Common request processing logic for both chat completions and responses API endpoints

Calculates the cost per token for a given model, prompt tokens, and completion tokens. Handles context caching as well. Input:

Factory for test handlers with standard defaults.

Raises ProxyException (OpenAI API compatible) if an exception is raised

Check if the request body is safe. A malicious user can set the api_base to their own domain and invoke POST /chat/completions to intercept

Read response content.

Safely read the request body and parse it as JSON. Parameters: - request: The request object to read the body from Returns: - d

List all API keys with optional filtering and pagination. Args: page (Optional[int]): Page number for pagination

Convenience wrapper around update_environment_variables that automatically extracts metadata/litellm_metadata from kwargs, so

Percent-encode one user-controlled URL path segment. ``urllib.parse.quote(..., safe="")`` intentionally leaves RFC 3986 unreserved characters

Create a virtual key; delete it on teardown. `models` restricts which models the key may call (None/[] means all). `user_id` is required for

Start the logging worker. Idempotent - safe to call multiple times.

A call that should have succeeded but didn't is a hard failure, never a skip: if the proxy can't make a call it's expected to, the test must fail.

Allows users to raise custom exceptions/log when a call fails, without having to deal with parsing Request body. Callbacks can return

Pre-routing hook called before the routing decision. When `session_affinity` is enabled and a session_id is resolvable on the reques

Returns an Exception as ProxyException, this ensures all exceptions are OpenAI API compatible

Build a human-readable error string from a non-zero `claude` CLI run. The CLI is annoying to debug because the most useful failure signal rar

Return the proxy config (base URL + master key), or hard-fail the test. ``env`` is injected for tests; production callers pass nothing and

Builds `StandardLoggingGuardrailInformation` and adds it to the request metadata so it can be used for logging to DataDog, Langfuse, etc.

Create a new record.

Includes router model_group_alias'es as well if team_id specified, returns matching team-specific models

Invoke `run_claude` for every `models[i]` concurrently and collect outcomes. Each `claude` CLI invocation is a long-lived subprocess that spends

Converts given messages from OpenAI format to Gemini format - Parts must be iterable - Roles must alternate b/w 'user' and 'model' (same

Get encoding, loading it lazily if needed.

Dispatch to the configured Bedrock guardrail API. ``checks`` selects the resource-less, detect-only InvokeGuardrailChecks API; otherw

Generate an API key based on the provided data. Docs: https://docs.litellm.ai/docs/proxy/virtual_keys Parameters: - duration: Optio

Allow users to create a new team. Apply user permissions to their team. 👉 [Detailed Doc on setting team budgets](https://docs.litellm.ai/doc

Common checks across jwt + key-based auth. 1. If team is blocked 1.1. If project is blocked 2. If team can call model 2.2 If pro

Maps an LLM Provider Exception to OpenAI Exception Format