Merge pull request #72220 from NousResearch/bb/tool-row-dedupe

fix(tools): stop the inline tool row stuttering its verb and repeating the command

apps/desktop/src/components/assistant-ui/tool/fallback-model.test.ts

@@ -340,6 +340,28 @@ describe('buildToolView title actions', () => {
     expect(view.titleAction).toEqual({ prefix: '', text: 'Running', suffix: ' pnpm run lint' })
   })
 
+  it('never stutters the verb or echoes the command when the backend context is a phrased label', () => {
+    // Older backends stamped tool.start with a *phrased* label
+    // ("Running sleep 70 + 2 commands") rather than a raw arg preview, and the
+    // desktop merges that into args.context. The row must still prepend its own
+    // verb exactly once, show the real command in the `$` transcript, and not
+    // repeat either string as detail.
+    const command = 'sleep 70; echo "a"; echo "b"'
+
+    const view = buildToolView(
+      part({
+        args: { command, context: 'Running sleep 70 + 2 commands' },
+        result: { exit_code: 0 },
+        toolName: 'terminal'
+      }),
+      ''
+    )
+
+    expect(view.title).toBe('Ran sleep 70 + 2 commands')
+    expect(view.terminalCommand).toBe(command)
+    expect(view.detail).toBe('')
+  })
+
   it('uses the runtime locale for title text and action placement', () => {
     setRuntimeI18nLocale('ja')
 

apps/desktop/src/components/assistant-ui/tool/fallback-model/index.ts

@@ -128,9 +128,14 @@ function readFileDisplayTarget(args: Record<string, unknown>, result: Record<str
   return [fileEditBasename(path), lineLabel].filter(Boolean).join(' ')
 }
 
+// The real command, preferring the actual argument over the backend's
+// display preview. `context` is a *summarized* preview ("sleep 70 + 2
+// commands") the gateway sends on tool.start before real args arrive — fine
+// as a placeholder for the live title, wrong for the `$` transcript, which
+// must show what actually ran.
 function shellCommand(args: Record<string, unknown>): string {
   return (
-    firstStringField(args, ['context', 'preview']) || firstStringField(args, ['command', 'code']) || contextValue(args)
+    firstStringField(args, ['command', 'code']) || firstStringField(args, ['context', 'preview']) || contextValue(args)
   )
 }
 
@@ -1079,6 +1084,13 @@ function toolDetailText(
     if (output || lines) {
       return [output, lines].filter(Boolean).join('\n')
     }
+
+    // A terminal row with no output already shows its command in the `$`
+    // transcript above; the generic fallback would print the same string a
+    // second time. `execute_code` has no transcript, so it keeps the fallback.
+    if (part.toolName === 'terminal') {
+      return ''
+    }
   }
 
   if (part.toolName === 'web_extract') {

tests/test_tui_gateway_server.py

@@ -1551,12 +1551,24 @@ def test_history_to_messages_preserves_tool_calls_for_resume_display():
 
     assert server._history_to_messages(history) == [
         {"role": "user", "text": "first prompt"},
-        {"context": "Searching files for resume", "name": "search_files", "role": "tool"},
+        {"context": "resume", "name": "search_files", "role": "tool"},
         {"role": "assistant", "text": "first answer"},
         {"role": "user", "text": "second prompt"},
     ]
 
 
+def test_tool_ctx_sends_an_arg_preview_not_a_phrased_label():
+    # Clients phrase their own verb around this string: the TUI renders
+    # `Terminal("<ctx>")` and the desktop prepends "Running"/"Ran". Sending a
+    # pre-phrased label made both stutter ("Ran Running sleep 70 + 2 commands")
+    # and stood in for the real command in the desktop's `$` transcript.
+    assert server._tool_ctx("terminal", {"command": 'sleep 70; echo "a"; echo "b"'}) == (
+        "sleep 70 + 2 commands"
+    )
+    assert server._tool_ctx("read_file", {"path": "/tmp/demo/package.json"}) == "package.json"
+    assert server._tool_ctx("web_search", {"query": "weather in NYC"}) == "weather in NYC"
+
+
 def test_history_to_messages_keeps_reasoning_only_assistant_turn():
     # A thinking-only assistant turn (reasoning present, no visible text) is
     # persisted and recallable, but was dropped from the resumed session view

tui_gateway/server.py

@@ -4505,10 +4505,20 @@ def _session_info(agent, session: dict | None = None) -> dict:
 
 
 def _tool_ctx(name: str, args: dict) -> str:
-    try:
-        from agent.display import build_tool_label
+    """Argument preview for a tool row — never a phrased label.
 
-        return build_tool_label(name, args, max_len=80) or ""
+    Clients own their own phrasing: the TUI wraps this as ``Terminal("...")``
+    and the desktop prepends its own localized verb ("Running"/"Ran"). Sending
+    ``build_tool_label`` here instead of the raw preview stutters the verb on
+    both surfaces ("Running Running sleep 70 + 2 commands") and leaks a display
+    label into the desktop's ``args.context``, where it stands in for the real
+    command. The friendly labels belong on the CLI spinner, which builds them
+    from ``build_tool_label`` at its own call sites.
+    """
+    try:
+        from agent.display import build_tool_preview
+
+        return build_tool_preview(name, args, max_len=80) or ""
     except Exception:
         return ""