Files
litellm/litellm/proxy/_lazy_features.py
T
Sameer Kankute e59e34bed3 Gemini managed agents support (#28270)
* Add support for environment variable in interactions api

* Add sdk  support for gemini create agent

* Add agents endpoint support via proxy

* Add outputs of each api

* Add routing for model and agents param

* Remove redundant condition in get_provider_agents_api_config

LlmProviders.GEMINI.value is literally the string "gemini", so the
second clause of the or was checking the exact same thing as the first.

Co-authored-by: Sameer Kankute <Sameerlite@users.noreply.github.com>

* fix: forward query-param credentials to list/get/delete/versions Gemini agent endpoints

The list_gemini_agents, get_gemini_agent, delete_gemini_agent, and
list_gemini_agent_versions endpoints previously constructed a hardcoded
data dict with no mechanism to pass provider credentials.  Unlike
create_gemini_agent (POST, reads litellm_params_template from body),
these GET/DELETE endpoints gave no way for multi-tenant callers to
supply a per-request api_key or other LiteLLM params.

Fix:
- Add _merge_query_params_into_data() helper that reads query parameters
  from the request and merges them into the data dict without overwriting
  already-set keys (e.g. path params like 'name').
- Support a JSON-encoded litellm_params_template query parameter
  (matching the POST body pattern) as well as flat key=value pairs
  (e.g. api_key=AIza...).
- Apply the helper in all four affected endpoints.
- Add 13 unit tests covering the helper and each endpoint.

Co-authored-by: Sameer Kankute <Sameerlite@users.noreply.github.com>

* fix: pass model=None for managed agent proxy endpoints to prevent agent name polluting data["model"]

Endpoints acreate_agent, aget_agent, adelete_agent, and alist_agent_versions
were passing model=<agent_name> to base_process_llm_request. This caused
common_processing_pre_call_logic to write the agent name into self.data["model"],
which then triggered spurious model-alias mapping, rate-limiting lookups, and
logging tied to a non-existent model deployment.

The agent name is already carried in data["name"] and is passed correctly to
the SDK functions (litellm.interactions.agents.*). There is no reason to also
set model=<agent_name>; the correct value is model=None for all five managed-agent
management routes.

Adds tests/test_litellm/proxy/google_endpoints/test_managed_agents_model_param.py
to verify all five managed-agent endpoints pass model=None.

Co-authored-by: Sameer Kankute <Sameerlite@users.noreply.github.com>

* fix: address greptile P1/P2 review comments

P1 (router.py): Restore fallback/retry support for acreate_interaction
and create_interaction. Both were silently moved to _init_interactions_api_endpoints
(direct call, no fallbacks). Moved them back to _ageneric_api_call_with_fallbacks
so users with configured fallback models keep retry behaviour.

P1 security (agents_endpoints.py): Remove flat query-param credential
path (e.g. ?api_key=AIza...) from _merge_query_params_into_data.
Credentials in URL query strings appear verbatim in server access logs,
CDN edge logs, and browser history. Only the JSON-encoded
litellm_params_template query param (matching the POST body pattern) is
retained.

P2 (interactions/http_handler.py): Extract _BaseHTTPHandler with shared
_handle_error, _sync_client, and _async_client helpers. InteractionsHTTPHandler
now extends _BaseHTTPHandler. The _async_client reads the provider from
litellm_params instead of hardcoding GEMINI.

P2 (interactions/agents/http_handler.py): AgentsHTTPHandler now extends
InteractionsHTTPHandler (which inherits _BaseHTTPHandler) so all shared
HTTP infrastructure is reused rather than duplicated. Removes the
hardcoded LlmProviders.GEMINI from the async client path.

Co-authored-by: Cursor <cursoragent@cursor.com>

* fix: address CI failures from greptile review fixes

- black: format interactions/agents/main.py and utils.py
- tests: update test_gemini_agents_endpoints.py to match new
  _merge_query_params_into_data behaviour (flat credential params are
  rejected; only JSON-encoded litellm_params_template is accepted)
- ci: add test_gemini_agents_endpoints.py to endpoints-and-responses
  shard in test-unit-proxy-db.yml so assert-shard-coverage passes
- tests: add _initialize_managed_agents_endpoints and
  _init_managed_agents_api_endpoints test coverage so router_code_coverage
  passes; also fix TestRouterCreateInteractionRouting to reflect that
  acreate_interaction now correctly routes through
  _ageneric_api_call_with_fallbacks (restoring fallback support)

Co-authored-by: Cursor <cursoragent@cursor.com>

* fix: remove InteractionsHTTPHandler._handle_error override to fix type errors

AgentsHTTPHandler extends InteractionsHTTPHandler and calls
self._handle_error(provider_config=agents_api_config) where
agents_api_config is BaseAgentsAPIConfig. Python MRO resolved _handle_error
to InteractionsHTTPHandler._handle_error which expected BaseInteractionsAPIConfig,
causing 10 mypy arg-type errors in interactions/agents/http_handler.py.

Removing the redundant override lets both classes inherit _BaseHTTPHandler._handle_error
(provider_config: Any) which is structurally correct for both config types.

Co-authored-by: Cursor <cursoragent@cursor.com>

* fix: agent-only interactions and managed agents provider routing

Resolve None custom_llm_provider in agents HTTP client lookup and set
custom_llm_provider on GenericLiteLLMParams for all agent CRUD paths.

Stop mapping agent names to proxy model routing; route interactions
through _init_interactions_api_endpoints with fallbacks only when model
is set. Consolidate duplicate router elif branches for interaction APIs.

Co-authored-by: Cursor <cursoragent@cursor.com>

* Fix greptile review

* test(agents): add unit tests for managed agents SDK and HTTP handler

Adds coverage for the new `litellm.interactions.agents` surface area:
- main.py: sync/async entry points (create/list/get/delete/list_versions),
  provider config lookup, logging-obj helper, async error wrapping
- http_handler.py: every CRUD method (sync + async paths), `_is_async`
  dispatch branches, and provider error mapping through GeminiAgentsConfig
- utils.py: get_provider_agents_api_config for supported / unsupported
  providers

Brings patch coverage on these files from <25% to ~100% so codecov/patch
is satisfied.

Co-authored-by: Mateo Wang <mateo-berri@users.noreply.github.com>

* docs(gemini-agents): fix misleading credential-passing examples in GET/DELETE docstrings (#28293)

The four GET/DELETE endpoint docstrings (list_gemini_agents,
get_gemini_agent, delete_gemini_agent, list_gemini_agent_versions)
documented passing per-request credentials as flat query parameters
(e.g. ?api_key=AIza...). However, _merge_query_params_into_data only
reads the JSON-encoded litellm_params_template query parameter and
intentionally ignores flat params (URL query strings appear verbatim
in access logs, browser history, and Referer headers).

Callers following the documented curl examples would have their
credentials silently dropped and hit auth failures against Gemini.

Update the examples to use the supported JSON-encoded
litellm_params_template query parameter, matching _merge_query_params_into_data's own docstring.

Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: Mateo Wang <mateo-berri@users.noreply.github.com>

* refactor(agents): rename provider-agnostic agent response types

Move GeminiAgent{ListResponse,DeleteResult,VersionsResponse} to
provider-neutral names (AgentListResponse, AgentDeleteResult,
AgentVersionsResponse) so the BaseAgentsAPIConfig interface no longer
references Gemini-specific type names.

* fix(gemini-agents): close veria-flagged credential-escalation gaps

Two high-severity findings from the veria-ai PR review are addressed:

1. **api_base override could leak the shared Gemini key**
   GeminiAgentsConfig.validate_environment falls back to GOOGLE_API_KEY /
   GEMINI_API_KEY when no api_key is supplied. Combined with caller-controlled
   api_base on the proxy CRUD endpoints, an authenticated user could redirect
   the outbound request to an attacker-controlled host and capture the
   operator's shared Gemini key from the x-goog-api-key header. The config
   now refuses env-fallback whenever api_base is explicitly overridden.

2. **Managed-agent CRUD exposed to ordinary LLM keys**
   The new /v1beta/agents routes live in google_routes (i.e. llm_api_routes),
   so any non-admin LLM key can reach them. Unlike /v1beta/models/...:
   generateContent these endpoints are NOT model-routed and have no
   model_list-supplied credentials, so env-fallback would let any LLM key
   list / create / delete agents inside the operator's Gemini project. Each
   endpoint now calls _enforce_caller_supplied_provider_key, which requires
   non-admin callers to supply their own Gemini api_key via
   litellm_params_template. Proxy admins keep the env-fallback convenience.

Tests cover non-admin rejection, admin allow-through, the api_base override
guard, and SDK env-fallback when api_base is not overridden.

Co-authored-by: Mateo Wang <mateo-berri@users.noreply.github.com>

* test(router): restore strict assert_called_once_with on interactions default-provider test

---------

Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: Sameer Kankute <Sameerlite@users.noreply.github.com>
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Co-authored-by: Mateo Wang <mateo-berri@users.noreply.github.com>
2026-05-19 16:02:03 -07:00

454 lines
17 KiB
Python

"""
Lazy registration for optional feature routers. Each LAZY_FEATURES entry
imports its module only on the first request matching its path prefix,
saving ~700 MB at idle for deployments that don't use these features.
First hit pays the import cost (1-3 s for heavy modules); /openapi.json
omits each feature's routes until the feature is warmed.
"""
import asyncio
import importlib
import sys
from dataclasses import dataclass, field
from typing import TYPE_CHECKING, Callable, Dict, Tuple
from starlette.types import Receive, Scope, Send
from litellm._logging import verbose_proxy_logger
if TYPE_CHECKING:
from fastapi import APIRouter, FastAPI
def _include_router(attr_name: str = "router") -> Callable[["FastAPI", object], None]:
def _register(app: "FastAPI", module: object) -> None:
app.include_router(getattr(module, attr_name))
return _register
def _mount_app(
prefix: str, attr_name: str = "app"
) -> Callable[["FastAPI", object], None]:
def _register(app: "FastAPI", module: object) -> None:
app.mount(path=prefix, app=getattr(module, attr_name))
return _register
@dataclass(frozen=True)
class LazyFeature:
name: str
module_path: str
path_prefixes: Tuple[str, ...]
register_fn: Callable[["FastAPI", object], None] = field(
default_factory=lambda: _include_router("router")
)
# For routes whose path has a leading parameter (e.g. /{server}/authorize)
# — startswith can't match those, so the matcher also checks endswith.
path_suffixes: Tuple[str, ...] = ()
# Keep the stub injected even after load — for mounted ASGI sub-apps
# whose routes don't appear in the parent app's openapi spec.
persistent_swagger_stub: bool = False
LAZY_FEATURES: Tuple[LazyFeature, ...] = (
LazyFeature(
name="guardrails",
module_path="litellm.proxy.guardrails.guardrail_endpoints",
path_prefixes=(
"/guardrails",
"/v2/guardrails",
"/apply_guardrail",
"/policies/usage",
),
),
LazyFeature(
name="policies",
module_path="litellm.proxy.management_endpoints.policy_endpoints",
# Trailing slash to avoid matching /policies/... (policy_engine).
path_prefixes=("/policy/", "/utils/test_policies_and_guardrails"),
),
LazyFeature(
name="policy_engine",
module_path="litellm.proxy.policy_engine.policy_endpoints",
path_prefixes=("/policies",),
),
LazyFeature(
name="policy_resolve",
module_path="litellm.proxy.policy_engine.policy_resolve_endpoints",
path_prefixes=("/policies/resolve", "/policies/attachments/estimate-impact"),
),
LazyFeature(
name="agents",
module_path="litellm.proxy.agent_endpoints.endpoints",
path_prefixes=("/v1/agents", "/agents", "/agent/"),
),
LazyFeature(
name="gemini_agents",
module_path="litellm.proxy.google_endpoints.agents_endpoints",
path_prefixes=("/v1beta/agents",),
),
LazyFeature(
name="a2a",
module_path="litellm.proxy.agent_endpoints.a2a_endpoints",
path_prefixes=("/a2a", "/v1/a2a"),
),
LazyFeature(
name="vector_stores",
module_path="litellm.proxy.vector_store_endpoints.endpoints",
path_prefixes=("/v1/vector_stores", "/vector_stores", "/v1/indexes"),
),
LazyFeature(
name="vector_store_management",
module_path="litellm.proxy.vector_store_endpoints.management_endpoints",
# Trailing slash to avoid matching /vector_stores/... (vector_stores).
path_prefixes=("/vector_store/", "/v1/vector_store/"),
),
LazyFeature(
name="vector_store_files",
# Routes appear under both /v1/vector_stores/{id}/files and the
# un-versioned form, so both prefixes must trigger the load.
module_path="litellm.proxy.vector_store_files_endpoints.endpoints",
path_prefixes=("/v1/vector_stores", "/vector_stores"),
),
LazyFeature(
name="tools",
module_path="litellm.proxy.management_endpoints.tool_management_endpoints",
path_prefixes=("/v1/tool", "/tool"),
),
LazyFeature(
name="search_tools",
module_path="litellm.proxy.search_endpoints.search_tool_management",
path_prefixes=("/search_tools",),
),
# mcp_management owns most /v1/mcp/* admin routes; mcp_app is the mounted
# streaming sub-app at /mcp.
LazyFeature(
name="mcp_management",
module_path="litellm.proxy.management_endpoints.mcp_management_endpoints",
path_prefixes=("/v1/mcp/",),
),
LazyFeature(
# Also serves /.well-known/oauth-* (OAuth metadata discovery).
# No /mcp/oauth prefix here: the mounted /mcp sub-app would
# shadow it, and there are no actual routes there anyway.
name="mcp_byok_oauth",
module_path="litellm.proxy._experimental.mcp_server.byok_oauth_endpoints",
path_prefixes=("/v1/mcp/oauth", "/.well-known/oauth-"),
),
LazyFeature(
# Serves OAuth dance endpoints (/authorize, /token, /callback,
# /register) plus several /.well-known/ discovery URLs at the proxy
# root — needed for MCP-over-OAuth flows even before /mcp is hit.
name="mcp_discoverable",
module_path="litellm.proxy._experimental.mcp_server.discoverable_endpoints",
path_prefixes=(
"/.well-known/oauth-",
"/.well-known/openid-configuration",
"/.well-known/jwks.json",
"/authorize",
"/token",
"/callback",
"/register",
),
# Catches the /{mcp_server_name}/authorize|token|register variants.
path_suffixes=("/authorize", "/token", "/register"),
),
LazyFeature(
name="mcp_rest",
module_path="litellm.proxy._experimental.mcp_server.rest_endpoints",
path_prefixes=("/mcp-rest",),
),
LazyFeature(
# Hardcoded /mcp matches BASE_MCP_ROUTE; importing the constant
# here would defeat lazy loading.
name="mcp_app",
module_path="litellm.proxy._experimental.mcp_server.server",
path_prefixes=("/mcp",),
register_fn=_mount_app("/mcp", attr_name="app"),
persistent_swagger_stub=True,
),
LazyFeature(
name="config_overrides",
module_path="litellm.proxy.management_endpoints.config_override_endpoints",
path_prefixes=("/config_overrides",),
),
LazyFeature(
name="realtime",
module_path="litellm.proxy.realtime_endpoints.endpoints",
path_prefixes=("/openai/v1/realtime", "/v1/realtime", "/realtime"),
),
LazyFeature(
name="anthropic_passthrough",
module_path="litellm.proxy.anthropic_endpoints.endpoints",
path_prefixes=("/v1/messages", "/anthropic", "/api/event_logging"),
),
LazyFeature(
name="anthropic_skills",
module_path="litellm.proxy.anthropic_endpoints.skills_endpoints",
path_prefixes=("/v1/skills", "/skills"),
),
LazyFeature(
name="langfuse_passthrough",
module_path="litellm.proxy.vertex_ai_endpoints.langfuse_endpoints",
path_prefixes=("/langfuse",),
),
LazyFeature(
name="evals",
module_path="litellm.proxy.openai_evals_endpoints.endpoints",
path_prefixes=("/v1/evals", "/evals"),
),
LazyFeature(
name="claude_code_marketplace",
module_path="litellm.proxy.anthropic_endpoints.claude_code_endpoints",
path_prefixes=("/claude-code",),
register_fn=_include_router("claude_code_marketplace_router"),
),
LazyFeature(
name="scim",
module_path="litellm.proxy.management_endpoints.scim.scim_v2",
path_prefixes=("/scim",),
register_fn=_include_router("scim_router"),
),
LazyFeature(
name="cloudzero",
module_path="litellm.proxy.spend_tracking.cloudzero_endpoints",
path_prefixes=("/cloudzero",),
),
LazyFeature(
name="vantage",
module_path="litellm.proxy.spend_tracking.vantage_endpoints",
path_prefixes=("/vantage",),
),
LazyFeature(
name="usage_ai",
module_path="litellm.proxy.management_endpoints.usage_endpoints",
path_prefixes=("/usage/ai",),
),
LazyFeature(
name="prompts",
module_path="litellm.proxy.prompts.prompt_endpoints",
path_prefixes=("/prompts", "/utils/dotprompt_json_converter"),
),
LazyFeature(
name="jwt_mappings",
module_path="litellm.proxy.management_endpoints.jwt_key_mapping_endpoints",
path_prefixes=("/jwt/key/mapping",),
),
LazyFeature(
name="compliance",
module_path="litellm.proxy.management_endpoints.compliance_endpoints",
path_prefixes=("/compliance",),
),
LazyFeature(
name="access_groups",
module_path="litellm.proxy.management_endpoints.access_group_endpoints",
path_prefixes=("/access_group", "/v1/access_group", "/v1/unified_access_group"),
),
)
class LazyFeatureMiddleware:
"""ASGI middleware that imports + registers a feature router on first
matching request. Idempotent; once loaded, subsequent requests skip."""
def __init__(
self,
app,
fastapi_app: "FastAPI",
features: Tuple[LazyFeature, ...] = LAZY_FEATURES,
):
self.app = app
self._fastapi_app = fastapi_app
self._features = features
# SERVER_ROOT_PATH is a process-startup env var, cache the normalized
# form once instead of recomputing per request. Lazy import to avoid
# pulling proxy.utils into this module's import graph at startup
# (proxy_server imports both).
from litellm.proxy.utils import get_server_root_path
self._root_path = get_server_root_path().rstrip("/")
# Loaded set / per-feature locks live on app.state so the warm endpoint
# and the middleware share them — preventing duplicate registrations
# when both paths fire for the same feature.
if not hasattr(fastapi_app.state, "lazy_loaded"):
fastapi_app.state.lazy_loaded = set()
fastapi_app.state.lazy_locks = {}
@property
def _loaded(self) -> set:
return self._fastapi_app.state.lazy_loaded
async def __call__(self, scope: Scope, receive: Receive, send: Send) -> None:
# Short-circuit once every feature has loaded.
if scope["type"] in ("http", "websocket") and len(self._loaded) < len(
self._features
):
path = scope.get("path", "")
# Strip SERVER_ROOT_PATH so prefix matching works under a server
# root path. Without this, requests like /api/v1/policies/... never
# match the registered prefixes (/policies/...) and lazy features
# stay unloaded — every endpoint under them returns 404. The
# `+ "/"` boundary prevents false-positive matches (e.g. /apiv2
# against root /api). If the path doesn't start with the prefix
# (e.g. a reverse proxy already stripped it), we leave it alone.
if self._root_path and path.startswith(self._root_path + "/"):
path = path[len(self._root_path) :]
for feat in self._features:
if feat.module_path in self._loaded:
continue
if any(path.startswith(p) for p in feat.path_prefixes) or any(
path.endswith(s) for s in feat.path_suffixes
):
await _force_load(self._fastapi_app, feat)
await self.app(scope, receive, send)
async def _force_load(app: "FastAPI", feat: LazyFeature) -> bool:
"""Import + register a lazy feature exactly once per (app, module).
Shared by the middleware and the /lazy/warm endpoint."""
if not hasattr(app.state, "lazy_loaded"):
app.state.lazy_loaded = set()
app.state.lazy_locks = {}
lock = app.state.lazy_locks.setdefault(feat.module_path, asyncio.Lock())
async with lock:
if feat.module_path in app.state.lazy_loaded:
return False
try:
# Import on a thread (heavy modules take 1-3 s). register_fn
# mutates app.router.routes, so it stays on the loop thread.
loop = asyncio.get_running_loop()
module = await loop.run_in_executor(
None, importlib.import_module, feat.module_path
)
feat.register_fn(app, module)
app.state.lazy_loaded.add(feat.module_path)
app.openapi_schema = None
verbose_proxy_logger.info(
"Lazy-loaded optional feature %r (module: %s)",
feat.name,
feat.module_path,
)
return True
except Exception as exc:
# Mark loaded anyway so we don't retry on every request.
app.state.lazy_loaded.add(feat.module_path)
verbose_proxy_logger.warning(
"Failed to lazy-load optional feature %r (module: %s): %s. "
"This feature's endpoints will return 404 until restart.",
feat.name,
feat.module_path,
exc,
)
return False
def attach_lazy_features(app: "FastAPI") -> None:
app.include_router(_make_warmup_router(app))
app.add_middleware(LazyFeatureMiddleware, fastapi_app=app)
def _make_warmup_router(app: "FastAPI") -> "APIRouter":
"""POST /lazy/warm/{name}: load a feature and return its partial openapi
so the Swagger plugin can merge in-place without a full /openapi.json refetch.
Requires auth — anyone who can hit the proxy can already trigger the same
imports by sending a real request to a feature's prefix, but gating this
debug endpoint avoids unauthenticated callers forcing the import chain."""
from fastapi import APIRouter, Depends, HTTPException
from fastapi.openapi.utils import get_openapi
from litellm.proxy.auth.user_api_key_auth import user_api_key_auth
router = APIRouter()
@router.post(
"/lazy/warm/{name}",
include_in_schema=False,
dependencies=[Depends(user_api_key_auth)],
)
async def warm(name: str):
feat = next((f for f in LAZY_FEATURES if f.name == name), None)
if feat is None:
raise HTTPException(404, f"unknown lazy feature: {name}")
if feat.persistent_swagger_stub:
return {"stub_path": None, "paths": {}, "components": {"schemas": {}}}
await _force_load(app, feat)
feat_routes = [
r
for r in app.routes
if any(getattr(r, "path", "").startswith(p) for p in feat.path_prefixes)
]
full = get_openapi(title=app.title, version=app.version, routes=feat_routes)
# Force all operations under one tag so they group under a single Swagger
# section — many lazy modules tag routes inconsistently.
for path_ops in full.get("paths", {}).values():
for op in path_ops.values():
if isinstance(op, dict):
op["tags"] = [feat.name]
return {
"stub_path": feat.path_prefixes[0],
"paths": full.get("paths", {}),
"components": {"schemas": full.get("components", {}).get("schemas", {})},
}
return router
def inject_lazy_stubs(schema: Dict) -> Dict:
"""Inject openapi entries for unloaded features. Uses the snapshot file
when available (full route info), otherwise falls back to a single
placeholder per feature. Any failure logs and returns the schema unchanged
so /openapi.json never 500s on a cosmetic injection bug."""
try:
from litellm.proxy._lazy_openapi_snapshot import load_snapshot
snapshot = load_snapshot()
paths = schema.setdefault("paths", {})
schemas = schema.setdefault("components", {}).setdefault("schemas", {})
for feat in LAZY_FEATURES:
if feat.module_path in sys.modules and not feat.persistent_swagger_stub:
continue
fragment = (snapshot or {}).get(feat.name)
if fragment:
for p, ops in fragment.get("paths", {}).items():
paths.setdefault(p, ops)
for name, sch in (
fragment.get("components", {}).get("schemas", {}).items()
):
schemas.setdefault(name, sch)
continue
prefix = feat.path_prefixes[0]
if prefix in paths:
continue
paths[prefix] = {
"get": {
"tags": [feat.name],
"summary": feat.name,
"responses": {"200": {"description": "OK"}},
}
}
except Exception as exc:
verbose_proxy_logger.warning("inject_lazy_stubs failed: %s", exc)
return schema
def lazy_tag_to_prefix() -> Dict[str, str]:
"""feature.name -> first prefix, used by the Swagger warmup JS plugin.
Returns empty when the snapshot is loaded — the plugin is unnecessary
because /openapi.json already has full route info."""
from litellm.proxy._lazy_openapi_snapshot import load_snapshot
if load_snapshot():
return {}
return {
feat.name: feat.path_prefixes[0]
for feat in LAZY_FEATURES
if not feat.persistent_swagger_stub
}