From b15884357f821bad65ef5f82748ab7806bdeafa6 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Thu, 28 May 2026 07:47:50 +0000
Subject: [PATCH] fix(compact_20260112): bound _call_summary_model with timeout

A slow or unresponsive summary model previously hung the parent
/v1/messages request with no escape hatch. Pass a 60s timeout on the
litellm.acompletion / llm_router.acompletion subrequest; on timeout the
existing summary_call_failed path forwards the request without
compaction rather than blocking indefinitely.
---
 .../context_management/constants.py           |  5 ++++
 .../context_management/editors/compact.py     |  6 +++++
 .../context_management/test_compact.py        | 27 +++++++++++++++++++
 3 files changed, 38 insertions(+)

diff --git a/litellm/llms/anthropic/experimental_pass_through/context_management/constants.py b/litellm/llms/anthropic/experimental_pass_through/context_management/constants.py
index eb099407f00..ebbc182c427 100644
--- a/litellm/llms/anthropic/experimental_pass_through/context_management/constants.py
+++ b/litellm/llms/anthropic/experimental_pass_through/context_management/constants.py
@@ -18,6 +18,11 @@ COMPACT_MIN_TRIGGER_TOKENS = 50_000
 # ``general_settings.context_management_summary_max_tokens``.
 COMPACT_SUMMARY_MAX_TOKENS = 4096
 COMPACT_SUMMARY_MAX_TOKENS_SETTING_KEY = "context_management_summary_max_tokens"
+# Wall-clock bound for the summary sub-call. Without this a slow or
+# unresponsive summary model would hang the parent ``/v1/messages`` request
+# with no escape hatch; on timeout the editor falls into the standard
+# ``summary_call_failed`` path and forwards the request without compaction.
+COMPACT_SUMMARY_TIMEOUT_SECONDS = 60.0
 COMPACT_SUMMARY_MODEL_SETTING_KEY = "context_management_summary_model"
 COMPACT_SUMMARY_SYSTEM_PREFIX = "Previous conversation summary: "
 
diff --git a/litellm/llms/anthropic/experimental_pass_through/context_management/editors/compact.py b/litellm/llms/anthropic/experimental_pass_through/context_management/editors/compact.py
index 1fe8072d3d2..a85427a9180 100644
--- a/litellm/llms/anthropic/experimental_pass_through/context_management/editors/compact.py
+++ b/litellm/llms/anthropic/experimental_pass_through/context_management/editors/compact.py
@@ -33,6 +33,7 @@ from ..constants import (
     COMPACT_SUMMARY_MAX_TOKENS_SETTING_KEY,
     COMPACT_SUMMARY_MODEL_SETTING_KEY,
     COMPACT_SUMMARY_SYSTEM_PREFIX,
+    COMPACT_SUMMARY_TIMEOUT_SECONDS,
 )
 from ..errors import AnthropicContextManagementError
 from ..result import PolyfillResult
@@ -569,10 +570,15 @@ async def _call_summary_model(
     # (see ``Router._common_checks_available_deployment``); without this the
     # summary subrequest could be routed to a deployment outside the caller's
     # permitted region.
+    # ``timeout`` bounds how long a slow/unresponsive summary model can stall
+    # the parent ``/v1/messages`` request. On timeout the caller catches the
+    # exception and surfaces ``applied_edits[0].error = "summary_call_failed"``,
+    # forwarding the request without compaction rather than hanging.
     call_kwargs: Dict[str, Any] = {
         "model": summary_model,
         "messages": summary_messages,
         "max_tokens": max_tokens,
+        "timeout": COMPACT_SUMMARY_TIMEOUT_SECONDS,
         "litellm_metadata": metadata,
     }
     if allowed_model_region is not None:
diff --git a/tests/test_litellm/llms/anthropic/experimental_pass_through/context_management/test_compact.py b/tests/test_litellm/llms/anthropic/experimental_pass_through/context_management/test_compact.py
index 02b229ad0a1..580e3d37af8 100644
--- a/tests/test_litellm/llms/anthropic/experimental_pass_through/context_management/test_compact.py
+++ b/tests/test_litellm/llms/anthropic/experimental_pass_through/context_management/test_compact.py
@@ -1105,6 +1105,33 @@ def test_summary_max_tokens_setting_falls_back_for_invalid_values():
             ), f"expected default for invalid override {bad!r}"
 
 
+async def test_summary_call_sends_default_timeout():
+    """``timeout`` is set on the summary call so a slow or unresponsive summary
+    model cannot hang the parent ``/v1/messages`` request indefinitely."""
+    from litellm.llms.anthropic.experimental_pass_through.context_management.constants import (
+        COMPACT_SUMMARY_TIMEOUT_SECONDS,
+    )
+    from litellm.llms.anthropic.experimental_pass_through.context_management.editors.compact import (
+        _call_summary_model,
+    )
+
+    captured_kwargs: dict = {}
+
+    class _FakeRouter:
+        async def acompletion(self, **kwargs):
+            captured_kwargs.update(kwargs)
+            return _make_mock_response("<summary>x</summary>")
+
+    await _call_summary_model(
+        summary_model="claude-haiku-4-5",
+        summary_messages=[{"role": "user", "content": "hi"}],
+        metadata={},
+        llm_router=_FakeRouter(),
+    )
+
+    assert captured_kwargs.get("timeout") == COMPACT_SUMMARY_TIMEOUT_SECONDS
+
+
 # ---------------------------------------------------------------------------
 # Editor: summary model key/team access gate
 # ---------------------------------------------------------------------------