fix(event-service): refcount subscribe/unsubscribe by context#2927
Merged
iscekic merged 31 commits intofeat/kilo-chat-migration-pr1from Apr 30, 2026
Merged
fix(event-service): refcount subscribe/unsubscribe by context#2927iscekic merged 31 commits intofeat/kilo-chat-migration-pr1from
iscekic merged 31 commits intofeat/kilo-chat-migration-pr1from
Conversation
…rpose); update all consumers
The badge_counts.badge_bucket column is a free-form string. To prevent namespace collisions as more surfaces start emitting badge updates (per-instance today, per-conversation later), centralize bucket-key derivation in @kilocode/notifications and route NotificationChannelDO through it. Mirrors the presence-context builders in @kilocode/event-service. Safe to introduce now without a data migration because PR 2's migration already wipes badge_counts.
…-chat producer Adds kiloclawInstanceContext and kiloclawConversationContext path builders to @kilocode/event-service, replacing hardcoded template literals in kilo-chat's event-push.ts and its test so all callers share a single source of truth.
When a chat message is persisted, fire-and-forget a call to NOTIFICATIONS.sendPushForConversation so non-sender human members of the conversation receive a push. Runs after realtime/event-service delivery inside postCommitFanOut, with errors swallowed so push failures cannot fail the send. - Skip when there are no other human recipients or no sandboxId. - senderUserId = callerId for human senders, null for bot senders. - title is "<sandboxLabel> · <conversationTitle>"; bodyPreview is the first 200 chars of the concatenated text blocks. - Add @kilocode/notifications workspace dep and layer the RPC method shape into Env via bindings.d.ts. - Add a notifications-stub worker to the vitest config so tests can spy on env.NOTIFICATIONS.sendPushForConversation, and globally mock sandbox-lookup in setup.ts (it imports pg via @kilocode/db).
…es, fix test mock - Remove `stream-chat` from `services/notifications/package.json`; the Stream webhook (its only consumer) was deleted earlier in the stack. - Regenerate `worker-configuration.d.ts` so the workerd runtime types match the current toolchain (sibling services were on `1.20260312.1`; this one had drifted to `1.20251217.0` from a stale local cache). - Fix the global test mock to reference the renamed `badge_counts` table; the setup file was authored against the pre-rename name and never matched. - Tidy two pre-existing lint nits in the new test files (`import type` for type-only import, drop unused `cols` parameter).
…leak - Switch `NotificationsService` from default-only to a named class export with a separate default. `services/kilo-chat/wrangler.jsonc` binds via `entrypoint: "NotificationsService"`, which resolves named module exports. The default-only form (`export default class NotificationsService`) exports under the `default` key — kilo-chat's RPC binding would not have resolved at deploy. Mirrors the existing pattern in `services/kilo-chat/src/index.ts` (`KiloChatService`). - `dispatchPush` now uses a two-stage idempotency record (`pending` → `delivered`). The badge increment was previously non-idempotent: an Expo failure returned `failed` without writing the idempotency key, so upstream retries (which the design explicitly invites) re-ran the increment before the next send and inflated the badge by one per retry. The `pending` marker is written before the increment and short-circuits the increment on retry; the `delivered` marker is only written on success. - `setAlarm` is now gated on `getAlarm() === null`. Calling `setAlarm` unconditionally on each successful push — as the previous code did — replaces the pending alarm and pushes the cleanup forward indefinitely on a conversation receiving more than one push per `IDEM_TTL_MS`, leaking expired idempotency entries. Adds two test cases covering the badge-retry and alarm-reset paths.
- Schedule the cleanup alarm when writing the `pending` marker, not only on `delivered`. Without this, an Expo failure followed by no further push activity for the conversation leaves the `pending` record in DO storage forever (no alarm was ever set to prune it). - After the alarm fires, reschedule for the earliest remaining record's expiry instead of leaving the alarm slot empty. Otherwise a quiet conversation strands its younger entries until some unrelated future dispatch wakes the DO up. Both paths go through a small `ensureCleanupAlarm` helper that gates on `getAlarm() === null` so a busy conversation still doesn't push the alarm forward on every call.
The kiloclaw-scoped presence paths are literally `/presence` prefixed
onto the kiloclaw event-context paths. Build them by composition so the
`/kiloclaw/{sandboxId}[/{conversationId}]` segment shape is defined in
exactly one place — `kiloclaw-contexts.ts`.
Pure refactor; same string output, template-literal types still narrow
to the same shape.
Introduces a single app-shell EventServiceProvider that owns the EventServiceClient and KiloChatClient for all authenticated routes. Mounted in (app)/layout.tsx so platform/instance/conversation presence subscriptions and the kilo-chat UI share one WebSocket. KiloChatLayout now consumes the global clients via useEventServiceClient() instead of spinning up its own pair, and the getToken prop is removed from KiloChatLayoutProps (along with both call sites). The local useEventService(getToken) factory is dead code and has been deleted; useInstanceContext / useConversationContext stay since they take EventServiceClient as a parameter.
Thin hook that subscribes the global EventServiceClient to a single context for the lifetime of the calling component, gated by an `active` flag. Will back upcoming platform- and instance-level presence indicators.
…eSubscription - Drop dead getToken field from KiloChatContextValue (no consumers). - Remove useInstanceContext / useConversationContext hooks; both call sites now use the shared usePresenceSubscription primitive directly. - Harden usePresenceSubscription against empty-string contexts.
- usePresenceSubscription: accept 'string | null' instead of empty-string sentinel; update call sites (KiloChatLayout, MessageArea, useInstancePresence) - kilo-chat router: validate expiresAt with z.iso.datetime() - kilo-chat-router test: verify the JWT payload (kiloUserId, tokenSource, version) and that expiresAt lands in the expected ~1h window - MessageArea: comment distinguishing the always-on chat-event subscription from the visibility-gated presence subscription
Multiple consumers can now independently hold the same context without trampling each other. The wire context.subscribe/context.unsubscribe messages are only sent on the 0->1 and 1->0 refcount transitions; the intermediate churn stays client-side. Resubscribe-on-reconnect dedupes by context key. Tests cover: double-subscribe collapses to a single wire send, partial unsubscribe keeps the context alive, last-consumer-out releases it, mixed batches only send newly-active contexts, unknown-context unsubscribes are no-ops, and reconnect resubscribes each context once.
Contributor
Code Review SummaryStatus: No Issues Found | Recommendation: Merge Files Reviewed (2 files)
Reviewed by gpt-5.5-2026-04-23 · 116,399 tokens |
3 tasks
Base automatically changed from
feat/kilo-chat-migration-pr4
to
feat/kilo-chat-migration-pr1
April 30, 2026 13:05
…to feat/kilo-chat-migration-pr4-5 # Conflicts: # packages/notifications/src/badge-buckets.ts # services/kilo-chat/wrangler.jsonc
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Small SDK hardening pulled out of the kilo-chat migration stack as PR 4.5. Stacks on top of #2924.
EventServiceClient.subscribe/unsubscribepreviously stored active contexts in aSet<string>, so two consumers subscribing to the same context and one of them unsubscribing would yank the subscription out from under the other. We don't hit this today (PR 4's three presence tiers each use a distinct context) but PR 5+ will start composing subscriptions over the same context (e.g., a sidebar unread-counts hook subscribing tokiloclawInstanceContextwhileKiloChatLayoutalready does).Fix: change
activeContextstoMap<string, number>. Wirecontext.subscribeis sent only on the 0→1 transition;context.unsubscribeonly on 1→0.resubscribeContexts()dedupes by context key on reconnect.This makes
subscribe/unsubscribeidempotent and composable across consumers — the SDK owns the contract so callers (webusePresenceSubscription, the upcoming mobile equivalents) stay naive about it.Plan update
docs/superpowers/plans/2026-04-29-mobile-kilo-chat-migration.md(gitignored, local-only) updated:Test plan
cd packages/event-service && pnpm test— 17 passed (12 existing + 5 new refcount tests + 1 sanity).cd packages/event-service && pnpm run typecheck— clean.cd packages/event-service && pnpm run lint— clean.cd services/event-service && pnpm test— 16 passed (consumer of the SDK; behavior unchanged for single-consumer call sites).pnpm run format:changed.Stack
PR #2907 → #2914 → #2918 → #2924 (PR 4) → this PR.