diff --git a/cmd/olliesrv/internal/agent/history.go b/cmd/olliesrv/internal/agent/history.go
index 5173e51..0fdbbfb 100644
--- a/cmd/olliesrv/internal/agent/history.go
+++ b/cmd/olliesrv/internal/agent/history.go
@@ -141,28 +141,6 @@ func (s *History) update(assistant backend.Message, results []toolResult) {
 	}
 }
 
-// llmSummarizeToolResult uses the LLM to produce a concise summary of a tool result.
-func llmSummarizeToolResult(ctx context.Context, b backend.Backend, name, content string) string {
-	prompt := fmt.Sprintf("Summarize this tool result from %q in 1-3 sentences, preserving key facts, file paths, values, and findings:\n\n%s", name, content)
-	ch, err := b.ChatStream(ctx, []backend.Message{{Role: "user", Content: prompt}}, nil, backend.GenerationParams{})
-	if err != nil {
-		return content
-	}
-	var sb strings.Builder
-	for ev := range ch {
-		if ev.Content != "" {
-			sb.WriteString(ev.Content)
-		}
-		if ev.Done {
-			break
-		}
-	}
-	if s := strings.TrimSpace(sb.String()); s != "" {
-		return fmt.Sprintf("[summary of %s result]: %s", name, s)
-	}
-	return content
-}
-
 // removeCancelledToolResults filters out tool results that were cancelled due
 // to interrupt, keeping completed work. Also removes the corresponding tool
 // calls from assistant messages to maintain a valid message sequence.
@@ -371,6 +349,21 @@ func flattenToolMessages(messages []backend.Message) []backend.Message {
 	return out
 }
 
+// injectPreamble ensures the preamble system message is at history position 0.
+// If a system message already exists there, its Content is replaced; otherwise
+// one is prepended.
+func (s *History) injectPreamble(preamble string) {
+	if preamble == "" {
+		return
+	}
+	msg := backend.Message{Role: "system", Content: preamble}
+	if len(s.messages) > 0 && s.messages[0].Role == "system" {
+		s.messages[0] = msg
+	} else {
+		s.messages = append([]backend.Message{msg}, s.messages...)
+	}
+}
+
 func (s *History) appendUserMessage(content string) {
 	s.messages = append(s.messages, backend.Message{Role: "user", Content: content})
 }
@@ -417,16 +410,62 @@ func (s *History) estimateTokens() int {
 }
 
 // stripCold summarizes large tool-result messages outside the hot tail
-// using an LLM call. Messages in the last hotTailSize slots are left verbatim.
+// using a single batched LLM call. Messages in the last hotTailSize slots
+// are left verbatim.
 func (s *History) stripCold(ctx context.Context, b backend.Backend) {
 	hot := len(s.messages) - hotTailSize
+	if hot < 0 {
+		return
+	}
+
+	// Collect cold-zone tool results exceeding 200 chars.
+	type item struct {
+		id     int
+		idx    int // index into s.messages
+		orig   string
+	}
+	var items []item
 	for i := range s.messages {
 		if i >= hot {
 			break
 		}
 		m := &s.messages[i]
 		if m.Role == "tool" && len(m.Content) > 200 {
-			m.Content = llmSummarizeToolResult(ctx, b, m.ToolCallID, m.Content)
+			items = append(items, item{id: len(items), idx: i, orig: m.Content})
+		}
+	}
+	if len(items) == 0 {
+		return
+	}
+
+	// Build a single prompt listing each item.
+	var sb strings.Builder
+	sb.WriteString("Compress each of the following tool results to 1-3 sentences preserving key facts, file paths, values, and findings.\n\nReturn a JSON object where each key is the id and the value is the summary, for example:\n{\"0\": \"summarized result 0\", \"1\": \"summarized result 1\"}\n\nOnly include keys that were in the input.\n\n")
+	for _, it := range items {
+		fmt.Fprintf(&sb, "---\nid=%d:\n%s\n", it.id, it.orig)
+	}
+
+	ch, err := b.ChatStream(ctx, []backend.Message{{Role: "user", Content: sb.String()}}, nil, backend.GenerationParams{})
+	if err != nil {
+		return
+	}
+	var resp strings.Builder
+	for ev := range ch {
+		if ev.Content != "" {
+			resp.WriteString(ev.Content)
+		}
+		if ev.Done {
+			break
+		}
+	}
+	summaries := make(map[string]string)
+	json.Unmarshal([]byte(resp.String()), &summaries)
+
+	// Apply summaries back into history.
+	for _, it := range items {
+		key := fmt.Sprintf("%d", it.id)
+		if summary, ok := summaries[key]; ok && summary != "" {
+			s.messages[it.idx].Content = summary
 		}
 	}
 }
diff --git a/cmd/olliesrv/internal/agent/loop.go b/cmd/olliesrv/internal/agent/loop.go
index db57535..0a8e2ab 100644
--- a/cmd/olliesrv/internal/agent/loop.go
+++ b/cmd/olliesrv/internal/agent/loop.go
@@ -126,9 +126,6 @@ func (ag *Agent) run(ctx context.Context) error {
 		}
 
 		history := h.history()
-		if preamble := rt.PreambleString(); preamble != "" {
-			history = append([]backend.Message{{Role: "system", Content: preamble}}, history...)
-		}
 
 		responseID := paths.NewUUID()
 		content, reasoning, toolCalls, stopReason, err := ag.streamResponse(ctx, history, responseID, step)
diff --git a/cmd/olliesrv/internal/agent/turn.go b/cmd/olliesrv/internal/agent/turn.go
index 2c1cc8e..4da4f58 100644
--- a/cmd/olliesrv/internal/agent/turn.go
+++ b/cmd/olliesrv/internal/agent/turn.go
@@ -197,6 +197,11 @@ func (ag *Agent) executeTurn(ctx context.Context, input string) string {
 		ag.history.resetTurnAccumulators()
 	}
 
+	// Inject the preamble system message into history before running.
+	if ag.history != nil {
+		ag.history.injectPreamble(ag.runtime.PreambleString())
+	}
+
 	// Run the turn, retrying once after compaction on context overflow.
 	var (
 		overflowRetried bool
@@ -229,6 +234,8 @@ func (ag *Agent) executeTurn(ctx context.Context, input string) string {
 			actCtx, actCancel = context.WithCancelCause(ctx)
 			handle = &actionHandle{cancel: actCancel}
 			ag.currentAction.Store(handle)
+			// Re-inject preamble — compaction strips it from history.
+			ag.history.injectPreamble(ag.runtime.PreambleString())
 			continue
 		}
 		break