Output Events

Claude Code emits one JSON object per line on stdout. Each has a top-level type field.

system Events

system/init — Session Start

Emitted once when the process starts.

{
  "type": "system",
  "subtype": "init",
  "session_id": "abc-123-def-456",
  "model": "claude-sonnet-4-6-20250514",
  "tools": ["Read", "Write", "Edit", "Bash", "Glob", "Grep", "Agent"]
}

Field	Type	Description
session_id	string	Claude Code's internal session ID
model	string	Model being used (e.g., claude-sonnet-4-6-20250514)
tools	string[]	Available tool names (when present)

system/result — Legacy Turn Complete

Older Claude Code versions emit this instead of a top-level result event.

{
  "type": "system",
  "subtype": "result",
  "session_id": "abc-123",
  "result": "\"Here is the summary...\"",
  "is_error": false
}

TIP

The result field is double-encoded. JSON.parse(event.result) gives you the actual string.

assistant Events

Emitted whenever the assistant's message state changes. In --verbose mode, this contains the full message snapshot — all content blocks accumulated so far.

{
  "type": "assistant",
  "message": {
    "role": "assistant",
    "content": [
      { "type": "thinking", "thinking": "Let me analyze this code..." },
      { "type": "text", "text": "Here's what I found:" },
      {
        "type": "tool_use",
        "id": "toolu_abc123",
        "name": "Read",
        "input": { "file_path": "/src/main.ts" }
      }
    ],
    "stop_reason": "tool_use"
  }
}

Content Block Types

text — Markdown Output

{ "type": "text", "text": "Here's the analysis..." }

thinking — Extended Thinking

{ "type": "thinking", "thinking": "Let me consider the architecture..." }

WARNING

Some Claude Code versions use "text" instead of "thinking" for the content field. Always check both:

const text = block.thinking ?? block.text ?? ''

tool_use — Tool Invocation

{
  "type": "tool_use",
  "id": "toolu_abc123",
  "name": "Read",
  "input": { "file_path": "/src/main.ts" }
}

The id field uniquely identifies this tool call — used for approval/denial and matching with tool_result.

tool_result — Tool Result (in assistant)

Can also appear in assistant events (less common). See user events below for the primary path.

{
  "type": "tool_result",
  "tool_use_id": "toolu_abc123",
  "content": "file contents here...",
  "is_error": false
}

Common Tool Names

Tool	Typical Input	Description
Read	{ "file_path": "..." }	Read file contents
Write	{ "file_path": "...", "content": "..." }	Create/overwrite file
Edit	{ "file_path": "...", "old_string": "...", "new_string": "..." }	String replacement
Bash	{ "command": "..." }	Execute shell command
Glob	{ "pattern": "..." }	Find files by pattern
Grep	{ "pattern": "...", "path": "..." }	Search file contents
Agent	{ "prompt": "...", "description": "..." }	Spawn sub-agent
AskUserQuestion	{ "question": "..." }	Ask user for input
WebSearch	{ "query": "..." }	Search the web
WebFetch	{ "url": "..." }	Fetch URL contents

user Events

Emitted when Claude Code processes a tool result. This is the primary path for tool results flowing through the conversation.

{
  "type": "user",
  "message": {
    "role": "user",
    "content": [
      {
        "type": "tool_result",
        "tool_use_id": "toolu_abc123",
        "content": "file contents...",
        "is_error": false
      }
    ]
  }
}

The content field is polymorphic — see Gotchas.

result Events

The primary turn-completion event in current Claude Code versions.

{
  "type": "result",
  "subtype": "success",
  "session_id": "abc-123",
  "result": "\"Task completed successfully.\"",
  "is_error": false,
  "total_cost_usd": 0.042,
  "duration_ms": 12500,
  "duration_api_ms": 8200,
  "num_turns": 3,
  "modelUsage": {
    "claude-sonnet-4-6-20250514": {
      "inputTokens": 1200,
      "outputTokens": 450,
      "cacheReadInputTokens": 8500,
      "cacheCreationInputTokens": 0,
      "contextWindow": 200000
    }
  }
}

Field	Type	Description
subtype	"success" | "error"	Whether the turn succeeded
total_cost_usd	number	Cumulative session cost in USD
duration_ms	number	Total wall time for this turn
duration_api_ms	number	Time spent in API calls
num_turns	number	Number of agentic loops in this interaction
modelUsage	Record<string, ModelUsage>	Per-model token breakdown

modelUsage Structure

Keyed by model ID. Each entry:

Field	Description
inputTokens	Fresh input tokens sent
outputTokens	Tokens generated
cacheReadInputTokens	Tokens served from prompt cache
cacheCreationInputTokens	Tokens written to prompt cache
contextWindow	Model's maximum context window

Total input tokens = inputTokens + cacheReadInputTokens + cacheCreationInputTokens

Both result paths

Both system/result (legacy) and top-level result can appear depending on Claude Code version. The translator handles both.

Ignored Event Types

Type	Purpose	Action
progress	Internal progress updates	Silently skipped
rate_limit_event	Rate limit notifications	Silently skipped