowerCase();
+            const stmt = this.db.prepare(`
+        SELECT nv.node_id AS node_id, nv.description_hash AS description_hash, nv.vector AS vector
+        FROM node_vectors nv
+        JOIN nodes n ON n.id = nv.node_id
+        WHERE n.deprecated = 0 AND (
+          LOWER(n.file_path) = ? OR
+          LOWER(n.file_path) LIKE ? ESCAPE '\\' OR
+          LOWER(n.file_path) LIKE ? ESCAPE '\\' OR
+          LOWER(n.file_path) LIKE ? ESCAPE '\\'
+        ) AND nv.model_id = ?
+      `);
+            const rows = stmt.all(absLower, `${absEscLower}, %`, `%, ${absEscLower}`, `%, ${absEscLower}, %`, embedder_1.EMBEDDING_MODEL_ID);
+            if (rows.length === 0) {
+                if (fs.existsSync(vectorsJsonPath))
+                    fs.unlinkSync(vectorsJsonPath);
+                return;
+            }
+            const vectors = {};
+            for (const r of rows) {
+                vectors[r.node_id] = { h: r.description_hash, v: Buffer.from(r.vector).toString('base64') };
+            }
+            const data = { file_path: repoRelPath, model_id: embedder_1.EMBEDDING_MODEL_ID, dim: embedder_1.EMBEDDING_DIM, vectors };
+            fs.mkdirSync(path.dirname(vectorsJsonPath), { recursive: true });
+            fs.writeFileSync(vectorsJsonPath, JSON.stringify(data, null, 2), 'utf-8');
+        }
+        catch (err) {
+            console.warn('⚠️ SQLite warning: Failed to write vectors JSON to disk:', err);
+        }
+    }
     /** Force-syncs all database nodes and workflows to disk JSON files. */
     syncToDisk() {
         try {