Repository navigation
feat(agents): publish synthesized agent events through canonical - #1778
Merged
Merged
Conversation
A half-open socket answers nothing and never fires close, so every recovery mechanism (reconnect, hydration, subscription replay) waited forever while the UI showed stale streaming state. A 30s interval now issues the existing app.version RPC with a 10s deadline; an unanswered probe closes the socket locally and the established onclose ladder owns recovery. agent.send gains a 20s timeout so a composer submit cannot park forever on a dead socket. The wedged-socket test now answers the liveness probe, since a socket that ignores it is exactly what the watchdog closes.
The reducer now writes routing.executionId onto the stored turn so clients can correlate a canonical turn to the local runtime without threading the event envelope through state. A payload-supplied executionId that disagrees with routing is rejected as a routing conflict.
Composer, project tree, and running-thread indicators now follow the canonical replica once a turn correlates by execution identity, so a dropped legacy terminal event can no longer strand runningThreadIds. The optimistic window before turnStarted and the finalizing phase stay on the legacy path until a canonical turn proves correlation. The narrative projection keeps its existing child-only lifecycle gate.
The relay opened one socket and never recovered: a dropped connection silently degraded the renderer to WebSocket-only delivery for the rest of the session. Reconnect in the main process with bounded exponential backoff, reset on successful connect, and stop when the window dies or the relay is torn down. The renderer re-suppresses push channels as frames resume, so no reconnect handshake is needed.
No caller remains: consumers read runningThreadIds directly.
Parent narrative recovery wrote items directly into canonical_agent_items, so subscribed replicas only learned about them through snapshots. Each mutation now commits an item.recorded envelope inside the existing elapsed-bounded batch writer, discards become narrativeRecoveryDiscarded tombstones, and the semantic writer flushes the envelopes through its buffered publication channel. The web store reruns recoverParentNarrative when a canonical batch touches narrative items, letting parent tool calls and segments reach the chat through the canonical stream.
Numbered livePublication intents are the enriched, sanitized events the agent.event broadcast sends today. Recording each one as a canonical publication.recorded envelope inside the same operation transaction makes the canonical stream carry the same renderer-facing payload, so replay and gap recovery deliver identical effects without the legacy channel. The reducer treats the envelope as a passthrough: identity and dedup bookkeeping apply, but no thread or item state mutates. On the client, canonical pushes and reconnect deltas dispatch the embedded event through handleAgentEvent with its publicationId, so the existing stable publication cursor accepts whichever copy arrives first and drops the other. Envelopes committed by unbuffered operation paths (control, stage-terminal, assistant text, terminal batches) queue on the writer and flush after the outer transaction commits, matching the buffered flush used by live operations.
Live publication intents embed provider events before the public enrichment stage, so canonical copies would persist raw tool inputs that the legacy wire path strips. Sanitize the embedded event the same way AgentEventPublicationService does before recording publication.recorded.
Goals, provider-unavailable notices, post-terminal hooks, control replies, and materialized assistant messages bypassed canonical durability and reached clients only through fire-and-forget agent.event broadcasts. Route them through recordSynthesizedPublications, which commits publication.recorded envelopes on the shared thread-scoped publication head so replay and reconnect recovery deliver them. Synthesized events carry no execution identity, so they publish under a reserved sentinel execution with persistCheckpoint disabled, and the event store seeds accepted sequences from prior events when no checkpoint exists. The client publication cursor dedups on the thread-scoped publicationId alone since turnExecutionId never participated in reserve. The legacy agent.event copy still emits with the same publicationId so whichever copy arrives first wins while the client migrates.
…hase The runtime reconcile stamps runtimePhase from a correlated canonical turn but never correlated the record's execution identity, so a canonically owned "running" phase made getCanonicalLifecycleTurn treat the provider-owned child turn as a local execution and suppress its projection. A later terminal canonical event could then no longer clear the stamped phase either. reconcileCanonicalRuntime now stamps turnExecutionId from the claiming turn's executionId on live claims, and the lifecycle gate only treats runtimePhase as locally owned when the canonical runtime turn is not the latest turn.
# Conflicts: # apps/web/src/stores/threadStore.ts
assertActiveTurnRecoveryRetention measured the serialized canonical envelope drafts, letting transport overhead trip the cap before the retained narrative itself did. Measure the persisted item payloads and discarded item ids again, matching the pre-envelope accounting.
The byte-size computation pushed prepareParentNarrativeRecovery over the complexity cap. Hoisting it keeps the retention accounting intact.
chuks-qua
changed the base branch from
feat/canonical-publication-events
to
main
September 29, 2026 21:53
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.
What
Server-synthesized
AgentEvents (goal updates and clears,/goalcontrol replies, provider-unavailable notices, post-terminal hook completions, and materialized assistant message fallbacks) reached clients only through fire-and-forgetagent.eventbroadcasts with no replay or reconnect recovery.They now publish through
CanonicalAgentBoundary.recordSynthesizedPublications, which commitspublication.recordedenvelopes on the same thread-scopedcanonical_writer_live_publication_headssequence the semantic writer uses, so writer and synthesized publication ids cannot collide. A legacyagent.eventcopy still emits with the samepublicationId; the client dedups on whichever arrives first.Synthesized events belong to no provider execution, so:
CANONICAL_SYNTHESIZED_EXECUTION_ID) withpersistCheckpoint: false; the event store seeds the accepted sequence from prior event rows when no checkpoint exists.AgentEventis stamped withpublicationIdonly. A sentinelturnExecutionIdwould fail the client's execution-correlation preflight and drop the event mid-turn.StableAgentEventPublications.acceptno longer requiresturnExecutionId; the thread-scopedpublicationIdwas already the sole dedup key inreserve.Why
Every renderer-facing publication must flow through the durable canonical stream so a lost or replayed connection rebuilds the same state the legacy broadcast would have shown. Synthesized events were the last emitters with no canonical counterpart; without this, deleting
agent.eventwould lose goals, banners, and control replies on any dropped event.Stacked on
feat/canonical-publication-events(#1777), which adds thepublication.recordedvariant and client dispatch this unit relies on.UI Changes
None.
Config Changes
None.
Review Notes
apps/serverorchestration + turns suites (33 files, 459 tests), canonical boundary suite including a new synthesized-sequencing test,bun run typecheckclean inapps/serverandapps/web, and web cursor/dispatch tests (12 tests).persistCheckpoint: falsekeeps commit bookkeeping out of turn state.HookCompleted) keep their realturnExecutionIdwhen available.agent.eventremoval stays in the final stack PR once all emitters are canonical.Generated with Devin
Need help on this PR? Tag
@codesmith-botwith what you need. Autofix is disabled.