From bae87bfc47b6f76c3eb00611e2769c4d0d3d8391 Mon Sep 17 00:00:00 2001
From: Grant Doyle <grant.doyle1@outlook.com>
Date: Wed, 20 May 2026 16:39:24 -0500
Subject: [PATCH] =?UTF-8?q?fix:=20keep=20"Building=E2=80=A6"=20indicator?=
 =?UTF-8?q?=20visible=20during=20tool=20calls=20+=20recover=20from=20empty?=
 =?UTF-8?q?=20responses?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Two real symptoms users hit during agentic loops:

1. The brick animation went away while tools were running. addMessageEl
   for the "Calling tool: X" line removed the indicator and it didn't
   come back until the whole iteration's tools finished. For slow
   stdio MCP calls (3-10s), users saw "Calling tool:" then static
   silence and assumed Mason was wedged.

2. Sometimes Opus 4.7 returned an *empty* text response after a tool
   loop — extended thinking burned through the 4096 max_tokens budget
   before any visible output. Mason silently rendered an empty
   assistant bubble and exited the loop, looking dead.

Changes:
- src/chat.ts: call showThinking() right after the "Calling tool:"
  message, so the brick stays visible while the tool actually runs.
- src/chat.ts: detect empty text response in result.type === "text"
  and surface a clear in-chat error explaining the likely cause
  ("Model returned an empty response — likely hit its token budget").
  Different message if there were prior tool calls (suggesting
  'continue' to resume) vs not (suggesting retry or switch models).
- src/main.ts: bump max_tokens 4096 -> 16384 (both chat-completions
  max_tokens and Responses API max_output_tokens). Gives extended-
  thinking models like Opus 4.7 headroom before content output, and
  reduces tool-call argument truncation (see prior sanitize fix).

Co-authored-by: Isaac
---
 src/chat.ts | 32 +++++++++++++++++++++++++++++---
 src/main.ts |  6 +++---
 2 files changed, 32 insertions(+), 6 deletions(-)

diff --git a/src/chat.ts b/src/chat.ts
index f99cbbb..8ecc8d8 100644
--- a/src/chat.ts
+++ b/src/chat.ts
@@ -317,14 +317,34 @@ async function chatLoop(_profile: { host?: string }): Promise<void> {
 
     if (result.type === "text") {
       const messagesEl = mason.el.messages as HTMLElement | null;
-      (mason.history as any[]).push({ role: "assistant", content: result.content });
+      const content = result.content || "";
+      // If we get back an empty text response in the middle of a multi-turn
+      // tool loop, the model most likely truncated (e.g. Opus 4.7 burned its
+      // max_tokens budget on extended thinking). Surface a clear hint
+      // instead of silently rendering an empty bubble and exiting.
+      if (!content.trim()) {
+        const hasPriorTools = (mason.history as any[]).some(
+          (m: any) => m.role === "tool" || (m.role === "assistant" && Array.isArray(m.tool_calls))
+        );
+        const hint = hasPriorTools
+          ? "Model returned an empty response — likely hit its token budget mid-thinking. Try sending 'continue' to resume, or rephrase your last request to reduce upstream context."
+          : "Model returned an empty response. Try sending the message again or switching to a different model.";
+        addMessageEl("error", hint);
+        (mason.history as any[]).push({ role: "assistant", content: "" });
+        // Remove the empty streaming bubble if one was created.
+        const sel = streamingEl as HTMLElement | null;
+        if (sel) sel.remove();
+        await saveCurrentChat();
+        return;
+      }
+      (mason.history as any[]).push({ role: "assistant", content });
       const sel = streamingEl as HTMLElement | null;
       if (sel) {
         sel.style.whiteSpace = "";
-        sel.innerHTML = renderMarkdown(result.content || "");
+        sel.innerHTML = renderMarkdown(content);
         if (messagesEl) messagesEl.scrollTop = messagesEl.scrollHeight;
       } else {
-        addMessageEl("assistant", result.content || "");
+        addMessageEl("assistant", content);
       }
       await saveCurrentChat();
       return;
@@ -435,6 +455,12 @@ async function chatLoop(_profile: { host?: string }): Promise<void> {
         }
 
         addMessageEl("tool-call", `Calling tool: ${toolName}`);
+        // Show the building-bricks indicator while the tool runs. addMessageEl
+        // above removed any existing indicator, and tool execution can take
+        // multiple seconds for stdio MCP / external API calls. Without this,
+        // users see "Calling tool:" then static silence and assume Mason is
+        // stuck.
+        showThinking();
 
         if (BUILTIN_TOOL_NAMES.has(toolName)) {
           try {
diff --git a/src/main.ts b/src/main.ts
index 2c1a176..94eae85 100644
--- a/src/main.ts
+++ b/src/main.ts
@@ -1982,7 +1982,7 @@ ipcMain.handle(
         }
       }
 
-      body = { model, max_output_tokens: 4096, input };
+      body = { model, max_output_tokens: 16384, input };
 
       if (tools && tools.length > 0) {
         body.tools = tools.map((t: any) => ({
@@ -2007,14 +2007,14 @@ ipcMain.handle(
         const hasSystem = messages.some((m: any) => m.role === "system");
         body = {
           model,
-          max_tokens: 4096,
+          max_tokens: 16384,
           messages: hasSystem ? messages : [systemMsg, ...messages],
           tools,
           tool_choice: "auto",
         };
         console.log(`[CHAT] Sending ${tools.length} tools: ${toolNames}`);
       } else {
-        body = { model, max_tokens: 4096, messages };
+        body = { model, max_tokens: 16384, messages };
       }
     }