fix(channels): add periodic sweep for pending message compaction

Auto-compaction stopped working after server restarts because provider
resolution failure was silent and there was no retry mechanism. Add a
10-minute periodic sweep in the flusher loop that checks DB for groups
exceeding the compaction threshold, and warn when compaction provider
cannot be resolved at startup.

internal/channels/history.go

@@ -27,8 +27,9 @@ const maxHistoryKeys = 1000
 const DefaultGroupHistoryLimit = 100
 
 const (
-	flushInterval = 3 * time.Second // periodic flush interval
-	flushBatchMax = 20              // flush when buffer reaches this size
+	flushInterval        = 3 * time.Second  // periodic flush interval
+	flushBatchMax        = 20               // flush when buffer reaches this size
+	compactSweepInterval = 10 * time.Minute // periodic compaction sweep for post-restart safety
 )
 
 // HistoryEntry represents a single tracked group message.

internal/channels/history_compaction.go

@@ -129,6 +129,37 @@ func CompactGroup(ctx context.Context, s store.PendingMessageStore, channelName,
 	return remaining, nil
 }
 
+// sweepCompaction checks DB for groups exceeding compaction threshold.
+// Called periodically from flushLoop as a safety net for post-restart scenarios
+// where RAM is empty but DB has accumulated messages.
+func (ph *PendingHistory) sweepCompaction() {
+	cfg := ph.compactionCfg
+	if cfg == nil || cfg.Provider == nil || ph.store == nil {
+		return
+	}
+	threshold := cfg.Threshold
+	if threshold <= 0 {
+		threshold = DefaultGroupHistoryLimit
+	}
+
+	ctx, cancel := context.WithTimeout(ph.tenantCtx(), 30*time.Second)
+	defer cancel()
+
+	groups, err := ph.store.ListGroups(ctx)
+	if err != nil {
+		slog.Warn("compaction.sweep_failed", "channel", ph.channelName, "error", err)
+		return
+	}
+	for _, g := range groups {
+		if g.ChannelName != ph.channelName {
+			continue
+		}
+		if g.MessageCount > threshold {
+			ph.MaybeCompact(g.HistoryKey, g.MessageCount, cfg)
+		}
+	}
+}
+
 // runCompaction performs LLM-based summarization of old messages.
 // Wraps CompactGroup with flush + RAM update + threshold check.
 func (ph *PendingHistory) runCompaction(historyKey string, cfg *CompactionConfig) {

internal/channels/history_flush.go

@@ -73,11 +73,15 @@ func (ph *PendingHistory) signalFlush() {
 }
 
 // flushLoop runs in background, flushing buffer periodically or when signaled.
+// Also runs periodic compaction sweep as a safety net for post-restart scenarios.
 func (ph *PendingHistory) flushLoop() {
 	defer close(ph.stopped)
 	ticker := time.NewTicker(flushInterval)
 	defer ticker.Stop()
 
+	compactTicker := time.NewTicker(compactSweepInterval)
+	defer compactTicker.Stop()
+
 	for {
 		select {
 		case <-ph.stopCh:
@@ -87,6 +91,8 @@ func (ph *PendingHistory) flushLoop() {
 			ph.flushNow()
 		case <-ticker.C:
 			ph.flushNow()
+		case <-compactTicker.C:
+			ph.sweepCompaction()
 		}
 	}
 }

internal/channels/instance_loader.go

@@ -282,6 +282,16 @@ func (l *InstanceLoader) loadInstance(ctx context.Context, inst store.ChannelIns
 			pc.SetPendingCompaction(cc)
 			slog.Debug("pending compaction configured", "channel", inst.Name, "provider", p.Name(), "model", model,
 				"threshold", cc.Threshold, "keep_recent", cc.KeepRecent, "max_tokens", cc.MaxTokens)
+		} else {
+			attemptedProvider := ""
+			if l.pendingCompactCfg != nil {
+				attemptedProvider = l.pendingCompactCfg.Provider
+			}
+			if attemptedProvider == "" && ag != nil {
+				attemptedProvider = ag.Provider
+			}
+			slog.Warn("pending compaction not configured: provider/model unavailable",
+				"channel", inst.Name, "agent_id", inst.AgentID, "attempted_provider", attemptedProvider)
 		}
 	}
 	l.manager.RegisterChannel(inst.Name, ch)