Garry's Opinionated OpenClaw/Hermes Agent Brain
TypeScript
/**
* #5890 — think/synthesize evidence gather must not be trimmed by adaptive
* return. With `search.adaptive_return` on, the reader-facing intent cap
* (entity 2 pages, other 6) used to cut a breadth-sized gather of 12
* matching meetings to 2 or 6 pages before synthesis saw them.
*
* Real PGLite + real runGather. Embeddings are mocked deterministically so
* hybridSearch takes the vector path (the keyword-only short-circuit skips
* adaptive return and would hide the bug).
*
* Serial: mock.module + process-global gateway config.
*/
import { afterAll, beforeAll, describe, expect, mock, test } from 'bun:test';
import { mkdtempSync, rmSync } from 'node:fs';
import { tmpdir } from 'node:os';
import { join } from 'node:path';
import * as realEmbedding from '../src/core/embedding.ts';
import { installFixtureChunks } from './helpers/page-projection.ts';
function unitVector(seed: number): Float32Array {
const arr = new Float32Array(1536);
for (let i = 0; i < arr.length; i++) arr[i] = Math.sin(1 + i * 0.001) + seed * 0.0001 * Math.cos(i);
const norm = Math.sqrt(arr.reduce((s, v) => s + v * v, 0));
return arr.map((v) => v / norm);
}
mock.module('../src/core/embedding.ts', () => ({
...realEmbedding,
embed: async () => unitVector(0),
embedQuery: async () => unitVector(0),
}));
const { runGather } = await import('../src/core/think/gather.ts');
const { configureGateway, resetGateway } = await import('../src/core/ai/gateway.ts');
const { PGLiteEngine } = await import('../src/core/pglite-engine.ts');
let engine: InstanceType<typeof PGLiteEngine>;
let tmpHome: string;
const savedHome = process.env.GBRAIN_HOME;
const MEETINGS = 12;
beforeAll(async () => {
tmpHome = mkdtempSync(join(tmpdir(), 'gbrain-5890-'));
process.env.GBRAIN_HOME = tmpHome;
resetGateway();
configureGateway({ embedding_model: 'openai:text-embedding-3-large', embedding_dimensions: 1536, env: { OPENAI_API_KEY: 'sk-fake' } });
engine = new PGLiteEngine();
await engine.connect({});
await engine.initSchema();
for (let i = 1; i <= MEETINGS; i++) {
const slug = `meetings/zorblax-review-${i}`;
const body = `Zorblax review ${i}: the team discussed the Zorblax rollout, risks and next steps.`;
await engine.putPage(slug, { type: 'meeting', title: `Zorblax review ${i}`, compiled_truth: body });
await installFixtureChunks(engine, slug, [{ chunk_index: 0, chunk_text: body, chunk_source: 'compiled_truth', embedding: unitVector(i) }]);
}
await engine.setConfig('search.adaptive_return', 'true');
}, 60_000);
afterAll(async () => {
await engine.disconnect();
resetGateway();
if (savedHome === undefined) delete process.env.GBRAIN_HOME;
else process.env.GBRAIN_HOME = savedHome;
rmSync(tmpHome, { recursive: true, force: true });
});
describe('#5890 gather keeps breadth with adaptive return on', () => {
for (const question of ['Zorblax', 'what did the team decide about the zorblax rollout']) {
test(`"${question}" gathers all ${MEETINGS} meetings`, async () => {
const result = await runGather(engine, { question, remote: false });
const meetings = result.pages.filter((p) => p.slug.startsWith('meetings/zorblax-review-'));
expect(new Set(meetings.map((p) => p.slug)).size).toBe(MEETINGS);
}, 30_000);
}
});