diff --git a/examples/fastapi_server/_app.py b/examples/fastapi_server/_app.py
index 0b8d672..d178491 100644
--- a/examples/fastapi_server/_app.py
+++ b/examples/fastapi_server/_app.py
@@ -137,7 +137,7 @@ async def chat(req: ChatRequest) -> ChatResponse:  # pylint: disable=unused-vari
                             ))
 
         except Exception as exc:
-            logger.exception("Error during agent run (session=%s)", session_id)
+            logger.error("Error during agent run (session=%s): %s", session_id, exc)
             raise HTTPException(status_code=500, detail=str(exc)) from exc
 
         return ChatResponse(
@@ -214,7 +214,7 @@ async def _event_generator() -> AsyncGenerator[str, None]:
                 yield _sse(StreamChunk(type="done", session_id=session_id))
 
             except Exception as exc:
-                logger.exception("Error during streaming run (session=%s)", session_id)
+                logger.error("Error during streaming run (session=%s): %s", session_id, exc)
                 yield _sse(StreamChunk(type="error", data=str(exc), session_id=session_id))
 
         return StreamingResponse(
diff --git a/examples/session_summarizer/run_agent.py b/examples/session_summarizer/run_agent.py
index 3d69308..d653a1a 100644
--- a/examples/session_summarizer/run_agent.py
+++ b/examples/session_summarizer/run_agent.py
@@ -127,7 +127,7 @@ async def summarize_session(session_service: InMemorySessionService, app_name: s
             print(f"   - Compression ratio: {summary.get_compression_ratio():.1f}%")
 
 
-SUMMARIZER_COUNT = 3  # Run summarization every SUMMARIZER_COUNT turns (e.g. 3 => every 3 turns)
+SUMMARIZER_COUNT = 2  # Keep the example short: summarize after a couple of turns.
 
 
 def create_summarizer_manager(model: OpenAIModel) -> SummarizerSessionManager:
@@ -154,8 +154,8 @@ def create_summarizer_manager(model: OpenAIModel) -> SummarizerSessionManager:
             #     set_summarizer_time_interval_threshold(10),
             # )
         ],
-        max_summary_length=600,  # Max summary length kept; default 1000; beyond shows ...
-        keep_recent_count=4,  # How many recent turns to keep; default 10
+        max_summary_length=300,  # Max summary length kept; default 1000; beyond shows ...
+        keep_recent_count=2,  # Keep only the latest turns so compression is easy to observe.
     )
     # Create SummarizerSessionManager
     summarizer_manager = SummarizerSessionManager(
@@ -169,7 +169,7 @@ def create_summarizer_manager(model: OpenAIModel) -> SummarizerSessionManager:
 async def llm_agent_summarizer():
     """Demo LlmAgent integrated with SummarizerSessionManager."""
     print("=" * 60)
-    print("Example 2: LlmAgent + SummarizerSessionManager demo")
+    print("Example: LlmAgent + SummarizerSessionManager demo")
     print("=" * 60)
     app_name = "llm_summarizer_manager_demo"
 
@@ -183,22 +183,13 @@ async def llm_agent_summarizer():
     current_session_id = str(uuid.uuid4())
     print(f"📊 Session: {app_name}/{user_id}/{current_session_id}")
 
-    # Demo conversation turns
+    # Short demo conversation.  Four turns are enough to trigger automatic
+    # summarization while keeping the example quick to run.
     conversations = [
         "Hello! I want to learn Python programming. Can you help me?",
         "What is a variable? Can you give an example?",
-        "Got it! What data types are there?",
-        "What does control flow mean?",
-        "I understand those ideas. I'd like a small project to practice.",
-        "OK! How do I build this calculator?",
-        "The calculator looks good—I ran it successfully. I'd like to learn more advanced Python.",
-        "I'd like to start with functions—I think they're central to programming.",
-        "I see—functions make code modular and reusable. I'd like to learn OOP next.",
-        "I get OOP now. I'd like to learn exception handling.",
-        "I've learned these advanced topics. I'd like a bigger project that ties them together.",
-        "Yes! How do I implement this library system?",
-        "The structure looks good. How do I persist data to files?",
-        "Great! I've covered basics and advanced topics including files. I'd like a recap of what I learned.",
+        "Please give me a tiny calculator example.",
+        "Can you recap what I learned so far?",
     ]
 
     print(f"\n💬 Multi-turn dialogue ({len(conversations)} turns)...")
@@ -230,18 +221,18 @@ async def llm_agent_summarizer():
                 # elif part.text:
                 #     print(f"\n✅ {part.text}")
 
-        # After every SUMMARIZER_COUNT turns, inspect session state
-        if index % SUMMARIZER_COUNT == 0:  # summarizer should fire around this cadence
-            if session:
-                print(f"\n📊 Session state after turn {index + 1}:")
-                summary = await session_service.summarizer_manager.get_session_summary(session)
+        # Inspect the summary after the threshold cadence.
+        if (index + 1) % SUMMARIZER_COUNT == 0 and session:
+            print(f"\n📊 Session state after turn {index + 1}:")
+            summary = await session_service.summarizer_manager.get_session_summary(session)
+            if summary:
                 print(f"   - Summary text: {summary.summary_text[:100]}...")
                 print(f"   - Original event count: {summary.original_event_count}")
                 print(f"   - Compressed event count: {summary.compressed_event_count}")
                 print(f"   - Compression ratio: {summary.get_compression_ratio()}")
+            else:
+                print("   - Summary not created yet.")
         print("\n" + "-" * 40)
-    # Manual forced summary test
-    await summarize_session(session_service, app_name, user_id, current_session_id)
 
 
 if __name__ == "__main__":
diff --git a/lint_flake8.sh b/lint_flake8.sh
new file mode 100644
index 0000000..1db04c7
--- /dev/null
+++ b/lint_flake8.sh
@@ -0,0 +1,21 @@
+#!/usr/bin/env bash
+
+set -euo pipefail
+
+# Usage:
+#   bash lint_flake8.sh                # check current project
+#   bash lint_flake8.sh path/to/check  # check a specific path
+
+TARGET_PATH="${1:-.}"
+
+if ! command -v flake8 >/dev/null 2>&1; then
+  echo "flake8 is not installed. Install it first:"
+  echo "  python3 -m pip install flake8"
+  exit 1
+fi
+
+echo "Running flake8 on: ${TARGET_PATH}"
+
+flake8 "${TARGET_PATH}" \
+  --max-line-length=120 \
+  --extend-exclude=".git,__pycache__,.pytest_cache,.mypy_cache,.ruff_cache,venv,.venv,build,dist,node_modules"
diff --git a/pyproject.toml b/pyproject.toml
index 5fea878..7f1145a 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -76,12 +76,12 @@ knowledge = [
 ]
 
 a2a = [
-    "a2a-sdk>=0.2.0",
+    "a2a-sdk<1.0.0,>=0.3.22",
     "protobuf>=5.29.5",
 ]
 
 agent-claude = [
-    "claude-agent-sdk>=0.1.3",
+    "claude-agent-sdk>=0.1.3,<0.1.64",
     "cloudpickle>=2.0.0",
 ]
 
@@ -115,20 +115,19 @@ dev = [
     "langchain_community>=0.3.27",
     "langchain_huggingface>=0.1.0",
     "ag-ui-protocol>=0.1.8",
-    "claude-agent-sdk>=0.1.3",
+    "claude-agent-sdk>=0.1.3,<0.1.64",
     "cloudpickle>=2.0.0",
     "typer>=0.9.0",
 ]
 
 all = [
-    "a2a-sdk>=0.2.0",
     "protobuf>=5.29.5",
     "numpy>=2.2.5",
     "langchain_community>=0.3.27",
     "langchain_huggingface>=0.1.0",
     "langchain_tavily",
     "ag-ui-protocol>=0.1.8",
-    "claude-agent-sdk>=0.1.3",
+    "claude-agent-sdk>=0.1.3,<0.1.64",
     "pytest",
     "pytest-asyncio",
     "rouge-score",
@@ -140,6 +139,7 @@ all = [
     "nanobot-ai>=0.1.4.post6",
     "aiofiles",
     "wecom-aibot-sdk-python>=0.1.5",
+    "a2a-sdk<1.0.0,>=0.3.22",
 ]
 
 [project.scripts]
diff --git a/tests/evaluation/test_eval_session_service.py b/tests/evaluation/test_eval_session_service.py
index 0ec0f6b..f55f2b2 100644
--- a/tests/evaluation/test_eval_session_service.py
+++ b/tests/evaluation/test_eval_session_service.py
@@ -18,6 +18,10 @@ def _make_session():
     """Create a mock session with required attributes."""
     session = MagicMock()
     session.events = []
+    def _insert_events(events, idx=None):
+        insert_idx = 0 if idx is None else idx
+        session.events[insert_idx:insert_idx] = list(events)
+    session.insert_events = MagicMock(side_effect=_insert_events)
     return session
 
 
diff --git a/tests/events/test_event.py b/tests/events/test_event.py
index 8587826..b28d852 100644
--- a/tests/events/test_event.py
+++ b/tests/events/test_event.py
@@ -42,6 +42,9 @@ def test_default_fields(self):
         assert event.tag is None
         assert event.filter_key is None
         assert event.object is None
+        assert event.model_flags > 0
+        assert event.is_model_visible() is True
+        assert event.is_summary_event() is False
 
     def test_auto_generated_id_is_valid_uuid(self):
         event = Event(invocation_id="inv-1", author="a")
@@ -308,6 +311,33 @@ def test_error_message_without_code_not_error(self):
         assert event.is_error() is False
 
 
+# ---------------------------------------------------------------------------
+# Event model visibility / summary flags
+# ---------------------------------------------------------------------------
+
+
+class TestEventModelFlags:
+    def test_set_model_visible_false(self):
+        event = Event(invocation_id="inv-1", author="a")
+        event.set_model_visible(False)
+        assert event.is_model_visible() is False
+
+    def test_visible_field_does_not_control_model_visibility(self):
+        event = Event(invocation_id="inv-1", author="a", visible=False)
+        assert event.is_model_visible() is True
+
+    def test_set_summary_event_true(self):
+        event = Event(invocation_id="inv-1", author="a")
+        event.set_summary_event(True)
+        assert event.is_summary_event() is True
+
+    def test_clear_summary_event(self):
+        event = Event(invocation_id="inv-1", author="a")
+        event.set_summary_event(True)
+        event.set_summary_event(False)
+        assert event.is_summary_event() is False
+
+
 # ---------------------------------------------------------------------------
 # Event.has_trailing_code_execution_result
 # ---------------------------------------------------------------------------
diff --git a/tests/server/a2a/converters/test_event_converter.py b/tests/server/a2a/converters/test_event_converter.py
index 571d2e3..3ce2024 100644
--- a/tests/server/a2a/converters/test_event_converter.py
+++ b/tests/server/a2a/converters/test_event_converter.py
@@ -11,19 +11,25 @@
 from unittest.mock import MagicMock, patch
 
 import pytest
-from a2a.types import (
-    Artifact,
-    DataPart,
-    Message,
-    Part as A2APart,
-    Role,
-    Task,
-    TaskArtifactUpdateEvent,
-    TaskState,
-    TaskStatus,
-    TaskStatusUpdateEvent,
-    TextPart,
-)
+try:
+    from a2a.types import (
+        Artifact,
+        DataPart,
+        Message,
+        Part as A2APart,
+        Role,
+        Task,
+        TaskArtifactUpdateEvent,
+        TaskState,
+        TaskStatus,
+        TaskStatusUpdateEvent,
+        TextPart,
+    )
+except ImportError:
+    pytest.skip(
+        "Installed a2a.types does not export DataPart/TextPart; skip legacy A2A tests.",
+        allow_module_level=True,
+    )
 from google.genai import types as genai_types
 
 from trpc_agent_sdk.context import InvocationContext
diff --git a/tests/server/a2a/converters/test_part_converter.py b/tests/server/a2a/converters/test_part_converter.py
index c8bcb58..865e631 100644
--- a/tests/server/a2a/converters/test_part_converter.py
+++ b/tests/server/a2a/converters/test_part_converter.py
@@ -13,7 +13,15 @@
 from unittest.mock import MagicMock
 
 import pytest
-from a2a import types as a2a_types
+try:
+    from a2a import types as a2a_types
+    _ = a2a_types.DataPart
+    _ = a2a_types.TextPart
+except (ImportError, AttributeError):
+    pytest.skip(
+        "Installed a2a.types does not export DataPart/TextPart; skip legacy A2A tests.",
+        allow_module_level=True,
+    )
 from google.genai import types as genai_types
 
 from trpc_agent_sdk.models import TOOL_STREAMING_ARGS
diff --git a/tests/server/a2a/converters/test_request_converter.py b/tests/server/a2a/converters/test_request_converter.py
index d4603cb..377b426 100644
--- a/tests/server/a2a/converters/test_request_converter.py
+++ b/tests/server/a2a/converters/test_request_converter.py
@@ -11,7 +11,13 @@
 
 import pytest
 from a2a.server.agent_execution.context import RequestContext
-from a2a.types import Message, Part, Role, TextPart
+try:
+    from a2a.types import Message, Part, Role, TextPart
+except ImportError:
+    pytest.skip(
+        "Installed a2a.types does not export TextPart; skip legacy A2A tests.",
+        allow_module_level=True,
+    )
 
 from trpc_agent_sdk.server.a2a.converters._request_converter import (
     _get_user_id_default,
diff --git a/tests/server/openclaw/agent/test__agent.py b/tests/server/openclaw/agent/test_agent.py
similarity index 100%
rename from tests/server/openclaw/agent/test__agent.py
rename to tests/server/openclaw/agent/test_agent.py
diff --git a/tests/server/openclaw/agent/test__prompts.py b/tests/server/openclaw/agent/test_prompts.py
similarity index 100%
rename from tests/server/openclaw/agent/test__prompts.py
rename to tests/server/openclaw/agent/test_prompts.py
diff --git a/tests/server/openclaw/channels/test__command_handler.py b/tests/server/openclaw/channels/test_command_handler.py
similarity index 100%
rename from tests/server/openclaw/channels/test__command_handler.py
rename to tests/server/openclaw/channels/test_command_handler.py
diff --git a/tests/server/openclaw/channels/test__repair.py b/tests/server/openclaw/channels/test_repair.py
similarity index 100%
rename from tests/server/openclaw/channels/test__repair.py
rename to tests/server/openclaw/channels/test_repair.py
diff --git a/tests/server/openclaw/channels/test__wecom.py b/tests/server/openclaw/channels/test_wecom.py
similarity index 100%
rename from tests/server/openclaw/channels/test__wecom.py
rename to tests/server/openclaw/channels/test_wecom.py
diff --git a/tests/server/openclaw/config/test__config.py b/tests/server/openclaw/config/test_config.py
similarity index 99%
rename from tests/server/openclaw/config/test__config.py
rename to tests/server/openclaw/config/test_config.py
index 88537e3..22f6123 100644
--- a/tests/server/openclaw/config/test__config.py
+++ b/tests/server/openclaw/config/test_config.py
@@ -243,7 +243,7 @@ def test_env_var_path(self, mock_set_config, tmp_path, monkeypatch):
     @patch("trpc_agent_sdk.server.openclaw.config._config.DEFAULT_CONFIG_PATH")
     def test_default_path_fallback(self, mock_default_path, mock_default_dir, mock_set_config, tmp_path, monkeypatch):
         monkeypatch.delenv(TRPC_CLAW_CONFIG, raising=False)
-        default_dir = tmp_path / ".trpc_agent_claw"
+        default_dir = tmp_path / ".trpc_claw"
         default_dir.mkdir()
         mock_default_dir.exists.return_value = True
         cfg_file = default_dir / "config.yaml"
diff --git a/tests/server/openclaw/metrics/test__langfuse.py b/tests/server/openclaw/metrics/test_langfuse.py
similarity index 100%
rename from tests/server/openclaw/metrics/test__langfuse.py
rename to tests/server/openclaw/metrics/test_langfuse.py
diff --git a/tests/server/openclaw/metrics/test__metrics.py b/tests/server/openclaw/metrics/test_metrics.py
similarity index 100%
rename from tests/server/openclaw/metrics/test__metrics.py
rename to tests/server/openclaw/metrics/test_metrics.py
diff --git a/tests/server/openclaw/service/test__heart_service.py b/tests/server/openclaw/service/test_heart_service.py
similarity index 100%
rename from tests/server/openclaw/service/test__heart_service.py
rename to tests/server/openclaw/service/test_heart_service.py
diff --git a/tests/server/openclaw/session_memory/test__claw_memory_service.py b/tests/server/openclaw/session_memory/test_claw_memory_service.py
similarity index 100%
rename from tests/server/openclaw/session_memory/test__claw_memory_service.py
rename to tests/server/openclaw/session_memory/test_claw_memory_service.py
diff --git a/tests/server/openclaw/session_memory/test__claw_session_service.py b/tests/server/openclaw/session_memory/test_claw_session_service.py
similarity index 100%
rename from tests/server/openclaw/session_memory/test__claw_session_service.py
rename to tests/server/openclaw/session_memory/test_claw_session_service.py
diff --git a/tests/server/openclaw/session_memory/test__claw_summarizer.py b/tests/server/openclaw/session_memory/test_claw_summarizer.py
similarity index 100%
rename from tests/server/openclaw/session_memory/test__claw_summarizer.py
rename to tests/server/openclaw/session_memory/test_claw_summarizer.py
diff --git a/tests/server/openclaw/skill/test__deps.py b/tests/server/openclaw/skill/test_deps.py
similarity index 100%
rename from tests/server/openclaw/skill/test__deps.py
rename to tests/server/openclaw/skill/test_deps.py
diff --git a/tests/server/openclaw/skill/test__skill_loader.py b/tests/server/openclaw/skill/test_skill_loader.py
similarity index 100%
rename from tests/server/openclaw/skill/test__skill_loader.py
rename to tests/server/openclaw/skill/test_skill_loader.py
diff --git a/tests/server/openclaw/skill/test__skill_parser.py b/tests/server/openclaw/skill/test_skill_parser.py
similarity index 100%
rename from tests/server/openclaw/skill/test__skill_parser.py
rename to tests/server/openclaw/skill/test_skill_parser.py
diff --git a/tests/server/openclaw/skill/test__skill_tool.py b/tests/server/openclaw/skill/test_skill_tool.py
similarity index 100%
rename from tests/server/openclaw/skill/test__skill_tool.py
rename to tests/server/openclaw/skill/test_skill_tool.py
diff --git a/tests/server/openclaw/skill/test__utils.py b/tests/server/openclaw/skill/test_utils.py
similarity index 100%
rename from tests/server/openclaw/skill/test__utils.py
rename to tests/server/openclaw/skill/test_utils.py
diff --git a/tests/server/openclaw/storage/test__aiofile_storage.py b/tests/server/openclaw/storage/test_aiofile_storage.py
similarity index 100%
rename from tests/server/openclaw/storage/test__aiofile_storage.py
rename to tests/server/openclaw/storage/test_aiofile_storage.py
diff --git a/tests/server/openclaw/storage/test__manager.py b/tests/server/openclaw/storage/test_manager.py
similarity index 100%
rename from tests/server/openclaw/storage/test__manager.py
rename to tests/server/openclaw/storage/test_manager.py
diff --git a/tests/server/openclaw/storage/test__utils.py b/tests/server/openclaw/storage/test_utils.py
similarity index 100%
rename from tests/server/openclaw/storage/test__utils.py
rename to tests/server/openclaw/storage/test_utils.py
diff --git a/tests/server/openclaw/test__cli.py b/tests/server/openclaw/test_cli.py
similarity index 100%
rename from tests/server/openclaw/test__cli.py
rename to tests/server/openclaw/test_cli.py
diff --git a/tests/server/openclaw/test__logger.py b/tests/server/openclaw/test_logger.py
similarity index 100%
rename from tests/server/openclaw/test__logger.py
rename to tests/server/openclaw/test_logger.py
diff --git a/tests/server/openclaw/test__utils.py b/tests/server/openclaw/test_utils.py
similarity index 100%
rename from tests/server/openclaw/test__utils.py
rename to tests/server/openclaw/test_utils.py
diff --git a/tests/sessions/test_base_session_service.py b/tests/sessions/test_base_session_service.py
index fbad9a3..ef898f8 100644
--- a/tests/sessions/test_base_session_service.py
+++ b/tests/sessions/test_base_session_service.py
@@ -172,7 +172,9 @@ def test_filter_by_num_recent_events(self):
         for i in range(10):
             session.events.append(_make_event(text=f"msg{i}"))
         svc.filter_events(session)
-        assert len(session.events) == 3
+        assert len(session.events) == 10
+        visible_events = [event for event in session.events if event.is_model_visible()]
+        assert [event.get_text() for event in visible_events] == ["msg7", "msg8", "msg9"]
 
     def test_filter_by_event_ttl(self):
         config = SessionServiceConfig(event_ttl_seconds=5.0)
@@ -188,8 +190,10 @@ def test_filter_by_event_ttl(self):
         session.events.append(new_event)
 
         svc.filter_events(session)
-        assert len(session.events) == 1
-        assert session.events[0].get_text() == "new"
+        assert len(session.events) == 2
+        visible_events = [event for event in session.events if event.is_model_visible()]
+        assert len(visible_events) == 1
+        assert visible_events[0].get_text() == "new"
 
     def test_filter_no_config(self):
         svc = ConcreteSessionService()
@@ -208,7 +212,8 @@ def test_filter_ttl_removes_all_old(self):
             e.timestamp = time.time() - 100
             session.events.append(e)
         svc.filter_events(session)
-        assert len(session.events) == 0
+        assert len(session.events) == 5
+        assert all(not event.is_model_visible() for event in session.events)
 
 
 class TestBaseSessionServiceSetSummarizerManager:
diff --git a/tests/sessions/test_session.py b/tests/sessions/test_session.py
index 4f072d5..39c83bd 100644
--- a/tests/sessions/test_session.py
+++ b/tests/sessions/test_session.py
@@ -79,11 +79,12 @@ def test_apply_event_filtering_max_events(self):
         # Apply filtering with max_events=5
         session.apply_event_filtering(max_events=5)
 
-        # Should keep last 5 events (Message 5-9), then find first user message (Message 6)
-        # and keep from there, resulting in Message 6-9 (4 events)
-        assert len(session.events) == 4
-        assert session.events[0].get_text() == "Message 6"
-        assert session.events[-1].get_text() == "Message 9"
+        # Filtering hides model-invisible events instead of deleting them.
+        visible_events = [event for event in session.events if event.is_model_visible()]
+        assert len(session.events) == 10
+        assert len(visible_events) == 4
+        assert visible_events[0].get_text() == "Message 6"
+        assert visible_events[-1].get_text() == "Message 9"
 
     def test_apply_event_filtering_ttl(self):
         """Test event filtering with TTL."""
@@ -116,13 +117,12 @@ def test_apply_event_filtering_ttl(self):
         # Apply TTL filtering with 2 seconds
         session.apply_event_filtering(event_ttl_seconds=2.0)
 
-        # TTL filter removes old events, leaving only recent 3 agent messages
-        # Since no user message in remaining events, last user message is inserted at beginning
-        # Result: 1 user message + 3 recent agent messages = 4 events
-        assert len(session.events) == 4
-        assert session.events[0].author == "user"
-        assert "Old user message 1" in session.events[0].get_text()
-        assert all("Recent" in e.get_text() for e in session.events[1:])
+        # TTL + user-anchor fallback keeps only the last user message visible.
+        visible_events = [event for event in session.events if event.is_model_visible()]
+        assert len(session.events) == 6
+        assert len(visible_events) == 1
+        assert visible_events[0].author == "user"
+        assert "Old user message 1" in visible_events[0].get_text()
 
     def test_apply_event_filtering_ttl_and_max_events(self):
         """Test event filtering with both TTL and max_events."""
@@ -152,12 +152,12 @@ def test_apply_event_filtering_ttl_and_max_events(self):
         # Apply both filters
         session.apply_event_filtering(event_ttl_seconds=5.0, max_events=5)
 
-        # TTL filters out old events, max_events limits to 5, then finds first user message
-        # Recent events 5-9 would be kept (5 events), first user message is at index with i%3==0
-        # So Recent 6, 7, 8, 9 (4 events) starting from first user (Recent 6)
-        assert len(session.events) == 4
-        assert session.events[0].get_text() == "Recent 6"
-        assert session.events[-1].get_text() == "Recent 9"
+        # Filtering hides model-invisible events instead of deleting them.
+        visible_events = [event for event in session.events if event.is_model_visible()]
+        assert len(session.events) == 20
+        assert len(visible_events) == 4
+        assert visible_events[0].get_text() == "Recent 6"
+        assert visible_events[-1].get_text() == "Recent 9"
 
     def test_apply_event_filtering_preserves_last_user_message(self):
         """Test that filtering preserves the last user message when all events are filtered."""
@@ -188,10 +188,12 @@ def test_apply_event_filtering_preserves_last_user_message(self):
         # Apply strict TTL filter that would remove all events
         session.apply_event_filtering(event_ttl_seconds=2.0)
 
-        # All events are old, but should preserve the last user message
-        assert len(session.events) == 1
-        assert session.events[0].author == "user"
-        assert session.events[0].get_text() == "Last user message"
+        # All events are old, but last user message remains model-visible.
+        visible_events = [event for event in session.events if event.is_model_visible()]
+        assert len(session.events) == 7
+        assert len(visible_events) == 1
+        assert visible_events[0].author == "user"
+        assert visible_events[0].get_text() == "Last user message"
 
     def test_apply_event_filtering_empty_events(self):
         """Test event filtering with no events."""
@@ -226,8 +228,10 @@ def test_apply_event_filtering_all_filtered_no_user_message(self):
         # Apply strict TTL filter
         session.apply_event_filtering(event_ttl_seconds=2.0)
 
-        # Should be empty since all are old and none are user messages
-        assert len(session.events) == 0
+        # All events are hidden from model history; raw events remain.
+        visible_events = [event for event in session.events if event.is_model_visible()]
+        assert len(session.events) == 5
+        assert len(visible_events) == 0
 
     def test_apply_event_filtering_case_insensitive_user(self):
         """Test that user detection is case-insensitive."""
@@ -254,10 +258,12 @@ def test_apply_event_filtering_case_insensitive_user(self):
         # Apply strict TTL filter
         session.apply_event_filtering(event_ttl_seconds=2.0)
 
-        # Should preserve the last user message (case-insensitive)
-        assert len(session.events) == 1
-        assert session.events[0].author.lower() == "user"
-        assert session.events[0].get_text() == "Message from uSeR"
+        # Last user message is preserved as model-visible (case-insensitive).
+        visible_events = [event for event in session.events if event.is_model_visible()]
+        assert len(session.events) == 5
+        assert len(visible_events) == 1
+        assert visible_events[0].author.lower() == "user"
+        assert visible_events[0].get_text() == "Message from uSeR"
 
     def test_apply_event_filtering_max_events_less_than_one(self):
         """Test that max_events <= 0 is treated as no limit."""
@@ -317,11 +323,12 @@ def test_add_event_with_filtering(self):
                           content=Content(parts=[Part.from_text(text=f"Message {i}")]))
             session.add_event(event, max_events=5)
 
-        # After adding 10 events with max_events=5, filtering finds first user message
-        # in the last 5 events (Messages 5-9), which is Message 6, keeps Message 6-9 (4 events)
-        assert len(session.events) == 4
-        assert session.events[0].get_text() == "Message 6"
-        assert session.events[-1].get_text() == "Message 9"
+        # Raw events remain, while only the model-visible window is trimmed.
+        visible_events = [event for event in session.events if event.is_model_visible()]
+        assert len(session.events) == 10
+        assert len(visible_events) == 4
+        assert visible_events[0].get_text() == "Message 6"
+        assert visible_events[-1].get_text() == "Message 9"
 
     def test_apply_event_filtering_keeps_first_user_message_and_after(self):
         """Test that filtering keeps the first user message and all events after it."""
@@ -354,7 +361,9 @@ def test_apply_event_filtering_keeps_first_user_message_and_after(self):
         # Apply max_events filter with small limit
         session.apply_event_filtering(max_events=3)
 
-        # Should keep user message and events after it
-        assert len(session.events) == 4  # 1 user + 3 agent responses
-        assert session.events[0].author == "user"
-        assert session.events[0].get_text() == "User question"
+        # When the retained tail has no user message, fallback keeps the last user message only.
+        visible_events = [event for event in session.events if event.is_model_visible()]
+        assert len(session.events) == 7
+        assert len(visible_events) == 1
+        assert visible_events[0].author == "user"
+        assert visible_events[0].get_text() == "User question"
diff --git a/tests/sessions/test_session_summarizer.py b/tests/sessions/test_session_summarizer.py
index 3077213..5cfc98b 100644
--- a/tests/sessions/test_session_summarizer.py
+++ b/tests/sessions/test_session_summarizer.py
@@ -340,7 +340,10 @@ async def mock_generate(request, stream=False, ctx=None):
         summary_text, result_events = await summarizer.create_session_summary_by_events(
             events, "s1", keep_recent_count=3)
         assert summary_text is not None
-        assert len(result_events) == 4  # 1 summary + 3 recent
+        assert len(result_events) == 11  # preserve all original events + 1 summary
+        visible_events = [event for event in result_events if event.is_model_visible()]
+        assert len(visible_events) == 4  # 1 summary + 3 recent
+        assert any(event.is_summary_event() for event in result_events)
 
     async def test_summary_without_keep_recent(self):
         model = _make_model_mock()
@@ -356,7 +359,9 @@ async def mock_generate(request, stream=False, ctx=None):
         summary_text, result_events = await summarizer.create_session_summary_by_events(
             events, "s1", keep_recent_count=None)
         assert summary_text is not None
-        assert len(result_events) == 1  # only summary event
+        assert len(result_events) == 6  # preserve all original events + 1 summary
+        visible_events = [event for event in result_events if event.is_model_visible()]
+        assert len(visible_events) == 1  # only summary event remains model-visible
 
     async def test_summary_no_events(self):
         model = _make_model_mock()
@@ -397,7 +402,10 @@ async def mock_generate(request, stream=False, ctx=None):
         session = _make_session(events=[_make_event(text=f"msg{i}") for i in range(10)])
         result = await summarizer.create_session_summary(session)
         assert result is not None
-        assert len(session.events) == 3  # 1 summary + 2 recent
+        assert len(session.events) == 11  # preserve all original events + 1 summary
+        visible_events = [event for event in session.events if event.is_model_visible()]
+        assert len(visible_events) == 3  # 1 summary + 2 recent
+        assert any(event.is_summary_event() for event in session.events)
 
     async def test_summary_no_update_on_failure(self):
         model = _make_model_mock()
diff --git a/trpc_agent_sdk/agents/_langgraph_agent.py b/trpc_agent_sdk/agents/_langgraph_agent.py
index 8fbc923..446895f 100644
--- a/trpc_agent_sdk/agents/_langgraph_agent.py
+++ b/trpc_agent_sdk/agents/_langgraph_agent.py
@@ -420,7 +420,9 @@ def _extract_resume_command(self, events: list[Event]) -> Optional[Command]:
         # Must use checkpointer to resume
         if not self.graph.checkpointer or len(events) == 0:
             return None
-        last_event = events[-1]
+        last_event = next((event for event in reversed(events) if event.is_model_visible()), None)
+        if not last_event:
+            return None
         if last_event.author == "user" and last_event.content and last_event.content.parts:
             part = last_event.content.parts[0]
             fc_rsp = part.function_response
@@ -676,6 +678,8 @@ def _get_last_human_messages(self, events: list[Event]) -> list[HumanMessage]:
         """
         messages = []
         for event in reversed(events):
+            if not event.is_model_visible():
+                continue
             if messages and event.author != "user":
                 break
             if event.author == "user" and event.content and event.content.parts:
@@ -718,6 +722,8 @@ def _get_conversation_with_agent(self, events: list[Event]) -> list[Union[HumanM
 
         messages = []
         for event in events:
+            if not event.is_model_visible():
+                continue
             if not event.content or not event.content.parts:
                 continue
 
diff --git a/trpc_agent_sdk/agents/core/_history_processor.py b/trpc_agent_sdk/agents/core/_history_processor.py
index 96c559f..3b45641 100644
--- a/trpc_agent_sdk/agents/core/_history_processor.py
+++ b/trpc_agent_sdk/agents/core/_history_processor.py
@@ -114,6 +114,10 @@ def filter_events(
 
         filtered_events = []
         for event in events:
+            # Step 0.5: Model visibility filtering
+            if not event.is_model_visible():
+                continue
+
             # Step 1: Timeline filtering
             if not self._should_include_event_by_timeline(event, self.timeline_filter_mode, ctx):
                 continue
diff --git a/trpc_agent_sdk/agents/core/_skill_processor.py b/trpc_agent_sdk/agents/core/_skill_processor.py
index 988eacb..f702130 100644
--- a/trpc_agent_sdk/agents/core/_skill_processor.py
+++ b/trpc_agent_sdk/agents/core/_skill_processor.py
@@ -9,7 +9,6 @@
 
 import json
 from typing import Any
-from typing import Callable
 from typing import List
 from typing import Optional
 
@@ -22,6 +21,7 @@
 from trpc_agent_sdk.skills import SkillLoadModeNames
 from trpc_agent_sdk.skills import SkillProfileFlags
 from trpc_agent_sdk.skills import SkillProfileNames
+from trpc_agent_sdk.skills import SkillRepositoryResolver
 from trpc_agent_sdk.skills import SkillToolsNames
 from trpc_agent_sdk.skills import docs_scan_prefix
 from trpc_agent_sdk.skills import docs_state_key
@@ -305,7 +305,7 @@ def __init__(
         forbidden_tools: Optional[list[str]] = None,
         tool_flags: Optional[SkillProfileFlags] = None,
         exec_tools_disabled: bool = False,
-        repo_resolver: Optional[Callable[[InvocationContext], BaseSkillRepository]] = None,
+        repo_resolver: Optional[SkillRepositoryResolver] = None,
         max_loaded_skills: int = 0,
     ) -> None:
         self._skill_repository = skill_repository
diff --git a/trpc_agent_sdk/code_executors/__init__.py b/trpc_agent_sdk/code_executors/__init__.py
index 276f4db..75c2f1f 100644
--- a/trpc_agent_sdk/code_executors/__init__.py
+++ b/trpc_agent_sdk/code_executors/__init__.py
@@ -19,6 +19,7 @@
 from ._base_workspace_runtime import BaseWorkspaceRuntime
 from ._base_workspace_runtime import DefaultWorkspace
 from ._base_workspace_runtime import new_default_workspace_runtime
+from ._base_workspace_runtime import WorkspaceRuntimeResolver
 from ._code_executor_context import CodeExecutorContext
 from ._constants import DEFAULT_CREATE_TIMEOUT_SEC
 from ._constants import DEFAULT_FILE_MODE
@@ -102,6 +103,7 @@
     "BaseWorkspaceRuntime",
     "DefaultWorkspace",
     "new_default_workspace_runtime",
+    "WorkspaceRuntimeResolver",
     "CodeExecutorContext",
     "DEFAULT_CREATE_TIMEOUT_SEC",
     "DEFAULT_FILE_MODE",
diff --git a/trpc_agent_sdk/code_executors/_base_workspace_runtime.py b/trpc_agent_sdk/code_executors/_base_workspace_runtime.py
index 3ee1e1e..d1a43ec 100644
--- a/trpc_agent_sdk/code_executors/_base_workspace_runtime.py
+++ b/trpc_agent_sdk/code_executors/_base_workspace_runtime.py
@@ -13,6 +13,7 @@
 from abc import ABC
 from abc import abstractmethod
 from typing import Callable
+from typing import TypeAlias
 from typing import List
 from typing import Optional
 
@@ -308,3 +309,7 @@ def new_default_workspace_runtime(
         DefaultWorkspace
     """
     return DefaultWorkspace(manager=manager, fs=fs, runner=runner)
+
+
+WorkspaceRuntimeResolver: TypeAlias = Callable[[InvocationContext], BaseWorkspaceRuntime]
+"""Callback to resolve a workspace runtime."""
diff --git a/trpc_agent_sdk/dsl/graph/_graph_agent.py b/trpc_agent_sdk/dsl/graph/_graph_agent.py
index 12e05c1..4e4ec5c 100644
--- a/trpc_agent_sdk/dsl/graph/_graph_agent.py
+++ b/trpc_agent_sdk/dsl/graph/_graph_agent.py
@@ -413,6 +413,8 @@ def _build_initial_state(self, ctx: InvocationContext) -> GraphState:
         user_input = ""
         user_input_event = None
         for event in reversed(ctx.session.events):
+            if not event.is_model_visible():
+                continue
             if event.author == "user" and event.content and event.content.parts:
                 for part in event.content.parts:
                     if part.text:
@@ -430,6 +432,8 @@ def _build_initial_state(self, ctx: InvocationContext) -> GraphState:
         messages = []
         if not has_saved_checkpoint:
             for event in ctx.session.events:
+                if not event.is_model_visible():
+                    continue
                 if event.content:
                     # Skip the user input event - it will be added via STATE_KEY_USER_INPUT
                     if event is user_input_event:
diff --git a/trpc_agent_sdk/dsl/graph/_node_action/_agent.py b/trpc_agent_sdk/dsl/graph/_node_action/_agent.py
index adf1e7a..eacad52 100644
--- a/trpc_agent_sdk/dsl/graph/_node_action/_agent.py
+++ b/trpc_agent_sdk/dsl/graph/_node_action/_agent.py
@@ -187,7 +187,7 @@ async def execute(self, state: State) -> dict[str, Any]:
                         if text_parts:
                             last_response = text_parts[-1]
 
-                    if not event.visible:
+                    if not event.visible or not event.is_model_visible():
                         if event.actions and event.actions.transfer_to_agent:
                             raise ValueError("Agent transfer requested but invisible is not allowed.")
                         continue
@@ -221,7 +221,7 @@ async def execute(self, state: State) -> dict[str, Any]:
                     await self._run_agent_event_callbacks(state, error_event)
                     if hasattr(child_session, "events"):
                         child_session.events.append(error_event.model_copy(deep=True))
-                    if error_event.visible:
+                    if error_event.visible or error_event.is_model_visible():
                         self.writer.write_event(error_event)
                     break
 
diff --git a/trpc_agent_sdk/evaluation/_eval_criterion.py b/trpc_agent_sdk/evaluation/_eval_criterion.py
index 4d8549e..6e4cbc2 100644
--- a/trpc_agent_sdk/evaluation/_eval_criterion.py
+++ b/trpc_agent_sdk/evaluation/_eval_criterion.py
@@ -196,7 +196,7 @@ def _json_deep_equal(self, actual: Any, expected: Any) -> bool:
         """Recursive equality using self.number_tolerance for numeric comparison."""
         if actual is None and expected is None:
             return True
-        if type(actual) != type(expected):
+        if type(actual) is not type(expected):
             return False
         tol = 1e-6 if self.number_tolerance is None else self.number_tolerance
         if isinstance(actual, (int, float)) and isinstance(expected, (int, float)):
diff --git a/trpc_agent_sdk/evaluation/_eval_session_service.py b/trpc_agent_sdk/evaluation/_eval_session_service.py
index b5bce27..d8c97b8 100644
--- a/trpc_agent_sdk/evaluation/_eval_session_service.py
+++ b/trpc_agent_sdk/evaluation/_eval_session_service.py
@@ -44,9 +44,13 @@ async def create_session(
             agent_context=agent_context,
         )
         if context_messages:
+            user_messages = []
             for content in reversed(context_messages):
                 author = content.role or "user"
-                session.events.insert(0, Event(author=author, content=content))
+                user_messages.append(Event(author=author, content=content))
+            if user_messages:
+                user_messages.reverse()
+                session.insert_events(user_messages)
             await self._inner.update_session(session)
         return session
 
diff --git a/trpc_agent_sdk/evaluation/_llm_judge.py b/trpc_agent_sdk/evaluation/_llm_judge.py
index 32f4452..b68ebbe 100644
--- a/trpc_agent_sdk/evaluation/_llm_judge.py
+++ b/trpc_agent_sdk/evaluation/_llm_judge.py
@@ -284,8 +284,8 @@ def _parse_final_response(self, response_text: str) -> ScoreResult:
         try:
             obj = FinalResponseOutput.model_validate_json(self._extract_json(response_text))
         except Exception as e:
-            raise ValueError(f"failed to parse final response JSON: {e}" +
-                             (f"; got: {(response_text or '')[:200]!r}" if response_text else "")) from e
+            response_preview = f"; got: {(response_text or '')[:200]!r}" if response_text else ""
+            raise ValueError(f"failed to parse final response JSON: {e}{response_preview}") from e
         label = obj.is_the_agent_response_valid.strip().lower()
         score = 1.0 if label == "valid" else 0.0
         return ScoreResult(score=score, reason=obj.reasoning.strip())
@@ -294,8 +294,8 @@ def _parse_rubric_response(self, response_text: str) -> ScoreResult:
         try:
             obj = RubricJudgeOutput.model_validate_json(self._extract_json(response_text))
         except Exception as e:
-            raise ValueError(f"failed to parse rubric response JSON: {e}" +
-                             (f"; got: {(response_text or '')[:500]!r}" if response_text else "")) from e
+            response_preview = f"; got: {(response_text or '')[:500]!r}" if response_text else ""
+            raise ValueError(f"failed to parse rubric response JSON: {e}{response_preview}") from e
         if not obj.items:
             raise ValueError("rubric response JSON contains empty items array")
         rubric_scores: list[RubricScore] = []
diff --git a/trpc_agent_sdk/evaluation/_trajectory_evaluator.py b/trpc_agent_sdk/evaluation/_trajectory_evaluator.py
index 605c1a7..e776b6f 100644
--- a/trpc_agent_sdk/evaluation/_trajectory_evaluator.py
+++ b/trpc_agent_sdk/evaluation/_trajectory_evaluator.py
@@ -37,7 +37,6 @@
 from ._criterion_registry import CRITERION_REGISTRY
 from ._eval_case import Invocation
 from ._eval_case import get_all_tool_calls
-from ._eval_criterion import ToolTrajectoryCriterion
 from ._eval_metrics import EvalMetric
 from ._eval_metrics import EvalStatus
 from ._eval_metrics import Interval
diff --git a/trpc_agent_sdk/evaluation/_utils.py b/trpc_agent_sdk/evaluation/_utils.py
index d127640..605bba3 100644
--- a/trpc_agent_sdk/evaluation/_utils.py
+++ b/trpc_agent_sdk/evaluation/_utils.py
@@ -37,7 +37,7 @@
 
 def _result_label_width() -> int:
     """Width for aligned result labels (Agent Name, Eval Set, etc.)."""
-    return max(len(l) for l in RESULT_LABELS)
+    return max(len(label) for label in RESULT_LABELS)
 
 
 class MetricRunRecord:
diff --git a/trpc_agent_sdk/events/_event.py b/trpc_agent_sdk/events/_event.py
index d6e9169..ed5537d 100644
--- a/trpc_agent_sdk/events/_event.py
+++ b/trpc_agent_sdk/events/_event.py
@@ -37,6 +37,9 @@
 from trpc_agent_sdk.types import FunctionCall
 from trpc_agent_sdk.types import FunctionResponse
 
+_EVENT_FLAG_MODEL_VISIBLE = 1 << 0
+_EVENT_FLAG_SUMMARY = 1 << 1
+
 
 class Event(LlmResponse):
     """Represents an event in a conversation between agents and users.
@@ -55,6 +58,7 @@ class Event(LlmResponse):
       id: The unique identifier of the event.
       timestamp: The timestamp of the event.
       visible: Whether the event is visible to outside observers. Default is True.
+      model_flags: Bit flags controlling model visibility and summary state.
       request_id: Optional request ID for tracking across system boundaries.
       parent_invocation_id: Optional parent invocation ID for nested agent executions.
       tag: Optional business-specific labels for filtering/routing.
@@ -160,6 +164,13 @@ class Event(LlmResponse):
     Provides a standardized way to identify event types alongside the legacy event_type in state_delta.
     """
 
+    model_flags: int = _EVENT_FLAG_MODEL_VISIBLE
+    """Bit flags for event model-state control.
+
+    - MODEL_VISIBLE flag controls whether this event can be seen by model history builders.
+    - SUMMARY flag marks this event as a summary-generated event.
+    """
+
     def model_post_init(self, __context):
         """Post initialization logic for the event."""
         # Generates a random ID for the event.
@@ -173,6 +184,28 @@ def is_final_response(self) -> bool:
         return (not self.get_function_calls() and not self.get_function_responses() and not self.partial
                 and not self.has_trailing_code_execution_result() and not self.has_trailing_executable_code())
 
+    def is_model_visible(self) -> bool:
+        """Returns whether the event should be visible to model history."""
+        return bool(self.model_flags & _EVENT_FLAG_MODEL_VISIBLE)
+
+    def is_summary_event(self) -> bool:
+        """Returns whether the event is generated as a summary event."""
+        return bool(self.model_flags & _EVENT_FLAG_SUMMARY)
+
+    def set_model_visible(self, model_visible: bool) -> None:
+        """Set whether this event can be seen by model history builders."""
+        if model_visible:
+            self.model_flags |= _EVENT_FLAG_MODEL_VISIBLE
+        else:
+            self.model_flags &= ~_EVENT_FLAG_MODEL_VISIBLE
+
+    def set_summary_event(self, is_summary: bool = True) -> None:
+        """Set whether this event is marked as a summary event."""
+        if is_summary:
+            self.model_flags |= _EVENT_FLAG_SUMMARY
+        else:
+            self.model_flags &= ~_EVENT_FLAG_SUMMARY
+
     def get_function_calls(self) -> list[FunctionCall]:
         """Returns the function calls in the event."""
         func_calls = []
diff --git a/trpc_agent_sdk/models/_anthropic_model.py b/trpc_agent_sdk/models/_anthropic_model.py
index 0f41175..845d960 100644
--- a/trpc_agent_sdk/models/_anthropic_model.py
+++ b/trpc_agent_sdk/models/_anthropic_model.py
@@ -353,7 +353,6 @@ def _create_streaming_tool_call_response(
         current_tool = accumulated_tool_uses[-1]
         tool_name = current_tool.get("name", "")
         tool_id = current_tool.get("id", "")
-        accumulated_args = current_tool.get("accumulated_input", "")
 
         if not tool_name:
             return None
@@ -517,7 +516,6 @@ async def _generate_stream(
         accumulated_tool_uses: list[dict] = []
         # Map content block index to tool use index in accumulated_tool_uses
         block_index_to_tool_index: dict[int, int] = {}
-        last_usage = None
 
         # Get the set of tool names that should stream
         streaming_tool_names = getattr(request, 'streaming_tool_names', None) or set()
@@ -590,16 +588,6 @@ async def _generate_stream(
                                 accumulated_tool_uses.append(tool_entry)
                                 block_index_to_tool_index[block_index] = tool_idx
 
-                    # Handle message delta events (for usage)
-                    elif hasattr(event, "type") and event.type == "message_delta":
-                        if hasattr(event, "usage"):
-                            usage = event.usage
-                            last_usage = GenerateContentResponseUsageMetadata(
-                                prompt_token_count=0,
-                                candidates_token_count=usage.output_tokens,
-                                total_token_count=usage.output_tokens,
-                            )
-
             # Get the final message for complete usage stats
             final_message = await stream.get_final_message()
 
diff --git a/trpc_agent_sdk/models/_openai_model.py b/trpc_agent_sdk/models/_openai_model.py
index bba4e59..d1c9e2e 100644
--- a/trpc_agent_sdk/models/_openai_model.py
+++ b/trpc_agent_sdk/models/_openai_model.py
@@ -175,6 +175,9 @@ def __init__(
 
         # Default generation config that can be overridden per request
         self.generate_content_config = generate_content_config
+        # Optional hard cap for tool-response payload injected into model
+        # context. Disabled by default; callers (e.g. OpenClaw) can opt in.
+        self._tool_response_clip_chars = int(kwargs.get("tool_response_clip_chars", 0) or 0)
 
         # Validate tool_prompt parameter
         if isinstance(self.tool_prompt, str):
@@ -379,6 +382,7 @@ def _format_messages(self, request: LlmRequest) -> List[Dict[str, Any]]:
                         else:
                             content += str(func_response.response)
                         content += "\n"
+                    content = self._clip_tool_response_text(content, "tool_response_merged")
                     if len(content) > 0:
                         tool_message = {
                             const.ROLE: const.USER,
@@ -388,13 +392,16 @@ def _format_messages(self, request: LlmRequest) -> List[Dict[str, Any]]:
                 else:
                     for func_response in function_responses:
                         # Standard tool message format for OpenAI API
+                        raw_text = (json.dumps(func_response.response, ensure_ascii=False) if isinstance(
+                            func_response.response, dict) else str(func_response.response))
+                        clipped_text = self._clip_tool_response_text(
+                            raw_text,
+                            getattr(func_response, "name", "tool"),
+                        )
                         tool_message = {
-                            const.ROLE:
-                            const.TOOL,
-                            const.TOOL_CALL_ID:
-                            getattr(func_response, "id", "unknown"),
-                            const.CONTENT: (json.dumps(func_response.response, ensure_ascii=False) if isinstance(
-                                func_response.response, dict) else str(func_response.response)),
+                            const.ROLE: const.TOOL,
+                            const.TOOL_CALL_ID: getattr(func_response, "id", "unknown"),
+                            const.CONTENT: clipped_text,
                         }
                         formatted_messages.append(tool_message)
 
@@ -860,7 +867,6 @@ def _create_streaming_tool_call_response(
         for idx, tool_call_data in enumerate(accumulated_tool_calls):
             function_map: dict = tool_call_data.get(ToolKey.FUNCTION, {})
             name = function_map.get(ToolKey.NAME, "")
-            accumulated_args = function_map.get(ToolKey.ARGUMENTS, "")
             tool_call_id = tool_call_data.get(ToolKey.ID, "")
 
             if not name:
@@ -1132,6 +1138,16 @@ def _convert_tools_to_openai_format(self, tools: List[Tool]) -> List[Dict[str, A
 
         return openai_tools
 
+    def _clip_tool_response_text(self, text: str, tool_name: str) -> str:
+        """Hard-clip tool response text to protect model context budget."""
+        limit = self._tool_response_clip_chars
+        if limit <= 0 or len(text) <= limit:
+            return text
+        truncated = len(text) - limit
+        suffix = f"\n...[TRUNCATED {truncated} CHARS FROM TOOL RESPONSE: {tool_name}]"
+        keep = max(0, limit - len(suffix))
+        return text[:keep] + suffix
+
     def _convert_schema_to_openai_format(self, schema: Schema) -> Dict[str, Any]:
         """Convert Google GenAI Schema to OpenAI parameters format.
 
diff --git a/trpc_agent_sdk/models/tool_prompt/_factory.py b/trpc_agent_sdk/models/tool_prompt/_factory.py
index 1fab498..6c0f259 100644
--- a/trpc_agent_sdk/models/tool_prompt/_factory.py
+++ b/trpc_agent_sdk/models/tool_prompt/_factory.py
@@ -80,7 +80,6 @@ def get_factory() -> ToolPromptFactory:
     Raises:
         RuntimeError: If factory is not initialized
     """
-    global _factory  # pylint: disable=invalid-name
     if _factory is None:
         raise RuntimeError("Factory is not initialized. Call initialize() first.")
     return _factory
diff --git a/trpc_agent_sdk/runners.py b/trpc_agent_sdk/runners.py
index 430c9a1..b8fe307 100644
--- a/trpc_agent_sdk/runners.py
+++ b/trpc_agent_sdk/runners.py
@@ -12,7 +12,11 @@
 from __future__ import annotations
 
 import asyncio
+import concurrent.futures
+import threading
 from typing import AsyncGenerator
+from typing import Awaitable
+from typing import Callable
 from typing import Optional
 
 from trpc_agent_sdk import cancel
@@ -38,6 +42,121 @@
 from trpc_agent_sdk.types import Part
 
 
+class _PostTurnWorkerThread:
+    """Dedicated thread + event loop for deferred post-turn processing."""
+
+    def __init__(
+        self,
+        *,
+        name: str,
+        run_job: Callable[[InvocationContext], Awaitable[None]],
+        maxsize: int,
+    ) -> None:
+        self._name = name
+        self._run_job = run_job
+        self._maxsize = max(1, int(maxsize))
+        self._thread: threading.Thread | None = None
+        self._loop: asyncio.AbstractEventLoop | None = None
+        self._queue: asyncio.Queue[InvocationContext | None] | None = None
+        self._ready = threading.Event()
+        self._closed = False
+        self._state_lock = threading.Lock()
+
+    def start(self) -> None:
+        if self._thread is not None and self._thread.is_alive():
+            return
+        with self._state_lock:
+            self._closed = False
+        self._ready.clear()
+        self._thread = threading.Thread(target=self._run, name=self._name, daemon=True)
+        self._thread.start()
+        self._ready.wait()
+
+    def submit(self, job: InvocationContext) -> bool:
+        self.start()
+        with self._state_lock:
+            if self._closed:
+                return False
+        loop = self._loop
+        queue = self._queue
+        if loop is None or queue is None or not loop.is_running():
+            return False
+        loop.call_soon_threadsafe(self._put_nowait, queue, job)
+        return True
+
+    def stop(self, *, drain_timeout: float = 10.0, join_timeout: float = 5.0) -> bool:
+        with self._state_lock:
+            self._closed = True
+        loop = self._loop
+        queue = self._queue
+        thread = self._thread
+        if loop is not None and queue is not None and loop.is_running():
+            future = asyncio.run_coroutine_threadsafe(
+                self._stop_after_drain(queue, drain_timeout),
+                loop,
+            )
+            try:
+                future.result(timeout=drain_timeout + 1.0)
+            except concurrent.futures.TimeoutError:
+                pass
+        if thread is not None:
+            thread.join(timeout=join_timeout)
+            return not thread.is_alive()
+        return True
+
+    def _run(self) -> None:
+        loop = asyncio.new_event_loop()
+        asyncio.set_event_loop(loop)
+        self._loop = loop
+        self._queue = asyncio.Queue(maxsize=self._maxsize)
+        self._ready.set()
+        try:
+            loop.run_until_complete(self._worker())
+        finally:
+            try:
+                pending = asyncio.all_tasks(loop)
+                for task in pending:
+                    task.cancel()
+                if pending:
+                    loop.run_until_complete(asyncio.gather(*pending, return_exceptions=True))
+            finally:
+                loop.close()
+                self._loop = None
+                self._queue = None
+
+    async def _worker(self) -> None:
+        assert self._queue is not None
+        while True:
+            item = await self._queue.get()
+            try:
+                if item is None:
+                    return
+                await self._run_job(invocation_context=item)
+            finally:
+                self._queue.task_done()
+
+    @staticmethod
+    def _put_nowait(
+        queue: asyncio.Queue[InvocationContext | None],
+        job: InvocationContext,
+    ) -> None:
+        try:
+            queue.put_nowait(job)
+        except asyncio.QueueFull:
+            logger.warning("Post-turn thread queue full; dropping deferred job")
+
+    @staticmethod
+    async def _stop_after_drain(
+        queue: asyncio.Queue[InvocationContext | None],
+        drain_timeout: float,
+    ) -> None:
+        try:
+            await asyncio.wait_for(queue.join(), timeout=drain_timeout)
+        except asyncio.TimeoutError:
+            pass
+        await queue.put(None)
+
+
 class Runner:
     """The Runner class is used to run agents.
 
@@ -72,6 +191,9 @@ def __init__(
         session_service: BaseSessionService,
         artifact_service: Optional[BaseArtifactService] = None,
         memory_service: Optional[BaseMemoryService] = None,
+        enable_post_turn_processing: bool = True,
+        defer_post_turn_processing: bool = False,
+        post_turn_queue_maxsize: int = 256,
     ):
         """Initializes the Runner.
 
@@ -81,12 +203,85 @@ def __init__(
             artifact_service: The artifact service for the runner.
             session_service: The session service for the runner.
             memory_service: The memory service for the runner.
+            enable_post_turn_processing: If False, skip post-turn summarization
+                and memory persistence entirely for this runner.
+            defer_post_turn_processing: If True, session summarization + memory
+                persistence run in a dedicated thread/event loop so request
+                completion is not blocked by post-turn I/O/LLM latency.
+            post_turn_queue_maxsize: Max buffered post-turn jobs per runner.
         """
         self.app_name = app_name
         self.agent = agent
         self.artifact_service = artifact_service
         self.session_service = session_service
         self.memory_service = memory_service
+        self._enable_post_turn_processing = enable_post_turn_processing
+        self._defer_post_turn_processing = defer_post_turn_processing
+        self._post_turn_thread: _PostTurnWorkerThread | None = None
+        self._post_turn_queue_maxsize = max(1, int(post_turn_queue_maxsize))
+
+    async def _run_post_turn_processing(
+        self,
+        *,
+        invocation_context: InvocationContext,
+    ) -> None:
+        """Run post-turn summarization + memory persistence."""
+        session = invocation_context.session
+        try:
+            await self.session_service.create_session_summary(session, ctx=invocation_context)
+            if self.memory_service and self.memory_service.enabled:
+                await self.memory_service.store_session(session, agent_context=invocation_context.agent_context)
+        except Exception as exc:  # noqa: BLE001
+            logger.error(
+                "Post-turn processing failed for session %s: %s",
+                getattr(session, "id", "?"),
+                exc,
+            )
+
+    def _ensure_post_turn_worker(self) -> None:
+        """Start post-turn worker lazily on first enqueue."""
+        if not self._defer_post_turn_processing:
+            return
+        if self._post_turn_thread is None:
+            self._post_turn_thread = _PostTurnWorkerThread(
+                name=f"runner-post-turn:{self.app_name}",
+                run_job=self._run_post_turn_processing,
+                maxsize=self._post_turn_queue_maxsize,
+            )
+        self._post_turn_thread.start()
+
+    async def _schedule_post_turn_processing(
+        self,
+        *,
+        invocation_context: InvocationContext,
+    ) -> None:
+        """Schedule post-turn work; non-blocking when deferred mode is on."""
+        if not self._enable_post_turn_processing:
+            return
+        if not self._defer_post_turn_processing:
+            await self._run_post_turn_processing(invocation_context=invocation_context, )
+            return
+
+        self._ensure_post_turn_worker()
+        if self._post_turn_thread is None:
+            return
+        if not self._post_turn_thread.submit(invocation_context):
+            logger.warning(
+                "Post-turn thread unavailable for app %s; dropping deferred job",
+                self.app_name,
+            )
+
+    async def _shutdown_post_turn_worker(self) -> None:
+        """Gracefully flush and stop deferred post-turn processing."""
+        if self._post_turn_thread is not None:
+            stopped = await asyncio.to_thread(
+                self._post_turn_thread.stop,
+                drain_timeout=10.0,
+                join_timeout=5.0,
+            )
+            if not stopped:
+                logger.warning("Post-turn thread shutdown timed out for app %s", self.app_name)
+            self._post_turn_thread = None
 
     async def cancel_run_async(
         self,
@@ -291,8 +486,8 @@ async def run_async(
                             # Check if transferring to the same agent
                             if transfer_target == current_agent.name:
                                 logger.warning(
-                                    "Transfer to same agent '%s' detected, add 'already on agent' message to let agent continue",
-                                    transfer_target)
+                                    "Transfer to same agent '%s' detected, add 'already on agent'"
+                                    "message to let agent continue", transfer_target)
                                 already_in_event = Event(
                                     invocation_id=invocation_context.invocation_id,
                                     author=current_agent.name,
@@ -345,14 +540,14 @@ async def run_async(
                         logger.debug("No transfer requested by %s, ending execution", current_agent.name)
                         break
 
-                # Trigger summarization if enabled
-                await self.session_service.create_session_summary(session, ctx=invocation_context)
-                if self.memory_service and self.memory_service.enabled:
-                    await self.memory_service.store_session(session, agent_context=agent_context)
+                # Trigger summarization/memory persistence. Can be deferred to a
+                # background worker to avoid blocking request completion.
+                await self._schedule_post_turn_processing(invocation_context=invocation_context, )
 
                 # Compute state after runner execution
                 state_end = dict(session.state)
-                if last_non_streaming_event and last_non_streaming_event.actions and last_non_streaming_event.actions.state_delta:
+                if (last_non_streaming_event and last_non_streaming_event.actions
+                        and last_non_streaming_event.actions.state_delta):
                     state_end.update(last_non_streaming_event.actions.state_delta)
 
                 # Call trace function with runner execution details
@@ -626,6 +821,7 @@ async def close(self):
         2. Close each toolset with proper error handling
         3. Ensure all resources are released before shutdown
         """
+        await self._shutdown_post_turn_worker()
         await self._cleanup_toolsets(self._collect_toolset(self.agent))
         if self.session_service:
             await self.session_service.close()
diff --git a/trpc_agent_sdk/server/a2a/_remote_a2a_agent.py b/trpc_agent_sdk/server/a2a/_remote_a2a_agent.py
index e76540c..52d818c 100644
--- a/trpc_agent_sdk/server/a2a/_remote_a2a_agent.py
+++ b/trpc_agent_sdk/server/a2a/_remote_a2a_agent.py
@@ -323,6 +323,8 @@ def _build_outgoing_message(self, ctx: InvocationContext) -> Optional[Message]:
 
         user_event = None
         for event in reversed(ctx.session.events):
+            if not event.is_model_visible():
+                continue
             if event.author == "user" and event.content:
                 user_event = event
                 break
diff --git a/trpc_agent_sdk/server/ag_ui/_core/_event_translator.py b/trpc_agent_sdk/server/ag_ui/_core/_event_translator.py
index 442efb8..747fbab 100644
--- a/trpc_agent_sdk/server/ag_ui/_core/_event_translator.py
+++ b/trpc_agent_sdk/server/ag_ui/_core/_event_translator.py
@@ -197,7 +197,6 @@ async def translate(self, trpc_event: TRPCEvent, thread_id: str, run_id: str) ->
             if trpc_event.actions and trpc_event.actions.state_delta:
                 yield self._create_state_delta_event(trpc_event.actions.state_delta, trpc_event.timestamp)
 
-
             # Handle error events - distinguish recoverable tool errors from fatal system errors.
             # Tool execution errors (with function_response) are recoverable: the error is already
             # passed back to the LLM as a tool result, so the LLM can retry or adjust its approach.
@@ -205,13 +204,12 @@ async def translate(self, trpc_event: TRPCEvent, thread_id: str, run_id: str) ->
             # emit RunErrorEvent to terminate the run.
             if trpc_event.is_error() and not function_responses:
                 # Fatal system/LLM error - emit RunErrorEvent to terminate the run
-                logger.error("Fatal error (non-recoverable), error_code=%s, error_message=%s",
-                             trpc_event.error_code, trpc_event.error_message)
+                logger.error("Fatal error (non-recoverable), error_code=%s, error_message=%s", trpc_event.error_code,
+                             trpc_event.error_message)
                 # Force close any streaming message before emitting error
                 async for close_event in self.force_close_streaming_message():
                     yield close_event
-                error_msg = (trpc_event.error_message
-                             or (trpc_event.custom_metadata or {}).get("error")
+                error_msg = (trpc_event.error_message or (trpc_event.custom_metadata or {}).get("error")
                              or "Unknown error")
                 yield RunErrorEvent(
                     type=EventType.RUN_ERROR,
diff --git a/trpc_agent_sdk/server/ag_ui/_plugin/_manager.py b/trpc_agent_sdk/server/ag_ui/_plugin/_manager.py
index 5d812a8..cd922e0 100644
--- a/trpc_agent_sdk/server/ag_ui/_plugin/_manager.py
+++ b/trpc_agent_sdk/server/ag_ui/_plugin/_manager.py
@@ -47,7 +47,7 @@ def get_service(self, service_name: str) -> Optional[AgUiService]:
 
         Args:
           service_name: The name of the service.
-        
+
         Returns:
           The AG-UI service.
         """
diff --git a/trpc_agent_sdk/server/ag_ui/_plugin/_service.py b/trpc_agent_sdk/server/ag_ui/_plugin/_service.py
index b0e7810..5cb5c13 100644
--- a/trpc_agent_sdk/server/ag_ui/_plugin/_service.py
+++ b/trpc_agent_sdk/server/ag_ui/_plugin/_service.py
@@ -35,7 +35,8 @@ def __init__(self, service_name: str, app: Optional[FastAPI] = None, agents: Dic
 
         Args:
             service_name: Name of the service used for route registration
-            app: Optional FastAPI app instance. If not provided, the service will not be registered with the FastAPI app.
+            app: Optional FastAPI app instance. If not provided, the service
+                 will not be registered with the FastAPI app.
             agents: Optional dictionary of agents keyed by URI path. If not provided,
                    an empty dictionary will be used.
         """
diff --git a/trpc_agent_sdk/server/agents/claude/_claude_agent.py b/trpc_agent_sdk/server/agents/claude/_claude_agent.py
index d745c61..9327399 100644
--- a/trpc_agent_sdk/server/agents/claude/_claude_agent.py
+++ b/trpc_agent_sdk/server/agents/claude/_claude_agent.py
@@ -832,6 +832,8 @@ def _extract_latest_user_message(self, ctx: InvocationContext) -> Optional[str]:
 
         # Look through events in reverse to find latest user message
         for event in reversed(ctx.session.events):
+            if not event.is_model_visible():
+                continue
             if event.author == "user" and event.content and event.content.parts:
                 for part in event.content.parts:
                     if part.text:
@@ -902,6 +904,8 @@ def _build_prompt_with_history(self, ctx: InvocationContext) -> Optional[str]:
 
         # Iterate through events to build conversation history
         for event in ctx.session.events:
+            if not event.is_model_visible():
+                continue
             if not event.content or not event.content.parts:
                 continue
 
diff --git a/trpc_agent_sdk/server/knowledge/langchain_knowledge.py b/trpc_agent_sdk/server/knowledge/langchain_knowledge.py
index c1095a1..c75be51 100644
--- a/trpc_agent_sdk/server/knowledge/langchain_knowledge.py
+++ b/trpc_agent_sdk/server/knowledge/langchain_knowledge.py
@@ -8,8 +8,9 @@
 from typing import Any
 from typing import Dict
 from typing import List
-from pydantic import BaseModel
+from typing_extensions import override
 
+from pydantic import BaseModel
 
 from langchain_core.document_loaders import BaseLoader
 from langchain_core.documents import BaseDocumentTransformer
@@ -31,15 +32,11 @@
     from langchain.chains.base import Chain
 
 from trpc_agent_sdk.context import AgentContext
-from trpc_agent_sdk.knowledge import (
-    KnowledgeBase,
-    KnowledgeFilterExpr,
-    SearchDocument,
-    SearchRequest,
-    SearchResult,
-)
+from trpc_agent_sdk.knowledge import KnowledgeBase
+from trpc_agent_sdk.knowledge import SearchDocument
+from trpc_agent_sdk.knowledge import SearchRequest
+from trpc_agent_sdk.knowledge import SearchResult
 from trpc_agent_sdk.log import logger
-from typing_extensions import override
 
 
 # SearchType vector retrieval type
@@ -73,7 +70,8 @@ def __init__(self,
         """Implement the default logic for RAG, integrate with tRPC-Agent framework, support Langchain ecosystem
 
         :params:
-        chain: complete Langchain chain; if a complete Langchain chain is already available, it can be directly called, other configurations are ignored
+        chain: complete Langchain chain; if available, it can be called directly
+            and other configurations are ignored
         prompt_template: Langchain Prompt template
         document_loader: Langchain document loader
         document_transformer: Langchain document transformer
@@ -81,7 +79,8 @@ def __init__(self,
         vectorstore: Langchain vector database
         retriever: Langchain retriever
         """
-        # Initialize embedder etc. components, if chain is not empty, other configurations will be ignored and Chain will be called directly
+        # Initialize embedder etc. components. If chain is set, other
+        # configurations are ignored and the chain is called directly.
         self.chain = chain
         self.prompt_template = prompt_template
         self.document_loader = document_loader
@@ -145,7 +144,7 @@ def _get_history_message(self, ctx: AgentContext, req: SearchRequest) -> str:
         if assistant_name:
             context += f"assistant: {assistant_name}\n"
         context += f"session_id: {session_id}\n"
-        context += f"content: "
+        context += "content: "
         for msg in history:
             context += msg.text()
 
@@ -164,7 +163,7 @@ async def _run_chain(self, context: str, req: SearchRequest, kwargs):
         if req.query.text:
             query = req.query.text
         else:
-            raise ValueError(f"query should be text, but got None")
+            raise ValueError("query should be text, but got None")
 
         chain_runnable_config = None
         if self.common_runnable_config:
@@ -255,7 +254,7 @@ async def search(self, ctx: AgentContext, req: SearchRequest) -> SearchResult:
         if req.query.text:
             query = req.query.text
         else:
-            raise ValueError(f"query should be text, but got None")
+            raise ValueError("query should be text, but got None")
 
         if self.prompt_template:
             query_runnable_config = None
diff --git a/trpc_agent_sdk/server/knowledge/tools/langchain_knowledge_searchtool.py b/trpc_agent_sdk/server/knowledge/tools/langchain_knowledge_searchtool.py
index b87398f..53dacab 100644
--- a/trpc_agent_sdk/server/knowledge/tools/langchain_knowledge_searchtool.py
+++ b/trpc_agent_sdk/server/knowledge/tools/langchain_knowledge_searchtool.py
@@ -161,12 +161,13 @@ def _get_declaration(self) -> Optional[FunctionDeclaration]:
             "dynamic_filter: KnowledgeFilterExpr object.\n"
             "Fields:\n"
             "- field: metadata field path, e.g. metadata.category\n"
-            "- operator: eq, ne, gt, gte, lt, lte, in, not in, like, not like, between, and, or\n"
+            "- operator: eq, ne, gt, gte, lt, lte, in, not in, like, "
+            "not like, between, and, or\n"
             "- value: comparison value, or an array of sub-conditions for and/or\n"
             "Examples:\n"
             "1) {\"field\":\"metadata.category\",\"operator\":\"eq\",\"value\":\"machine-learning\"}\n"
-            "2) {\"operator\":\"and\",\"value\":[{\"field\":\"metadata.status\",\"operator\":\"eq\",\"value\":\"active\"}]}"
-        )
+            "2) {\"operator\":\"and\",\"value\":[{\"field\":\"metadata.status\","
+            "\"operator\":\"eq\",\"value\":\"active\"}]}")
         return FunctionDeclaration(
             name=self.name,
             description=self.description,
diff --git a/trpc_agent_sdk/server/openclaw/_utils.py b/trpc_agent_sdk/server/openclaw/_utils.py
index 4fef5a5..4b4c782 100644
--- a/trpc_agent_sdk/server/openclaw/_utils.py
+++ b/trpc_agent_sdk/server/openclaw/_utils.py
@@ -57,7 +57,7 @@ def register_channel_without_stream_progress(channel: str) -> None:
 
 def merge_assistant_text(current: str, incoming: str) -> str:
     """Merge assistant text chunks while avoiding cumulative duplicates.
-    
+
     Args:
         current: The current text.
         incoming: The incoming text.
diff --git a/trpc_agent_sdk/server/openclaw/config/_config.py b/trpc_agent_sdk/server/openclaw/config/_config.py
index c638516..87884a6 100644
--- a/trpc_agent_sdk/server/openclaw/config/_config.py
+++ b/trpc_agent_sdk/server/openclaw/config/_config.py
@@ -28,7 +28,6 @@
 from pydantic import ConfigDict
 from pydantic import Field
 from trpc_agent_sdk.abc import MemoryServiceConfig
-from trpc_agent_sdk.code_executors import CodeBlockDelimiter
 from trpc_agent_sdk.server.langfuse.tracing.opentelemetry import LangfuseConfig
 
 from ._constants import AGENT_FILE_NAME
diff --git a/trpc_agent_sdk/server/openclaw/session_memory/_claw_summarizer.py b/trpc_agent_sdk/server/openclaw/session_memory/_claw_summarizer.py
index a53ebc6..0b15433 100644
--- a/trpc_agent_sdk/server/openclaw/session_memory/_claw_summarizer.py
+++ b/trpc_agent_sdk/server/openclaw/session_memory/_claw_summarizer.py
@@ -293,10 +293,10 @@ async def _call_llm_for_memory(
         If the first response is empty or fails XML validation a second attempt is made before giving up.
         ``memory_update`` must be at least 20 characters long (guards against the LLM echoing only the XML tags
             or returning a blank body).
-        A response that passes regex parsing but is too short is treated the same as a parse failure and 
+        A response that passes regex parsing but is too short is treated the same as a parse failure and
         triggers a retry.
 
-        Returns ``("", "")`` when all attempts fail (the caller's failure counter will decide whether to raw-archive or 
+        Returns ``("", "")`` when all attempts fail (the caller's failure counter will decide whether to raw-archive or
         simply skip this round).
 
         Args:
@@ -353,11 +353,11 @@ def _find_safe_split(self, events: List[Event], keep_n: int) -> tuple[List[Event
         tool response in the recent window and its matching call in the
         to-be-summarized bucket.
 
-        Starting from ``len(events) - keep_n`` (the ideal split), scan backward until an event authored by ``"user"`` 
+        Starting from ``len(events) - keep_n`` (the ideal split), scan backward until an event authored by ``"user"``
         is found.  Split immediately before that event so the entire user turn (plus any following model / tool events)
         stays in the recent bucket intact.
 
-        If no user event is found before the ideal split (e.g. the session starts with a long system preamble) 
+        If no user event is found before the ideal split (e.g. the session starts with a long system preamble)
         the ideal split is used as-is.
 
         Args:
@@ -385,8 +385,8 @@ def _find_safe_split(self, events: List[Event], keep_n: int) -> tuple[List[Event
     def _make_raw_archive(self, events: List[Event], current_memory: str) -> tuple[str, str]:
         """Build a plain-text ``(history_entry, memory_update)`` without LLM.
 
-        Called after ``MAX_FAILURES_BEFORE_RAW_ARCHIVE`` consecutive LLM failures.  The raw conversation text is 
-        appended to the existing long-term memory so that the session always makes progress even 
+        Called after ``MAX_FAILURES_BEFORE_RAW_ARCHIVE`` consecutive LLM failures.  The raw conversation text is
+        appended to the existing long-term memory so that the session always makes progress even
         when the summarization LLM is unavailable.
 
         Args:
@@ -424,7 +424,7 @@ class ClawSummarizerSessionManager(SummarizerSessionManager):
     """trpc_claw-style summarizer manager that stores dual-layer :class:`SessionSummary` objects.
 
     Uses :class:`ClawSessionSummarizer` by default.
-    ``create_session_summary`` syncs ``raw_events`` before summarizing and stores a :class:`SessionSummary` 
+    ``create_session_summary`` syncs ``raw_events`` before summarizing and stores a :class:`SessionSummary`
     (with accumulated ``history_entries``) in the in-memory cache.
     ``get_session_summary`` returns a :class:`SessionSummary`.
 
diff --git a/trpc_agent_sdk/server/openclaw/ui.py b/trpc_agent_sdk/server/openclaw/ui.py
index 8a249e5..0f5c1a1 100644
--- a/trpc_agent_sdk/server/openclaw/ui.py
+++ b/trpc_agent_sdk/server/openclaw/ui.py
@@ -38,8 +38,8 @@
 
 def _load_browser_html() -> str:
     """Load browser UI HTML template from package data."""
-    return resources.files("trpc_agent_sdk.server.openclaw").joinpath("templates", "ui.html").read_text(
-        encoding="utf-8")
+    return resources.files("trpc_agent_sdk.server.openclaw").joinpath("templates",
+                                                                      "ui.html").read_text(encoding="utf-8")
 
 
 _BROWSER_HTML = _load_browser_html()
diff --git a/trpc_agent_sdk/sessions/_base_session_service.py b/trpc_agent_sdk/sessions/_base_session_service.py
index dd41fb7..cba2d8f 100644
--- a/trpc_agent_sdk/sessions/_base_session_service.py
+++ b/trpc_agent_sdk/sessions/_base_session_service.py
@@ -183,17 +183,17 @@ async def get_session_summary(self, session: Session) -> Optional[str]:
 
     def filter_events(self, session: Session) -> None:
         """Filter events based on the session config."""
+        visible_events = [event for event in session.events if event.is_model_visible()]
         if self._session_config.num_recent_events > 0:
-            session.events = session.events[-self._session_config.num_recent_events:]
+            if len(visible_events) > self._session_config.num_recent_events:
+                hide_count = len(visible_events) - self._session_config.num_recent_events
+                for event in visible_events[:hide_count]:
+                    event.set_model_visible(False)
         if self._session_config.event_ttl_seconds > 0:
             cutoff_timestamp = time.time() - self._session_config.event_ttl_seconds
-            i = len(session.events) - 1
-            while i >= 0:
-                if session.events[i].timestamp <= cutoff_timestamp:
-                    break
-                i -= 1
-            if i >= 0:
-                session.events = session.events[i + 1:]
+            for event in visible_events:
+                if event.timestamp <= cutoff_timestamp:
+                    event.set_model_visible(False)
 
     @override
     async def close(self) -> None:
diff --git a/trpc_agent_sdk/sessions/_in_memory_session_service.py b/trpc_agent_sdk/sessions/_in_memory_session_service.py
index 55dfaa1..938e202 100644
--- a/trpc_agent_sdk/sessions/_in_memory_session_service.py
+++ b/trpc_agent_sdk/sessions/_in_memory_session_service.py
@@ -240,7 +240,7 @@ def _warning(message: str) -> None:
         # Get session with TTL wrapper
         storage_session = self._get_session(app_name, user_id, session_id)
         if storage_session is None:
-            _warning(f"session not found")
+            _warning("session not found")
             return event
 
         # Add event to storage session
diff --git a/trpc_agent_sdk/sessions/_session.py b/trpc_agent_sdk/sessions/_session.py
index 53cde22..d227309 100644
--- a/trpc_agent_sdk/sessions/_session.py
+++ b/trpc_agent_sdk/sessions/_session.py
@@ -8,7 +8,7 @@
 from __future__ import annotations
 
 import time
-from typing import List
+from typing import List, Optional
 
 from pydantic import Field
 from trpc_agent_sdk.abc import SessionABC
@@ -66,33 +66,57 @@ def apply_event_filtering(self, event_ttl_seconds: float = 0.0, max_events: int
         if event_ttl_seconds <= 0 and max_events <= 0:
             return
 
-        # Save original events for potential user message recovery
-        original_events = self.events.copy()
+        # Apply filtering only to the currently model-visible events.  Raw
+        # session events stay in place; events filtered out of this visible
+        # window are hidden from model history.
+        visible_events = [event for event in self.events if event.is_model_visible()]
+        if not visible_events:
+            return
+        retained_events = visible_events.copy()
 
         # Step 1: Apply TTL filtering if configured
         if event_ttl_seconds > 0:
             cutoff_time = time.time() - event_ttl_seconds
-            self.events = [e for e in self.events if e.timestamp >= cutoff_time]
+            retained_events = [e for e in retained_events if e.timestamp >= cutoff_time]
 
         # Step 2: Apply count filtering if configured
         if max_events > 0:
-            if len(self.events) > max_events:
-                self.events = self.events[-max_events:]
-
-        for i, event in enumerate(self.events):
-            if self._is_user_message(event):
-                self.events = self.events[i:]
-                return
+            if len(retained_events) > max_events:
+                retained_events = retained_events[-max_events:]
 
-        # Step 3: If all events were filtered out, insert the first user message at the beginning
-        # Find the last user message from original events
-        for event in reversed(original_events):
+        for i, event in enumerate(retained_events):
             if self._is_user_message(event):
-                self.events.insert(0, event)
-                return
-
-        # If no user message found, keep events empty
-        self.events = []
+                retained_events = retained_events[i:]
+                break
+        else:
+            # Step 3: If all visible events were filtered out, retain the
+            # first user message that the original behavior would have
+            # re-inserted, but only from the already-visible subset.
+            retained_events = []
+            for event in reversed(visible_events):
+                if self._is_user_message(event):
+                    retained_events.insert(0, event)
+                    break
+
+        retained_ids = {id(event) for event in retained_events}
+        for event in visible_events:
+            if id(event) not in retained_ids:
+                event.set_model_visible(False)
+
+    def get_first_visible_event_idx(self) -> int:
+        """Get the first visible event index in the session."""
+        first_visible_idx = 0
+        for idx, event in enumerate(self.events):
+            if event.is_model_visible():
+                first_visible_idx = idx
+                break
+        return first_visible_idx
+
+    def insert_events(self, events: List[Event], idx: Optional[int] = None) -> None:
+        """Insert events at the given index, replacing the existing events."""
+        if idx is None:
+            idx = self.get_first_visible_event_idx()
+        self.events[idx:idx] = events
 
     def _is_user_message(self, event: Event) -> bool:
         """Check if an event is a user message.
diff --git a/trpc_agent_sdk/sessions/_session_summarizer.py b/trpc_agent_sdk/sessions/_session_summarizer.py
index af84b7d..c1e7054 100644
--- a/trpc_agent_sdk/sessions/_session_summarizer.py
+++ b/trpc_agent_sdk/sessions/_session_summarizer.py
@@ -58,7 +58,8 @@
 3. Actions taken or planned
 4. Context that should be remembered for future interactions
 
-Keep the summary concise but comprehensive. Focus on what would be most important to remember for continuing the conversation.
+Keep the summary concise but comprehensive. Focus on what would be most important to remember
+for continuing the conversation.
 
 Conversation:
 {conversation_text}
@@ -246,6 +247,8 @@ def _extract_conversation_text(self, events: List[Event]) -> str:
         current_text = ""
 
         for event in events:
+            if not event.is_model_visible():
+                continue
             if not event.content or not event.content.parts:
                 continue
 
@@ -373,16 +376,17 @@ async def create_session_summary_by_events(self,
             Events after compression
         """
         if keep_recent_count is None:
-            recent_events = []
             old_events = events
         else:
-            recent_events = events[-keep_recent_count:]
             old_events = events[:-keep_recent_count]
         try:
-            original_count = len(events)
+            original_count = sum(1 for event in events if event.is_model_visible())
+            old_visible_events = [event for event in old_events if event.is_model_visible()]
+            if not old_visible_events:
+                return None, events
 
             # Generate summary of old events
-            summary_text = await self._compress_session_to_summary(old_events, session_id, ctx)
+            summary_text = await self._compress_session_to_summary(old_visible_events, session_id, ctx)
 
             if summary_text:
                 # Create summary event
@@ -392,11 +396,18 @@ async def create_session_summary_by_events(self,
                                           parts=[Part.from_text(text=f"Previous conversation summary: {summary_text}")],
                                           role="system"),
                                       timestamp=time.time())
+                summary_event.set_summary_event(True)
+                summary_event.set_model_visible(True)
+
+                # Hide old visible events from model history without dropping raw data.
+                for event in old_visible_events:
+                    event.set_model_visible(False)
 
-                # Replace old events with summary and keep recent events
-                events = [summary_event] + recent_events
+                # Insert summary near the old/recent boundary while preserving all events.
+                insert_index = len(old_events)
+                events.insert(insert_index, summary_event)
 
-                compressed_count = len(events)
+                compressed_count = sum(1 for event in events if event.is_model_visible())
                 logger.info("Compressed session %s: %s events -> %s events", session_id, original_count,
                             compressed_count)
 
@@ -416,10 +427,8 @@ async def create_session_summary(self, session: Session, ctx: InvocationContext
             Summary text if successful, None otherwise
             Events after compression
         """
-        summary_text, events = await self.create_session_summary_by_events(session.events, session.id,
-                                                                           self.__keep_recent_count, ctx)
-        if summary_text:
-            session.events = events
+        summary_text, _ = await self.create_session_summary_by_events(session.events, session.id,
+                                                                      self.__keep_recent_count, ctx)
         return summary_text
 
     def get_summary_metadata(self) -> Dict[str, Any]:
diff --git a/trpc_agent_sdk/sessions/_sql_session_service.py b/trpc_agent_sdk/sessions/_sql_session_service.py
index 779604a..b7e715a 100644
--- a/trpc_agent_sdk/sessions/_sql_session_service.py
+++ b/trpc_agent_sdk/sessions/_sql_session_service.py
@@ -162,18 +162,27 @@ class SessionStorageEvent(SessionStorageBase):
     actions: Mapped[MutableDict[str, Any]] = mapped_column(DynamicPickleType)
     long_running_tool_ids_json: Mapped[Optional[str]] = mapped_column(Text, nullable=True)
     branch: Mapped[str] = mapped_column(UTF8MB4String(DEFAULT_MAX_VARCHAR_LENGTH), nullable=True)
+    request_id: Mapped[str] = mapped_column(UTF8MB4String(DEFAULT_MAX_VARCHAR_LENGTH), nullable=True)
+    parent_invocation_id: Mapped[str] = mapped_column(UTF8MB4String(DEFAULT_MAX_VARCHAR_LENGTH), nullable=True)
+    tag: Mapped[str] = mapped_column(UTF8MB4String(DEFAULT_MAX_VARCHAR_LENGTH), nullable=True)
+    filter_key: Mapped[str] = mapped_column(UTF8MB4String(DEFAULT_MAX_VARCHAR_LENGTH), nullable=True)
+    requires_completion: Mapped[bool] = mapped_column(Boolean, nullable=True)
+    version: Mapped[int] = mapped_column(Integer, nullable=False, default=0)
     timestamp: Mapped[PreciseTimestamp] = mapped_column(PreciseTimestamp, default=func.now())
-
-    content: Mapped[dict[str, Any]] = mapped_column(DynamicJSON, nullable=True)
-    grounding_metadata: Mapped[dict[str, Any]] = mapped_column(DynamicJSON, nullable=True)
-    usage_metadata: Mapped[dict[str, Any]] = mapped_column(DynamicJSON, nullable=True)
-    custom_metadata: Mapped[dict[str, Any]] = mapped_column(DynamicJSON, nullable=True)
+    visible: Mapped[bool] = mapped_column(Boolean, nullable=True)
+    object: Mapped[str] = mapped_column(UTF8MB4String(DEFAULT_MAX_VARCHAR_LENGTH), nullable=False, default="")
+    model_flags: Mapped[int] = mapped_column(Integer, nullable=False, default=1)
 
     partial: Mapped[bool] = mapped_column(Boolean, nullable=True)
     turn_complete: Mapped[bool] = mapped_column(Boolean, nullable=True)
     error_code: Mapped[str] = mapped_column(UTF8MB4String(DEFAULT_MAX_VARCHAR_LENGTH), nullable=True)
     error_message: Mapped[str] = mapped_column(UTF8MB4String(1024), nullable=True)
     interrupted: Mapped[bool] = mapped_column(Boolean, nullable=True)
+    response_id: Mapped[str] = mapped_column(UTF8MB4String(DEFAULT_MAX_VARCHAR_LENGTH), nullable=True)
+    content: Mapped[dict[str, Any]] = mapped_column(DynamicJSON, nullable=True)
+    grounding_metadata: Mapped[dict[str, Any]] = mapped_column(DynamicJSON, nullable=True)
+    custom_metadata: Mapped[dict[str, Any]] = mapped_column(DynamicJSON, nullable=True)
+    usage_metadata: Mapped[dict[str, Any]] = mapped_column(DynamicJSON, nullable=True)
 
     storage_session: Mapped[StorageSession] = relationship(
         "StorageSession",
@@ -201,29 +210,39 @@ def long_running_tool_ids(self, value: set[str]):
     def from_event(cls, session: Session, event: Event) -> SessionStorageEvent:
         storage_event = SessionStorageEvent(
             id=event.id,
+            app_name=session.app_name,
+            user_id=session.user_id,
+            session_id=session.id,
             invocation_id=event.invocation_id,
             author=event.author,
-            branch=event.branch,
             actions=event.actions,
-            session_id=session.id,
-            app_name=session.app_name,
-            user_id=session.user_id,
-            timestamp=datetime.fromtimestamp(event.timestamp),
             long_running_tool_ids=event.long_running_tool_ids,
+            branch=event.branch,
+            request_id=event.request_id,
+            parent_invocation_id=event.parent_invocation_id,
+            tag=event.tag,
+            filter_key=event.filter_key,
+            requires_completion=event.requires_completion,
+            version=event.version,
+            timestamp=datetime.fromtimestamp(event.timestamp),
+            visible=event.visible,
+            object=event.object,
+            model_flags=event.model_flags,
             partial=event.partial,
             turn_complete=event.turn_complete,
             error_code=event.error_code,
             error_message=event.error_message,
             interrupted=event.interrupted,
+            response_id=event.response_id,
         )
         if event.content:
             storage_event.content = event.content.model_dump(exclude_none=True, mode="json")
         if event.grounding_metadata:
             storage_event.grounding_metadata = event.grounding_metadata.model_dump(exclude_none=True, mode="json")
-        if event.usage_metadata:
-            storage_event.usage_metadata = event.usage_metadata.model_dump(exclude_none=True, mode="json")
         if event.custom_metadata:
             storage_event.custom_metadata = event.custom_metadata
+        if event.usage_metadata:
+            storage_event.usage_metadata = event.usage_metadata.model_dump(exclude_none=True, mode="json")
         return storage_event
 
     def to_event(self) -> Event:
@@ -231,19 +250,29 @@ def to_event(self) -> Event:
             id=self.id,
             invocation_id=self.invocation_id,
             author=self.author,
-            branch=self.branch,
             actions=self.actions,  # type: ignore
-            timestamp=self.timestamp.timestamp(),
-            content=decode_content(self.content),
             long_running_tool_ids=self.long_running_tool_ids,
+            branch=self.branch,
+            request_id=self.request_id,
+            parent_invocation_id=self.parent_invocation_id,
+            tag=self.tag,
+            filter_key=self.filter_key,
+            requires_completion=self.requires_completion,
+            version=self.version,
+            visible=self.visible,
+            object=self.object,
+            model_flags=self.model_flags,
+            timestamp=self.timestamp.timestamp(),
             partial=self.partial,
             turn_complete=self.turn_complete,
             error_code=self.error_code,
             error_message=self.error_message,
             interrupted=self.interrupted,
+            response_id=self.response_id,
+            content=decode_content(self.content),
             grounding_metadata=decode_grounding_metadata(self.grounding_metadata),
-            usage_metadata=decode_usage_metadata(self.usage_metadata),
             custom_metadata=self.custom_metadata,
+            usage_metadata=decode_usage_metadata(self.usage_metadata),
         )
 
 
diff --git a/trpc_agent_sdk/sessions/_summarizer_manager.py b/trpc_agent_sdk/sessions/_summarizer_manager.py
index 0906c2d..e938d25 100644
--- a/trpc_agent_sdk/sessions/_summarizer_manager.py
+++ b/trpc_agent_sdk/sessions/_summarizer_manager.py
@@ -105,8 +105,9 @@ async def create_session_summary(self,
         if is_should_summarize:
             logger.debug("Summarizing session %s", session.id)
 
-            # Compress the session
-            original_event_count = len(session.events)
+            # Compress the session. Invisible events are treated as already
+            # compressed/deleted for summary metrics; raw events remain stored.
+            original_event_count = self._count_visible_events(session)
             summary_text = await self._summarizer.create_session_summary(session, ctx)
             if summary_text:
                 app_name = session.app_name
@@ -119,13 +120,17 @@ async def create_session_summary(self,
                     session_id=session.id,
                     summary_text=summary_text,
                     original_event_count=original_event_count,
-                    compressed_event_count=len(session.events),
+                    compressed_event_count=self._count_visible_events(session),
                     summary_timestamp=time.time(),
                 )
             # Update the stored session
             if self._base_service:
                 await self._base_service.update_session(session)
 
+    @staticmethod
+    def _count_visible_events(session: Session) -> int:
+        return sum(1 for event in session.events if event.is_model_visible())
+
     async def get_session_summary(self, session: Session) -> Optional[SessionSummary]:
         """Get a summary of a session.
 
diff --git a/trpc_agent_sdk/skills/__init__.py b/trpc_agent_sdk/skills/__init__.py
index d457bd6..2f429d4 100644
--- a/trpc_agent_sdk/skills/__init__.py
+++ b/trpc_agent_sdk/skills/__init__.py
@@ -59,6 +59,7 @@
 from ._repository import FsSkillRepository
 from ._repository import VisibilityFilter
 from ._repository import create_default_skill_repository
+from ._repository import SkillRepositoryResolver
 from ._skill_config import get_skill_config
 from ._skill_config import get_skill_load_mode
 from ._skill_config import set_skill_config
@@ -95,6 +96,7 @@
 from ._utils import set_state_delta
 from .tools import SkillLoadTool
 from .tools import SkillRunTool
+from .tools import SkillExecTool
 from .tools import skill_list
 from .tools import skill_list_docs
 from .tools import skill_list_tools
@@ -136,6 +138,7 @@
     "FsSkillRepository",
     "VisibilityFilter",
     "create_default_skill_repository",
+    "SkillRepositoryResolver",
     "get_skill_config",
     "get_skill_load_mode",
     "set_skill_config",
@@ -172,6 +175,7 @@
     "set_state_delta",
     "SkillLoadTool",
     "SkillRunTool",
+    "SkillExecTool",
     "skill_list",
     "skill_list_docs",
     "skill_list_tools",
diff --git a/trpc_agent_sdk/skills/_repository.py b/trpc_agent_sdk/skills/_repository.py
index 08189bf..3658287 100644
--- a/trpc_agent_sdk/skills/_repository.py
+++ b/trpc_agent_sdk/skills/_repository.py
@@ -21,9 +21,11 @@
 from typing import Callable
 from typing import List
 from typing import Optional
+from typing import TypeAlias
 from typing_extensions import override
 
 import yaml
+from trpc_agent_sdk.context import InvocationContext
 from trpc_agent_sdk.code_executors import BaseWorkspaceRuntime
 from trpc_agent_sdk.code_executors import create_local_workspace_runtime
 from trpc_agent_sdk.log import logger
@@ -535,3 +537,7 @@ def create_default_skill_repository(
         workspace_runtime=workspace_runtime,
         enable_hot_reload=enable_hot_reload,
     )
+
+
+SkillRepositoryResolver: TypeAlias = Callable[[InvocationContext], BaseSkillRepository]
+"""Callback to resolve a skill repository."""
diff --git a/trpc_agent_sdk/skills/_toolset.py b/trpc_agent_sdk/skills/_toolset.py
index f276322..cdaab4d 100644
--- a/trpc_agent_sdk/skills/_toolset.py
+++ b/trpc_agent_sdk/skills/_toolset.py
@@ -20,6 +20,7 @@
 from trpc_agent_sdk.abc import ToolSetABC
 from trpc_agent_sdk.abc import ToolPredicate
 from trpc_agent_sdk.abc import ToolABC
+from trpc_agent_sdk.code_executors import WorkspaceRuntimeResolver
 from trpc_agent_sdk.context import InvocationContext
 from trpc_agent_sdk.context import get_invocation_ctx
 from trpc_agent_sdk.tools import FunctionTool
@@ -30,6 +31,7 @@
 from ._repository import FsSkillRepository
 from ._repository import BaseSkillRepository
 from ._registry import SKILL_REGISTRY
+from ._repository import SkillRepositoryResolver
 from ._registry import SkillToolFunction
 from ._skill_config import DEFAULT_SKILL_CONFIG
 from ._skill_config import set_skill_config
@@ -69,6 +71,8 @@ class SkillToolSet(ToolSetABC):
     def __init__(self,
                  paths: Optional[List[str]] = None,
                  repository: BaseSkillRepository = None,
+                 repo_resolver: Optional[SkillRepositoryResolver] = None,
+                 workspace_runtime_resolver: Optional[WorkspaceRuntimeResolver] = None,
                  enable_hot_reload: bool = False,
                  tool_filter: Optional[Union[ToolPredicate, List[str]]] = None,
                  is_include_all_tools: bool = True,
@@ -91,7 +95,8 @@ def __init__(self,
         """
         super().__init__(tool_filter=tool_filter, is_include_all_tools=is_include_all_tools)
         self.name = "skill_toolset"
-
+        self._repo_resolver: Optional[SkillRepositoryResolver] = repo_resolver
+        self._workspace_runtime_resolver: Optional[WorkspaceRuntimeResolver] = workspace_runtime_resolver
         self._repository = repository or FsSkillRepository(
             *(paths or []),
             enable_hot_reload=enable_hot_reload,
@@ -100,9 +105,11 @@ def __init__(self,
         self._create_ws_name_cb = create_ws_name_cb or default_create_ws_name_callback
         self._skill_stager = skill_stager or CopySkillStager()
         self._load_tool = SkillLoadTool(repository=self._repository,
+                                        repo_resolver=repo_resolver,
                                         skill_stager=self._skill_stager,
                                         create_ws_name_cb=self._create_ws_name_cb)
         self._run_tool = SkillRunTool(repository=self._repository,
+                                      repo_resolver=repo_resolver,
                                       create_ws_name_cb=self._create_ws_name_cb,
                                       skill_stager=self._skill_stager,
                                       **run_tool_kwargs)
@@ -118,9 +125,11 @@ def __init__(self,
             self._runtime_tools = runtime_tools
         else:
             workspace_exec_tool = WorkspaceExecTool(workspace_runtime=self._repository.workspace_runtime,
+                                                    workspace_runtime_resolver=self._workspace_runtime_resolver,
                                                     create_ws_name_cb=self._create_ws_name_cb)
             self._runtime_tools: List[ToolABC] = [
                 SaveArtifactTool(workspace_runtime=self._repository.workspace_runtime,
+                                 workspace_runtime_resolver=self._workspace_runtime_resolver,
                                  create_ws_name_cb=self._create_ws_name_cb),
                 workspace_exec_tool,
                 WorkspaceWriteStdinTool(workspace_exec_tool),
@@ -145,12 +154,16 @@ async def get_tools(self, invocation_context: Optional[InvocationContext] = None
         tools: List[ToolABC] = []
         skill_functions: List[SkillToolFunction] = SKILL_REGISTRY.get_all()
         skill_functions.extend(self._function_tools)
+        if self._repo_resolver is not None:
+            repository = self._repo_resolver(invocation_context)
+        else:
+            repository = self._repository
         if not invocation_context:
             invocation_context = get_invocation_ctx()
         if invocation_context:
             agent_context = invocation_context.agent_context
             agent_context.with_metadata(SKILL_REGISTRY_KEY, SKILL_REGISTRY)
-            agent_context.with_metadata(SKILL_REPOSITORY_KEY, self._repository)
+            agent_context.with_metadata(SKILL_REPOSITORY_KEY, repository)
             if not is_exist_skill_config(agent_context):
                 set_skill_config(agent_context, self._skill_config)
         tools.append(self._load_tool)
diff --git a/trpc_agent_sdk/skills/tools/_save_artifact.py b/trpc_agent_sdk/skills/tools/_save_artifact.py
index 17cb70c..3511a47 100644
--- a/trpc_agent_sdk/skills/tools/_save_artifact.py
+++ b/trpc_agent_sdk/skills/tools/_save_artifact.py
@@ -19,6 +19,7 @@
 from trpc_agent_sdk.code_executors import DIR_RUNS
 from trpc_agent_sdk.code_executors import DIR_WORK
 from trpc_agent_sdk.code_executors import WORKSPACE_ENV_DIR_KEY
+from trpc_agent_sdk.code_executors import WorkspaceRuntimeResolver
 from trpc_agent_sdk.code_executors.utils import normalize_globs
 from trpc_agent_sdk.context import InvocationContext
 from trpc_agent_sdk.filter import BaseFilter
@@ -164,6 +165,7 @@ def __init__(
         self,
         max_file_bytes: int = _DEFAULT_MAX_BYTES,
         workspace_runtime: Optional[BaseWorkspaceRuntime] = None,
+        workspace_runtime_resolver: Optional[WorkspaceRuntimeResolver] = None,
         create_ws_name_cb: Optional[CreateWorkspaceNameCallback] = None,
         filters_name: Optional[list[str]] = None,
         filters: Optional[list[BaseFilter]] = None,
@@ -175,13 +177,19 @@ def __init__(
             filters_name=filters_name,
             filters=filters,
         )
+        self._workspace_runtime_resolver = workspace_runtime_resolver
         self._max_file_bytes = max_file_bytes
         self._workspace_runtime = workspace_runtime
         self._create_ws_name_cb = create_ws_name_cb or default_create_ws_name_callback
 
+    def _get_workspace_runtime(self, ctx: InvocationContext) -> BaseWorkspaceRuntime:
+        if self._workspace_runtime_resolver is not None:
+            return self._workspace_runtime_resolver(ctx)
+        return self._workspace_runtime
+
     async def _resolve_workspace_root(self, ctx: InvocationContext) -> str:
         """Resolve workspace root, preferring the shared workspace_exec workspace."""
-        runtime = self._workspace_runtime
+        runtime = self._get_workspace_runtime(ctx)
         if runtime is not None:
             workspace_id = self._create_ws_name_cb(ctx)
             ws = await runtime.manager(ctx).create_workspace(workspace_id, ctx)
diff --git a/trpc_agent_sdk/skills/tools/_skill_load.py b/trpc_agent_sdk/skills/tools/_skill_load.py
index 3579d31..869f90d 100644
--- a/trpc_agent_sdk/skills/tools/_skill_load.py
+++ b/trpc_agent_sdk/skills/tools/_skill_load.py
@@ -36,6 +36,7 @@
 from ._common import default_create_ws_name_callback
 from ._common import set_staged_workspace_dir
 from ._copy_stager import CopySkillStager
+from .._repository import SkillRepositoryResolver
 
 
 class SkillLoadTool(BaseTool):
@@ -44,6 +45,7 @@ class SkillLoadTool(BaseTool):
     def __init__(
         self,
         repository: BaseSkillRepository,
+        repo_resolver: Optional[SkillRepositoryResolver] = None,
         skill_stager: Optional[Stager] = None,
         create_ws_name_cb: Optional[CreateWorkspaceNameCallback] = None,
         filters: Optional[List[BaseFilter]] = None,
@@ -53,6 +55,7 @@ def __init__(
         self._skill_stager: Stager = skill_stager or CopySkillStager()
         self._create_ws_name_cb: Optional[
             CreateWorkspaceNameCallback] = create_ws_name_cb or default_create_ws_name_callback
+        self._repo_resolver: Optional[SkillRepositoryResolver] = repo_resolver
 
     @override
     def _get_declaration(self) -> Optional[FunctionDeclaration]:
@@ -80,6 +83,11 @@ def _get_declaration(self) -> Optional[FunctionDeclaration]:
                             description="Result of skill_load. message is a string indicating the skill was loaded."),
         )
 
+    def _get_repository(self, ctx: InvocationContext) -> Optional[BaseSkillRepository]:
+        if self._repo_resolver is not None:
+            return self._repo_resolver(ctx)
+        return self._repository
+
     @override
     async def _run_async_impl(self, *, tool_context: InvocationContext, args: dict[str, Any]) -> str:
         if not (args["skill_name"] or "").strip():
@@ -88,7 +96,8 @@ async def _run_async_impl(self, *, tool_context: InvocationContext, args: dict[s
         docs = args.get("docs", [])
         include_all_docs = args.get("include_all_docs", False)
         normalized_skill = skill_name.strip()
-        skill = self._repository.get(normalized_skill)
+        repository = self._get_repository(tool_context)
+        skill = repository.get(normalized_skill)
         await self._ensure_staged(ctx=tool_context, skill_name=skill_name)
         clean_docs = [doc.strip() for doc in (docs or []) if isinstance(doc, str) and doc.strip()]
         self.__set_state_delta_for_skill_load(tool_context, skill_name, clean_docs, include_all_docs)
@@ -97,12 +106,13 @@ async def _run_async_impl(self, *, tool_context: InvocationContext, args: dict[s
         return f"skill {skill_name!r} loaded"
 
     async def _ensure_staged(self, *, ctx: InvocationContext, skill_name: str) -> None:
-        runtime = self._repository.workspace_runtime
+        repository = self._get_repository(ctx)
+        runtime = repository.workspace_runtime
         manager = runtime.manager(ctx)
         ws_id = self._create_ws_name_cb(ctx)
         ws = await manager.create_workspace(ws_id, ctx)
         result = await self._skill_stager.stage_skill(
-            SkillStageRequest(skill_name=skill_name, repository=self._repository, workspace=ws, ctx=ctx))
+            SkillStageRequest(skill_name=skill_name, repository=repository, workspace=ws, ctx=ctx))
         set_staged_workspace_dir(ctx, skill_name, result.workspace_skill_dir)
 
     def __set_state_delta_for_skill_load(self,
diff --git a/trpc_agent_sdk/skills/tools/_skill_run.py b/trpc_agent_sdk/skills/tools/_skill_run.py
index 01dcdc8..8760f82 100644
--- a/trpc_agent_sdk/skills/tools/_skill_run.py
+++ b/trpc_agent_sdk/skills/tools/_skill_run.py
@@ -41,8 +41,8 @@
 from .._common import get_state_delta_value
 from .._common import loaded_state_key
 from .._constants import SKILL_ARTIFACTS_STATE_KEY
-from .._constants import SKILL_REPOSITORY_KEY
 from .._repository import BaseSkillRepository
+from .._repository import SkillRepositoryResolver
 from .._utils import shell_quote
 from ..stager import SkillStageRequest
 from ..stager import Stager
@@ -378,6 +378,7 @@ class SkillRunTool(BaseTool):
     def __init__(
         self,
         repository: BaseSkillRepository,
+        repo_resolver: Optional[SkillRepositoryResolver] = None,
         filters: Optional[List[BaseFilter]] = None,
         *,
         require_skill_loaded: bool = False,
@@ -392,6 +393,7 @@ def __init__(
 
         Args:
             repository: Skill repository.
+            repo_resolver: Skill repository resolver.
             filters: Optional tool filters.
             require_skill_loaded: When True, skill_run raises unless skill_load was called first
                                   for this skill in the current session.
@@ -416,6 +418,7 @@ def __init__(
             filters=filters,
         )
         self._repository = repository
+        self._repo_resolver: Optional[SkillRepositoryResolver] = repo_resolver
         self._require_skill_loaded = require_skill_loaded
         self._force_save_artifacts = force_save_artifacts
         self._allowed_cmds: frozenset[str] = frozenset(c.strip() for c in (allowed_cmds or []) if c.strip())
@@ -479,10 +482,10 @@ def _get_declaration(self) -> FunctionDeclaration:
     # Repository access
     # ------------------------------------------------------------------
 
-    def _get_repository(self, context: InvocationContext) -> BaseSkillRepository:
-        if self._repository:
-            return self._repository
-        return context.agent_context.get_metadata(SKILL_REPOSITORY_KEY)
+    def _get_repository(self, context: InvocationContext) -> Optional[BaseSkillRepository]:
+        if self._repo_resolver is not None:
+            return self._repo_resolver(context)
+        return self._repository
 
     # ------------------------------------------------------------------
     # Skill-loaded check
@@ -802,7 +805,8 @@ async def _run_program(
         )
         if ret.exit_code != 0:
             raw_stderr = ret.stderr or ""
-            logger.info("Failed to run program: exit_code=%s, stderr=%s", ret.exit_code, raw_stderr.strip())
+            logger.info("Failed to run program: cmd=%s, exit_code=%s, stderr=%s", cmd, ret.exit_code,
+                        raw_stderr.strip())
         return ret
 
     def _resolve_cwd(self, cwd: str, skill_dir: str) -> str:
diff --git a/trpc_agent_sdk/skills/tools/_workspace_exec.py b/trpc_agent_sdk/skills/tools/_workspace_exec.py
index 6597a1e..2d60830 100644
--- a/trpc_agent_sdk/skills/tools/_workspace_exec.py
+++ b/trpc_agent_sdk/skills/tools/_workspace_exec.py
@@ -19,6 +19,7 @@
 from trpc_agent_sdk.code_executors import BaseCodeExecutor
 from trpc_agent_sdk.code_executors import BaseProgramSession
 from trpc_agent_sdk.code_executors import BaseWorkspaceRuntime
+from trpc_agent_sdk.code_executors import WorkspaceRuntimeResolver
 from trpc_agent_sdk.code_executors import DEFAULT_EXEC_YIELD_MS
 from trpc_agent_sdk.code_executors import DEFAULT_SESSION_KILL_SEC
 from trpc_agent_sdk.code_executors import DEFAULT_SESSION_TTL_SEC
@@ -31,6 +32,7 @@
 from trpc_agent_sdk.code_executors import ProgramPoll
 from trpc_agent_sdk.code_executors import WorkspaceRunProgramSpec
 from trpc_agent_sdk.code_executors import poll_line_limit
+from trpc_agent_sdk.code_executors import WorkspaceInfo
 from trpc_agent_sdk.code_executors import wait_for_program_output
 from trpc_agent_sdk.code_executors import yield_duration_ms
 from trpc_agent_sdk.code_executors.utils import normalize_globs
@@ -156,10 +158,6 @@ class _WriteInput(BaseModel):
     append_newline: bool = Field(default=False)
 
 
-class _KillInput(BaseModel):
-    session_id: str = Field(default="")
-
-
 @dataclass
 class _ExecSession:
     proc: BaseProgramSession
@@ -174,6 +172,7 @@ class WorkspaceExecTool(BaseTool):
     def __init__(
         self,
         workspace_runtime: BaseWorkspaceRuntime,
+        workspace_runtime_resolver: Optional[WorkspaceRuntimeResolver] = None,
         create_ws_name_cb: Optional[CreateWorkspaceNameCallback] = None,
         session_ttl: float = DEFAULT_SESSION_TTL_SEC,
         filters_name: Optional[list[str]] = None,
@@ -187,25 +186,25 @@ def __init__(
             filters=filters,
         )
         self._workspace_runtime = workspace_runtime
+        self._workspace_runtime_resolver = workspace_runtime_resolver
         self._create_ws_name_cb = create_ws_name_cb or default_create_ws_name_callback
         self._ttl = session_ttl
         self._sessions: dict[str, _ExecSession] = {}
 
-    def _runtime(self) -> BaseWorkspaceRuntime:
-        runtime = self._workspace_runtime
-        if runtime is None:
-            raise ValueError("workspace_exec requires an executor with live workspace support")
-        return runtime
+    def _runtime(self, ctx: InvocationContext) -> BaseWorkspaceRuntime:
+        if self._workspace_runtime_resolver is not None:
+            return self._workspace_runtime_resolver(ctx)
+        return self._workspace_runtime
 
-    async def _workspace(self, ctx: InvocationContext):
-        runtime = self._runtime()
+    async def _workspace(self, ctx: InvocationContext) -> tuple[BaseWorkspaceRuntime, WorkspaceInfo]:
+        runtime = self._runtime(ctx)
         manager = runtime.manager(ctx)
         workspace_id = self._create_ws_name_cb(ctx)
         ws = await manager.create_workspace(workspace_id, ctx)
         return runtime, ws
 
     def _supports_interactive(self, ctx: InvocationContext) -> bool:
-        runner = self._runtime().runner(ctx)
+        runner = self._runtime(ctx).runner(ctx)
         start_program = getattr(runner, "start_program", None)
         return start_program is not None
 
@@ -305,7 +304,7 @@ async def _run_async_impl(self, *, tool_context: InvocationContext, args: dict[s
         if (not inputs.background) and (not tty) and yield_ms <= 0:
             return await _run_one_shot(runtime, ws, spec, tool_context)
 
-        runner = runtime.runner(tool_context)
+        runner = self._runtime(tool_context).runner(tool_context)
         interactive_spec = WorkspaceRunProgramSpec(
             cmd=spec.cmd,
             args=spec.args,
@@ -506,6 +505,7 @@ def create_workspace_exec_tools(
     code_executor: BaseCodeExecutor,
     *,
     workspace_runtime: Optional[BaseWorkspaceRuntime] = None,
+    workspace_runtime_resolver: Optional[WorkspaceRuntimeResolver] = None,
     session_ttl: float = DEFAULT_SESSION_TTL_SEC,
     filters_name: Optional[list[str]] = None,
     filters: Optional[list[BaseFilter]] = None,
@@ -514,6 +514,7 @@ def create_workspace_exec_tools(
     exec_tool = WorkspaceExecTool(
         code_executor=code_executor,
         workspace_runtime=workspace_runtime,
+        workspace_runtime_resolver=workspace_runtime_resolver,
         session_ttl=session_ttl,
         filters_name=filters_name,
         filters=filters,
diff --git a/trpc_agent_sdk/tools/_preload_memory_tool.py b/trpc_agent_sdk/tools/_preload_memory_tool.py
index eb032dd..3a2a75d 100644
--- a/trpc_agent_sdk/tools/_preload_memory_tool.py
+++ b/trpc_agent_sdk/tools/_preload_memory_tool.py
@@ -21,7 +21,6 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 #
-
 """Preload memory tool for TRPC Agent framework."""
 
 from __future__ import annotations
diff --git a/trpc_agent_sdk/tools/mem0_tool.py b/trpc_agent_sdk/tools/mem0_tool.py
index 4f48bdb..1655504 100644
--- a/trpc_agent_sdk/tools/mem0_tool.py
+++ b/trpc_agent_sdk/tools/mem0_tool.py
@@ -102,12 +102,14 @@ async def _run_async_impl(self, *, tool_context: InvocationContext, args: dict[s
         """Save important information to memory"""
         user_id = tool_context.user_id
         try:
-            result = await self.client.add([{
-                "role": "user",
-                "content": args["content"]
-            }],
-                                           user_id=user_id,
-                                           **self.kwargs)
+            result = await self.client.add(
+                [{
+                    "role": "user",
+                    "content": args["content"]
+                }],
+                user_id=user_id,
+                **self.kwargs,
+            )
             return {"status": "success", "message": "Information saved to memory", "result": result, "user_id": user_id}
         except Exception as e:
             return {"status": "error", "message": f"Failed to save memory: {str(e)}", "user_id": user_id}
diff --git a/trpc_agent_sdk/types/_event_actions.py b/trpc_agent_sdk/types/_event_actions.py
index 2e5235f..1b2f815 100644
--- a/trpc_agent_sdk/types/_event_actions.py
+++ b/trpc_agent_sdk/types/_event_actions.py
@@ -46,15 +46,15 @@ class EventActions(BaseModel):
     skip_summarization: Optional[bool] = None
     """If true, it won't call model to summarize function response.
 
-  Only used for function_response event.
-  """
+    Only used for function_response event.
+    """
 
     state_delta: dict[str, object] = Field(default_factory=dict)
     """Indicates that the event is updating the state with the given delta."""
 
     artifact_delta: dict[str, int] = Field(default_factory=dict)
     """Indicates that the event is updating an artifact. key is the filename,
-  value is the version."""
+    value is the version."""
 
     transfer_to_agent: Optional[str] = None
     """If set, the event transfers to the specified agent."""