fix(tools/fetch): cap fetch tool at 50K tokens

Christian Rocha created 

This prevents the fetch tool from blowing through context windows. Prior
to this the limit was 5MB, or about 390K tokens.

For larger fetches, the agent is urged to use agentic_fetch, which
performs the fetch in an isolated context window.

Change summary

internal/agent/tools/fetch.go | 12 +++++++++---
1 file changed, 9 insertions(+), 3 deletions(-)

Detailed changes

internal/agent/tools/fetch.go 🔗 

@@ -160,10 +160,16 @@ func NewFetchTool(permissions permission.Service, workingDir string, client *htt
 					content = "<html>\n<body>\n" + body + "\n</body>\n</html>"
 				}
 			}
+			// maxFetchContentSize is the maximum size of content to return
+			// (200KB). This limit ensures content fits within model context
+			// windows. 200KB is about 50K tokens, which should leave more room
+			// for system prompts and conversation history.
+			const maxFetchContentSize = 200 * 1024
+
 			// truncate content if it exceeds max read size
-			if int64(len(content)) > MaxReadSize {
-				content = content[:MaxReadSize]
-				content += fmt.Sprintf("\n\n[Content truncated to %d bytes]", MaxReadSize)
+			if len(content) > maxFetchContentSize {
+				content = content[:maxFetchContentSize]
+				content += fmt.Sprintf("\n\n[Content truncated to %d bytes. The original page was larger: consider using agentic_fetch for large pages.]", maxFetchContentSize)
 			}
 
 			return fantasy.NewTextResponse(content), nil