• v1.21.0 5360ec731b

    v1.21.0 — a missed MCP server is retried
    All checks were successful
    CI / build-and-test (push) Successful in 2m8s
    CI / docker-build (push) Successful in 37s
    Stable

    joeyr released this 2026-09-16 13:35:29 -05:00 | 51 commits to main since this release

    Changed

    • A missed MCP server is retried instead of given up on. The bot asks
      every server in MCP_SERVERS for its tools once, at startup, and a server
      that did not answer was then missing for the life of the process. The case
      that hits is the cheap one: the bot and the espnff server are containers
      on the same host and they start together, so the bot can lose the race by a
      few seconds and then run for days with /aufhocker espnff mcp refusing
      every question. Nothing in Discord announced it — the first sign was a
      fantasy question refused, and the fix was a restart nobody knew to do.

      A server that misses the boot is now retried in the background: after a
      minute, then doubling to a two-hour ceiling. Each reconnect registers that
      server's tools on the live tool executor, so the fantasy command starts
      working mid-process and /aufhocker espnff help flips from "not connected"
      on its own. Both messages now say to wait rather than to restart.

      Backoff rather than a fixed interval because the two failures underneath
      have different lengths: a co-boot race is over inside a minute, while a
      server that is genuinely down is down for hours and polling it every minute
      buys nothing. Only the servers that are actually missing are retried — a
      session that connected is held open as before, since re-handshaking a
      working server would drop its tools mid-flight. Shutdown cancels the timer.

    Downloads