fix: preserve tool failures under noisy output

tests/test_agent_loop.py

@@ -120,6 +120,59 @@ class ParallelRegistry:
         return {"ok": True, "path": arguments["path"]}
 
 
+class CountingRegistry:
+    def __init__(self) -> None:
+        self.calls = 0
+
+    def specs(self, *, read_only=False, allow_delegate=True):
+        return [TOOL_METADATA["read_file"].openai_spec()]
+
+    async def execute(self, name, arguments, context):
+        self.calls += 1
+        return {"ok": True, "path": arguments["path"]}
+
+
+async def test_tool_call_batch_is_bounded(settings) -> None:
+    crowded = {
+        "choices": [
+            {
+                "message": {
+                    "role": "assistant",
+                    "content": None,
+                    "tool_calls": [tool_call(str(index), "read_file", {"path": f"{index}.txt"}) for index in range(20)],
+                },
+                "finish_reason": "tool_calls",
+            }
+        ]
+    }
+    final = {"choices": [{"message": {"role": "assistant", "content": "done"}, "finish_reason": "stop"}]}
+    store = RuntimeStore(settings.database_url)
+    await store.initialize()
+    registry = CountingRegistry()
+    loop = AgentLoop(ScriptedProvider([crowded, final]), registry, store, max_tool_output_chars=10_000)
+    events = []
+
+    async def callback(event_type, payload):
+        events.append((event_type, payload))
+
+    try:
+        answer = await loop.run(
+            spec=get_model_spec("work-medium"),
+            messages=[{"role": "user", "content": "检查当前状态"}],
+            identity=UserIdentity("u1", "u1@example.test", "U1", "user"),
+            raw_user_jwt="jwt",
+            chat_id="bounded-batch",
+            callback=callback,
+        )
+    finally:
+        await store.close()
+
+    assert answer == "done"
+    assert registry.calls == 8
+    limited = next(payload for event_type, payload in events if event_type == "tool.batch_limited")
+    assert limited == {"requested": 20, "accepted": 8, "iteration": 1, "depth": 0}
+
+
 async def test_read_only_tool_calls_run_in_parallel(settings) -> None:
     first = {
         "choices": [

tests/test_text.py

@@ -0,0 +1,26 @@
+from __future__ import annotations
+
+from agent_platform.text import truncate_middle
+
+
+def test_truncate_middle_preserves_error_at_end() -> None:
+    text = "stdout:" + ("x" * 1000) + "\nNameError: start is not defined"
+
+    truncated = truncate_middle(text, 120)
+
+    assert len(truncated) == 120
+    assert truncated.startswith("stdout:")
+    assert "middle omitted" in truncated
+    assert truncated.endswith("NameError: start is not defined")
+
+
+def test_truncate_middle_leaves_short_text_unchanged() -> None:
+    assert truncate_middle("short", 20) == "short"
+
+
+def test_truncate_middle_honors_tiny_limits() -> None:
+    marker = "<omitted>"
+
+    assert truncate_middle("abcdef", 0, marker) == ""
+    assert truncate_middle("abcdef", 3, marker) == "abc"
+    assert truncate_middle("abcdefghijklmnopqrstuvwxyz", len(marker) + 1, marker) == "a<omitted>"