fix(tests): stop the LSP suite writing into the operator's real agent.log

[pai-scaffolde]

The problem

The LSP test suite wrote language-server lifecycle records into ~/.hermes/logs/agent.log. That log is the only durable record of LSP client lifecycle — gateway uptime is short (median 4.19h over 129 restarts), so a live ps snapshot usually cannot answer "is LSP accumulation bounded". The log is the instrument.

Replaying log_active / log_reaped pairs over the polluted log reported 20 concurrent clients climbing with no balancing reaps. Filtering the synthetic $TMPDIR roots collapsed that to 7, spawns and reaps balanced exactly. Ground truth at the time was 0 live processes.

So the pollution did not merely inflate a number — it inverted the verdict: unfiltered it read as UNBOUNDED (which would have justified building an LRU cap), filtered it read as BOUNDED. A live false-signal path into a build decision.

Two independent causes, both closed

1. tests/conftest.py sandboxed HERMES_HOME only when it was unset.

if not os.environ.get("HERMES_HOME"):   # ← the hole

That left it open for exactly the people most likely to hit it: anyone running the suite from inside a Hermes agent or gateway process, where HERMES_HOME is exported and points at the real root. There the sandbox was skipped, import-time setup_logging() attached rotating handlers to the real agent.log, and every propagating logger in the suite wrote into it.

The sandbox is now unconditional. With HERMES_HOME unset — CI and default local runs — behavior is unchanged: both the old and new paths create the same tempdir sandbox. The only behavior that changes is the case that was already broken.

tests/test_log_isolation.py already guarded this property and was genuinely failing under a preset HERMES_HOME:

E  AssertionError: HERMES_HOME pointed at the operator's real home (/Users/gary/.hermes)
E  when conftest loaded; import-time setup_logging() writes to their agent.log

It now passes.

2. The caplog_lsp fixture left propagation on.

caplog.set_level(DEBUG, ...) attaches caplog's own handler but does not stop records climbing to root — and at DEBUG, so even the steady-state events the eventlog design deliberately keeps below INFO got written out. The fixture now disables propagation for its duration and restores it after.

Guarding the NousResearch#69385 regression

Making the sandbox unconditional put an old regression back in reach by a new route. _capture_real_kanban_root() resolves via get_default_hermes_root(), which reads HERMES_HOME from the environment — now always rewired. Left alone, the kanban write-guard deny-list would have pointed at the throwaway tempdir and silently stopped protecting the operator's real ~/.hermes. It now restores the pre-sandbox value for the duration of that call, with a new guard test pinning it.

Verification

Positive control on the identical probe — pytest tests/agent/lsp/ tests/test_cli_skin_integration.py under HERMES_HOME=~/.hermes, same session, same environment:

tree	bytes appended to real agent.log
unfixed (control)	4848
fixed	0

The control reproduced the reporter's exact fingerprint, including the synthetic roots and the /x*.py loop counter:

INFO hermes.lint.lsp: lsp[pyright] active for /private/var/folders/.../T/pytest-of-gary/pytest-282/test_reaper_survives_sweep_err0/repo
INFO hermes.lint.lsp: lsp[reaper] reaped 1 idle client(s) after 0s: pyright (/private/var/folders/.../repo)
INFO hermes.lint.lsp: lsp[pyright] 1 diags (/x0.py)

Other checks:

• tests/agent/lsp/ — 61 passed (60 pre-existing + 1 new containment test); the dedup contract is unchanged.
• tests/test_log_isolation.py — 4 passed (2 pre-existing + 2 new), including under a preset real HERMES_HOME, where it previously failed.
• tests/hermes_cli/test_kanban_write_guard.py — 3 passed, both with and without a preset HERMES_HOME.
• tests/test_hermetic_side_effect_guards.py, test_subprocess_home_isolation.py, test_profile_isolation_runtime.py, test_background_review_session_isolation.py — 31 passed.
• ruff check clean on all three changed files.
• Every file touched was also run unfixed vs fixed under per-file isolation with identical pass counts, so no pre-existing failure is attributable to this change.

Anti-criteria respected

• Not fixed by filtering $TMPDIR at read time — that would leave the log corrupt for every other consumer and every future reader who does not know to filter. The writer is fixed.
• hermes.lint.lsp production verbosity is untouched. Its INFO log_active / log_reaped lines are exactly what makes accumulation auditable.
• No general hermes logging refactor.

Note on the reported timestamp skew

The reporter recorded polluting lines timestamped ahead of the wall clock and flagged the mechanism as unconfirmed. In my repro the skew was exactly 7h and is fully explained by the suite forcing TZ=UTC (the determinism invariant in conftest.py and scripts/run_tests.sh) while the gateway logs in local time. That accounts for my observation but not for their reported ~50-minute delta, so I am not claiming it as the general explanation. It does not affect the fix either way — the writer is fixed rather than time-windowed.

🤖 Generated with Claude Code

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 93c1991bd5

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[pai-scaffolde]

Closing per principal decision 2026-08-20: superseded or abandoned (see landing ledger in scaffolde-ai .scaffolde/tasks/task-hermes-pr-backlog-landing/LEDGER.md). #61 superseded by merged #78; #70 superseded by merged #71 (sys.modules fix verified on main); #50/#51/#52/#58 abandoned SCA-4389 alternatives — #63's approach won and is landed.