mohitsoni48/turbollm is a project that states it automatically tunes local LLM engines for the GPU and serves them through a web UI and OpenAI- and Anthropic-compatible APIs. The commit 632fae6 of October 31, 2026 is a test-only change that touches just two files: turbollm/src/chat/chat-routes.detached.test.ts and turbollm/src/chat/chat-routes.stop.test.ts. A total of 31 lines changed, with 21 additions and 10 deletions.
What the two files share is the change to hold the conversation store in memory rather than on disk. Previously, each test created a temporary directory and opened ConversationStore as a disk-based SQLite store; now IN_MEMORY_DATA_DIR is passed in. At the same time, the test timeout guard was raised from 20 seconds to 60 seconds (GUARD_MS). The commit message explains that on the hosted Windows runner each test takes around 1 second while stalling for several seconds, and that in the full suite there was a case where two adjacent tests timed out at the 20-second guard. Only the single test that reads data back from disk, (q), keeps the disk store via the persist option.
That said, this is test harness cleanup that does not touch user-facing features. How much of the storage layer is verified by the remaining tests after moving the store to memory is not stated in the commit. The commit message notes that even the guard raised to 60 seconds still catches actual hangs as failures, but that comes with the cost of slower failure signals. It is closer to a conservative maintenance change addressing test run time than to a functional improvement.