perf(sessions): push new-session detection over SSE #181

Merged

dries merged 1 commit from feat/push-new-sessions into main

2026-06-29 23:14:31 +02:00

dries commented

2026-06-29 19:07:00 +02:00

Owner

Problem

New sessions only appeared on the next list-poll tick — ~9s worst case (4s background refresher + 3–5s frontend poll) — because session creation was never broadcast. Detection relied entirely on polling OpenCode's SQLite DB.

Approach

Reuse infrastructure that already exists:

The auto-approve watcher already holds a long-lived connection to every running OpenCode instance's /event SSE stream.
The /api/events broadcast hub and the frontend useGlobalEvents listener already exist.

On the first sighting of a session.updated event for a given session ID, the watcher invalidates the sessions cache and broadcasts a new ocman.session.changed event. The frontend invalidates the ['sessions'] query, so the list refetches immediately instead of waiting for the next poll.

A seen-set dedupes session.updated (which fires per turn/token) so only the first sighting — i.e. "a session appeared" — triggers work. Per-keystroke updates are dropped.

Result

New live session: ~9s → sub-second.
More decoupled from the DB, not less: the trigger is OpenCode's own event; the DB is read once on demand instead of every 4s. The 4s refresher stays as the fallback for dead/never-running instances.

Changes

Backend

autoapprove.go — session.updated dispatch case + dispatchSessionChanged → onSessionChanged tee callback.
autoapprove_watcher.go — handleSessionChanged with seen-set dedup; busts cache + broadcasts on first sight.
broadcast.go — broadcastSessionChanged → ocman.session.changed.
models_cache.go — InvalidateSessionsCache() (expires entries, keeps last-good for stale-on-busy).

Frontend

useGlobalEvents.ts — ocman.session.changed listener + onSessionChanged registry.
App.tsx — invalidates ['sessions'] on the event.

Tests

Tee dispatch of session.updated (envelope/flat/both casings/missing-id).
Watcher dedup (first sight acts, duplicates no-op, empty id ignored).
InvalidateSessionsCache forces a refetch.
Frontend listener register/notify/unsubscribe.

All Go + frontend suites, go vet, tsc, ESLint, and the platform-branching check pass.

Known limitations

Push only covers currently-running OpenCode instances (lsof-discovered). Dead/never-running instances still surface via the existing poll.
Seen-set is unbounded (marked with a ponytail: comment); swap to LRU if per-machine session count ever grows large.

## Problem New sessions only appeared on the next list-poll tick — ~9s worst case (4s background refresher + 3–5s frontend poll) — because session creation was never broadcast. Detection relied entirely on polling OpenCode's SQLite DB. ## Approach Reuse infrastructure that already exists: - The auto-approve watcher already holds a long-lived connection to every running OpenCode instance's `/event` SSE stream. - The `/api/events` broadcast hub and the frontend `useGlobalEvents` listener already exist. On the **first** sighting of a `session.updated` event for a given session ID, the watcher invalidates the sessions cache and broadcasts a new `ocman.session.changed` event. The frontend invalidates the `['sessions']` query, so the list refetches immediately instead of waiting for the next poll. A **seen-set** dedupes `session.updated` (which fires per turn/token) so only the first sighting — i.e. "a session appeared" — triggers work. Per-keystroke updates are dropped. ## Result - New live session: **~9s → sub-second**. - More decoupled from the DB, not less: the trigger is OpenCode's own event; the DB is read once on demand instead of every 4s. The 4s refresher stays as the fallback for dead/never-running instances. ## Changes **Backend** - `autoapprove.go` — `session.updated` dispatch case + `dispatchSessionChanged` → `onSessionChanged` tee callback. - `autoapprove_watcher.go` — `handleSessionChanged` with seen-set dedup; busts cache + broadcasts on first sight. - `broadcast.go` — `broadcastSessionChanged` → `ocman.session.changed`. - `models_cache.go` — `InvalidateSessionsCache()` (expires entries, keeps last-good for stale-on-busy). **Frontend** - `useGlobalEvents.ts` — `ocman.session.changed` listener + `onSessionChanged` registry. - `App.tsx` — invalidates `['sessions']` on the event. ## Tests - Tee dispatch of `session.updated` (envelope/flat/both casings/missing-id). - Watcher dedup (first sight acts, duplicates no-op, empty id ignored). - `InvalidateSessionsCache` forces a refetch. - Frontend listener register/notify/unsubscribe. All Go + frontend suites, `go vet`, `tsc`, ESLint, and the platform-branching check pass. ## Known limitations - Push only covers currently-running OpenCode instances (lsof-discovered). Dead/never-running instances still surface via the existing poll. - Seen-set is unbounded (marked with a `ponytail:` comment); swap to LRU if per-machine session count ever grows large.

dries added 1 commit

2026-06-29 19:07:00 +02:00

perf(sessions): push new-session detection over SSE

CI / Frontend (pull_request) Successful in 4m13s

Details

CI / Backend (pull_request) Successful in 5m2s

Details

CI / Playwright E2E (pull_request) Successful in 6m33s

Details

CI / Build Desktop (macOS arm64) (pull_request) Successful in 1m34s

Details

CI / Build (pull_request) Successful in 3m18s

Details

CI / Semantic Tag (pull_request) Has been skipped

Details

CI / Coverage Ratchet (pull_request) Successful in 4m27s

Details

57e3c368d1

New sessions appeared only on the next list-poll tick (~9s worst case)
because creation was never broadcast. Tap the OpenCode /event stream
the auto-approve watcher already consumes: on the first sighting of a
session.updated event for a session ID, invalidate the sessions cache
and broadcast ocman.session.changed so clients refetch immediately.

A seen-set dedupes per-turn/per-token session.updated events so only
the first sighting (session created) triggers a refresh. Live sessions
now surface sub-second while the background refresher remains the
fallback for non-live instances.

forgejo-actions commented

2026-06-29 19:16:28 +02:00

Coverage ratchet: ✅ pass

Suite	Baseline	This PR	Δ
go	69.40%	69.50%	+0.10	✅
frontend	57.39%	57.41%	+0.02	✅

Tolerance: -0.1%. Baseline stored on gh-pages.

### Coverage ratchet: ✅ pass | Suite | Baseline | This PR | Δ | | |---|---|---|---|---| | go | 69.40% | 69.50% | +0.10 | ✅ | | frontend | 57.39% | 57.41% | +0.02 | ✅ | _Tolerance: -0.1%. Baseline stored on `gh-pages`._