Format single-model capture results.

Format single-model capture results as HTML.

Format single-model capture results.

Get the Arcade API base URL, checking env var at runtime.

Return complete HTML document for capture results.

Get the status of a deployment. Args: engine_url: The base URL of the Arcade Engine server_name: The name of the server to g

Get JSON-RPC error code for exception.

Extract header from JWT without verification. Args: token: JWT token string Returns: Header dictionary conta

Get org_id and project_id from config if available.

Extract URL and method from the __cause__ exception.

Get the class name of the resource server validator. Args: resource_server_validator: The resource server validator instance or N

Handle APIError and its subclasses.

Handle tasks/cancel -- cancel a task (flat shape).

Convert exception to appropriate error response.

Handle tasks/get -- returns current task state immediately (flat shape).

Handle tasks/result -- blocks until task terminal, returns underlying result.

Handle initialize request.

Handle tasks/list -- returns tasks scoped to authorization context.

Handle list tools request.

Handle non-HTTP Google API errors.

Handle ping request.

Handle TransportQueryError (GraphQL errors in response body).

Handle read resource request.

Process request in stateful mode - maintain session state.

Process request in stateless mode - new transport per request.

Handle unsupported HTTP methods.

Convert hex trace/span IDs to base64 for protobuf ``ParseDict``.

Initialize Arcade client for runtime authorization.

Initialize middleware chain.

Build a JSON-RPC initialize request.

Load disabled toolkits from the environment variable. The ARCADE_DISABLED_TOOLKITS environment variable should contain a comma-separa

Load disabled tools from the environment variable. The ARCADE_DISABLED_TOOLS environment variable should contain a comma-separated li

Import gql transport exceptions lazily and cache the result.

Load org/project context from the shared Arcade config (same source as the CLI). Returns (org_id, project_id) when both are available

Log error with timing.

Log incoming request.

Log response with timing.

Create a cache key from URL and headers.

Create a failing multi-run case with failure_reason in run details.

Map Microsoft Graph APIError to appropriate ToolRuntimeError.

Called by the companion HTTP thread when the user clicks a button.

Read a resource (for Context.read_resource).

Monitor deployment with live status and streaming logs display. Args: engine_url: The base URL of the Arcade Engine server_n

Normalize critic weights when FuzzyWeight is used. Filters out placeholder critics (like NoneCritic, which always has weight=0) and norm

Normalize additional_messages by parsing JSON strings into proper objects. OpenAI returns: - Tool call arguments as JSON str

Return (key, value) if the line looks like KEY=VALUE, else None. Handles quotes and escaped chars via shlex.

Parse raw message dict into typed message based on method.

Extract retry-after from Google API errors. Returns milliseconds to wait before retry. Defaults to 1000ms if not found.

Poll deployment status until it's running or error.

Extract server spans from response ``_meta`` and ingest into collectors.

Attach tool error details to the active telemetry span when present.

Register method handlers.

Replay events after the specified event ID.

Resolve the best interpreter for detached usage tracking.

Best-effort resolution of the PRM (Protected Resource Metadata) URL. Inspects session init_options for a canonical_url set by the transport,

Choose stdout/stderr targets for the temporary validation server process. ``arcade deploy`` starts a short-lived child server to validate the

Returns (evaluation_result, actual_calls, failure_detail). failure_detail is non-None when execution failed before evaluate() could run.

Run the lifespan context manager.

Run evaluation using OpenAI client. Args: client: The OpenAI client. model: The model name. case: The eva

Run evaluation suite with Anthropic client. Returns: For regular evaluations: A single result dict. For comparative evaluations:

Strip query params and fragments from URI for privacy.

Strip query params and fragments from URI for privacy.

Strip query params and fragments from URI for privacy.

Mock docs lookup — returns the canned paragraph most similar to the query.

Pre-populate with stable demo data so evals have predictable IDs.

Select the user_id for the tool's context. User ID selection priority: - Authenticated user from front-door auth - Configured

Serialize a single evaluation case.

Serialize the input to a tool function.

Serialize the output of a tool function.

Restrict a file so only the current user can read/write it on Windows. On POSIX systems ``chmod 600`` removes group/other access. On Windows,

Start server components (called by MCPComponent.start).

Start the companion webapp on a background thread.