fix(agent): prevent "..." fallback when iteration budget exhausted

Add early detection and graceful degradation when the model exhausts its
iteration budget without producing a text response:

- Inject 75% budget nudge telling model to start summarizing
- Strip all tool schemas on final iteration to force text-only response
- Adaptive web_fetch maxChars (60K→20K→10K) via iteration progress context
- Bump DefaultMaxIterations 20→30 for web-research workloads

internal/agent/loop.go

@@ -584,6 +584,21 @@ func (l *Loop) runLoop(ctx context.Context, req RunRequest) (*RunResult, error)
 			messages = append(messages, providers.Message{Role: "user", Content: nudge})
 		}
 
+		// Iteration budget nudge: when model has used 75% of iterations without
+		// producing any text response, warn it to start summarizing.
+		if maxIter > 0 && iteration > 1 && iteration == maxIter*3/4 && finalContent == "" {
+			messages = append(messages, providers.Message{
+				Role:    "user",
+				Content: "[System] You have used 75% of your iteration budget without providing a text response. Start summarizing your findings and respond to the user within the next few iterations.",
+			})
+		}
+
+		// Inject iteration progress into context so tools can adapt (e.g. web_fetch reduces maxChars).
+		iterCtx := tools.WithIterationProgress(ctx, tools.IterationProgress{
+			Current: iteration,
+			Max:     maxIter,
+		})
+
 		// Emit activity event: thinking phase
 		emitRun(AgentEvent{
 			Type:    protocol.AgentEventActivity,
@@ -645,6 +660,16 @@ func (l *Loop) runLoop(ctx context.Context, req RunRequest) (*RunResult, error)
 			toolDefs = filtered
 		}
 
+		// Final iteration: strip all tools to force a text-only response.
+		// Without this the model may keep requesting tools and exit with "...".
+		if iteration == maxIter {
+			toolDefs = nil
+			messages = append(messages, providers.Message{
+				Role:    "user",
+				Content: "[System] Final iteration reached. Summarize all findings and respond to the user now. No more tool calls allowed.",
+			})
+		}
+
 		// Use per-request model override if set (e.g. heartbeat uses cheaper model).
 		model := l.model
 		if req.ModelOverride != "" {
@@ -924,7 +949,7 @@ func (l *Loop) runLoop(ctx context.Context, req RunRequest) (*RunResult, error)
 				}
 			}
 			if result == nil {
-				result = l.tools.ExecuteWithContext(ctx, registryName, tc.Arguments, req.Channel, req.ChatID, req.PeerKind, req.SessionKey, nil)
+				result = l.tools.ExecuteWithContext(iterCtx, registryName, tc.Arguments, req.Channel, req.ChatID, req.PeerKind, req.SessionKey, nil)
 			}
 			stopSlowTimer()
 
@@ -1100,7 +1125,7 @@ func (l *Loop) runLoop(ctx context.Context, req RunRequest) (*RunResult, error)
 						}
 					}
 					if result == nil {
-						result = l.tools.ExecuteWithContext(ctx, registryName, tc.Arguments, req.Channel, req.ChatID, req.PeerKind, req.SessionKey, nil)
+						result = l.tools.ExecuteWithContext(iterCtx, registryName, tc.Arguments, req.Channel, req.ChatID, req.PeerKind, req.SessionKey, nil)
 					}
 					stopSlowTimer()
 					l.emitToolSpanEnd(ctx, spanID, spanStart, result)

internal/config/defaults.go

@@ -7,7 +7,7 @@ const (
 	DefaultContextWindow   = 200000
 	DefaultMaxTokens       = 8192
 	DefaultMaxMessageChars = 32000
-	DefaultMaxIterations   = 20
+	DefaultMaxIterations   = 30
 	DefaultTemperature     = 0.7
 	DefaultHistoryShare    = 0.75
 )

internal/tools/context_keys.go

@@ -412,6 +412,26 @@ func RunMediaNamesFromCtx(ctx context.Context) map[string]string {
 	return v
 }
 
+// --- Iteration progress (loop → tools) ---
+
+const ctxIterProgress toolContextKey = "tool_iter_progress"
+
+// IterationProgress carries the agent loop's current iteration state
+// so tools can adapt behaviour (e.g. reduce output size) as the budget shrinks.
+type IterationProgress struct {
+	Current int
+	Max     int
+}
+
+func WithIterationProgress(ctx context.Context, p IterationProgress) context.Context {
+	return context.WithValue(ctx, ctxIterProgress, p)
+}
+
+func IterationProgressFromCtx(ctx context.Context) (IterationProgress, bool) {
+	v, ok := ctx.Value(ctxIterProgress).(IterationProgress)
+	return v, ok
+}
+
 // --- Per-agent sandbox config override ---
 
 const ctxSandboxCfg toolContextKey = "tool_sandbox_config"

internal/tools/web_fetch.go

@@ -192,6 +192,17 @@ func (t *WebFetchTool) Execute(ctx context.Context, args map[string]any) *Result
 		maxChars = int(mc)
 	}
 
+	// Adaptive maxChars: reduce as iterations progress to prevent context bloat.
+	if prog, ok := IterationProgressFromCtx(ctx); ok && prog.Max > 0 {
+		ratio := float64(prog.Current) / float64(prog.Max)
+		switch {
+		case ratio >= 0.75:
+			maxChars = min(maxChars, 10000)
+		case ratio >= 0.50:
+			maxChars = min(maxChars, 20000)
+		}
+	}
+
 	// Check cache (scoped per channel to prevent cross-channel cache poisoning)
 	channel := ToolChannelFromCtx(ctx)
 	cacheKey := fmt.Sprintf("fetch:%s:%s:%s:%d", channel, rawURL, extractMode, maxChars)