fix(cli): announce "preparing <tool>…" once per tool per batch

The tool-gen callback fires once per tool CALL, so a model issuing three
parallel terminal calls printed three identical status lines before the
spinner took over (#10478). Coalesce repeats of the same tool name until a
tool actually starts (tool.started), which marks the next generation batch.

The line itself stays: it is the only feedback while a large argument
payload (e.g. a 45 KB write_file) streams, before the spinner exists.
PR #10598 removed it entirely; this keeps the first line and drops only
the duplicates.

Live A/B (in-process HTTP mock streaming 3 parallel terminal calls into a
real HermesCLI with display.streaming on): before 3 lines, after 1.

Fixes #10478

hermes_cli/cli_stream_mixin.py

@@ -621,12 +621,20 @@ class CLIStreamMixin:
 
     def _on_tool_gen_start(self, tool_name: str) -> None:
         """Model began generating tool-call arguments: close open boxes once, then print a status
-        line so a large payload (e.g. 45 KB write_file) doesn't look like a frozen screen."""
+        line so a large payload (e.g. 45 KB write_file) doesn't look like a frozen screen.
+
+        Fires once per tool CALL, so a batch of parallel calls to the same tool printed the same
+        line N times (#10478); repeats within one generation batch are coalesced. The set is
+        cleared when a tool actually starts (``tool.started``), i.e. on the next batch."""
         from cli import _cprint
-        if getattr(self, "_stream_box_opened", False):
+        if getattr(self, '_stream_box_opened', False):
             self._flush_stream()
             self._stream_box_opened = False
         self._close_reasoning_box()
+        announced = self.__dict__.setdefault("_tool_gen_announced", set())
+        if tool_name in announced:
+            return
+        announced.add(tool_name)
         from agent.display import get_tool_emoji
         _cprint(f"  ┊ {get_tool_emoji(tool_name, default='⚡')} preparing {tool_name}…")
 
@@ -666,6 +674,7 @@ class CLIStreamMixin:
         # Feed the pet: tools mean "running"; a failed tool latches the turn to end on a sulk.
         if event_type == "tool.started":
             self._pet_reasoning = False
+            self.__dict__.pop("_tool_gen_announced", None)
         elif event_type == "tool.completed" and kwargs.get("is_error"):
             self._pet_turn_error = True
         elif event_type and event_type.startswith("reasoning"):

tests/cli/test_tool_gen_start_dedupe.py

@@ -0,0 +1,25 @@
+"""A batch of parallel tool calls announces "preparing <tool>…" once per tool, not once per call (#10478)."""
+
+from unittest.mock import patch
+
+from tests.cli.test_tool_progress_scrollback import _make_cli
+import tests.cli.test_tool_progress_scrollback as _scrollback
+
+
+def _announce(cli, names):
+    printed = []
+    with patch.object(_scrollback._cli_mod, "_cprint", lambda line: printed.append(line)):
+        for n in names:
+            cli._on_tool_gen_start(n)
+    return printed
+
+
+def test_repeated_tool_in_one_batch_prints_once():
+    cli = _make_cli(tool_progress="off")
+    printed = _announce(cli, ["terminal", "terminal", "terminal", "read_file"])
+    assert sum("preparing terminal" in p for p in printed) == 1
+    assert sum("preparing read_file" in p for p in printed) == 1
+    # A tool actually starting closes the batch; the next generation announces again.
+    with patch.object(_scrollback._cli_mod, "_cprint", lambda line: None):
+        cli._on_tool_progress("tool.started", "terminal", "ls", {"command": "ls"})
+    assert sum("preparing terminal" in p for p in _announce(cli, ["terminal"])) == 1