From deffec5b89922f25cf1eb34a2475b1664384f07f Mon Sep 17 00:00:00 2001
From: voidborne-d <voidborne-d@users.noreply.github.com>
Date: Tue, 31 Mar 2026 22:13:26 +0000
Subject: [PATCH] fix: colon-safe endpoint ID extraction + dedup stale cleanup

Address review feedback from greptile-apps[bot]:
- Use regex to anchor on :exact: / :subpath: delimiters instead of
  naive split(':', 1)[0], so endpoint IDs containing colons (e.g.
  'svc:v2') are extracted correctly.
- Deduplicate remove_endpoint_routes() calls when an endpoint has
  both exact and subpath routes (prevents redundant dict scans).
- Add test for colon-safe ID extraction.
---
 .../pass_through_endpoints.py                 | 20 +++++++++--
 .../test_passthrough_registry_leak.py         | 36 +++++++++++++++++++
 2 files changed, 54 insertions(+), 2 deletions(-)

diff --git a/litellm/proxy/pass_through_endpoints/pass_through_endpoints.py b/litellm/proxy/pass_through_endpoints/pass_through_endpoints.py
index ed9a3d820fd..4de1e7e6de0 100644
--- a/litellm/proxy/pass_through_endpoints/pass_through_endpoints.py
+++ b/litellm/proxy/pass_through_endpoints/pass_through_endpoints.py
@@ -2301,6 +2301,9 @@ async def initialize_pass_through_endpoints(
         )
 
     # remove the ones that are not visited from the list
+    import re
+
+    removed_endpoint_ids: set = set()
     for endpoint_key in registered_pass_through_endpoints:
         if endpoint_key not in visited_endpoints:
             # Route keys are formatted as "{endpoint_id}:exact:{path}:{methods}"
@@ -2309,8 +2312,21 @@ async def initialize_pass_through_endpoints(
             # we must split it out here.  Previously the full key was passed
             # verbatim, which never matched any stored endpoint_id and silently
             # left stale routes in the registry forever.
-            stale_endpoint_id = endpoint_key.split(":", 1)[0]
-            InitPassThroughEndpointHelpers.remove_endpoint_routes(stale_endpoint_id)
+            #
+            # Use regex to anchor on `:exact:` or `:subpath:` delimiters so
+            # endpoint IDs that contain colons (e.g. "svc:v2") are handled
+            # correctly instead of being truncated at the first colon.
+            _match = re.match(r"^(.+?):(?:exact|subpath):", endpoint_key)
+            stale_endpoint_id = (
+                _match.group(1) if _match else endpoint_key.split(":", 1)[0]
+            )
+            # Deduplicate: an endpoint with include_subpath=True produces two
+            # registry keys (exact + subpath).  Only call remove once.
+            if stale_endpoint_id not in removed_endpoint_ids:
+                InitPassThroughEndpointHelpers.remove_endpoint_routes(
+                    stale_endpoint_id
+                )
+                removed_endpoint_ids.add(stale_endpoint_id)
 
 
 def _get_pass_through_endpoints_from_config() -> List[PassThroughGenericEndpoint]:
diff --git a/tests/pass_through_unit_tests/test_passthrough_registry_leak.py b/tests/pass_through_unit_tests/test_passthrough_registry_leak.py
index a299a56e6e1..0e89d09c02e 100644
--- a/tests/pass_through_unit_tests/test_passthrough_registry_leak.py
+++ b/tests/pass_through_unit_tests/test_passthrough_registry_leak.py
@@ -253,3 +253,39 @@ def test_explicit_id_is_preserved():
             del _registered_pass_through_routes[k]
 
     asyncio.run(_run())
+
+
+# ---------------------------------------------------------------------------
+# Test 6 – stale endpoint cleanup: colon-safe ID extraction
+# ---------------------------------------------------------------------------
+
+def test_stale_endpoint_cleanup_colon_safe_split():
+    """
+    Verify that the cleanup loop correctly extracts the endpoint_id from
+    a registry key even when the ID contains a colon (e.g. 'svc:v2'),
+    and that deduplication prevents redundant remove calls.
+    """
+    import re
+
+    # Simulate registry keys with a colon-containing ID
+    colon_id = "svc:v2"
+    route_key_exact = f"{colon_id}:exact:/some/path:GET,POST"
+    route_key_subpath = f"{colon_id}:subpath:/some/path:GET,POST"
+
+    # Verify regex-based extraction works correctly for both key types
+    for key in [route_key_exact, route_key_subpath]:
+        _match = re.match(r"^(.+?):(?:exact|subpath):", key)
+        extracted = _match.group(1) if _match else key.split(":", 1)[0]
+        assert extracted == colon_id, (
+            f"Expected {colon_id!r} but got {extracted!r} from key {key!r}. "
+            "The colon-safe split is broken."
+        )
+
+    # Also verify standard auto-generated IDs still work
+    auto_id = "auto-abc123def456"
+    auto_key = f"{auto_id}:exact:/test:GET"
+    _match = re.match(r"^(.+?):(?:exact|subpath):", auto_key)
+    extracted = _match.group(1) if _match else auto_key.split(":", 1)[0]
+    assert extracted == auto_id, (
+        f"Expected {auto_id!r} but got {extracted!r} for auto-generated ID"
+    )