From 32180bee4d4ab5673e59f7aa160fb3d3f24621ba Mon Sep 17 00:00:00 2001
From: Ollie Agent <ollie@example.com>
Date: Sat, 15 Aug 2026 13:27:16 +0200
Subject: [PATCH] defer cold summaries until needed

---
 cmd/olliesrv/internal/agent/history.go        | 27 +++++++++++-
 .../internal/agent/history_trigger_test.go    | 44 +++++++++++++++++++
 cmd/olliesrv/internal/agent/turn.go           | 17 +++++--
 3 files changed, 83 insertions(+), 5 deletions(-)
 create mode 100644 cmd/olliesrv/internal/agent/history_trigger_test.go

diff --git a/cmd/olliesrv/internal/agent/history.go b/cmd/olliesrv/internal/agent/history.go
index 39cbb09..5eb1746 100644
--- a/cmd/olliesrv/internal/agent/history.go
+++ b/cmd/olliesrv/internal/agent/history.go
@@ -413,7 +413,32 @@ func (s *History) estimateTokens() int {
 	return chars / 4
 }
 
-const maxColdSummariesPerCall = 8
+const (
+	maxColdSummariesPerCall = 8
+	minColdSummaries        = 3
+	minColdSummaryTokens    = 32_000
+)
+
+func (s *History) pendingColdSummaryStats() (count, tokens int) {
+	hot := len(s.messages) - hotTailSize
+	if hot <= 0 {
+		return 0, 0
+	}
+	for i := 0; i < hot; i++ {
+		m := &s.messages[i]
+		if m.Role != "tool" || len(m.Content) <= 200 || m.SummaryApplied {
+			continue
+		}
+		if s.summaryCache != nil {
+			if _, ok := s.summaryCache[toolSummaryHash(*m)]; ok {
+				continue
+			}
+		}
+		count++
+		tokens += len(m.Content) / 4
+	}
+	return count, tokens
+}
 
 // stripCold summarizes large tool-result messages outside the hot tail
 // using a single batched LLM call. Messages in the last hotTailSize slots
diff --git a/cmd/olliesrv/internal/agent/history_trigger_test.go b/cmd/olliesrv/internal/agent/history_trigger_test.go
new file mode 100644
index 0000000..0c79160
--- /dev/null
+++ b/cmd/olliesrv/internal/agent/history_trigger_test.go
@@ -0,0 +1,44 @@
+package agent
+
+import (
+	"strings"
+	"testing"
+
+	"ollie/cmd/olliesrv/internal/backend"
+)
+
+func TestPendingColdSummaryStatsRequiresCountAndTokens(t *testing.T) {
+	h := newHistory("test")
+	for i := 0; i < minColdSummaries; i++ {
+		h.messages = append(h.messages, backend.Message{Role: "tool", ToolCallID: "tc", Content: strings.Repeat("result ", minColdSummaryTokens/minColdSummaries/2+1000)})
+	}
+	for i := 0; i < hotTailSize; i++ {
+		h.messages = append(h.messages, backend.Message{Role: "user", Content: "hot"})
+	}
+	count, tokens := h.pendingColdSummaryStats()
+	if count != minColdSummaries {
+		t.Fatalf("pending count = %d, want %d", count, minColdSummaries)
+	}
+	if tokens < minColdSummaryTokens {
+		t.Fatalf("pending tokens = %d, want at least %d", tokens, minColdSummaryTokens)
+	}
+}
+
+func TestPendingColdSummaryStatsSkipsCacheAndApplied(t *testing.T) {
+	h := newHistory("test")
+	for i := 0; i < hotTailSize; i++ {
+		h.messages = append(h.messages, backend.Message{Role: "user", Content: "hot"})
+	}
+	cached := backend.Message{Role: "tool", ToolCallID: "cached", Content: strings.Repeat("cached ", 100)}
+	applied := backend.Message{Role: "tool", ToolCallID: "applied", Content: strings.Repeat("applied ", 100), SummaryApplied: true}
+	pending := backend.Message{Role: "tool", ToolCallID: "pending", Content: strings.Repeat("pending ", 100)}
+	h.messages = append(h.messages, cached, applied, pending)
+	for i := 0; i < hotTailSize; i++ {
+		h.messages = append(h.messages, backend.Message{Role: "user", Content: "hot"})
+	}
+	h.summaryCache[toolSummaryHash(cached)] = "cached summary"
+	count, _ := h.pendingColdSummaryStats()
+	if count != 1 {
+		t.Fatalf("pending count = %d, want 1", count)
+	}
+}
diff --git a/cmd/olliesrv/internal/agent/turn.go b/cmd/olliesrv/internal/agent/turn.go
index a002d85..303064b 100644
--- a/cmd/olliesrv/internal/agent/turn.go
+++ b/cmd/olliesrv/internal/agent/turn.go
@@ -204,10 +204,19 @@ func (ag *Agent) executeTurn(ctx context.Context, input string) string {
 
 	// Strip cold-zone tool results (one batched LLM call per turn).
 	if ag.history != nil {
-		usage, estimated := ag.history.stripCold(actCtx, ag.runtime.Backend)
-		if usage.InputTokens > 0 || usage.OutputTokens > 0 {
-			ag.history.addUsage(usage, estimated)
-			ag.notifyChange()
+		pendingCount, pendingTokens := ag.history.pendingColdSummaryStats()
+		ctxLen := ag.runtime.Backend.ContextLength(actCtx)
+		if ctxLen <= 0 {
+			ctxLen = defaultContextLength
+		}
+		ctxTokens := ag.history.estimateTokens()
+		forceStripCold := ctxTokens*100 >= ctxLen*40
+		if (pendingCount >= minColdSummaries && pendingTokens >= minColdSummaryTokens) || forceStripCold {
+			usage, estimated := ag.history.stripCold(actCtx, ag.runtime.Backend)
+			if usage.InputTokens > 0 || usage.OutputTokens > 0 {
+				ag.history.addUsage(usage, estimated)
+				ag.notifyChange()
+			}
 		}
 	}