refactor: move stripCold to executeTurn, replace clone with rollback pointer

- Remove per-step stripCold from run(); call once per turn in executeTurn
  after accumulator reset and before preamble injection.
- Replace expensive deep-clone snapshot with pre-turn length truncation.
- Delete unused cloneMessages() helper.

cmd/olliesrv/internal/agent/history.go

@@ -368,35 +368,6 @@ func (s *History) appendUserMessage(content string) {
 	s.messages = append(s.messages, backend.Message{Role: "user", Content: content})
 }
 
-// cloneMessages returns a deep copy of the message slice.
-func cloneMessages(msgs []backend.Message) []backend.Message {
-	out := make([]backend.Message, len(msgs))
-	for i, m := range msgs {
-		out[i] = backend.Message{
-			ID:         m.ID,
-			Role:       m.Role,
-			Content:    m.Content,
-			Reasoning:  m.Reasoning,
-			ToolCallID: m.ToolCallID,
-		}
-		if len(m.ContentBlocks) > 0 {
-			out[i].ContentBlocks = make([]backend.ContentBlock, len(m.ContentBlocks))
-			copy(out[i].ContentBlocks, m.ContentBlocks)
-		}
-		if len(m.ToolCalls) > 0 {
-			out[i].ToolCalls = make([]backend.ToolCall, len(m.ToolCalls))
-			for j, tc := range m.ToolCalls {
-				out[i].ToolCalls[j] = backend.ToolCall{
-					ID:        tc.ID,
-					Name:      tc.Name,
-					Arguments: append(json.RawMessage(nil), tc.Arguments...),
-				}
-			}
-		}
-	}
-	return out
-}
-
 // estimateTokens returns a rough token count (~4 chars per token).
 func (s *History) estimateTokens() int {
 	chars := 0

cmd/olliesrv/internal/agent/loop.go

@@ -121,9 +121,6 @@ func (ag *Agent) run(ctx context.Context) error {
 
 	for {
 		ag.emit(Event{Role: "state", Content: "thinking"})
-		if budget := contextBudget(ctx, rt.Backend); budget > 0 && h.estimateTokens() > budget {
-			h.stripCold(ctx, rt.Backend)
-		}
 
 		history := h.history()
 

cmd/olliesrv/internal/agent/turn.go

@@ -88,9 +88,9 @@ func (ag *Agent) executeTurn(ctx context.Context, input string) string {
 
 	// Snapshot messages before this turn modifies them. Restored on context
 	// overflow so the retry starts from a clean state.
-	var snapMessages []backend.Message
+	var preTurnLen int
 	if ag.history != nil {
-		snapMessages = cloneMessages(ag.history.messages)
+		preTurnLen = len(ag.history.messages)
 	}
 
 	if ag.history == nil {
@@ -197,6 +197,11 @@ func (ag *Agent) executeTurn(ctx context.Context, input string) string {
 		ag.history.resetTurnAccumulators()
 	}
 
+	// Strip cold-zone tool results (one batched LLM call per turn).
+	if ag.history != nil {
+		ag.history.stripCold(ctx, ag.runtime.Backend)
+	}
+
 	// Inject the preamble system message into history before running.
 	if ag.history != nil {
 		ag.history.injectPreamble(ag.runtime.PreambleString())
@@ -221,7 +226,7 @@ func (ag *Agent) executeTurn(ctx context.Context, input string) string {
 		var ctxErr *backend.ContextOverflowError
 		if !overflowRetried && errors.As(err, &ctxErr) && ag.history != nil {
 			overflowRetried = true
-			ag.history.messages = snapMessages
+			ag.history.messages = ag.history.messages[:preTurnLen]
 			ag.emit(Event{Role: "info", Content: "context overflow — compacting and retrying...\n"})
 			ag.SetState("compacting")
 			if _, cerr := ag.runCompact(ctx, "overflow"); cerr != nil {