Build stable span attributes for failed tool-call diagnostics.

Compute which fields differ between tracks for a given case. Compares each track against the first track (baseline). Args: case

Create a new MCP gateway on Arcade Cloud. Args: access_token: OAuth access token for the Engine API. name: Human-readable gateway

Create a Pydantic model from a TypedDict class. When strict=True, the generated model (and all nested TypedDict models) forbid unknown k

Create a tar.gz archive of the package directory. Args: package_dir: Path to the package directory to archive Returns:

Keep only the root span and its direct children (depth <= 1).

Find and load as Toolkits all installed packages in the current Python interpreter's environment that have a registered entry

Find the Python interpreter in the virtual environment. NOTE: This function assumes it is called from the project root directory (where .ven

Spawn a detached process that checks PyPI and writes the cache.

Create a failed result from an exception.

Get the current model context if available.

Get distinct_id based on authentication state. We use principal_id for authenticated users and anon_id for anonymous users. Returns:

Get the progress token for a task, or None.

Get the redirect URI for this server.

Get a tool from the catalog by name. Args: name: The name of the tool, potentially including the toolkit name separated by the `s

Build an HTTP/SSE config entry pointing at a local ``arcade mcp http`` server. The server must be started separately, e.g.:: arcade mcp

Get all registered track names. Returns: List of track names in registration order.

Reorganize evaluation results for cross-model comparison. Groups results by suite -> case -> model, enabling side-by-side tables. Args:

Return True if an error was stored for this task via ``set_error``. Callers should use this to disambiguate error results from successful

Check if a task is in a terminal state (without lock).

Create comparative results simulating 3 different MCP servers for Linear. This represents a realistic use case: comparing how well an LLM works w

Apply schema defaults to arguments. Fills in default values from the tool schema for: - Missing parameters (key not in args)

Perform the complete OAuth login flow. This function: 1. Fetches OAuth config from the Coordinator 2. Starts a local callback server

Interactively prompt the user to select toolkits. Returns a list of selected toolkit names.

Send a notifications/elicitation/complete notification to this session's client. This MUST be sent only to the client that initiated the elic

Set client initialization parameters.

Determine whether a background PyPI check is needed.

Shut down the callback server.

Converts a snake_case name to PascalCase.

Convert ReadableSpan objects to OTLP JSON (ExportTraceServiceRequest).

Stop collecting and return the collected spans.

Convert to dictionary for JSON serialization.

Validates the configuration, user, and returns the Config object.

Wait for the OAuth callback to complete.

Add NoneCritics for any fields in the expected tool calls that are not already in the critics list. Args: expected_tool_

Assert that ``updated`` equals ``original`` except for a single new or replaced entry at ``updated[parent_key][server_name]``. Every other key

Convert Arcade ValueSchema to JSON Schema format.

Convert Arcade ValueSchema to JSON Schema format for Anthropic.

Factory method to create EvalCase instances. Used by the comparative mixin to create EvalCase without circular imports.

Return paths in order, removing duplicates (case-insensitive on Windows).

Send all queued notifications.

Format conversation messages as rich Markdown.

Format evaluation details as markdown.

Return HTML header with embedded CSS for styling.

Get all secrets from the engine.

Get status string from evaluation.

Return known Windows Cursor config locations (primary first).

Create a ResourceOwner for test isolation checks.

Create a MaterializedTool and MCP tool definition for a function returning the given TypedDict.

Parse incoming data into a typed MCPMessage. Accepts a raw JSON string, already-parsed dict, or an existing MCPMessage.

Return cached toolkit map if the cache file exists, is fresh, and matches the active context.

Resolve a gateway name or slug to the actual slug. The user may pass a name (``opencode``) or a slug (``pascal_opencode``). We look up existi

Run a single capture task with error handling. Returns CaptureTaskResult with success/failure info instead of raising.

Run a single evaluation task with error handling. Returns EvalTaskResult with success/failure info instead of raising.

Sanitize-and-omit variant of ``_validate_rfc6750_quoted_value``. Returns ``value`` when it conforms to the RFC 6750 ``error-char`` grammar. R

Serialize critic results for detailed output.

Convert a URI template to a compiled regex with named groups. ``{param}`` -> ``(?P<param>[^/]+)`` ``{param*}`` -> ``(?P<param>.+)`` (wildca

Replace template variables with dummy values to produce a concrete sample URI. ``{param}`` -> ``__param__`` ``{param*}`` -> ``__param__/nest

Validate the critics. Args: critics: The list of critics. name: The name of the evaluation case. Ra

Validate that the name follows the required pattern: - Alphanumeric characters and underscores only - Must end with alphanume

Validate a canonical URL or PRM resource_metadata URL. Four layers of enforcement applied in order: 1. RFC 6750 ``error-char`` (no DQUOTE /

Validate a single scope token against the RFC 6750 grammar. The grammar is ``scope-token = 1*( %x21 / %x23-5B / %x5D-7E )`` — at least one pr

Build the ``WWW-Authenticate`` value for a 403 insufficient_scope. Order: ``resource_metadata`` (if provided), ``error="insufficient_scope"``,

Remove expired tasks based on TTL. Expiry = createdAt + ttl. Tasks with ttl=None never expire.

Deploy the server to Arcade Engine. Args: engine_url: The base URL of the Arcade Engine deployment_request: The deployment r

Extract properties from TypedDict, Pydantic models, or other structured types. Returns (properties, required_keys). For a known structured t

Load a Toolkit from an entrypoint. The entrypoint value is used as the toolkit name, while the package name is extracted fro

Get models, using default if none specified.

Extract required secrets from the /worker/tools endpoint. Args: base_url: Base URL of the server server_name: Name of the se

Get the org to use: default if available, otherwise first in list.

Get all registered track names. Returns: List of track names in registration order.

Mapping between Python types and HTTP/JSON types

Get the wire type information for a given type.

Group capture results by suite, case, then track for tab-based display. Args: captures: List of CaptureResult objects. Returns:

Return True if a successful result was stored for this task. Used by the background-task cancellation handler to avoid overwriting a

List existing MCP gateways from the user's project. Returns a list of gateway dicts (each with ``id``, ``slug``, ``name``, ``tool_filter``, e

Create empty evaluation results.

Resolve tool name and apply schema defaults in one step. This combines name resolution (for Anthropic underscore -> dot conversion) w

Read lines until we get a JSON-RPC message with the expected id.