fix(schema): skip unsupported large-dim HNSW indexes (#1734)

Takeover of #2510: migrations v40 (facts) and v55 (query_cache)
unconditionally created HNSW indexes with the configured embedding
dimension, so `gbrain init` with embedding_dimensions above pgvector's
per-type HNSW caps (vector 2000 / halfvec 4000) failed with
"column cannot have more than 4000 dimensions for hnsw index".

- vector-index.ts: add PGVECTOR_HNSW_HALFVEC_MAX_DIMS + hnswMaxDimsForType
- migrate.ts v40/v55: emit the HNSW index only when dims fit the cap,
  otherwise a comment noting exact scans remain available
- embedding-dim-check.ts: buildFactsAlterRecipe skips the reindex step
  above the cap for the same reason
- tests: 4096d init round-trip on PGLite (columns exist, indexes
  skipped) + recipe-skip unit test

Drops the unrelated context-engine.ts interface change and the
tsconfig.json strictFunctionTypes=false hunk from #2510; typecheck is
clean without them.

Co-authored-by: javieraldape <javieraldape@users.noreply.github.com>
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

.github/workflows/test.yml

@@ -206,11 +206,7 @@ jobs:
     needs: cache-check
     if: needs.cache-check.outputs.hit != 'true'
     runs-on: ubuntu-latest
-    # 20 (was 15): shard 4 runs ~14.5 min on master (dream.test.ts ~29s/test
-    # dominates it) and hits the 15-min ceiling on slower runners, cancelling
-    # mid-run with 0 test failures. Rebalancing via
-    # scripts/mine-shard-weights.ts is the real fix; this stops the bleeding.
-    timeout-minutes: 20
+    timeout-minutes: 15
     strategy:
       fail-fast: false
       matrix:

src/core/embedding-dim-check.ts

@@ -14,7 +14,7 @@
  */
 
 import type { BrainEngine } from './engine.ts';
-import { PGVECTOR_HNSW_VECTOR_MAX_DIMS } from './vector-index.ts';
+import { PGVECTOR_HNSW_VECTOR_MAX_DIMS, hnswMaxDimsForType } from './vector-index.ts';
 import { gbrainPath } from './config.ts';
 import { resolveRecipe } from './ai/model-resolver.ts';
 import type { Recipe } from './ai/types.ts';
@@ -609,6 +609,17 @@ export function buildFactsAlterRecipe(
   const opclass = columnType === 'halfvec' ? 'halfvec_cosine_ops' : 'vector_cosine_ops';
   const targetType = columnType === 'halfvec' ? `halfvec(${configuredDims})` : `vector(${configuredDims})`;
   const dimsChanged = columnDims !== configuredDims;
+  const hnswMaxDims = hnswMaxDimsForType(columnType);
+  const indexLines = configuredDims <= hnswMaxDims
+    ? [
+        `CREATE INDEX idx_facts_embedding_hnsw`,
+        `  ON facts USING hnsw (embedding ${opclass})`,
+        `  WHERE embedding IS NOT NULL AND expired_at IS NULL;`,
+      ]
+    : [
+        `-- Skip reindex. ${columnType}(${configuredDims}) exceeds pgvector's HNSW cap of ${hnswMaxDims};`,
+        `-- fact similarity falls back to exact scans.`,
+      ];
   return [
     `-- ALTER ${columnType}(${columnDims}) → ${columnType}(${configuredDims}) on indexed column.`,
     `-- HOLD a maintenance window: this rewrites every row's embedding.`,
@@ -629,9 +640,7 @@ export function buildFactsAlterRecipe(
       : []),
     `ALTER TABLE facts ALTER COLUMN embedding TYPE ${targetType}`,
     `  USING embedding::${targetType};`,
-    `CREATE INDEX idx_facts_embedding_hnsw`,
-    `  ON facts USING hnsw (embedding ${opclass})`,
-    `  WHERE embedding IS NOT NULL AND expired_at IS NULL;`,
+    ...indexLines,
   ].join('\n');
 }
 

src/core/migrate.ts

@@ -1,6 +1,7 @@
 import type { BrainEngine } from './engine.ts';
 import { slugifyPath } from './sync.ts';
 import { getFtsLanguage } from './fts-language.ts';
+import { hnswMaxDimsForType } from './vector-index.ts';
 
 /**
  * Schema migrations — run automatically on initSchema().
@@ -2276,11 +2277,19 @@ export const MIGRATIONS: Migration[] = [
         useHalfvec = true;
       }
 
-      const vecType = useHalfvec ? 'HALFVEC' : 'VECTOR';
+      const columnType = useHalfvec ? 'halfvec' : 'vector';
+      const vecType = columnType.toUpperCase();
       // HNSW operator class must match the column type:
       //   VECTOR(n)  → vector_cosine_ops
       //   HALFVEC(n) → halfvec_cosine_ops
       const opclass = useHalfvec ? 'halfvec_cosine_ops' : 'vector_cosine_ops';
+      const hnswMaxDims = hnswMaxDimsForType(columnType);
+      const factsEmbeddingIndexSql = embeddingDim <= hnswMaxDims
+        ? `CREATE INDEX IF NOT EXISTS idx_facts_embedding_hnsw
+          ON facts USING hnsw (embedding ${opclass})
+          WHERE embedding IS NOT NULL AND expired_at IS NULL;`
+        : `-- idx_facts_embedding_hnsw skipped: pgvector HNSW ${columnType} indexes support
+        -- at most ${hnswMaxDims} dimensions; exact vector scans remain available.`;
       // FK to sources is added in a separate ALTER TABLE rather than inline
       // on the column. Inline `REFERENCES` worked on PGLite but silently
       // got dropped by postgres.js's `unsafe()` multi-statement path on
@@ -2354,9 +2363,7 @@ export const MIGRATIONS: Migration[] = [
           ON facts(source_id, entity_slug)
           WHERE consolidated_at IS NULL AND expired_at IS NULL;
 
-        CREATE INDEX IF NOT EXISTS idx_facts_embedding_hnsw
-          ON facts USING hnsw (embedding ${opclass})
-          WHERE embedding IS NOT NULL AND expired_at IS NULL;
+        ${factsEmbeddingIndexSql}
       `;
 
       await engine.runMigration(40, factsDDL);
@@ -2870,8 +2877,16 @@ export const MIGRATIONS: Migration[] = [
         useHalfvec = true;
       }
 
-      const vecType = useHalfvec ? 'HALFVEC' : 'VECTOR';
+      const columnType = useHalfvec ? 'halfvec' : 'vector';
+      const vecType = columnType.toUpperCase();
       const opclass = useHalfvec ? 'halfvec_cosine_ops' : 'vector_cosine_ops';
+      const hnswMaxDims = hnswMaxDimsForType(columnType);
+      const queryCacheEmbeddingIndexSql = embeddingDim <= hnswMaxDims
+        ? `CREATE INDEX IF NOT EXISTS idx_query_cache_embedding_hnsw
+          ON query_cache USING hnsw (embedding ${opclass})
+          WHERE embedding IS NOT NULL;`
+        : `-- idx_query_cache_embedding_hnsw skipped: pgvector HNSW ${columnType} indexes support
+        -- at most ${hnswMaxDims} dimensions; exact vector scans remain available.`;
 
       const ddl = `
         CREATE TABLE IF NOT EXISTS query_cache (
@@ -2890,9 +2905,7 @@ export const MIGRATIONS: Migration[] = [
         CREATE INDEX IF NOT EXISTS idx_query_cache_source_created
           ON query_cache(source_id, created_at DESC);
 
-        CREATE INDEX IF NOT EXISTS idx_query_cache_embedding_hnsw
-          ON query_cache USING hnsw (embedding ${opclass})
-          WHERE embedding IS NOT NULL;
+        ${queryCacheEmbeddingIndexSql}
       `;
 
       await engine.runMigration(55, ddl);

src/core/skill-manifest.ts

@@ -23,15 +23,6 @@
  * hold conventions and shared rule files, not skills. Files like
  * `_brain-filing-rules.md` live at the root and are not considered
  * skills by either loader.
- *
- * ClawHub-installed workspace skills (#1767): a skill dir carrying
- * `.clawhub/origin.json` is an externally-managed runtime integration
- * (e.g. an email or catalog skill), not a gbrain-routable skill. The
- * derive path SKIPS those so `gbrain doctor` resolver_health doesn't
- * hard-fail on them — UNLESS the skill's SKILL.md frontmatter declares
- * `triggers:`, which is the explicit opt-in to gbrain routing (and the
- * same surface that makes it reachable). An explicit manifest.json that
- * lists a ClawHub skill also keeps strict checking (verbatim path).
  */
 
 import { existsSync, readFileSync, readdirSync, statSync } from 'fs';
@@ -69,27 +60,9 @@ function parseSkillName(skillMdPath: string): string | null {
   }
 }
 
-/**
- * Does the SKILL.md frontmatter declare a `triggers:` key? A ClawHub-
- * installed skill that ships gbrain `triggers:` has explicitly opted in
- * to gbrain routing and gets full resolver checks (#1767).
- */
-function declaresTriggers(skillMdPath: string): boolean {
-  try {
-    const content = readFileSync(skillMdPath, 'utf-8');
-    const fmMatch = content.match(/^---\n([\s\S]*?)\n---/);
-    if (!fmMatch) return false;
-    return /^triggers:/m.test(fmMatch[1]);
-  } catch {
-    return false;
-  }
-}
-
 /**
  * Walk skillsDir, return every `<skillsDir>/<dir>/SKILL.md` as a
- * ManifestEntry. Dotfile and underscore-prefixed dirs are skipped, as
- * are ClawHub-installed external skills that haven't opted in to gbrain
- * routing via `triggers:` frontmatter (#1767).
+ * ManifestEntry. Dotfile and underscore-prefixed dirs are skipped.
  */
 function deriveManifest(skillsDir: string): ManifestEntry[] {
   const out: ManifestEntry[] = [];
@@ -120,12 +93,6 @@ function deriveManifest(skillsDir: string): ManifestEntry[] {
     const skillMd = join(subdirAbs, 'SKILL.md');
     if (!existsSync(skillMd)) continue;
 
-    // ClawHub-installed external skill (#1767): skip unless it opts in
-    // to gbrain routing by declaring `triggers:` in its frontmatter.
-    if (existsSync(join(subdirAbs, '.clawhub', 'origin.json')) && !declaresTriggers(skillMd)) {
-      continue;
-    }
-
     const frontmatterName = parseSkillName(skillMd);
     const name = frontmatterName && frontmatterName !== '' ? frontmatterName : entry;
     out.push({ name, path: `${entry}/SKILL.md` });

src/core/vector-index.ts

@@ -17,6 +17,7 @@
 import type { BrainEngine } from './engine.ts';
 
 export const PGVECTOR_HNSW_VECTOR_MAX_DIMS = 2000;
+export const PGVECTOR_HNSW_HALFVEC_MAX_DIMS = 4000;
 
 const CHUNK_EMBEDDING_HNSW_INDEX =
   'CREATE INDEX IF NOT EXISTS idx_chunks_embedding ON content_chunks USING hnsw (embedding vector_cosine_ops);';
@@ -29,6 +30,10 @@ export function chunkEmbeddingIndexSql(dims: number): string {
   ].join('\n');
 }
 
+export function hnswMaxDimsForType(columnType: 'vector' | 'halfvec'): number {
+  return columnType === 'halfvec' ? PGVECTOR_HNSW_HALFVEC_MAX_DIMS : PGVECTOR_HNSW_VECTOR_MAX_DIMS;
+}
+
 export function applyChunkEmbeddingIndexPolicy(sql: string, dims: number): string {
   return sql.replaceAll(CHUNK_EMBEDDING_HNSW_INDEX, chunkEmbeddingIndexSql(dims));
 }

test/check-resolvable.test.ts

@@ -382,35 +382,6 @@ describe("DRY detection — checkResolvable", () => {
   });
 });
 
-describe("#1767 — ClawHub workspace skills are not resolver-required", () => {
-  let dir: string;
-  afterEachCleanup(() => dir && rmSync(dir, { recursive: true, force: true }));
-
-  test("ClawHub skill without gbrain metadata produces no unreachable/mece_gap", () => {
-    dir = mkdtempSync(join(tmpdir(), "gbrain-clawhub-"));
-    // Native gbrain skill: routable via frontmatter triggers. No manifest.json
-    // (the OpenClaw derive path from the issue repro).
-    mkdirSync(join(dir, "query"), { recursive: true });
-    writeFileSync(
-      join(dir, "query", "SKILL.md"),
-      `---\nname: query\ndescription: test\ntriggers:\n  - "what do we know"\n---\n\n# query\n`
-    );
-    // ClawHub-installed integration: no triggers, no resolver row.
-    mkdirSync(join(dir, "agentmail", ".clawhub"), { recursive: true });
-    writeFileSync(
-      join(dir, "agentmail", ".clawhub", "origin.json"),
-      JSON.stringify({ registry: "https://clawhub.ai", slug: "agentmail" })
-    );
-    writeFileSync(join(dir, "agentmail", "SKILL.md"), `---\nname: agentmail\ndescription: email integration\n---\n\n# agentmail\n`);
-
-    const report = checkResolvable(dir);
-    const agentmailIssues = report.issues.filter(i => i.skill === "agentmail");
-    expect(agentmailIssues).toEqual([]);
-    expect(report.ok).toBe(true);
-    expect(report.summary.total_skills).toBe(1);
-  });
-});
-
 describe("v0.22.4 regression — actual repo skills/ has 0 errors", () => {
   test("repo skills/ pass check-resolvable cleanly (zero errors AND zero warnings)", () => {
     // The v0.22.4 (Part A) contract was zero warnings AND zero errors.

test/embedding-dim-check-facts.test.ts

@@ -122,9 +122,9 @@ describe('buildFactsAlterRecipe', () => {
   });
 
   test('vector recipe uses vector_cosine_ops + vector(N) USING cast', () => {
-    const recipe = buildFactsAlterRecipe(1024, 2048, 'vector');
-    expect(recipe).toContain('vector(2048)');
-    expect(recipe).toContain('USING embedding::vector(2048)');
+    const recipe = buildFactsAlterRecipe(1024, 1536, 'vector');
+    expect(recipe).toContain('vector(1536)');
+    expect(recipe).toContain('USING embedding::vector(1536)');
     expect(recipe).toContain('vector_cosine_ops');
     expect(recipe).not.toContain('halfvec_cosine_ops');
   });
@@ -163,6 +163,14 @@ describe('buildFactsAlterRecipe', () => {
     expect(recipe).not.toContain('UPDATE facts SET embedding = NULL');
     expect(recipe).toContain('USING embedding::vector(1536)');
   });
+
+  test('halfvec recipe skips HNSW rebuild above pgvector cap', () => {
+    const recipe = buildFactsAlterRecipe(1536, 4096, 'halfvec');
+    expect(recipe).toContain('halfvec(4096)');
+    expect(recipe).toContain('Skip reindex');
+    expect(recipe).toContain("exceeds pgvector's HNSW cap of 4000");
+    expect(recipe).not.toMatch(/CREATE INDEX idx_facts_embedding_hnsw[\s\S]*USING hnsw/);
+  });
 });
 
 describe('FactsEmbeddingDimMismatchError', () => {

test/facts-migration-dim.test.ts

@@ -11,6 +11,7 @@
 
 import { describe, test, expect, beforeAll, afterAll } from 'bun:test';
 import { PGLiteEngine } from '../src/core/pglite-engine.ts';
+import { configureGateway, resetGateway } from '../src/core/ai/gateway.ts';
 
 let engine: PGLiteEngine;
 
@@ -93,4 +94,60 @@ describe('migration v45 facts column shape', () => {
     );
     expect(after[0].udt_name).toBe(before[0].udt_name);
   });
+
+});
+
+describe('migration v45/v55 large-dim HNSW policy', () => {
+  let largeDimEngine: PGLiteEngine;
+
+  beforeAll(async () => {
+    configureGateway({
+      embedding_model: 'litellm:custom-4096d',
+      embedding_dimensions: 4096,
+      env: { ...process.env },
+    });
+
+    largeDimEngine = new PGLiteEngine();
+    await largeDimEngine.connect({});
+    await largeDimEngine.initSchema();
+  });
+
+  afterAll(async () => {
+    await largeDimEngine.disconnect();
+    resetGateway();
+  });
+
+  test('4096d init skips unsupported HNSW indexes but keeps vector columns', async () => {
+    const formatRows = await largeDimEngine.executeRaw<{ format_type: string }>(
+      `SELECT format_type(atttypid, atttypmod) AS format_type
+       FROM pg_attribute
+       WHERE attrelid = 'facts'::regclass AND attname = 'embedding'`,
+    );
+    expect(formatRows[0]?.format_type).toMatch(/(halfvec|vector)\(4096\)/);
+
+    const indexRows = await largeDimEngine.executeRaw<{ exists: boolean }>(
+      `SELECT EXISTS (
+         SELECT 1 FROM pg_indexes
+         WHERE tablename = 'facts'
+           AND indexname = 'idx_facts_embedding_hnsw'
+       ) AS exists`,
+    );
+    expect(indexRows[0]?.exists).toBe(false);
+
+    const queryCacheFormatRows = await largeDimEngine.executeRaw<{ format_type: string }>(
+      `SELECT format_type(atttypid, atttypmod) AS format_type
+       FROM pg_attribute
+       WHERE attrelid = 'query_cache'::regclass AND attname = 'embedding'`,
+    );
+    expect(queryCacheFormatRows[0]?.format_type).toMatch(/(halfvec|vector)\(4096\)/);
+
+    const queryCacheIndexRows = await largeDimEngine.executeRaw<{ exists: boolean }>(
+      `SELECT EXISTS (
+         SELECT 1 FROM pg_indexes
+         WHERE tablename = 'query_cache'
+           AND indexname = 'idx_query_cache_embedding_hnsw'
+       ) AS exists`,
+    );
+    expect(queryCacheIndexRows[0]?.exists).toBe(false);
+  }, 60000);
 });

test/skill-manifest.test.ts

@@ -166,55 +166,6 @@ describe('loadOrDeriveManifest', () => {
     expect(r.skills.map(s => s.name)).toEqual(['apple', 'mango', 'zebra']);
   });
 
-  // #1767 — ClawHub-installed workspace skills are external integrations,
-  // not gbrain-routable skills. The derive path skips them unless they
-  // opt in via `triggers:` frontmatter.
-  it('skips ClawHub-origin skills without triggers frontmatter (#1767)', () => {
-    const dir = scratch();
-    writeSkill(dir, 'query', 'query');
-    writeSkill(dir, 'agentmail', 'agentmail');
-    mkdirSync(join(dir, 'agentmail', '.clawhub'), { recursive: true });
-    writeFileSync(
-      join(dir, 'agentmail', '.clawhub', 'origin.json'),
-      JSON.stringify({ registry: 'https://clawhub.ai', slug: 'agentmail' })
-    );
-    const r = loadOrDeriveManifest(dir);
-    expect(r.derived).toBe(true);
-    expect(r.skills.map(s => s.name)).toEqual(['query']);
-  });
-
-  it('includes ClawHub-origin skills that opt in via triggers frontmatter (#1767)', () => {
-    const dir = scratch();
-    writeSkill(dir, 'agentmail', 'agentmail');
-    mkdirSync(join(dir, 'agentmail', '.clawhub'), { recursive: true });
-    writeFileSync(
-      join(dir, 'agentmail', '.clawhub', 'origin.json'),
-      JSON.stringify({ registry: 'https://clawhub.ai', slug: 'agentmail' })
-    );
-    writeFileSync(
-      join(dir, 'agentmail', 'SKILL.md'),
-      `---\nname: agentmail\ndescription: test\ntriggers:\n  - "send email"\n---\n\n# agentmail\n`
-    );
-    const r = loadOrDeriveManifest(dir);
-    expect(r.derived).toBe(true);
-    expect(r.skills.map(s => s.name)).toEqual(['agentmail']);
-  });
-
-  it('keeps ClawHub-origin skills listed in an explicit manifest.json (#1767)', () => {
-    // Explicit manifest.json is a deliberate declaration — strict checking stays.
-    const dir = scratch();
-    writeSkill(dir, 'agentmail', 'agentmail');
-    mkdirSync(join(dir, 'agentmail', '.clawhub'), { recursive: true });
-    writeFileSync(
-      join(dir, 'agentmail', '.clawhub', 'origin.json'),
-      JSON.stringify({ registry: 'https://clawhub.ai', slug: 'agentmail' })
-    );
-    writeManifest(dir, { skills: [{ name: 'agentmail', path: 'agentmail/SKILL.md' }] });
-    const r = loadOrDeriveManifest(dir);
-    expect(r.derived).toBe(false);
-    expect(r.skills.map(s => s.name)).toEqual(['agentmail']);
-  });
-
   it('treats dirs without SKILL.md as not-a-skill', () => {
     const dir = scratch();
     writeSkill(dir, 'query', 'query');