fix(logging): stop pytest runs from writing to the live agent.log

[lEWFkRAD]

What does this PR do?

Stops pytest tests/ from appending mock-provider log records to the real user's agent.log (%LOCALAPPDATA%\hermes\logs\agent.log on native Windows, ~/.hermes/logs/agent.log on POSIX).

run_agent._hermes_home is resolved once at import time. Under pytest, test modules are imported during collection — before the per-test _hermetic_environment fixture redirects HERMES_HOME — so agent_init's setup_logging(hermes_home=_ra()._hermes_home) attached root-logger file handlers to the real install's logs/agent.log. Every mock-provider test in the process then wrote to the live log, producing entries like provider=xai-oauth base_url=https://api.x.ai/v1 model=grok-4.3 ... msg='Say OK' that look exactly like the agent silently calling real provider APIs.

The fix is at the callsite (per the guidance already in the tests/conftest.py docstring) plus defense-in-depth in the conftest itself:

1. agent_init now calls setup_logging() with no hermes_home= override — the default resolves get_hermes_home() at call time, which reads the env var live. Behavior for real CLI/gateway runs is unchanged (env doesn't change between import and agent init there).
2. tests/conftest.py exports a sandbox HERMES_HOME at conftest import time, before pytest imports any test module, so every other import-time get_hermes_home() consumer (cli._hermes_home, dotenv loading) freezes a sandbox path instead of the real install.

Related Issue

Fixes #57118

Type of Change

• 🐛 Bug fix (non-breaking change that fixes an issue)

Changes Made

• agent/agent_init.py — drop the hermes_home=_ra()._hermes_home override on the setup_logging() call; resolve the log directory from the live environment instead of run_agent's import-time frozen path.
• tests/conftest.py — set HERMES_HOME to a session tempdir at module import time (unconditionally, so a HERMES_HOME inherited from a developer shell or Docker deployment can't leak into tests either). The existing per-test fixture still re-points it per test.
• tests/run_agent/test_logging_home_isolation.py (new) — two regression tests:
  • test_agent_logging_ignores_import_time_frozen_home: simulates a stale run_agent._hermes_home and asserts AIAgent.__init__ attaches agent.log under the live HERMES_HOME, with nothing under the stale path. Fails if the agent_init fix is reverted.
  • test_conftest_sandboxes_hermes_home_at_import_time: asserts run_agent._hermes_home did not freeze the platform-default (real) home. Fails if the conftest sandbox is reverted.

How to Test

1. On a machine with a live Hermes install, note the line count of <hermes-home>/logs/agent.log.
2. On main, run python -m pytest tests/run_agent/test_run_agent_codex_responses.py -q — grep 'api.x.ai' <hermes-home>/logs/agent.log shows new mock records (13 new lines in my repro, Windows 11 native).
3. On this branch, repeat — the live log gains zero test records (verified: 93 tests across test_run_agent_codex_responses.py + test_auxiliary_client_xai_oauth_recovery.py, live-log signature count unchanged).
4. python -m pytest tests/run_agent/test_logging_home_isolation.py -q — both regression tests pass; each fails with its half of the fix reverted.

Checklist

Code

• I've read the Contributing Guide
• My commit messages follow Conventional Commits (fix(scope):, feat(scope):, etc.)
• I searched for existing PRs to make sure this isn't a duplicate (fix(windows): native Windows correctness fixes + green CLI test suite #57016 touches tests/cli/conftest.py reload pollution but not this)
• My PR contains only changes related to this fix/feature (no unrelated commits)
• I've run pytest tests/ -q and all tests pass (per-file runs of the affected areas: tests/run_agent/test_run_agent.py 414 passed, tests/run_agent/test_run_agent_codex_responses.py + xai recovery 93 passed, tests/cron/test_cron_profile_isolation.py, tests/cron/test_cron_no_agent.py, tests/cli/test_cli_approval_ui.py, tests/agent/test_auxiliary_client.py all green; tests/test_hermes_logging.py / tests/test_hermes_constants.py have pre-existing Windows-native chmod/symlink failures identical on unmodified main)
• I've added tests for my changes (required for bug fixes, strongly encouraged for features)
• I've tested on my platform: Windows 11 native (Python 3.12)

Documentation & Housekeeping

• I've updated relevant documentation (README, docs/, docstrings) — comments at both changed sites explain the import-time-freeze trap
• I've updated cli-config.yaml.example if I added/changed config keys — N/A
• I've updated CONTRIBUTING.md or AGENTS.md if I changed architecture or workflows — N/A
• I've considered cross-platform impact (Windows, macOS) per the compatibility guide — the mechanism is platform-independent (POSIX pytest tests/ wrote to the real ~/.hermes/logs/agent.log the same way); fix uses only tempfile/env, no platform branches
• I've updated tool descriptions/schemas if I changed tool behavior — N/A

Screenshots / Logs

Pollution on main (live %LOCALAPPDATA%\hermes\logs\agent.log after running one test file):

2026-07-02 09:59:12,610 INFO [...] run_agent: OpenAI client created (agent_init, shared=True) thread=MainThread:11288 provider=xai-oauth base_url=https://api.x.ai/v1 model=grok-4.3
2026-07-02 09:59:13,113 INFO [...] agent.turn_context: conversation turn: session=... model=gpt-5.4 provider=copilot platform=unknown history=0 msg='Say OK'

After this branch: same run adds zero records to the live log.

[teknium1]

Thanks for isolating the import-time HERMES_HOME path; the current defect is still present at agent/agent_init.py:706.

Problems

• The new helper in tests/run_agent/test_logging_home_isolation.py:25 only scans root handlers. Current main routes real file handlers through the async QueueListener; hermes_logging.rotating_file_handlers() is the supported accessor (hermes_logging.py:694-700). The proposed assertion will therefore not see agent.log.
• Resolving a per-test live home at agent/agent_init.py:706 needs handler lifecycle work. tests/conftest.py:355-361 changes HERMES_HOME every test, while setup_logging() registers per-path handlers before its initialized guard (hermes_logging.py:320-365) and only deduplicates identical paths (hermes_logging.py:740-746). That can retain queued handlers for prior test homes.

Suggested changes

• Make the new test use rotating_file_handlers() and reset queued logging state in its fixture/teardown.
• Ensure changing HERMES_HOME does not accumulate stale logging handlers, then cover that invariant.

Automated hermes-sweeper review.

[teknium1]

Current main keeps rotating file handlers on the QueueListener, not the root logger (hermes_logging.py:694-700). Please inspect hermes_logging.rotating_file_handlers() here and reset queued-handler state in teardown; otherwise this helper cannot see the agent.log handler and the test leaks listener/file-handler state.

[teknium1]

Thanks @lEWFkRAD — closing as resolved on main: PR #74553 landed a session-level HERMES_HOME sandbox at conftest import time (via #71271), which closes the import-time-frozen log path this PR targeted, and tests/test_log_isolation.py now pins the invariant. Your diagnosis of the class (also filed as #57118) was right — credited there too.