diff --git a/pkg/agent/context.go b/pkg/agent/context.go
index 6fccbaf53..ab9c36d05 100644
--- a/pkg/agent/context.go
+++ b/pkg/agent/context.go
@@ -333,6 +333,8 @@ func skillFilesModifiedSince(skillsDir string, t time.Time) bool {
 	return changed
 }
 
+const maxBootstrapFileChars = 4096
+
 func (cb *ContextBuilder) LoadBootstrapFiles() string {
 	bootstrapFiles := []string{
 		"AGENTS.md",
@@ -345,7 +347,11 @@ func (cb *ContextBuilder) LoadBootstrapFiles() string {
 	for _, filename := range bootstrapFiles {
 		filePath := filepath.Join(cb.workspace, filename)
 		if data, err := os.ReadFile(filePath); err == nil {
-			fmt.Fprintf(&sb, "## %s\n\n%s\n\n", filename, data)
+			content := string(data)
+			if len(content) > maxBootstrapFileChars {
+				content = content[:maxBootstrapFileChars] + "\n... (truncated)"
+			}
+			fmt.Fprintf(&sb, "## %s\n\n%s\n\n", filename, content)
 		}
 	}
 
diff --git a/pkg/agent/loop.go b/pkg/agent/loop.go
index 8f315ece9..973a1a45a 100644
--- a/pkg/agent/loop.go
+++ b/pkg/agent/loop.go
@@ -969,8 +969,19 @@ func (al *AgentLoop) runLLMIteration(
 			}
 			messages = append(messages, toolResultMsg)
 
-			// Save tool result message to session
-			agent.Sessions.AddFullMessage(opts.SessionKey, toolResultMsg)
+			// Save a truncated version of tool results to session history.
+			// Full results are already in `messages` for the current iteration;
+			// persisting truncated versions prevents context bloat on reload.
+			truncatedContent := contentForLLM
+			const maxToolResultHistory = 500
+			if len(truncatedContent) > maxToolResultHistory {
+				truncatedContent = truncatedContent[:maxToolResultHistory] + fmt.Sprintf("... (truncated, %d chars total)", len(contentForLLM))
+			}
+			agent.Sessions.AddFullMessage(opts.SessionKey, providers.Message{
+				Role:       "tool",
+				Content:    truncatedContent,
+				ToolCallID: tc.ID,
+			})
 		}
 	}
 
@@ -1014,7 +1025,7 @@ func (al *AgentLoop) maybeSummarize(agent *AgentInstance, sessionKey, channel, c
 	tokenEstimate := al.estimateTokens(newHistory)
 	threshold := agent.ContextWindow * 75 / 100
 
-	if len(newHistory) > 20 || tokenEstimate > threshold {
+	if len(newHistory) > 12 || tokenEstimate > threshold {
 		summarizeKey := agent.ID + ":" + sessionKey
 		if _, loading := al.summarizing.LoadOrStore(summarizeKey, true); !loading {
 			go func() {
diff --git a/pkg/tools/filesystem.go b/pkg/tools/filesystem.go
index 62eb703c6..2ba1f041e 100644
--- a/pkg/tools/filesystem.go
+++ b/pkg/tools/filesystem.go
@@ -9,6 +9,7 @@ import (
 	"regexp"
 	"strings"
 	"time"
+	"unicode/utf8"
 
 	"github.com/sipeed/picoclaw/pkg/fileutil"
 )
@@ -119,6 +120,26 @@ func (t *ReadFileTool) Parameters() map[string]any {
 
 const maxReadFileChars = 10000
 
+func isBinaryContent(data []byte) bool {
+	if len(data) == 0 {
+		return false
+	}
+	sample := data
+	if len(sample) > 512 {
+		sample = sample[:512]
+	}
+	if !utf8.Valid(sample) {
+		return true
+	}
+	nullCount := 0
+	for _, b := range sample {
+		if b == 0 {
+			nullCount++
+		}
+	}
+	return nullCount > 0
+}
+
 func (t *ReadFileTool) Execute(ctx context.Context, args map[string]any) *ToolResult {
 	path, ok := args["path"].(string)
 	if !ok {
@@ -130,6 +151,10 @@ func (t *ReadFileTool) Execute(ctx context.Context, args map[string]any) *ToolRe
 		return ErrorResult(err.Error())
 	}
 
+	if isBinaryContent(content) {
+		return ErrorResult(fmt.Sprintf("binary file detected (%d bytes). Use exec to inspect binary files (e.g. file, xxd, strings)", len(content)))
+	}
+
 	text := string(content)
 	if len(text) > maxReadFileChars {
 		text = text[:maxReadFileChars] + fmt.Sprintf("\n\n... (truncated, showing %d of %d chars. Use exec with head/tail/sed for specific sections)", maxReadFileChars, len(content))
@@ -238,15 +263,25 @@ func (t *ListDirTool) Execute(ctx context.Context, args map[string]any) *ToolRes
 	return formatDirEntries(entries)
 }
 
+const maxDirEntries = 200
+
 func formatDirEntries(entries []os.DirEntry) *ToolResult {
 	var result strings.Builder
-	for _, entry := range entries {
-		if entry.IsDir() {
-			result.WriteString("DIR:  " + entry.Name() + "\n")
+	total := len(entries)
+	limit := total
+	if limit > maxDirEntries {
+		limit = maxDirEntries
+	}
+	for i := 0; i < limit; i++ {
+		if entries[i].IsDir() {
+			result.WriteString("DIR:  " + entries[i].Name() + "\n")
 		} else {
-			result.WriteString("FILE: " + entry.Name() + "\n")
+			result.WriteString("FILE: " + entries[i].Name() + "\n")
 		}
 	}
+	if total > maxDirEntries {
+		fmt.Fprintf(&result, "\n... (%d more entries not shown, total: %d)\n", total-maxDirEntries, total)
+	}
 	return NewToolResult(result.String())
 }
 
diff --git a/pkg/tools/subagent.go b/pkg/tools/subagent.go
index 69f1a49a2..0f490cff2 100644
--- a/pkg/tools/subagent.go
+++ b/pkg/tools/subagent.go
@@ -359,13 +359,18 @@ func (t *SubagentTool) Execute(ctx context.Context, args map[string]any) *ToolRe
 		userContent = userContent[:maxUserLen] + "..."
 	}
 
-	// ForLLM: Full execution details
+	// ForLLM: Execution details (truncated to control token usage)
 	labelStr := label
 	if labelStr == "" {
 		labelStr = "(unnamed)"
 	}
+	resultContent := loopResult.Content
+	const maxSubagentLLM = 3000
+	if len(resultContent) > maxSubagentLLM {
+		resultContent = resultContent[:maxSubagentLLM] + fmt.Sprintf("... (truncated, %d chars total)", len(loopResult.Content))
+	}
 	llmContent := fmt.Sprintf("Subagent task completed:\nLabel: %s\nIterations: %d\nResult: %s",
-		labelStr, loopResult.Iterations, loopResult.Content)
+		labelStr, loopResult.Iterations, resultContent)
 
 	return &ToolResult{
 		ForLLM:  llmContent,
diff --git a/pkg/tools/web.go b/pkg/tools/web.go
index 30ccfe603..8470f29dc 100644
--- a/pkg/tools/web.go
+++ b/pkg/tools/web.go
@@ -661,21 +661,17 @@ func (t *WebFetchTool) Execute(ctx context.Context, args map[string]any) *ToolRe
 		text = text[:maxChars]
 	}
 
-	result := map[string]any{
-		"url":       urlStr,
-		"status":    resp.StatusCode,
-		"extractor": extractor,
-		"truncated": truncated,
-		"length":    len(text),
-		"text":      text,
+	var llmContent string
+	if truncated {
+		llmContent = fmt.Sprintf("[%s %d truncated=%v]\n%s", urlStr, resp.StatusCode, truncated, text)
+	} else {
+		llmContent = fmt.Sprintf("[%s %d]\n%s", urlStr, resp.StatusCode, text)
 	}
 
-	resultJSON, _ := json.MarshalIndent(result, "", "  ")
-
 	return &ToolResult{
-		ForLLM: string(resultJSON),
+		ForLLM: llmContent,
 		ForUser: fmt.Sprintf(
-			"Fetched %d bytes from %s (extractor: %s, truncated: %v)",
+			"Fetched %d chars from %s (extractor: %s, truncated: %v)",
 			len(text),
 			urlStr,
 			extractor,