0094 — Delete the conversation reconciler sweeps: the turn-settled stream is the sole path

• Status: Accepted
• Date: 2026-07-16
• Deciders: Gera
• Relates:
  • ADR-0088 — the durable conversation.turn_settled event stream. This ADR replaces 0088's PR-3 ("relax the reconcilers to hourly, retire after a further clean month") with an immediate full deletion.
  • ADR-0084 — inline topic stamping. Its reconciler half is deleted; its classify-and-stamp logic is unchanged.
  • ADR-0079 — single canonical writer; the D4 sweep-side regarding-reconcile rode the topics sweep and is deleted with it.

Decision

Delete BOTH conversation "reconciler" backstop sweeps entirely — the topic sweep (conversationTopicClassificationWorkflow / classifyRecentConversationTopics, Schedule conversation-topics-reconciler) and the grade sweep (conversationGradeSweepWorkflow / gradeRecentConversations, Schedule conversation-grader-reconciler) — while keeping the always-on real-time inline path untouched.

"No reconcilers, no arms, no flags — always stamp, always grade, always topic." The ADR-0088 event stream is verified live in prod: conversationTopicStampWorkflow + conversationGraderWorkflow execute per settled turn on ALL channels (SMS/email via the inbound-processor Lambda, voice call-ended, dashboard/simulator). The stream carries durability (at-least-once SQS FIFO + DLQ alarm), so the 15-minute sweeps were pure redundancy — the reconcilerCaughtMissed retirement instrument existed to prove exactly this.

What was deleted

• The sweep workflows (conversationGradeSweepWorkflow, conversationTopicClassificationWorkflow) + their worker/barrel registrations.
• The sweep activities (gradeRecentConversationsActivity, classifyRecentConversationTopicsActivity) + their withAutomationRun wraps and config/automations.toml job entries.
• The Temporal Schedule upsert halves of both clients (upsertConversationGradeSchedule, upsertConversationTopicsSchedule, CONVERSATION_GRADE_SCHEDULE_ID, CONVERSATION_TOPICS_SCHEDULE_ID, RECONCILER_INTERVAL, RECONCILER_WINDOW_HOURS, RETIRED_NIGHTLY_SCHEDULE_ID, deleteRetiredNightlySchedule) + the two operator upsert scripts, and the conversation-topics-reconciler entry in the ADR-0045 namespace-prep contract.
• The domain sweep functions (gradeRecentConversations, classifyRecentConversationTopics, conversationsToGrade, conversationsInRecencyWindow) and the reconcilerCaughtMissed / reconciler_caught_missed_conversation retirement counter everywhere.
• The D4 sweep-side refuse-to-guess regarding reconcile (reconcileConversationRegarding, decideRegardingReconcile) + its single_writer_regarding_unresolvable bake-alert kind. The INLINE backstamp (backstampWorkOrderReportSegment: bounded retry + the loud single_writer_regarding_tie_failed marker) is the sole tie path.

What was kept (100% intact)

• conversationTopicStampWorkflow + stampConversationTopicsActivity + classifyAndStampConversation (inline topic stamp, 10s debounce as of 2026-08-28 — see ADR-0084's amendment note).
• conversationGraderWorkflow + gradeConversationActivity + gradeAndAlertConversation + the RED→#alerts rubric (conversation_grade_red) + the redAlertedAt/gradedAt claim/watermark dedup.
• The bridge consumers (topic-stamp, grade, maintenance-dirty-bits) in lambda/conversation-events-bridge/registry.ts and the client signal fns they call (signalUntaggedMessagesArrived, signalConversationTurnSettled).

Accepted trades (named, not hidden)

1. Crash-before-publish drop is not retried. A settle edge lost before the awaited SQS publish (or swallowed after its one retry) gets no sweep re-cover. The thread's NEXT turn re-emits and the untagged rows / ungraded state are picked up then; a thread that never speaks again keeps its last state.
2. No historical backfill lane. The wide-window sweep invocation (manual windowHours override) is gone. A one-off backfill, if ever needed, is a script over classifyAndStampConversation — not a standing reconciler.
3. D4 sweep reconcile gone. Untied ≥24h rows that the inline backstamp's retry could not tie are found via the single_writer_regarding_tie_failed log marker, not an alert-driven sweep.

Ops follow-ups (not in this PR)

• Delete the two live Temporal Cloud Schedules (conversation-topics-reconciler, conversation-grader-reconciler) on propflow-renewal-prod — the code deletion makes them dead letters (their workflow types no longer register on the workers).
• tools/agent-smith's reconciler-caught-missed-clean-week bake checker is now dead (its counter no longer exists) — remove it in a Smith-side change.