service: event seq collision drops session events after config reload #51798

Summary

Editing opencode.jsonc while a session is actively running causes the service to lose session events. Two concurrent writers allocate the same event.seq for one session aggregate, and the second insert fails on UNIQUE(aggregate_id, seq). The affected turn is reported to the model as an error and the event is dropped, leaving the session unable to continue.

Environment

• opencode version: 2.0.18
• OS: Linux 6.18.33.2-microsoft-standard-WSL2 feat: compact and other improvements #1 SMP PREEMPT_DYNAMIC Thu Jun 18 21:54:43 UTC 2026 x86_64 (WSL2, networkingMode=mirrored)
• Terminal: TERM=xterm-256color, TERM_PROGRAM unset, COLORTERM unset, WT_SESSION set (Windows Terminal)
• Shell: /usr/bin/zsh
• Install/channel: curl install, latest channel
• Active plugins: none configured

Reproduction

1. Start a session and let it run a tool call so the session has in-flight event writes.
2. While that session is still running, edit the global config at ~/.config/opencode/opencode.jsonc (even a whitespace/comment-only edit).
3. The service watches ~/.config/opencode with inotify and reloads configuration.
4. Within ~20 seconds, the running session's event write fails and the turn errors.

Observed twice in a single session:

2026-09-28T07:42:08.607  mtime of ~/.config/opencode/opencode.jsonc
2026-09-28T07:42:09.950  violation #1
2026-09-28T07:42:26.539  violation #2

Both failures are seq=427 on the same session aggregate, from two different event types — which is what makes this look like a sequence-allocation race rather than a duplicate insert:

params: evt_0e6f6d01b001iSAQxlXQDgakGK,ses_f191622cbffe670EoOV4COHAGu,427,message.part.updated.1
params: evt_0e6f710ea001N6MKHZveJfATKn,ses_f191622cbffe670EoOV4COHAGu,427,message.updated.1

Expected Behavior

A configuration reload should not interfere with in-flight session writes. Either the sequence allocator should serialize per aggregate so the second writer takes seq=428, or the insert should retry on conflict.

Actual Behavior

level=ERROR message=process error="Failed query: insert into \"event\" (\"id\", \"aggregate_id\", \"seq\", \"type\", \"data\") values (?, ?, ?, ?, ?)"
  cause: EffectDrizzleQueryError ... (cause: SqlError ... SQLiteError: UNIQUE constraint failed: event.aggregate_id, event.seq)

The failure propagates into Session.updateMessage -> SessionProcessor.cleanup and is surfaced to the model as an UnknownError turn. The session stopped accepting prompts afterward.

Additional Context

• Reproduced 2 times, both within one session. I have not yet reproduced it from a clean start, so I cannot confirm the config edit is causal rather than coincidental with an unrelated write race — the reload and the collision are tightly correlated in both cases, but n=2 in a single session is weak evidence for causation.
• The service subscribes to the config directory, so this is reachable without any explicit action beyond editing a file:

  level=INFO message="watcher subscribe" path=/home/xylar/.config/opencode type=directory ignores=3 role=server
  level=INFO message="watcher started" path=/home/xylar/.config/opencode type=directory backend=inotify ignores=3 role=server
  

• Because sessions are event-sourced, a dropped event is not just a lost log line — it is a gap in the session's state history.
• No workaround found other than editing config only while no session is running. Restarting the service clears the stuck state.
• If a config edit legitimately needs to invalidate in-memory state, it would help to know whether that path is meant to be concurrency-safe.