fix(agents): derive config agent ids from the agent name

The id was a sha256 of the whole config entry with os.environ placeholders already
resolved, so rotating a token in static_headers minted a new agent id and silently
invalidated every permission granted on the old one; the secured keys lost access
while unrestricted keys kept theirs. Agent names are already unique in the registry,
so hash the name and leave the id alone when the entry's values change.

Config agent ids change once on upgrade, so existing grants on config-defined agents
need to be re-pointed at the new ids.

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

litellm/proxy/agent_endpoints/agent_registry.py

@@ -1,5 +1,4 @@
 import hashlib
-import json
 from collections.abc import Iterator, Mapping, Sequence
 from datetime import datetime, timezone
 from typing import Any, Protocol, TypedDict
@@ -114,8 +113,15 @@ class AgentRegistry:
                 public_agent_list.append(agent)
         return public_agent_list
 
-    def _create_agent_id(self, agent_config: AgentConfig) -> str:
-        return hashlib.sha256(json.dumps(agent_config, sort_keys=True).encode()).hexdigest()
+    def _create_agent_id(self, agent_name: str) -> str:
+        """
+        Derive a config agent's id from its name, which is unique across the registry.
+
+        Hashing the whole config entry instead would fold every mutable value into the
+        identity, so rotating a secret in ``static_headers`` would mint a new agent id
+        and silently invalidate the permissions granted on the old one.
+        """
+        return hashlib.sha256(agent_name.encode()).hexdigest()
 
     def load_agents_from_config(self, agent_config: Sequence[AgentConfig] | None = None):
         """
@@ -139,14 +145,14 @@ class AgentRegistry:
 
             agent_name = agent_config_item.get("agent_name")
             agent_card_params = agent_config_item.get("agent_card_params")
-            if not all([agent_name, agent_card_params]):
+            if not agent_name or not agent_card_params:
                 continue
 
             if any(agent.agent_name == agent_name for agent in self.agent_list):
                 continue
 
             # create a stable hash id for config item
-            config_hash = self._create_agent_id(agent_config_item)
+            config_hash = self._create_agent_id(agent_name)
 
             self.register_agent(agent_config=AgentResponse(agent_id=config_hash, **agent_config_item))  # type: ignore
 

tests/test_litellm/proxy/agent_endpoints/test_agent_registry.py

@@ -279,3 +279,54 @@ def test_load_agents_from_config_with_an_empty_list_clears_the_remembered_agents
 
     registry.load_agents_from_db_and_config(db_agents=None)
     assert registry.get_agent_list() == [], "a removed config agent must not come back on the next rebuild"
+
+
+def test_config_agent_id_survives_a_static_header_secret_rotation():
+    """
+    A config agent's id must not move when a value inside the entry changes.
+
+    The id used to be a sha256 of the whole config entry with `os.environ/...`
+    placeholders already resolved, so rotating an upstream token minted a brand new
+    agent id: every permission granted on the old id stopped matching and the keys
+    that had been secured lost access while unrestricted keys kept theirs.
+    """
+    registry_before = AgentRegistry()
+    registry_before.load_agents_from_config(
+        [
+            {
+                "agent_name": "rotating-agent",
+                "agent_card_params": _sample_agent_card_params(),
+                "static_headers": {"x-upstream-token": "token-v1"},
+            }
+        ]
+    )
+
+    registry_after = AgentRegistry()
+    registry_after.load_agents_from_config(
+        [
+            {
+                "agent_name": "rotating-agent",
+                "agent_card_params": _sample_agent_card_params(),
+                "static_headers": {"x-upstream-token": "token-v2"},
+            }
+        ]
+    )
+
+    agent_before = registry_before.get_agent_by_name("rotating-agent")
+    agent_after = registry_after.get_agent_by_name("rotating-agent")
+    assert agent_before is not None and agent_after is not None
+    assert agent_before.agent_id == agent_after.agent_id
+
+
+def test_config_agents_with_different_names_get_different_ids():
+    """Distinct config agents must stay distinguishable for permission grants."""
+    registry = AgentRegistry()
+    registry.load_agents_from_config(
+        [
+            {"agent_name": "agent-alpha", "agent_card_params": _sample_agent_card_params()},
+            {"agent_name": "agent-beta", "agent_card_params": _sample_agent_card_params()},
+        ]
+    )
+
+    agent_ids = {agent.agent_id for agent in registry.get_agent_list()}
+    assert len(agent_ids) == 2