claude-wire

Appearance

Events ​

All events are part of the TRelayEvent discriminated union. Switch on event.type to handle each one.

ts
type TRelayEvent =
  | TTextEvent
  | TThinkingEvent
  | TToolUseEvent
  | TToolResultEvent
  | TSessionMetaEvent
  | TTurnCompleteEvent
  | TErrorEvent
  | TStructuredOutputEvent;

TTextEvent ​

Text content from the assistant.

ts
{
  type: "text",
  content: "Here's the answer..."
}

TThinkingEvent ​

Internal reasoning (extended thinking / chain of thought).

ts
{
  type: "thinking",
  content: "Let me analyze the code..."
}

TToolUseEvent ​

Claude wants to use a tool. If you have a tool handler configured, it will be called automatically.

ts
{
  type: "tool_use",
  toolUseId: "toolu_abc123",
  toolName: "Read",
  input: { file_path: "main.ts" }
}

The input field is unknown -- it is the parsed JSON object from the wire protocol, passed through as-is. Use JSON.stringify(event.input) if you need the string form.

TToolResultEvent ​

Result of a tool execution, emitted from user messages in the wire protocol.

ts
{
  type: "tool_result",
  toolUseId: "toolu_abc123",
  output: "const x = 1;\n",
  isError: false
}

TSessionMetaEvent ​

Emitted once at the start of a session with metadata.

ts
{
  type: "session_meta",
  sessionId: "sess-abc123",
  model: "claude-sonnet-4-6",
  tools: ["Read", "Write", "Edit", "Bash", "Glob", "Grep"]
}

Session mode: re-emitted on respawn

In session mode, session_meta is re-emitted after each process respawn -- for example, on auto-retry after a transient failure or after a turn-limit recycle. It is not emitted exactly once per logical session; treat each emission as the start of a new underlying process within the same session.

TTurnCompleteEvent ​

Emitted at the end of each turn with cost, token, and timing data.

ts
{
  type: "turn_complete",
  sessionId: "sess-abc123",
  costUsd: 0.018,
  inputTokens: 3500,          // total input (base + cache read + cache creation)
  outputTokens: 120,
  cacheReadTokens: 3000,      // tokens read from prompt cache (~10% billing rate)
  cacheCreationTokens: 0,     // tokens written to prompt cache (~125% billing rate)
  contextWindow: 200000,
  durationMs: 8500
}

All fields except type are optional -- they may be absent on legacy system/result events or when the CLI omits modelUsage.

TErrorEvent ​

An error from the session.

ts
{
  type: "error",
  message: "Something went wrong",
  sessionId: "sess-abc123"
}

TStructuredOutputEvent ​

Emitted once per turn when the CLI was launched with --json-schema and produced a constrained value. The translator surfaces whichever channel fires first: a synthetic StructuredOutput tool_use block (preferred), or the terminal result event's structured_output field (fallback). Dedup ensures a single event per turn.

ts
{
  type: "structured_output",
  value: { greeting: "hello" }   // parsed JSON object, not a string
}

buildResult exposes the value on TAskResult.structuredOutput for one-shot consumers; streaming consumers can read it from this event directly. claude.askJson and session.askJson read raw.structuredOutput preferentially over raw.text so Stop-hook nag messages or other commentary in the same turn don't corrupt validation. See JSON: Standard Schema route for caller-facing usage.