mirror of
https://github.com/NousResearch/hermes-agent.git
synced 2026-11-17 04:11:35 +00:00
Independent review of #120220 found restored tests that could pass without reaching the code they name. Each fix below was A/B'd against a sabotaged production line (red) and the real one (green). - test_clarify_progress_leak: the fake agent slept 2x0.35 s as its only sync. It now waits (<= 5 s) for the progress task's first bubble; the queue is FIFO, so a leaked clarify bubble is what lands first. Reverting the clarify skip in run_turn_runner goes red on the leaked JSON. - test_desktop_lifecycle_windows_live holder-scan case: the kanban row was rebuilt after the serve kill, re-reading the dead pid; NoSuchProcess was swallowed into an empty scan and returned False without reaching the classifier. Rows are now snapshotted while both processes live; a classifier that accepts everything goes red (the old version stayed green). - *_windows_live (control socket, desktop lifecycle, plan reconciliation, shim fail-closed): fixed time.sleep settles become bounded waits on the real condition (argv visible to psutil, pipe gone, handle released) and child readline() calls get a 60 s bound. - test_gateway_proc_fallback BSD ps case: drop the is_windows stub and mark it linux_only like its neighbour; a revert to `ps -A eww` goes red. - test_delivery_ledger_single_connection counts only connections to the ledger's own database, not every sqlite3.connect in the process; pruning on a second connection goes red. - test_codex_models: the hardcoded gpt-5.6 list becomes a relationship to DEFAULT_CODEX_MODELS and the forward-compat table; adding a -pro slug to either goes red. - group-chat-view inline-code: the renderer stub wrapped the code in .aui-md, which the generic inline-code rule themes on its own, so the room's data-slot rule was never tested. Without the wrapper, renaming the room's data-slot goes red.
37 lines
1.5 KiB
Python
37 lines
1.5 KiB
Python
"""Recording a reply opens exactly one SQLite connection.
|
|
|
|
``record_obligation`` runs on every outbound final response. Its retention prune must run on the
|
|
recording connection, inside the recording transaction — the way the cron ledgers prune — not on a
|
|
second connection opened after the first one closed.
|
|
"""
|
|
from __future__ import annotations
|
|
|
|
import sqlite3
|
|
from pathlib import Path
|
|
|
|
from gateway import delivery_ledger as dl
|
|
|
|
|
|
def test_recording_a_reply_does_not_open_a_second_connection(tmp_path, monkeypatch):
|
|
db = tmp_path / "state.db"
|
|
monkeypatch.setattr(dl, "_db_path", lambda: db)
|
|
real_connect = sqlite3.connect
|
|
opened: list[str] = []
|
|
|
|
def counting_connect(*args, **kwargs):
|
|
target = args[0] if args else kwargs.get("database")
|
|
# Only the ledger's own database counts: unrelated sqlite users in the
|
|
# process (other modules, background threads) must not move the tally.
|
|
if Path(str(target)).resolve() == db.resolve():
|
|
opened.append(str(target))
|
|
return real_connect(*args, **kwargs)
|
|
|
|
monkeypatch.setattr(sqlite3, "connect", counting_connect)
|
|
dl.record_obligation(obligation_id="ob-warm", session_key="s", platform="p", chat_id="c",
|
|
thread_id=None, content="warm-up (schema)")
|
|
opened.clear()
|
|
|
|
dl.record_obligation(obligation_id="ob-1", session_key="s", platform="p", chat_id="c",
|
|
thread_id=None, content="hello")
|
|
|
|
assert len(opened) == 1, opened
|