From fcd2f4eb18a2bb7085cf227a106feed3d0b902ac Mon Sep 17 00:00:00 2001
From: pei711 <1259636180@qq.com>
Date: Wed, 14 Oct 2026 06:44:10 +0000
Subject: [PATCH] fix(agent): keep replayed tool call IDs stable (#43796)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Co-authored-by: zjp <你的邮箱>
---
 api/core/agent/base_agent_runner.py           |  6 ++--
 .../core/agent/test_base_agent_runner.py      | 35 +++++++++++++++++++
 2 files changed, 38 insertions(+), 3 deletions(-)

diff --git a/api/core/agent/base_agent_runner.py b/api/core/agent/base_agent_runner.py
index 806f7c6590f..0bbc3880c58 100644
--- a/api/core/agent/base_agent_runner.py
+++ b/api/core/agent/base_agent_runner.py
@@ -408,9 +408,9 @@ class BaseAgentRunner(AppRunner):
                         else:
                             tool_responses = dict.fromkeys(tool_names, observation_payload)
 
-                        for tool in tool_names:
-                            # generate a uuid for tool call
-                            tool_call_id = str(uuid.uuid4())
+                        for index, tool in enumerate(tool_names):
+                            # Keep IDs stable across turns so replayed history can use prompt caching.
+                            tool_call_id = str(uuid.uuid5(uuid.NAMESPACE_OID, f"{agent_thought.id}:{index}:{tool}"))
                             tool_calls.append(
                                 AssistantPromptMessage.ToolCall(
                                     id=tool_call_id,
diff --git a/api/tests/unit_tests/core/agent/test_base_agent_runner.py b/api/tests/unit_tests/core/agent/test_base_agent_runner.py
index 71ad572cce1..c1bbdf8e2d9 100644
--- a/api/tests/unit_tests/core/agent/test_base_agent_runner.py
+++ b/api/tests/unit_tests/core/agent/test_base_agent_runner.py
@@ -611,6 +611,41 @@ def test_organize_history_reconstructs_tool_flows(
     assert any(isinstance(item, module.AssistantPromptMessage) for item in result)
 
 
+def test_organize_history_reuses_stable_tool_call_ids(
+    runner: BaseAgentRunner,
+    sqlite_session: Session,
+) -> None:
+    message = _message(message_id="m2")
+    thought = _thought(
+        thought_id="thought-1",
+        message_id=message.id,
+        tool="tool1;tool2",
+        tool_input=json.dumps({"tool1": {}, "tool2": {}}),
+        observation=json.dumps({"tool1": "o1", "tool2": "o2"}),
+    )
+    _persist_history(sqlite_session, message, thoughts=[thought])
+
+    first = runner.organize_agent_history([], session=sqlite_session)
+    second = runner.organize_agent_history([], session=sqlite_session)
+
+    def tool_call_ids(messages: list[module.PromptMessage]) -> tuple[list[str], list[str]]:
+        call_ids = [
+            tool_call.id
+            for prompt in messages
+            if isinstance(prompt, module.AssistantPromptMessage)
+            for tool_call in prompt.tool_calls or []
+        ]
+        response_ids = [prompt.tool_call_id for prompt in messages if isinstance(prompt, module.ToolPromptMessage)]
+        return call_ids, response_ids
+
+    first_call_ids, first_response_ids = tool_call_ids(first)
+    second_call_ids, second_response_ids = tool_call_ids(second)
+
+    assert first_call_ids == second_call_ids
+    assert first_call_ids == first_response_ids == second_response_ids
+    assert len(set(first_call_ids)) == 2
+
+
 def test_organize_history_without_tool_name(runner: BaseAgentRunner, sqlite_session: Session) -> None:
     message = _message(message_id="m3")
     thought = _thought(thought_id="thought-1", message_id=message.id, tool=None)