Persistent PTY: every input lists all terminals, and the daemon sweeps /proc per terminal #60728

Activity

1. opencodejr commented on Oct 30, 2026

   

   opencodejrBot

   on Oct 30, 2026

   Contributor

   More actions

   Thanks for the detailed measurements. I could reproduce the main part of this.

   The script below creates persistent terminals through opencode serve and asks for the snapshot of one terminal. It counts the daemon's /proc opens for that single request:

   Terminals	/proc opens (2.2.4, daemon 0.4.0)	Median latency
   1	828	4 ms
   5	2,532	8 ms
   25	12,012	34 ms
   50	26,112	66 ms

   This still happens on the latest dev (171dbd6). PersistentPty.get(id) still sends the daemon's list op for each per-terminal operation, and list still sweeps all of /proc once per terminal to find its foreground process.

   The other two parts did not show up on the current daemon (0.4.0). With 50+ terminals and no requests, the idle daemon made no openat, accept4 or clock_nanosleep calls in 10 s. The polling was removed in #59701 and #59698, after the 0.3.1 release you measured. Upgrading should take care of the idle CPU and the 0–10 ms accept delay; the cost of each op still grows with the number of terminals.

   Reproduction script (repro.sh)

   #!/usr/bin/env bash
   # Reproduces #60728: every per-terminal op in `opencode serve` lists ALL terminals,
   # and the daemon's `list` walks /proc once per terminal, so op latency grows with
   # terminal count.
   # Usage: start `opencode serve --port 4199 --hostname 127.0.0.1` in a project dir,
   # then: PASS=<server password> ./repro.sh   (needs curl, strace, jq-less)
   set -euo pipefail
   URL=${URL:-http://127.0.0.1:4199}
   AUTH="opencode:${PASS:?set PASS to the server password}"
   SES=$(curl -s -u "$AUTH" -X POST -H 'content-type: application/json' -d '{}' "$URL/api/session" | grep -o '"id":"ses_[^"]*"' | head -1 | cut -d'"' -f4)
   create() {
     curl -s -u "$AUTH" -X POST -H 'content-type: application/json' -d '{"command":"/bin/sh","args":[],"title":"t","env":{},"cwd":"/tmp"}' \
       "$URL/api/experimental/session/$SES/terminal" | grep -o '"id":"pty_[^"]*"' | head -1 | cut -d'"' -f4
   }
   FIRST=$(create)
   DAEMON=$(pgrep -n -x opencode-pty)
   echo "session=$SES first=$FIRST daemon_pid=$DAEMON"
   measure() {
     local n=$1
     # one snapshot (per-terminal op) under strace: count /proc opens in the daemon
     timeout 5 strace -f -qq -e trace=openat -p "$DAEMON" -o /tmp/opencode/st.$n 2>/dev/null &
     local sp=$!; sleep 0.5
     local t=$(curl -s -o /dev/null -u "$AUTH" -w '%{time_total}' "$URL/api/experimental/persistent-pty/$FIRST/snapshot")
     sleep 0.5; kill $sp 2>/dev/null || true; wait $sp 2>/dev/null || true
     local procs=$(grep -c '"/proc/' /tmp/opencode/st.$n || true)
     # latency without strace, median of 10
     local lat=$(for i in $(seq 10); do curl -s -o /dev/null -u "$AUTH" -w '%{time_total}\n' "$URL/api/experimental/persistent-pty/$FIRST/snapshot"; done | sort -n | sed -n 5p)
     echo "terminals=$n  snapshot_of_ONE_terminal: daemon /proc opens=$procs  median latency=${lat}s"
   }
   measure 1
   for target in 5 25 50; do
     while [ "$(curl -s -u "$AUTH" "$URL/api/experimental/session/$SES/terminal" | grep -o '"id":"pty_' | wc -l)" -lt $target ]; do create >/dev/null; done
     measure $target
   done

   Run it: bash repro.sh
2. opencodejr

   added

   reproduced

   on Oct 30, 2026