* fix(proxy): make async_post_call_response_headers_hook consistent across all endpoints
The response headers hook had 5 gaps that prevented callbacks from
reliably extracting routing metadata across endpoint types:
1. Hook never fired for /audio/transcriptions (endpoint bypasses
base_process_llm_request)
2. custom_llm_provider not accessible in hook data for any endpoint
3. custom_llm_provider not stamped in ResponsesAPIResponse._hidden_params
(unlike chat completions)
4. model_info under inconsistent keys (metadata vs litellm_metadata)
5. request_headers always None at all call sites
This adds a litellm_call_info parameter to the hook that normalizes
routing metadata (custom_llm_provider, model_info, api_base, model_id)
regardless of endpoint type. Also stamps custom_llm_provider on
Responses API responses, adds the hook call to the transcription
handler, and passes request_headers at all call sites.
Supersedes PR #21385.
* fix(proxy): address review feedback — safer backwards compat and None guards
- Replace try/except TypeError with inspect.signature() check for
litellm_call_info backwards compatibility. This avoids masking real
TypeErrors inside callback implementations and prevents double
invocation with inconsistent parameters.
- Use (data.get("key") or {}) instead of data.get("key", {}) to guard
against keys that exist with an explicit None value, which would
cause AttributeError on the subsequent .get() call.
* fix(proxy): cache inspect.signature result for callback compat check
Move the inspect.signature() call into a module-level helper with a
dict cache keyed by callback identity. Avoids repeated introspection
per request per callback in the hot path.
* fix(proxy): use class identity for signature cache key
Key the _CALLBACK_ACCEPTS_CALL_INFO cache by id(type(cb)) instead of
id(cb) to avoid stale entries from Python address reuse after GC.
All instances of the same callback class share the same method
signature, so class identity is both safer and more cache-efficient.
- Thread api_version through HTTP handlers to Azure realtime endpoints
- Make expires_at optional in RealtimeClientSecretResponse
- Fix test token expiry times to be in the future
- Populate user_id and team_id in minimal_auth for spend tracking
Made-with: Cursor
- Azure Model Router transform_response: let parent extract actual model from raw response
- common_request_processing: skip model override for Azure Model Router requests
- proxy_server: skip streaming chunk model restamp for Azure Model Router
- Add _is_azure_model_router_request helper
- Add tests for non-streaming and streaming
Made-with: Cursor
* fix(mcp): OAuth2 chat connect - tools fetch, auth flow, and status fixes
- schema.prisma: add missing MCP table fields (approval_status, submitted_by, submitted_at, reviewed_at, review_notes) to prevent destructive migrations
- rest_endpoints.py: inject user OAuth token via extra_headers for OAuth2 servers so tools list is populated; add server name->UUID resolution so MCPConnectPicker name lookups work
- mcp_registry.json: fix Atlassian defaults (transport: http, url: .../v1/mcp)
- ChatPage.tsx: read mcpOauthReturn param to init sidebarView="apps" on OAuth return, clean up param after mount
- MCPAppsPanel.tsx: auto-add OAuth2 servers to selectedServers when credential detected; onConnect also enables server for chat; disconnect removes from selectedServers
- mcp_servers.tsx: sort servers by created_at DESC
- useUserMcpOAuthFlow.tsx: append mcpOauthReturn=apps to return URL so Apps panel is mounted on return
* address greptile review feedback (greploop iteration 1)
* fix(mcp): inject stored OAuth2 token when fetching tools via /responses API
When a user has connected an OAuth2 MCP server (e.g. Atlassian) and then
uses the /responses endpoint with that server, tool listing was failing
because the stored per-user OAuth token was never injected.
Two fixes:
1. server.py: add _get_user_oauth_extra_headers_from_db() helper; call it
in _get_tools_from_mcp_servers when oauth2_headers is None for an OAuth2
server, falling back to the user's stored token in LiteLLM_MCPUserCredentials
2. litellm_proxy_mcp_handler.py: also intercept MCP tools whose server_url
matches */mcp/<server_name> (e.g. http://localhost:4000/mcp/atlassian_test)
by rewriting them to litellm_proxy/mcp/<server_name> so they go through
the internal handler (and get the OAuth token injected) instead of being
forwarded to OpenAI raw where localhost is unreachable
* address greptile review feedback (greploop iteration 2)
* test(mcp): add unit test for OAuth2 token injection in _get_tools_from_mcp_servers
Verifies that when _get_tools_from_mcp_servers is called for an OAuth2 MCP
server without oauth2_headers in the request, the implementation:
- calls _prefetch_oauth_creds_for_user once (not per-server) to avoid N+1 queries
- passes the stored token as extra_headers={"Authorization": "Bearer ..."} to
_get_tools_from_server so the upstream OAuth2 MCP server authenticates correctly
* address greptile review feedback (greploop iteration 3)
* address greptile review feedback (greploop iteration 4)
* address greptile review feedback (greploop iteration 5)
* redesign credentials table to use Tremor table layout matching Keys page
* fix: /server/oauth authorize 422 - make client_id optional, fall back to real DB server
* fix: mcp_token client_id optional, resolve from server record
* fix: look up real server by UUID (get_mcp_server_by_id) before falling back to name
* Update litellm/responses/mcp/litellm_proxy_mcp_handler.py
Co-authored-by: greptile-apps[bot] <165735046+greptile-apps[bot]@users.noreply.github.com>
* fix: address greptile feedback - client_id guards, dict spread, helper refactor, tests
- mcp_management_endpoints: raise 400 when resolved_client_id is empty in
mcp_authorize and mcp_token instead of forwarding "" to upstream
- litellm_proxy_mcp_handler: use {**tool, "server_url": ...} spread instead
of dict(tool) + mutation for shallow copy safety
- rest_endpoints: extract _oauth2_server_ids set comprehension to a named
_get_oauth2_server_ids() helper for clarity; add Set to typing imports
- test_rest_endpoints: add tests for name→UUID resolution path,
access-denied when resolved UUID not in allowed list, and OAuth2 user
token injection for single-server requests; fix fake_get_tools signature
to accept extra_headers kwarg
---------
Co-authored-by: greptile-apps[bot] <165735046+greptile-apps[bot]@users.noreply.github.com>
Snowflake's Cortex LLM API (like Anthropic) requires tool_choice as an
object with a "type" field, not as a bare string. Passing tool_choice="auto"
(or "required"/"none") results in error 390142 "invalid payload".
This fix transforms OpenAI string tool_choice values to the Snowflake
object format:
- "auto" -> {"type": "auto"}
- "required" -> {"type": "any"} (Snowflake/Anthropic convention)
- "none" -> {"type": "none"}
The dict-to-dict transformation for specific function tool choices
({"type": "function", "function": {"name": "..."}} -> {"type": "tool",
"name": [...]}) remains unchanged.
Fixes#23284
Co-authored-by: gambletan <tan@echooo.com>
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Co-authored-by: Krish Dholakia <krrishdholakia@gmail.com>
* feat(ui): group MCP tools by CRUD risk category in tool permission panels
Adds a CRUD-classification layer to the MCP tool allowlist UI so admins
can allow/block an entire risk category (Read / Create / Update / Delete)
with a single toggle instead of managing a flat list of individual tools.
- New `mcpToolCrudClassification.ts` utility: regex-based classifier that
buckets tool names/descriptions into read/create/update/delete/unknown
- New `McpCrudPermissionPanel` component: collapsible sections per CRUD
group, group-level Switch toggle, individual tool checkboxes, risk badges
(green Safe / yellow Medium / red High Risk)
- `mcp_tool_configuration.tsx`: adds "Risk Groups / Flat List" radio toggle;
defaults to the CRUD-grouped view, flat list is still accessible
- `MCPToolPermissions.tsx` (key/team assignment): replaces flat checkboxes
with the CRUD panel; adds per-server view toggle; delete tools are blocked
by default for newly-added servers (safer default for key/team scoping)
No backend or schema changes — uses existing `allowed_tools` and
`mcp_tool_permissions` fields.
* fix(mcp): OAuth2 chat connect - tools fetch, auth flow, and status fixes
- schema.prisma: add missing MCP table fields (approval_status, submitted_by, submitted_at, reviewed_at, review_notes) to prevent destructive migrations
- rest_endpoints.py: inject user OAuth token via extra_headers for OAuth2 servers so tools list is populated; add server name->UUID resolution so MCPConnectPicker name lookups work
- mcp_registry.json: fix Atlassian defaults (transport: http, url: .../v1/mcp)
- ChatPage.tsx: read mcpOauthReturn param to init sidebarView="apps" on OAuth return, clean up param after mount
- MCPAppsPanel.tsx: auto-add OAuth2 servers to selectedServers when credential detected; onConnect also enables server for chat; disconnect removes from selectedServers
- mcp_servers.tsx: sort servers by created_at DESC
- useUserMcpOAuthFlow.tsx: append mcpOauthReturn=apps to return URL so Apps panel is mounted on return
* fix(mcp-crud-ui): address greptile review feedback
- use Checkbox (not Switch) for group toggle so indeterminate works
- add toolPermissionsRef to avoid stale closure race on concurrent server fetches
- remove unused blockDeleteByDefault prop from McpCrudPermissionPanel
- classify tools by name first; fall back to description only when name yields no match
- add Risk Groups / Flat List toggle to mcp_tool_configuration.tsx
* fix(mcp-crud-ui): address greptile 3/5 review
- remove non-functional XIcon remove-server button (no onRemoveServer prop wired)
- fix stale closure in MCPAppsPanel auto-enable effect: use serversRef/selectedServersRef
- remove utility re-export from McpCrudPermissionPanel (classifyToolOp, groupToolsByCrud)
- remove redundant selectedTools.length === 0 guard (always true when !toolPermissions[id])
* fix(mcp-crud-ui): address greptile 3/5 review round 2
- check READ_RE before DELETE_RE in classifyToolOp so tools like
get_removed_entries are not silently blocked by delete-by-default
- expand undefined (allow-all) to full tool name list instead of
collapsing to [] (allow-none) in MCPToolPermissions and mcp_tool_configuration
- log OAuth credential fetch failures instead of silently swallowing them
* fix: cursor-pointer on read-only rows, stable sort, simplify handleCrudPanelChange
* fix: sanitize user_id/server_id in log to prevent log injection
* fix: add OAuth headers to call_tool_rest_api, fix stale accessToken closure, fix group toggle on filtered subset
* fix: batch OAuth creds query, hide empty CRUD groups on search, onChange stability
* fix: double-add race, conditional bulk query, narrow DELETE_RE, hoist search input
* fix(mcp): clear oauthConnected on deselect; null guard on allowedTools prop
* fix(mcp): remove user-provided values from debug log to fix log-injection lint
* fix(mcp): fix allowedTools undefined semantics; remove unused import and color field
- Replace gemini-pro with gemini-3-pro-preview in test_cost_discount_vertex_ai
(gemini-pro removed from cost map)
- Replace github/claude-3-5-sonnet-latest with github/claude-3-7-sonnet-20250219
in test_supports_function_calling_github_anthropic_alias (model removed)
- Add supports_multimodal, uses_embed_content, input/output_cost_per_token_above_256k_tokens
to JSON schema in test_utils.py (new properties added to model cost map)
Co-authored-by: yuneng-jiang <yuneng-jiang@users.noreply.github.com>
yuneng-jiang