Errors

Thirteen codes. Each one has fixed copy in the widget.

Failures arrive as AgentError from actionwire, with a code field, or as a ToolResult with ok: false and a code.

import { AgentError } from 'actionwire';

try {
  await assistant.send('delete everything');
} catch (error) {
  if (error instanceof AgentError && error.code === 'BUSY') {
    // A turn is already running.
  }
}

Codes

Code	Shown to the person	Cause
UNSUPPORTED_WEBMCP	This browser does not support WebMCP.	document.modelContext is absent, incomplete, or not a secure context.
DISCOVERY_FAILED	The assistant could not read the tools on this page.	getTools() threw, or returned something that is not a list.
INVALID_SCHEMA	A tool definition is not valid.	A schema is not an object, or its dialect is unsupported.
TOOL_UNAVAILABLE	This tool is not available.	The chosen tool is not in the current list.
STALE_TOOLS	The tool list changed. Try again.	The revision moved after planning or after the prompt appeared.
INVALID_ARGUMENTS	The tool input is not valid.	Model arguments failed schema validation.
EXECUTION_FAILED	The tool failed.	The handler threw, or returned a result that is not JSON.
CONFIRMATION_DENIED	The action was cancelled.	A person selected Cancel, or pressed Escape.
ABORTED	The assistant was cancelled.	cancel(), clear(), or dispose() ran.
TIMEOUT	The assistant timed out.	The model call or a tool call passed timeoutMs.
MODEL_ERROR	The assistant could not complete the turn.	The model endpoint failed or returned an unusable body.
TURN_LIMIT	The assistant reached the turn limit.	The model asked for more than maxRounds rounds.
BUSY	The assistant is busy.	send() ran while a turn was already going.

Which ones end the turn

BUSY rejects the send() promise and changes nothing. CONFIRMATION_DENIED, TOOL_UNAVAILABLE, INVALID_ARGUMENTS, EXECUTION_FAILED, and STALE_TOOLS become a failed tool result: the model sees them and can try something else in the same turn.

The rest end the turn and land in state.error.

Tool results never throw

A tool that fails returns a ToolResult with ok: false, a code, and a text the model can read. The turn continues. This is intentional, so a model can recover from one bad call without losing the conversation.

Result size

A tool result is capped at 8192 characters of text. Longer text is cut. A result whose JSON exceeds that size keeps its text summary but drops the structured data field.

Read moreTroubleshooting