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
2.8 KB
Raw Blame
/**
 * Principle-text sanitizer (token-diet, t_40eb1de7): strips leaked
 * tool-call XML fragments and collapses multi-level backslash-quote
 * escaping at the write seam, so corrupted client payloads can no
 * longer poison `Brain/preferences/` frontmatter and every derived
 * view (active.md, digest, MCP resources).
 *
 * Fixture strings mirror the real corruption found on the live vault
 * on 2026-06-02 (see docs/brainstorm/token-diet/design.md).
 */

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

import { sanitisePrinciple } from "../../../src/core/brain/text/sanitize-principle.ts";

describe("sanitisePrinciple", () => {
  test("clean text passes through unchanged", () => {
    const s = 'Use "o2b search" before answering vault questions.';
    expect(sanitisePrinciple(s)).toBe(s);
  });

  test("collapses multi-level backslash-quote chains to a single quote", () => {
    const corrupted =
      'phrasing like \\\\\\"давай так:\\\\\\" or \\\\\\"пусть агент делает X\\\\\\"';
    expect(sanitisePrinciple(corrupted)).toBe(
      'phrasing like "давай так:" or "пусть агент делает X"',
    );
  });

  test("cuts a leaked closing tool-call tag and everything after it", () => {
    const corrupted =
      'everything else needs explicit per-edit approval.</principle>\\\\n<parameter name=\\\\\\"scope\\\\\\">collaboration';
    expect(sanitisePrinciple(corrupted)).toBe("everything else needs explicit per-edit approval.");
  });

  test("cuts at a leaked parameter-open fragment whose attributes carry escaped quotes", () => {
    const corrupted = 'real rule text<parameter name=\\"scope\\">writing';
    expect(sanitisePrinciple(corrupted)).toBe("real rule text");
  });

  test("legitimate prose mentioning <parameter> with plain quotes passes through", () => {
    const legit = 'Document tool calls as <parameter name="scope"> blocks in examples.';
    expect(sanitisePrinciple(legit)).toBe(legit);
  });

  test("a single escaped quote is legitimate prose and survives", () => {
    const legit = 'Inside JSON strings write \\" instead of a bare quote.';
    expect(sanitisePrinciple(legit)).toBe(legit);
  });

  test("a trailing literal newline escape outside a leak cut survives", () => {
    const legit = "End streamed lines with \\n";
    expect(sanitisePrinciple(legit)).toBe(legit);
  });

  test("idempotent: sanitizing twice equals sanitizing once", () => {
    const corrupted =
      'rule \\\\\\"with quotes\\\\\\" and tail</principle>\\\\n<parameter name=\\\\\\"scope\\\\\\">x';
    const once = sanitisePrinciple(corrupted);
    expect(sanitisePrinciple(once)).toBe(once);
  });

  test("non-string input coerces to empty string", () => {
    expect(sanitisePrinciple(undefined)).toBe("");
    expect(sanitisePrinciple(42)).toBe("");
  });
});