memind/capabilities.test.mjs

import test from 'node:test';
import assert from 'node:assert/strict';
import {
  buildAgentExtensionPolicy,
  CAPABILITY_CATALOG,
  clampUserCapabilities,
  DEFAULT_USER_CAPABILITIES,
  normalizeCapabilityPatch,
  sandboxDeveloperTools,
} from './capabilities.mjs';
import { applyPoliciesToCapabilities } from './policies.mjs';

test('default user policy blocks dangerous capabilities', () => {
  assert.equal(DEFAULT_USER_CAPABILITIES.shell, false);
  assert.equal(DEFAULT_USER_CAPABILITIES.filesystem, false);
  assert.equal(DEFAULT_USER_CAPABILITIES.extension_admin, false);
  assert.equal(DEFAULT_USER_CAPABILITIES.static_publish, false);
  assert.equal(DEFAULT_USER_CAPABILITIES.code_browse, false);
});

test('buildAgentExtensionPolicy returns null overrides and auto mode for unrestricted users', () => {
  const policy = buildAgentExtensionPolicy({}, { unrestricted: true });
  assert.equal(policy.extensionOverrides, null);
  assert.equal(policy.gooseMode, 'auto');
});

test('static_publish enables sandbox developer tools without shell by default', () => {
  const caps = { ...DEFAULT_USER_CAPABILITIES, static_publish: true };
  assert.deepEqual(sandboxDeveloperTools(caps), ['write', 'edit', 'read_image']);

  const policy = buildAgentExtensionPolicy(caps);
  const developer = policy.extensionOverrides.find((ext) => ext.name === 'developer');
  assert.deepEqual(developer?.available_tools, ['write', 'edit', 'read_image']);
  assert.equal(developer?.available_tools.includes('shell'), false);
  assert.equal(developer?.available_tools.includes('tree'), false);
  const summon = policy.extensionOverrides.find((ext) => ext.name === 'summon');
  assert.deepEqual(summon?.available_tools, ['load_skill']);
  assert.ok(policy.extensionOverrides.some((ext) => ext.name === 'projectmemory'));
  assert.equal(policy.enableContextMemory, true);
});

test('static_publish with shell survives network deny and includes tree', () => {
  const caps = applyPoliciesToCapabilities(
    { ...DEFAULT_USER_CAPABILITIES, static_publish: true, shell: true },
    { network_egress: 'deny' },
  );
  assert.equal(caps.shell, true);
  const policy = buildAgentExtensionPolicy(caps, {
    policies: { network_egress: 'deny', goose_mode: 'chat' },
  });
  const developer = policy.extensionOverrides.find((ext) => ext.name === 'developer');
  assert.deepEqual(developer?.available_tools, ['write', 'edit', 'shell', 'tree', 'read_image']);
});

test('buildAgentExtensionPolicy filters developer tools', () => {
  const policy = buildAgentExtensionPolicy({
    ...DEFAULT_USER_CAPABILITIES,
    shell: true,
    filesystem: false,
    code_browse: true,
  });
  const developer = policy.extensionOverrides.find((ext) => ext.name === 'developer');
  assert.ok(developer);
  assert.deepEqual(developer.available_tools, ['shell', 'tree', 'read_image']);
  assert.equal(
    policy.extensionOverrides.some((ext) => ext.name === 'extensionmanager'),
    false,
  );
});

test('normalizeCapabilityPatch ignores unknown keys', () => {
  assert.deepEqual(normalizeCapabilityPatch({ shell: true, unknown: true }), { shell: true });
});

test('clampUserCapabilities blocks extension_admin even when stored as true', () => {
  const clamped = clampUserCapabilities({
    ...DEFAULT_USER_CAPABILITIES,
    extension_admin: true,
    shell: true,
  });
  assert.equal(clamped.extension_admin, false);
  assert.equal(clamped.shell, true);
});

test('catalog keys are unique', () => {
  const keys = CAPABILITY_CATALOG.map((item) => item.key);
  assert.equal(keys.length, new Set(keys).size);
});