Skip to content

"Vibe-Trading: Your Personal Trading Agent"

fix(swarm): close per-task ChatLLM to stop pooled connection leak

cgycorey129cb25e86ac
diff --git a/agent/src/providers/chat.py b/agent/src/providers/chat.py
index dd6356810..377695374 100644
--- a/agent/src/providers/chat.py
+++ b/agent/src/providers/chat.py
@@ -296,6 +296,27 @@ class ChatLLM:
reasoning_effort=runtime_cfg.langchain_reasoning_effort.strip().lower(),
)
+ def close(self) -> None:
+ """Best-effort release of the underlying provider HTTP client.
+
+ The LangChain adapter (ChatOpenAI and its OpenAI-compatible
+ subclasses) owns a pooled ``httpx.Client`` that is not guaranteed to
+ be refcount-collected promptly — cyclic references can defer it to a
+ GC pass. Long-running callers (swarm workers build one ChatLLM per
+ task) must call this when the instance is done with, or sockets
+ accumulate in CLOSE-WAIT. Safe to call multiple times; providers
+ without a closeable client are a no-op.
+ """
+ llm = self._llm
+ for attr in ("root_client", "root_async_client", "client"):
+ client = getattr(llm, attr, None)
+ close_fn = getattr(client, "close", None)
+ if callable(close_fn):
+ try:
+ close_fn()
+ except Exception:
+ logger.debug("ChatLLM.close: failed to close %s", attr, exc_info=True)
+
def chat(self, messages: List[Dict[str, Any]], tools: Optional[List[Dict[str, Any]]] = None, timeout: Optional[int] = None) -> LLMResponse:
"""Call the LLM synchronously.
diff --git a/agent/src/swarm/worker.py b/agent/src/swarm/worker.py
index 8b58cf49f..29deb03f1 100644
--- a/agent/src/swarm/worker.py
+++ b/agent/src/swarm/worker.py
@@ -382,17 +382,75 @@ def run_worker(
include_shell_tools: bool = False,
grounding_block: str = "",
agent_config: AgentConfig | None = None,
+) -> WorkerResult:
+ """Run one worker task, releasing the per-task LLM client on exit.
+
+ Builds a fresh :class:`ChatLLM` for the task and guarantees it is closed
+ even on early returns (timeout, token limit, tool error). The underlying
+ LangChain adapter owns a pooled ``httpx.Client`` that is not promptly
+ refcount-collected; without this, long-running swarm deployments
+ accumulate one CLOSE-WAIT socket per task (#1141).
+
+ Args:
+ agent_spec: Agent role specification with tools/skills/model config.
+ task: The task to execute, including prompt template.
+ upstream_summaries: Summaries from upstream tasks keyed by input_from keys.
+ user_vars: User-provided variables for template rendering.
+ run_dir: Path to .swarm/runs/{run_id}/ directory.
+ event_callback: Optional callback for swarm events.
+ include_shell_tools: Whether this worker may register shell tools.
+ grounding_block: Optional pre-rendered "Ground Truth" markdown that
+ anchors the worker on real recent prices for symbols mentioned in
+ ``user_vars``. Forwarded verbatim to :func:`build_worker_prompt`.
+ agent_config: Optional resolved agent config carrying remote MCP
+ server definitions. Threaded from :class:`SwarmRuntime` and
+ consumed by :func:`build_swarm_registry` to merge remote MCP
+ tools with the local-tool pool before applying the agent's
+ whitelist. ``None`` preserves the prior local-only behavior.
+
+ Returns:
+ WorkerResult with status, summary, artifacts, and iteration count.
+ """
+ llm = ChatLLM(model_name=agent_spec.model_name)
+ try:
+ return _run_worker_impl(
+ agent_spec=agent_spec,
+ task=task,
+ upstream_summaries=upstream_summaries,
+ user_vars=user_vars,
+ run_dir=run_dir,
+ llm=llm,
+ event_callback=event_callback,
+ include_shell_tools=include_shell_tools,
+ grounding_block=grounding_block,
+ agent_config=agent_config,
+ )
+ finally:
+ llm.close()
+
+
+def _run_worker_impl(
+ agent_spec: SwarmAgentSpec,
+ task: SwarmTask,
+ upstream_summaries: dict[str, str],
+ user_vars: dict[str, str],
+ run_dir: Path,
+ event_callback: Callable[[SwarmEvent], None] | None = None,
+ include_shell_tools: bool = False,
+ grounding_block: str = "",
+ agent_config: AgentConfig | None = None,
+ *,
+ llm: ChatLLM,
) -> WorkerResult:
"""Execute a single worker task using a lightweight ReAct loop.
Steps:
1. Build filtered ToolRegistry from agent_spec.tools
- 2. Create ChatLLM with agent_spec.model_name
- 3. Build system prompt with role + upstream summaries + filtered skills
- 4. Resolve task.prompt_template with user_vars
- 5. Run ReAct loop (for iteration in range(max_iterations))
- 6. Write summary to artifacts/{agent_id}/summary.md
- 7. Return WorkerResult
+ 2. Build system prompt with role + upstream summaries + filtered skills
+ 3. Resolve task.prompt_template with user_vars
+ 4. Run ReAct loop (for iteration in range(max_iterations))
+ 5. Write summary to artifacts/{agent_id}/summary.md
+ 6. Return WorkerResult
Args:
agent_spec: Agent role specification with tools/skills/model config.
@@ -400,6 +458,8 @@ def run_worker(
upstream_summaries: Summaries from upstream tasks keyed by input_from keys.
user_vars: User-provided variables for template rendering.
run_dir: Path to .swarm/runs/{run_id}/ directory.
+ llm: Pre-built ChatLLM; ownership stays with the caller
+ (:func:`run_worker` closes it in a ``finally``).
event_callback: Optional callback for swarm events.
include_shell_tools: Whether this worker may register shell tools.
grounding_block: Optional pre-rendered "Ground Truth" markdown that
@@ -429,10 +489,7 @@ def run_worker(
include_shell_tools=include_shell_tools,
)
- # 2. Create LLM
- llm = ChatLLM(model_name=agent_spec.model_name)
-
- # 3. Build system prompt with filtered skills
+ # 2. Build system prompt with filtered skills
skills_loader = SkillsLoader()
skill_desc = _filter_skill_descriptions(skills_loader, agent_spec.skills)
system_prompt = build_worker_prompt(
diff --git a/agent/tests/test_swarm_m4_e2e.py b/agent/tests/test_swarm_m4_e2e.py
index b703e9a02..076f4db7d 100644
--- a/agent/tests/test_swarm_m4_e2e.py
+++ b/agent/tests/test_swarm_m4_e2e.py
@@ -143,6 +143,9 @@ class _StubChatLLM:
def chat(self, messages, tools=None, timeout=None): # pragma: no cover — fallback path
return self.stream_chat(messages, tools=tools)
+ def close(self) -> None:
+ """No-op: the stub owns no HTTP client."""
+
def _stub_llm_factory(responses: list[LLMResponse]):
"""Build a callable that reads as ``ChatLLM(model_name=...)`` would.
diff --git a/agent/tests/test_swarm_m5_trust_model.py b/agent/tests/test_swarm_m5_trust_model.py
index 634489dbc..382501002 100644
--- a/agent/tests/test_swarm_m5_trust_model.py
+++ b/agent/tests/test_swarm_m5_trust_model.py
@@ -297,6 +297,9 @@ class _StubChatLLM:
def chat(self, messages, tools=None, timeout=None): # pragma: no cover
return self.stream_chat(messages, tools=tools)
+ def close(self) -> None:
+ """No-op: the stub owns no HTTP client."""
+
def _stub_llm_factory(responses: list[LLMResponse]):
stub = _StubChatLLM(responses)
diff --git a/agent/tests/test_swarm_output_contract.py b/agent/tests/test_swarm_output_contract.py
index bbb56d1d2..0e3a26f71 100644
--- a/agent/tests/test_swarm_output_contract.py
+++ b/agent/tests/test_swarm_output_contract.py
@@ -346,6 +346,9 @@ class _ScriptedLLM:
def stream_chat(self, messages, tools=None, timeout=None, on_text_chunk=None):
return self._responses.pop(0)
+ def close(self) -> None:
+ """No-op: the scripted stub owns no HTTP client."""
+
@pytest.mark.parametrize(
("result", "raises", "expected_event_status", "expected_worker_status"),
@@ -428,3 +431,98 @@ def test_collect_artifacts_rejects_symlink_escape(tmp_path: Path) -> None:
pytest.skip("symlinks are not available on this platform")
assert _collect_artifacts(artifact_dir) == []
+
+
+# ---------------------------------------------------------------------------
+# ChatLLM lifecycle: run_worker must close the per-task LLM (regression #1141)
+# ---------------------------------------------------------------------------
+
+
+def test_run_worker_closes_llm_on_success(monkeypatch, tmp_path: Path) -> None:
+ """A successful run must close the per-task ChatLLM exactly once."""
+ import src.swarm.worker as worker_mod
+
+ registry = ToolRegistry()
+ registry.register(_ResultTool("done", raises=False))
+ monkeypatch.setattr(
+ worker_mod, "build_swarm_registry", lambda *args, **kwargs: registry
+ )
+ closed = []
+
+ class _TrackingLLM(_ScriptedLLM):
+ def close(self) -> None:
+ closed.append(1)
+
+ monkeypatch.setattr(
+ worker_mod, "ChatLLM", lambda *args, **kwargs: _TrackingLLM()
+ )
+
+ worker_result = run_worker(
+ agent_spec=SwarmAgentSpec(
+ id="analyst",
+ role="Analyst",
+ system_prompt="Analyse the result.",
+ tools=["market_probe"],
+ max_iterations=3,
+ ),
+ task=SwarmTask(
+ id="task", agent_id="analyst", prompt_template="Probe the market."
+ ),
+ upstream_summaries={},
+ user_vars={},
+ run_dir=tmp_path,
+ )
+ assert worker_result.status.value == "completed"
+ assert len(closed) == 1, f"close must be called exactly once, got {len(closed)}"
+
+
+def test_run_worker_closes_llm_on_tool_exception(monkeypatch, tmp_path: Path) -> None:
+ """close() must fire even when the run dies on a tool exception."""
+ import src.swarm.worker as worker_mod
+
+ registry = ToolRegistry()
+ registry.register(_ResultTool("boom", raises=True))
+ monkeypatch.setattr(
+ worker_mod, "build_swarm_registry", lambda *args, **kwargs: registry
+ )
+ closed = []
+
+ class _TrackingLLM(_ScriptedLLM):
+ def close(self) -> None:
+ closed.append(1)
+
+ monkeypatch.setattr(
+ worker_mod, "ChatLLM", lambda *args, **kwargs: _TrackingLLM()
+ )
+
+ worker_result = run_worker(
+ agent_spec=SwarmAgentSpec(
+ id="analyst",
+ role="Analyst",
+ system_prompt="Analyse the result.",
+ tools=["market_probe"],
+ max_iterations=3,
+ ),
+ task=SwarmTask(
+ id="task", agent_id="analyst", prompt_template="Probe the market."
+ ),
+ upstream_summaries={},
+ user_vars={},
+ run_dir=tmp_path,
+ )
+ assert worker_result.status.value == "incomplete"
+ assert len(closed) == 1, f"close must fire on failure path, got {len(closed)}"
+
+
+def test_chatllm_close_is_best_effort_noop_without_client() -> None:
+ """ChatLLM.close() must not raise when the provider adapter has no
+ closeable client (e.g. scripted/native adapters without root_client)."""
+ from src.providers.chat import ChatLLM
+
+ class _BareLLM:
+ def bind_tools(self, tools):
+ return self
+
+ llm = ChatLLM.__new__(ChatLLM)
+ llm._llm = _BareLLM() # no root_client / client attributes
+ llm.close() # must not raise
diff --git a/agent/tests/test_swarm_worker_content_filter.py b/agent/tests/test_swarm_worker_content_filter.py
index ff67eefbd..fbbf93d71 100644
--- a/agent/tests/test_swarm_worker_content_filter.py
+++ b/agent/tests/test_swarm_worker_content_filter.py
@@ -48,6 +48,9 @@ class _ScriptedChatLLM:
def __call__(self, *args, **kwargs) -> "_ScriptedChatLLM":
return self
+ def close(self) -> None:
+ """No-op: the scripted stub owns no HTTP client."""
+
def stream_chat(
self, messages, tools=None, on_text_chunk=None, timeout=None
) -> LLMResponse:
diff --git a/agent/tests/test_swarm_worker_stream_retry.py b/agent/tests/test_swarm_worker_stream_retry.py
index 2be91fdc0..96250899a 100644
--- a/agent/tests/test_swarm_worker_stream_retry.py
+++ b/agent/tests/test_swarm_worker_stream_retry.py
@@ -63,6 +63,10 @@ class _FlakyChatLLM:
"""
return self
+ def close(self) -> None:
+ """No-op: the stub owns no HTTP client."""
+ return self
+
def stream_chat(self, messages, tools=None, on_text_chunk=None, timeout=None) -> LLMResponse:
"""Raise the next queued error or return the final response.
diff --git a/agent/tests/test_swarm_worker_tool_result_truncation.py b/agent/tests/test_swarm_worker_tool_result_truncation.py
index 6d190883e..f883d95a9 100644
--- a/agent/tests/test_swarm_worker_tool_result_truncation.py
+++ b/agent/tests/test_swarm_worker_tool_result_truncation.py
@@ -60,6 +60,9 @@ class _ScriptedChatLLM:
def __call__(self, *args, **kwargs) -> "_ScriptedChatLLM":
return self
+ def close(self) -> None:
+ """No-op: the scripted stub owns no HTTP client."""
+
def stream_chat(
self, messages, tools=None, on_text_chunk=None, timeout=None
) -> LLMResponse: