fix(test): mock keychain lookup in anthropic oauth setup tests

[zons-zhaozhy]

Problem

3 tests in tests/agent/test_anthropic_adapter.py fail on macOS:

FAILED TestRunOauthSetupToken::test_returns_token_from_credential_files
FAILED TestRunOauthSetupToken::test_returns_token_from_env_var
FAILED TestRunOauthSetupToken::test_returns_none_when_no_creds_found

All fail with:

TypeError: the JSON object must be str, bytes or bytearray, not MagicMock

Root Cause

_read_claude_code_credentials_from_keychain() calls subprocess.run internally. The tests mock subprocess.run at the outer level via patch("subprocess.run"), but the MagicMock return value leaks into json.loads() inside the keychain helper.

This only manifests on macOS where keychain lookups are attempted. The upstream CI runs on Linux and skips the keychain path entirely, so it never hits this.

Fix

monkeypatch.setattr the keychain helper to return None in the three affected tests, preventing the MagicMock from reaching json.loads().

Testing

Before: 3 FAILED, 150 passed
After:  153 passed

[teknium1]

Thanks for isolating this macOS-specific test path. The premise remains valid on current main: run_oauth_setup_token() calls read_claude_code_credentials() after the test's mocked subprocess.run (agent/anthropic_adapter.py:1350-1355), and the Darwin keychain helper passes that mocked stdout to json.loads() while catching only JSONDecodeError (agent/anthropic_adapter.py:884-912). The three affected tests still install that global mock without keychain isolation (tests/agent/test_anthropic_adapter.py:684,702,715).

Automated hermes-sweeper review.

[teknium1]

Thanks @zons-zhaozhy — closing as resolved on main: the suite-wide autouse Keychain guard from PR #74517 (via #35464) supersedes the per-test mock, covering every test module. You were the earliest on this class (May 30) — credited. Verified post-merge.