Repository navigation
Read-idle timeout on the agenticd connection loop #119
Description
Question
The carved-out slice of audit finding #5 (the rest is deferred to the backlog issue #118, decided on #109): give the connection read/dispatch/write loop a read-idle deadline so a peer that opens a socket and never completes a frame can't hold a spawn_local task forever.
Scope:
- Apply a bounded read-idle timeout around
read_frame_bytesinhandle_connection(crates/agenticd/src/server.rs:140). On timeout, log (withpeer_uid) and close the connection — the same log-and-close shape theFrameError::TooLargearm already uses (no correlation_id available for a structured reply mid-frame). - Pick a generous default (a slow but legitimate first frame must not trip it); make it a const with a comment, tunable later.
- The idle timeout resets per successful frame so a long-lived connection issuing well-spaced requests is unaffected.
- Do NOT add connection/per-UID caps, rate budgets, or queue-depth metrics here — those are deferred to Socket availability hardening: connection caps, per-UID budgets, queue-depth metrics #118 by the Decide socket DoS budgets: fix in-map or defer #109 decision.
Tests: a peer that connects and sends nothing (or a partial frame) is dropped after the deadline; a peer sending well-spaced complete frames is not dropped.
Resolution records: PR link, the chosen timeout value, test names.
Activity
tonibergholm commentedon Jul 10, 2026
Resolution
Shipped as PR #123 (pending review/merge).
What landed:
handle_connectionwraps eachread_frame_bytesin a per-frameREAD_IDLE_TIMEOUT(30s); on elapse it log-and-closes withpeer_uid(same shape as the oversize-frame arm). The clock resets per frame, so well-spaced-within-window connections are unaffected. Injected viahandle_connection_with_idle_timeoutso tests use a short deadline.Chosen value: 30s — generous enough that a slow legitimate first frame won't trip it; idle-beyond-window connections just reconnect (cheap over a Unix socket).
Scope held: no connection/per-UID caps, rate budgets, or queue metrics — those stay in backlog #118 per the #109 decision.
Tests:
idle_connection_is_dropped_after_deadline(connect-and-send-nothing dropped with "idle" error),well_spaced_frames_are_not_dropped(Pings spaced within the window survive, clean close returns Ok). Both avoidtokio::spawnsince the handler future isn'tSend. Workspace tests + clippy + fmt green.Add a read-idle timeout to the agenticd connection loop (#123)