Fixed

  • MCP hub tool pruning learns from past transcripts again (17.17.1). On 17.17.0 the first-start backfill could stop after a few files and leave pruning without history; on one machine the service also stopped answering for about 98 minutes while it ran. Transcripts are now read in 1 MiB chunks with the time budget checked inside each file. Only lines naming a hub tool are examined, oversized tool-output lines are skipped, and sessions still running are left to live recording. On that machine: 971 transcripts in 0.7 s, longest event-loop stall 2 ms, peak memory 137 MB (was 522 MB). A store whose earlier run stopped early learns again once on the next start.

Update HarnessLink and every installed harness to its current release:

hnl update

New machine? Install with:

curl -fsSL https://harnesslink.sh/install.sh | bash

Newer: 17.17.2 · Older: 17.17.0 · All releases · Atom feed

Flag notifications