Helper util to add tools to optional_params.

Resolve and cache a collision-free short tool prefix on ``server``. Called at registration time for every MCP server entering the reg

Batch-fetch access groups in a single DB query and return a per-group resource map. Returns {ag_id: {"models": [...], "mcp_server_ids":

Build (forward, reverse) tool-name maps for a single request. forward[original] = sanitized -- only present when name was rewritten reverse

Builds complete response from raw chunks - Converts str chunks to generic chunks - Converts generic chunks to litellm chunks

Build a complete model cost entry from kwargs and model_info. Collects all CustomPricingLiteLLMParams fields present in kwargs and merges met

Build the Prisma DATABASE_URL query params controlling connection pool behavior. `connect_timeout` / `socket_timeout` map to the Prisma URL param

Build the audience/issuer/options kwargs for ``jwt.decode``. Setting ``JWT_AUDIENCE`` (and optionally ``JWT_ISSUER``) turns on the co

Build OAuth protected resource response with the appropriate URL pattern. For pass-through MCP servers (``MCPServer.is_oauth_passthrough``),

Produce SSE bytes for a simple text response (no tool calls).

Invoke the configured summary model. Prefers ``llm_router.acompletion`` so the model alias resolves against the proxy's ``model_list``; falls

Raise litellm.Timeout if the stream has exceeded LITELLM_MAX_STREAMING_DURATION_SECONDS.

Enforce model-permission checks for MCP sampling requests. Runs the same authorization checks as ``/chat/completions``: key-level, team-level

Only proxy admins may set `allowed_passthrough_routes` (top-level or under `metadata`) — it short-circuits the role-based route gate, so keys

Check rate limits using THREE-PHASE approach to prevent partial increments. Phase 1: Read-only check of ALL limits (no increments)

Check if team member is over their max budget within the team.

Convert a litellm completion response to MCP CreateMessageResult. Args: response: The litellm ModelResponse. model_name: The

Create a mock httpx response

Create prefixed tools and update tool mapping. Args: tools: List of original tools from server server: Serve

Decode a litellm-encoded encrypted-content item ID. Returns a dict with ``model_id`` and ``item_id`` keys, or ``None`` if the string

Decode realtime token payload; returns None for legacy/raw ephemeral tokens.

Decodes unified (LiteLLM-managed) vector_store_ids in file_search tools to provider-native IDs. Non-unified IDs are passed through unchanged

Drive the real boundary flow: open at ``pre_call`` then close at the async callback. ``ambient``, if given, is the span that is the active OTel co

Key-level model allowlist and client fallbacks (same as standard auth). Not included in common_checks — common_checks enforces team/user/proj

Ensure that tool_result messages have corresponding tool_calls in the previous assistant message. This is critical for Anthropic API

Open one upstream session and cache InitializeResult.instructions if missing. No-op when: - YAML/DB instructions are set o

Translate a single upstream Responses API chunk into the list of Interactions API events it should produce. Returning a list

Generate new token, update DB, invalidate cache, and return response.

Assert that the guardrail BLOCKS the sentence.

Extract pure base64 data from an image URL. If the URL is a data URL (e.g., "data:image/png;base64,iVBOR..."), extract and return only t

Extract the highest media resolution (detail) from image content in messages. This is used to set the global media_resolution in generation_

Pull the unified ``category`` / ``rate_limit_type`` fields off any exception that declares them (``litellm.RateLimitError`` and bare-

Extract and aggregate usage metadata from a list of WebSocket messages. Args: websocket_messages: List of WebSocket mess

Sync version of _async_filter_health_check_unhealthy_deployments.

Return allowed MCP servers for a request after applying filters. Args: user_api_key_auth: The authenticated user's API key info.

Get the key's own MCP ceiling from its object_permission (mcp_servers, tag-style mcp_access_groups, mcp_tool_permissions). U

Generate authentication headers based on auth type.

Get or lazy-load the batch rate limiter.

Get the header passed to LiteLLM to pass to downstream MCP servers By default litellm will check for the header `x-mcp-auth` by sett

Get the prisma client from proxy_server.

Return the list of SCIM ResourceType definitions per RFC 7643 Section 6.

Return the router's deployment-budget callback. Uses exact-type matching so proxy subclasses (e.g. virtual-key model budgets)

Handle hallucinated parallel tool call from openai - https://community.openai.com/t/model-tries-to-call-unknown-function-multi-tool-use-parallel/

Walk the exception chain to find an AttributeError at any depth. Checks __cause__, __context__, and the litellm-specific original_exception a

Check if a value has real content (not None, not empty list, not blank string).

True when a guardrail explicitly registers post_call. event_hook=None matches all hooks in should_run_guardrail but must not defer as

Filter out litellm params from the request body

Create a mock logging object.

Build a Router whose model_list looks like what the proxy creates for team-scoped BYOK DB models: the internal model_name is ``<public_na

Build a minimal UserAPIKeyAuth-like object for tests.

Check if a model string matches the Claude model naming pattern. Matches patterns like: - claude-opus-4-7 - claude-sonnet-4-6 -

Called on the event loop thread when the waitpid thread detects engine death.

Parse a boolean environment variable. Accepts true/false, 1/0, yes/no, on/off (case-insensitive, whitespace-tolerant) — matching the

Perform a health check for each model in the list. max_concurrency: Optional limit on concurrent health check requests.

Performs a direct OAuth token exchange including the PKCE code_verifier. fastapi-sso does not forward code_verifier, so when PKCE is