Skip to content

Local-first 🧠 memory for Hermes Agent that lives in your Obsidian vault and remembers project context. Nightly 😴 dream passes turn repeat corrections into confirmed preferences with measurable confidence. Adapters ship for Claude Code, Codex, and OpenClaw, with an MCP server for anything else.

TypeScript
1.2 KB
Raw Blame
/**
 * The `brain_search` when-to-search cue (t_cb1e9d3d): the tool description
 * carries exactly one sentence naming `brain_recall_gate` as the classifier
 * to consult before widening recall - calibration for the agent reading the
 * tool list, nothing more. No fan-out counts and no parallel-search
 * language: the description must not promise machinery it does not have.
 */

import { describe, expect, test } from "bun:test";

import { SEARCH_TOOLS } from "../../src/mcp/search-tools.ts";

const SEARCH_DESCRIPTION = SEARCH_TOOLS.find((t) => t.name === "brain_search")!.description;

describe("brain_search when-to-search cue", () => {
  test("the description names brain_recall_gate exactly once", () => {
    expect(SEARCH_DESCRIPTION).toContain("brain_recall_gate");
    expect(SEARCH_DESCRIPTION.split("brain_recall_gate")).toHaveLength(2);
  });

  test("the cue names widening recall, with no fan-out or parallel-search language", () => {
    expect(/widening recall/i.test(SEARCH_DESCRIPTION)).toBe(true);
    expect(/parallel/i.test(SEARCH_DESCRIPTION)).toBe(false);
    expect(/fan-out/i.test(SEARCH_DESCRIPTION)).toBe(false);
    expect(/\d+\s+search(es|es per)/i.test(SEARCH_DESCRIPTION)).toBe(false);
  });
});