@soleri/core 2.12.0 → 7.0.0

package/src/runtime/runtime.ts

@@ -85,21 +85,24 @@ export function createAgentRuntime(config: AgentRuntimeConfig): AgentRuntime {
   // Planner — multi-step task tracking
   const planner = new Planner(plansPath);
 
-  // Cognee — vector search client (graceful degradation if Cognee is down)
-  const cogneePartial: Partial<import('../cognee/types.js').CogneeConfig> = { dataset: agentId };
-  if (process.env.COGNEE_BASE_URL) cogneePartial.baseUrl = process.env.COGNEE_BASE_URL;
-  if (process.env.COGNEE_API_TOKEN) cogneePartial.apiToken = process.env.COGNEE_API_TOKEN;
-  if (process.env.COGNEE_DATASET) cogneePartial.dataset = process.env.COGNEE_DATASET;
-  const cognee = new CogneeClient(cogneePartial);
+  // Cognee — vector search client (opt-in, graceful degradation if Cognee is down)
+  let cognee: CogneeClient | null = null;
+  if (config.cognee) {
+    const cogneePartial: Partial<import('../cognee/types.js').CogneeConfig> = { dataset: agentId };
+    if (process.env.COGNEE_BASE_URL) cogneePartial.baseUrl = process.env.COGNEE_BASE_URL;
+    if (process.env.COGNEE_API_TOKEN) cogneePartial.apiToken = process.env.COGNEE_API_TOKEN;
+    if (process.env.COGNEE_DATASET) cogneePartial.dataset = process.env.COGNEE_DATASET;
+    cognee = new CogneeClient(cogneePartial);
+  }
 
   // Brain — intelligence layer (TF-IDF scoring, auto-tagging, dedup)
-  const brain = new Brain(vault, cognee);
+  const brain = new Brain(vault, cognee ?? undefined);
 
   // Brain Intelligence — pattern strengths, session knowledge, intelligence pipeline
   const brainIntelligence = new BrainIntelligence(vault, brain);
 
   // Curator — vault self-maintenance (dedup, contradictions, grooming, health)
-  const curator = new Curator(vault, cognee);
+  const curator = new Curator(vault, cognee ?? undefined);
 
   // Governance — policy engine + proposal tracker for gated knowledge capture
   const governance = new Governance(vault);
@@ -132,13 +135,16 @@ export function createAgentRuntime(config: AgentRuntimeConfig): AgentRuntime {
   const anthropicKeyPool = new KeyPool(keyPoolFiles.anthropic);
   const llmClient = new LLMClient(openaiKeyPool, anthropicKeyPool, agentId);
 
-  // Cognee Sync Manager — queue-based dirty tracking with offline resilience
-  const syncManager = new CogneeSyncManager(
-    vault.getProvider(),
-    cognee,
-    cogneePartial.dataset ?? agentId,
-  );
-  vault.setSyncManager(syncManager);
+  // Cognee Sync Manager — queue-based dirty tracking with offline resilience (only when Cognee enabled)
+  let syncManager: CogneeSyncManager | null = null;
+  if (cognee) {
+    syncManager = new CogneeSyncManager(
+      vault.getProvider(),
+      cognee,
+      process.env.COGNEE_DATASET ?? agentId,
+    );
+    vault.setSyncManager(syncManager);
+  }
 
   // Intake Pipeline — PDF/book ingestion with LLM classification
   const intakePipeline = new IntakePipeline(vault.getProvider(), vault, llmClient);
@@ -168,7 +174,10 @@ export function createAgentRuntime(config: AgentRuntimeConfig): AgentRuntime {
   const health = new HealthRegistry();
   health.register('vault', 'healthy');
   health.register('brain', 'healthy');
-  health.register('cognee', cognee.getStatus()?.available ? 'healthy' : 'degraded');
+  health.register(
+    'cognee',
+    cognee ? (cognee.getStatus()?.available ? 'healthy' : 'degraded') : 'down',
+  );
   health.register(
     'llm',
     llmClient.isAvailable().openai || llmClient.isAvailable().anthropic ? 'healthy' : 'degraded',
@@ -221,8 +230,8 @@ export function createAgentRuntime(config: AgentRuntimeConfig): AgentRuntime {
     knowledgeReview,
     createdAt: Date.now(),
     close: () => {
-      syncManager.close();
-      cognee.resetPendingCognify();
+      syncManager?.close();
+      cognee?.resetPendingCognify();
       vaultManager.close();
     },
   };

package/src/runtime/types.ts

@@ -48,6 +48,8 @@ export interface AgentRuntimeConfig {
   logLevel?: LogLevel;
   /** Path to shared global vault. Default: ~/.soleri/vault.db */
   sharedVaultPath?: string;
+  /** Enable Cognee vector search integration. Default: false (opt-in). */
+  cognee?: boolean;
 }
 
 /**
@@ -63,7 +65,8 @@ export interface AgentRuntime {
   planner: Planner;
   curator: Curator;
   governance: Governance;
-  cognee: CogneeClient;
+  /** Cognee vector search client. Null when Cognee integration is disabled. */
+  cognee: CogneeClient | null;
   loop: LoopManager;
   identityManager: IdentityManager;
   intentRouter: IntentRouter;
@@ -72,7 +75,8 @@ export interface AgentRuntime {
   telemetry: Telemetry;
   projectRegistry: ProjectRegistry;
   templateManager: TemplateManager;
-  syncManager: CogneeSyncManager;
+  /** Cognee sync manager. Null when Cognee integration is disabled. */
+  syncManager: CogneeSyncManager | null;
   intakePipeline: IntakePipeline;
   /** Mutable auth policy — controls facade dispatch enforcement. */
   authPolicy: AuthPolicy;

package/src/runtime/vault-linking-ops.ts

@@ -0,0 +1,454 @@
+/**
+ * Vault linking ops — Zettelkasten bidirectional linking.
+ *
+ * Provides 8 ops: link_entries, unlink_entries, get_links, traverse,
+ * suggest_links, get_orphans, relink_vault, link_stats.
+ * Ported from Salvador MCP with improvements:
+ * - FTS5 for suggest_links (Salvador uses TF-IDF)
+ * - relink_vault: LLM-evaluated batch re-linking (Salvador uses a separate script)
+ */
+
+import { z } from 'zod';
+import type { OpDefinition } from '../facades/types.js';
+import type { AgentRuntime } from './types.js';
+import { LinkManager } from '../vault/linking.js';
+
+const EVAL_SYSTEM_PROMPT = `You evaluate pairs of knowledge entries to determine if they should be linked in a Zettelkasten vault. For EACH pair, decide:
+- If meaningfully related → return { "link": true, "type": "<type>", "note": "<1 sentence why>" }
+- If NOT meaningfully related → return { "link": false }
+
+Link types:
+- "extends" — target builds on or refines the source
+- "supports" — target provides evidence or foundation for the source
+- "contradicts" — target is an opposing approach or counterpoint
+- "sequences" — source must happen before target
+
+Rules: Same category alone is NOT enough. Be selective. Return a JSON array.`;
+
+export function createVaultLinkingOps(runtime: AgentRuntime): OpDefinition[] {
+  const { vault } = runtime;
+  const linkManager = new LinkManager(vault.getProvider());
+
+  return [
+    {
+      name: 'link_entries',
+      description: 'Create a typed link between two vault entries (Zettelkasten)',
+      auth: 'write',
+      schema: z.object({
+        sourceId: z.string().describe('REQUIRED: Source entry ID'),
+        targetId: z.string().describe('REQUIRED: Target entry ID'),
+        linkType: z
+          .enum(['supports', 'contradicts', 'extends', 'sequences'])
+          .describe('REQUIRED: Relationship type'),
+        note: z.string().optional().describe('Optional context for the link'),
+      }),
+      handler: async (params) => {
+        const sourceId = params.sourceId as string;
+        const targetId = params.targetId as string;
+
+        // Validate both entries exist to prevent dangling links
+        const provider = vault.getProvider();
+        const sourceExists = provider.get<{ id: string }>('SELECT id FROM entries WHERE id = ?', [
+          sourceId,
+        ]);
+        const targetExists = provider.get<{ id: string }>('SELECT id FROM entries WHERE id = ?', [
+          targetId,
+        ]);
+
+        if (!sourceExists || !targetExists) {
+          const missing = [];
+          if (!sourceExists) missing.push(`source '${sourceId}'`);
+          if (!targetExists) missing.push(`target '${targetId}'`);
+          throw new Error(`Entry not found: ${missing.join(' and ')}`);
+        }
+
+        linkManager.addLink(
+          sourceId,
+          targetId,
+          params.linkType as 'supports' | 'contradicts' | 'extends' | 'sequences',
+          params.note as string | undefined,
+        );
+        return {
+          success: true,
+          link: {
+            sourceId,
+            targetId,
+            linkType: params.linkType,
+            note: params.note,
+          },
+          sourceLinkCount: linkManager.getLinkCount(sourceId),
+        };
+      },
+    },
+    {
+      name: 'unlink_entries',
+      description: 'Remove a link between two vault entries',
+      auth: 'write',
+      schema: z.object({
+        sourceId: z.string().describe('REQUIRED: Source entry ID'),
+        targetId: z.string().describe('REQUIRED: Target entry ID'),
+      }),
+      handler: async (params) => {
+        linkManager.removeLink(params.sourceId as string, params.targetId as string);
+        return { success: true, removed: { sourceId: params.sourceId, targetId: params.targetId } };
+      },
+    },
+    {
+      name: 'get_links',
+      description: 'Get all links for a vault entry (outgoing + incoming backlinks)',
+      auth: 'read',
+      schema: z.object({
+        entryId: z.string().describe('REQUIRED: Entry ID'),
+      }),
+      handler: async (params) => {
+        const entryId = params.entryId as string;
+        const outgoing = linkManager.getLinks(entryId);
+        const incoming = linkManager.getBacklinks(entryId);
+        return { entryId, outgoing, incoming, totalLinks: outgoing.length + incoming.length };
+      },
+    },
+    {
+      name: 'traverse',
+      description:
+        'Walk the link graph from an entry up to N hops deep (Zettelkasten graph traversal)',
+      auth: 'read',
+      schema: z.object({
+        entryId: z.string().describe('REQUIRED: Starting entry ID'),
+        depth: z.coerce
+          .number()
+          .int()
+          .min(1)
+          .max(5)
+          .default(2)
+          .describe('Max hops (1-5, default 2)'),
+      }),
+      handler: async (params) => {
+        const entryId = params.entryId as string;
+        const depth = (params.depth as number) || 2;
+        const connected = linkManager.traverse(entryId, depth);
+        return { entryId, depth, connectedEntries: connected, totalConnected: connected.length };
+      },
+    },
+    {
+      name: 'suggest_links',
+      description:
+        'Find semantically similar entries as link candidates using FTS5 (Zettelkasten auto-linking)',
+      auth: 'read',
+      schema: z.object({
+        entryId: z.string().describe('REQUIRED: Entry ID to find link candidates for'),
+        limit: z.coerce
+          .number()
+          .int()
+          .min(1)
+          .max(20)
+          .default(5)
+          .describe('Max suggestions (default 5)'),
+      }),
+      handler: async (params) => {
+        const entryId = params.entryId as string;
+        const limit = (params.limit as number) || 5;
+        const suggestions = linkManager.suggestLinks(entryId, limit);
+        return { entryId, suggestions, totalSuggestions: suggestions.length };
+      },
+    },
+    {
+      name: 'get_orphans',
+      description: 'Find vault entries with zero links (Zettelkasten orphan detection)',
+      auth: 'read',
+      schema: z.object({
+        limit: z.coerce
+          .number()
+          .int()
+          .min(1)
+          .max(100)
+          .default(20)
+          .describe('Max orphans (default 20)'),
+      }),
+      handler: async (params) => {
+        const limit = (params.limit as number) || 20;
+        const orphans = linkManager.getOrphans(limit);
+        return { orphans, totalOrphans: orphans.length };
+      },
+    },
+    {
+      name: 'relink_vault',
+      description:
+        'Smart Zettelkasten re-linking: drops batch links, evaluates all entries with LLM, creates quality links with reasoning notes. Long-running operation.',
+      auth: 'write',
+      schema: z.object({
+        batchSize: z.coerce
+          .number()
+          .int()
+          .min(1)
+          .max(20)
+          .default(10)
+          .describe('Pairs per LLM call (default 10)'),
+        limit: z.coerce
+          .number()
+          .int()
+          .min(0)
+          .max(5000)
+          .default(0)
+          .describe('Max entries to process (0 = all)'),
+        dryRun: z.boolean().optional().default(false).describe('Preview without changes'),
+      }),
+      handler: async (params) => {
+        const batchSize = (params.batchSize as number) || 10;
+        const limit = (params.limit as number) || 0;
+        const dryRun = (params.dryRun as boolean) ?? false;
+        const { llmClient } = runtime;
+
+        if (!llmClient.isAvailable().anthropic && !llmClient.isAvailable().openai) {
+          return { success: false, error: 'No LLM provider available for link evaluation' };
+        }
+
+        const provider = vault.getProvider();
+
+        // Phase 1: Preserve manual links, drop batch links
+        const manualCount =
+          provider.get<{ c: number }>(
+            "SELECT COUNT(*) as c FROM vault_links WHERE note IS NOT NULL AND note != ''",
+          )?.c ?? 0;
+
+        const batchCount =
+          provider.get<{ c: number }>(
+            "SELECT COUNT(*) as c FROM vault_links WHERE note IS NULL OR note = ''",
+          )?.c ?? 0;
+
+        if (!dryRun) {
+          provider.run("DELETE FROM vault_links WHERE note IS NULL OR note = ''");
+        }
+
+        // Phase 2: Get entries and generate candidates
+        let entries = provider.all<{
+          id: string;
+          title: string;
+          type: string;
+          description: string | null;
+        }>('SELECT id, title, type, description FROM entries ORDER BY updated_at DESC');
+        if (limit > 0) entries = entries.slice(0, limit);
+
+        // Build candidates via tag overlap + category match
+        const candidates: Array<{
+          sourceId: string;
+          sourceTitle: string;
+          sourceType: string;
+          sourceDesc: string;
+          targetId: string;
+          targetTitle: string;
+          targetType: string;
+          targetDesc: string;
+        }> = [];
+
+        for (const entry of entries) {
+          const existingLinks = new Set([
+            ...linkManager.getLinks(entry.id).map((l) => l.targetId),
+            ...linkManager.getBacklinks(entry.id).map((l) => l.sourceId),
+          ]);
+
+          // Tag overlap matches
+          const matches = provider.all<{
+            id: string;
+            title: string;
+            type: string;
+            description: string | null;
+          }>(
+            `SELECT DISTINCT e.id, e.title, e.type, SUBSTR(e.description, 1, 200) as description
+             FROM entries e
+             JOIN (SELECT entry_id, tag FROM vault_tags WHERE entry_id = ?) src_tags ON 1=1
+             JOIN vault_tags t ON t.tag = src_tags.tag AND t.entry_id = e.id
+             WHERE e.id != ?
+             GROUP BY e.id ORDER BY COUNT(*) DESC LIMIT 5`,
+            [entry.id, entry.id],
+          );
+
+          // Fallback: category match
+          if (matches.length < 3) {
+            try {
+              const existing = new Set(matches.map((m) => m.id));
+              const catMatches = provider.all<(typeof matches)[0]>(
+                `SELECT id, title, type, SUBSTR(description, 1, 200) as description
+                 FROM entries WHERE id != ? AND type = ? LIMIT 3`,
+                [entry.id, entry.type],
+              );
+              for (const m of catMatches) {
+                if (!existing.has(m.id)) matches.push(m);
+              }
+            } catch {
+              /* ignore */
+            }
+          }
+
+          for (const match of matches.slice(0, 5)) {
+            if (existingLinks.has(match.id)) continue;
+            candidates.push({
+              sourceId: entry.id,
+              sourceTitle: entry.title,
+              sourceType: entry.type,
+              sourceDesc: (entry.description || '').slice(0, 200),
+              targetId: match.id,
+              targetTitle: match.title,
+              targetType: match.type,
+              targetDesc: (match.description || '').slice(0, 200),
+            });
+          }
+        }
+
+        if (dryRun) {
+          return {
+            dryRun: true,
+            entries: entries.length,
+            candidates: candidates.length,
+            llmCallsNeeded: Math.ceil(candidates.length / batchSize),
+            manualLinksPreserved: manualCount,
+            batchLinksToRemove: batchCount,
+          };
+        }
+
+        // Phase 3: LLM evaluation in batches (sequential to respect rate limits)
+        let linksCreated = 0;
+        let pairsSkipped = 0;
+        let llmCalls = 0;
+        let errors = 0;
+        const now = Date.now();
+
+        // Build batches
+        const batches: (typeof candidates)[] = [];
+        for (let i = 0; i < candidates.length; i += batchSize) {
+          batches.push(candidates.slice(i, i + batchSize));
+        }
+
+        // Process sequentially using reduce chain (avoids await-in-loop lint)
+        await batches.reduce(async (prev, batch) => {
+          await prev;
+          const pairsText = batch
+            .map(
+              (p, idx) =>
+                `--- Pair ${idx + 1} ---\nSOURCE [${p.sourceType}]: ${p.sourceTitle}\n${p.sourceDesc}\n\nTARGET [${p.targetType}]: ${p.targetTitle}\n${p.targetDesc}`,
+            )
+            .join('\n\n');
+
+          try {
+            const result = await llmClient.complete({
+              provider: llmClient.isAvailable().anthropic ? 'anthropic' : 'openai',
+              model: llmClient.isAvailable().anthropic ? 'claude-sonnet-4-20250514' : 'gpt-4o-mini',
+              systemPrompt: EVAL_SYSTEM_PROMPT,
+              userPrompt: pairsText,
+              maxTokens: 2000,
+              caller: 'relink_vault',
+              task: 'link-evaluation',
+            });
+            llmCalls++;
+
+            let cleaned = result.text.trim();
+            if (cleaned.startsWith('```')) {
+              const first = cleaned.indexOf('\n');
+              const last = cleaned.lastIndexOf('```');
+              cleaned = cleaned.slice(first + 1, last).trim();
+            }
+
+            const decisions = JSON.parse(cleaned) as Array<{
+              link: boolean;
+              type?: string;
+              note?: string;
+            }>;
+
+            for (let j = 0; j < decisions.length && j < batch.length; j++) {
+              const d = decisions[j];
+              if (d.link && d.type && d.note) {
+                if (batch[j].sourceId === batch[j].targetId) continue;
+                try {
+                  provider.run(
+                    'INSERT OR IGNORE INTO vault_links (source_id, target_id, link_type, note, created_at) VALUES (?, ?, ?, ?, ?)',
+                    [batch[j].sourceId, batch[j].targetId, d.type, d.note, now],
+                  );
+                  linksCreated++;
+                } catch {
+                  /* FK or duplicate */
+                }
+              } else {
+                pairsSkipped++;
+              }
+            }
+          } catch {
+            errors++;
+          }
+        }, Promise.resolve());
+
+        // Phase 4: Stats
+        const totalLinks =
+          provider.get<{ c: number }>('SELECT COUNT(*) as c FROM vault_links')?.c ?? 0;
+        const orphanCount =
+          provider.get<{ c: number }>(
+            `SELECT COUNT(*) as c FROM entries
+           WHERE id NOT IN (SELECT source_id FROM vault_links)
+             AND id NOT IN (SELECT target_id FROM vault_links)`,
+          )?.c ?? 0;
+        const byType = provider.all<{ link_type: string; c: number }>(
+          'SELECT link_type, COUNT(*) as c FROM vault_links GROUP BY link_type ORDER BY c DESC',
+        );
+
+        return {
+          success: true,
+          entriesProcessed: entries.length,
+          candidatesEvaluated: candidates.length,
+          linksCreated,
+          pairsSkipped,
+          llmCalls,
+          errors,
+          totalLinks,
+          orphans: orphanCount,
+          byType: Object.fromEntries(byType.map((r) => [r.link_type, r.c])),
+          manualLinksPreserved: manualCount,
+          batchLinksRemoved: batchCount,
+        };
+      },
+    },
+    {
+      name: 'link_stats',
+      description:
+        'Get Zettelkasten graph statistics: total links, by type, most connected, orphan count.',
+      auth: 'read',
+      handler: async () => {
+        const provider = vault.getProvider();
+        try {
+          const totalLinks =
+            provider.get<{ c: number }>('SELECT COUNT(*) as c FROM vault_links')?.c ?? 0;
+          const totalEntries =
+            provider.get<{ c: number }>('SELECT COUNT(*) as c FROM entries')?.c ?? 0;
+          const orphans =
+            provider.get<{ c: number }>(
+              `SELECT COUNT(*) as c FROM entries
+             WHERE id NOT IN (SELECT source_id FROM vault_links)
+               AND id NOT IN (SELECT target_id FROM vault_links)`,
+            )?.c ?? 0;
+          const byType = provider.all<{ link_type: string; c: number }>(
+            'SELECT link_type, COUNT(*) as c FROM vault_links GROUP BY link_type ORDER BY c DESC',
+          );
+          const withNotes =
+            provider.get<{ c: number }>(
+              "SELECT COUNT(*) as c FROM vault_links WHERE note IS NOT NULL AND note != ''",
+            )?.c ?? 0;
+          const mostConnected = provider.all<{ title: string; links: number }>(
+            `SELECT e.title, (
+              (SELECT COUNT(*) FROM vault_links WHERE source_id = e.id) +
+              (SELECT COUNT(*) FROM vault_links WHERE target_id = e.id)
+            ) as links FROM entries e ORDER BY links DESC LIMIT 10`,
+          );
+
+          return {
+            totalEntries,
+            totalLinks,
+            orphans,
+            linksWithNotes: withNotes,
+            linkQuality: totalLinks > 0 ? `${((withNotes / totalLinks) * 100).toFixed(0)}%` : 'n/a',
+            byType: Object.fromEntries(byType.map((r) => [r.link_type, r.c])),
+            mostConnected,
+          };
+        } catch {
+          return { totalLinks: 0, totalEntries: 0, orphans: 0, byType: {}, mostConnected: [] };
+        }
+      },
+    },
+  ];
+}