Add a new team member permission `/team/daily/activity` that allows
non-admin team members to see all team usage data, not just their own.
## Changes
- Add `TEAM_DAILY_ACTIVITY` to `KeyManagementRoutes` enum and available
team member permissions
- Add `_team_member_has_permission` helper in `common_utils.py`
- Modify `/team/daily/activity` endpoint to skip API key filtering when
the member has this permission
- Add permission description and method detection in the UI
- Add backend and frontend tests
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Four tests were broken by commit e00c181f0c (Mcp user permissions #21462):
1. test_list_tools_single_server_unprefixed_names: The commit changed
_get_tools_from_mcp_servers to always add server prefixes (add_prefix=True),
removing the conditional that skipped prefixing for single servers.
Updated assertion from "toolA" → "zapier-toolA".
2. test_mcp_get_prompt_success: mcp_get_prompt now extracts the server name
from a prefixed prompt name via split_server_prefix_from_name(). Passing
unprefixed "hello" returns server_name="" which matches no server → 403.
Updated call to use "server_a-hello" so the server lookup succeeds.
3. test_e2e_jwt_team_mcp_permissions_enforced &
4. test_e2e_jwt_team_mcp_key_intersection:
The commit replaced `from typing import List` with
`from litellm.proxy.proxy_server import general_settings` in
MCPRequestHandler.get_allowed_mcp_servers(). Both tests mock
litellm.proxy.proxy_server with a types.ModuleType that lacked
general_settings, causing ImportError. Added general_settings={} to
both mock modules.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
test_complete_reload_flow and test_distributed_reload_check_function both
trigger code paths that assign a minimal stub dict to litellm.model_cost
(via the /reload/model_cost_map endpoint and _check_and_reload_model_cost_map).
Without restoring, subsequent tests in the same worker can't find gpt-4o
pricing and calculate spend=0.0 instead of the expected value.
Added try/finally save-and-restore of litellm.model_cost in both tests,
matching the pattern used in test_reload_model_cost_map_admin_access.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* feat(schema.prisma): add object permissions for end users
allows controlling if end user can call specific mcp servers
* feat: cleanup for customer_endpoints support of object permission id
* fix: cleanup str
* feat(customers/): enforce end user can only call allowed mcps - if configured
* docs: document customer/end user object permission usage
* feat: enforce end user permissions on MCP tool calls
This commit implements end user permission enforcement for MCP servers:
1. Always add server prefixes to MCP tool names
- Removed conditional logic that only added prefixes when multiple servers existed
- Now always adds server prefix for consistent tool naming across all scenarios
- Updated 5 locations in server.py (list_tools, get_prompts, get_resources,
get_resource_templates, get_prompt)
2. Created MCP End User Permission Guardrail Hook
- New guardrail hook: litellm/proxy/guardrails/guardrail_hooks/mcp_end_user_permission.py
- Runs on post_call to validate tool calls in LLM responses
- Extracts MCP server name from tool names (splits on first '-')
- Checks if end_user_id has permissions for the MCP server
- Raises GuardrailRaisedException if end user lacks permission
- Supports both streaming and non-streaming responses
3. Added comprehensive tests
- Test file: tests/test_litellm/proxy/guardrails/guardrail_hooks/test_mcp_end_user_permission.py
- Tests cover: authorized/unauthorized tools, non-MCP tools, no end_user scenarios
- Tests permission checking logic and exception raising
The hook integrates with the existing MCPRequestHandler._get_allowed_mcp_servers_for_end_user
to fetch end user permissions and enforce access control at the response level.
Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>
* refactor: remove redundant add_prefix variable assignments
Simplified the code by removing intermediate `add_prefix` variable
assignments and passing `True` directly to function calls since
we now always add server prefixes.
Changes:
- Removed `add_prefix = True` variable assignments in 5 locations
- Changed `add_prefix=add_prefix` to `add_prefix=True` in function calls
- Added inline comments to clarify the behavior
This makes the code more concise and clearer in intent.
Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>
* feat(auth_utils.py): support safety_identifier as a valid way of passing the end user id for responses api
* feat(llms): ensure 'tools' is correctly updated for responses api
* fix: fix greptile feedback
* feat: transformation.py
proper responses api tool handling for guardrail translation layer
---------
Co-authored-by: Claude Sonnet 4.5 <noreply@anthropic.com>
* feat(ui/): initial commit adding a compliance testing playground
allow proxy admins to test policies and guardrails against datasets
* feat(ui/): make score more friendly
* feat(policy_endpoints.py): new helper function for testing policies
* feat(policy_endpoints.py): expose new endpoint for testing policies and guardrails
enables compliance playground to work as expected
* feat(complianceui.tsx): show returned text
* docs: add DATABRICKS_API_KEY to environment settings reference
* fix: streaming test usage check on Pydantic model
* fix: mock litellm.proxy.proxy_server in test_skip_server_startup
Users had to set store_model_in_db in the config YAML and restart the proxy,
causing service downtime. This change allows the value to be written to the
LiteLLM_Config table and read from the database at runtime, with DB values
overriding config file values.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix(middleware): replace BaseHTTPMiddleware with pure ASGI middleware
BaseHTTPMiddleware wraps streaming responses with receive_or_disconnect
per chunk, blocking the event loop and causing severe throughput
degradation under concurrent streaming load (53% of CPU in profiling).
Converts PrometheusAuthMiddleware to a pure ASGI middleware using the
__call__(scope, receive, send) protocol.
* fix(streaming): remove expensive debug logging and optimize usage stripping
- Remove print_verbose calls that format chunk/response Pydantic objects,
triggering millions of __repr__ calls (8% of CPU in profiling)
- Guard remaining verbose_logger.debug with isEnabledFor(DEBUG) and use
lazy %s formatting instead of f-strings
- Replace usage stripping round-trip (model_dump + delete + reconstruct)
with a _usage_stripped flag, deferring exclusion to serialization time
* fix(proxy): remove per-chunk debug log and use _usage_stripped flag
- Remove verbose_proxy_logger.debug that formatted every streaming chunk
- Honor _usage_stripped flag from streaming handler to exclude usage
during model_dump_json serialization instead of reconstructing objects
* fix(proxy): remove per-chunk debug log in async_data_generator
Remove verbose_proxy_logger.debug that formatted every streaming chunk,
which triggered expensive Pydantic serialization on the hot path.
* fix indentation and add clarifying comment for usage stripping
* fix: guard calculate_total_usage against None usage in chunks
* fix: store chunk copy to preserve usage for calculate_total_usage
When `default_internal_user_params` was set, `insert_sso_user()` only
preserved SSO-provided roles if `role_mappings` was explicitly configured.
Roles from other valid SSO sources (Microsoft app_roles,
GENERIC_USER_ROLE_ATTRIBUTE, custom SSO handlers) were silently
overwritten with the default "internal_user" role, causing admin users
to be downgraded on first login or after user deletion.
Replace the `role_mappings_configured` gate with `_should_use_role_from_sso_response()`
which validates the role is a recognized LitellmUserRoles value regardless
of origin. Also removes an unnecessary DB round-trip to litellm_ssoconfig
on every new SSO user creation.
Fixes: admin users seeing internal-user UI after SSO login
Fixes: test_get_redirect_url_for_sso flaking due to local env vars
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- test_litellm_pre_call_utils.py: wrap test body in try/finally so
litellm.callbacks is always restored even when an assertion fails,
addressing greptile review comment
- test_langfuse_otel.py: resolve trivial merge conflict in comment
("unpatched" vs "unpatch-ed"), keeping correct spelling
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
test_reload_model_cost_map_admin_access calls the /reload/model_cost_map
HTTP endpoint with get_model_cost_map mocked to return a single-entry
dict. The endpoint handler does a direct module-level assignment
(litellm.model_cost = new_model_cost_map) which persists after the
patch context manager exits, stripping all models except gpt-3.5-turbo
from the in-memory cost map and causing subsequent tests that rely on
models like gemini-1.5-flash, multimodalembedding@001, and gpt-4o to
fail with "model not mapped" errors or zero-cost spend payloads.
Fix: save litellm.model_cost before the test and restore it (along with
invalidating the case-insensitive lookup cache) in a finally block.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Three test isolation issues fixed:
1. test_mcp_debug.py: Replace deprecated asyncio.get_event_loop().run_until_complete()
with asyncio.run() in TestWrapSendWithDebugHeaders. In Python 3.10+,
get_event_loop() raises RuntimeError when no event loop is set in the
current thread, causing test_injects_headers and test_body_messages_unchanged
to fail in isolation.
2. test_mcp_server_manager.py: After _reload_mcp_manager_module() creates a new
global_mcp_server_manager instance, server.py still holds a stale reference
to the old instance. Tests in test_mcp_server.py that populate the new
manager's registry and then call server.py functions (e.g. _get_tools_from_mcp_servers)
get empty results because server.py reads from the old manager. Fix: update
server.py's module-level reference after each reload.
3. test_litellm_pre_call_utils.py: test_add_litellm_metadata_from_request_headers
sets litellm.callbacks without restoring it afterward. Add cleanup to restore
original callbacks after the test to prevent state leaking to subsequent tests.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Four finally blocks in test_internal_user_endpoints.py and one in
test_ui_sso.py used the pattern:
if original_default_params is not None:
litellm.default_internal_user_params = original_default_params
else:
delattr(litellm, "default_internal_user_params")
Since the attribute is defined in litellm/__init__.py with a default of
None, `getattr(litellm, "default_internal_user_params", None)` returns
None. The else branch then calls delattr(), permanently removing the
attribute from the module for the rest of the process.
Subsequent tests in the same pytest-xdist worker (e.g.
test_add_new_member_* in test_management_helpers_utils.py) then fail
with: AttributeError: module 'litellm' has no attribute
'default_internal_user_params'
Fix: replace all five flawed finally blocks with a simple assignment:
litellm.default_internal_user_params = original_default_params
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
The test was using setattr() to set module-level attributes (including
proxy_logging_obj = MagicMock()) on the real litellm.proxy.proxy_server
module, but the finally block only had `pass` — no cleanup.
This left proxy_logging_obj as a MagicMock in subsequent tests running
in the same pytest-xdist worker, causing TypeError when log_db_metrics
decorator called asyncio.create_task(proxy_logging_obj.service_logging_obj
.async_service_success_hook(...)) — a MagicMock is not a coroutine.
Fix: save original attribute values before the test and restore them in
the finally block to ensure test isolation.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
test_role_mappings_override_default_internal_user_params was calling
delattr(litellm, 'default_internal_user_params') in its finally block
when the original value was None. This removes the attribute entirely from
the module, causing subsequent tests in the same xdist worker to get
AttributeError when accessing litellm.default_internal_user_params (because
litellm.__getattr__ has no handler for this name).
Fix: always restore the attribute by assignment (litellm.default_internal_user_params = original_default_params)
rather than deleting it.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- /spend/calculate: wrap response in proper OpenAPI 3.x content structure
- /credentials: split stacked route decorators into separate handlers to
eliminate path parameter conflict between by_name and by_model routes
* feat(schema.prisma): add object permissions for end users
allows controlling if end user can call specific mcp servers
* feat: cleanup for customer_endpoints support of object permission id
* fix: cleanup str
* feat(customers/): enforce end user can only call allowed mcps - if configured
* docs: document customer/end user object permission usage
* feat: address greptile comments
* Add MCP_SECURITY enum to SupportedGuardrailIntegrations
* Add MCP security guardrail initializer
* Add MCPSecurityGuardrail implementation
* Add MCP Security policy template
* Add Type filter to policy templates UI
* Add unit tests for MCP security guardrail
* fix(lint): remove unused Dict import from mcp_security_guardrail
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* Add French language support for EU AI Act Article 5 guardrail (#21427)
* Add French language support for EU AI Act Article 5 template
- Create eu_ai_act_article5_fr.yaml with comprehensive French keywords
- Includes identifier words: concevoir, créer, développer, noter, classer, etc.
- Includes block words: crédit social, comportement social, émotion des employés, etc.
- Includes always-block keywords for explicit prohibited practices
- Includes exceptions for research, compliance, and legitimate use cases
- Catches circumvention attempts with phrase variations
* Add comprehensive tests for French EU AI Act guardrail
- Test 3 critical scenarios: blocked query, circumvention attempt, safe query
- Test edge cases: case-insensitive, mixed language, research exceptions
- All 7 tests passing
- Validates both blocking and allowing behavior
* Fix content filter to support conditional matching without inherit_from
- Enable conditional matching when identifier_words + additional_block_words are present
- Previously required inherit_from, but EU AI Act templates are self-contained
- Fixes Greptile feedback: conditional matching now works as documented
* Add pure conditional matching test for French guardrail
- Test identifier + block word combinations not in always_block_keywords
- Verifies conditional matching works independently
- Addresses Greptile feedback about test coverage gap
* Fix exception word bypass risk in French template
- Replace short words (film, jeu, juste) with context-specific phrases
- Prevents substring matching bypasses (e.g., enjeu matching jeu)
- Add tests for bypass prevention and legitimate game context
- Addresses Greptile security feedback
* Make conditional match assertion more robust
- Use getattr to safely access exception detail field
- Check if detail is dict before calling .get()
- Addresses Greptile feedback about brittle string assertion
* Add French EU AI Act Article 5 policy template to registry
- Add eu-ai-act-article5-fr template for French language support
- Includes French description and guardrail info
- Matches structure of English template
* Address greptile review feedback (greploop iteration 1)
- Use status_code=400 instead of 403 to match guardrail logging convention
- Use prefix stripping instead of split('/')[-1] for robust server name extraction
* remove French EU AI Act template from policy_templates.json
---------
Co-authored-by: Julio Quinteros Pro <jquinter@gmail.com>
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
Use importlib.import_module + reload uniformly in both code paths
to ensure fresh module state regardless of whether litellm was
previously in sys.modules. This fixes the inconsistency where the
"not in sys.modules" branch didn't reload the module.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- test_pillar_guardrails.py: Fix fixture to properly update module-level
litellm reference using global keyword and assignment from reload
- test_anthropic_experimental_pass_through_messages_handler.py: Add missing
assert keywords to kwargs comparison statements (lines 36, 60-62)
- test_proxy_server.py: Replace silent pytest.skip with explicit assertion
to catch router initialization regressions
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
The _safe_get_request_headers caching uses request.state._cached_headers,
which returns a truthy MagicMock on bare MagicMock() objects instead of
None, breaking content-type detection for form-data tests.
