The connector hooks into Claude Code’s event system and captures structured data from each session. This page documents what events are captured and how they map to Infinium traces.

Hook Events

The connector listens for these Claude Code hook events:

Event	When It Fires	What’s Captured
SessionStart	Claude Code session begins	Session ID, model, working directory, source (startup/resume)
UserPromptSubmit	User sends a prompt	Prompt text
PreToolUse	Before a tool executes	Tool name, input parameters
PostToolUse	After a tool completes	Tool output, duration
PostToolUseFailure	Tool execution fails	Error message
Stop	Assistant finishes responding	Last assistant message, triggers trace send
SessionEnd	Session closes	Final cleanup, triggers trace send
SubagentStart	Subagent/cowork spawned	Subagent ID, type
SubagentStop	Subagent completes	Transcript path

How Events Become Traces

Events are grouped into turns — each turn is one prompt-response cycle:

Turn 1:
  UserPromptSubmit  → "Fix the login bug"
  PreToolUse        → Read(file="auth.py")
  PostToolUse       → [file contents]
  PreToolUse        → Edit(file="auth.py", ...)
  PostToolUse       → [edit applied]
  Stop              → "I've fixed the login bug by..."
                       ↓
                    Trace sent to Infinium

Turn 2:
  UserPromptSubmit  → "Now add tests"
  PreToolUse        → Write(file="test_auth.py", ...)
  PostToolUse       → [file created]
  Stop              → "I've added tests for..."
                       ↓
                    Trace sent to Infinium

Each Stop event triggers a trace send. On SessionEnd, a final trace is sent and the session file is cleaned up.

Trace Structure

Each trace sent to Infinium contains:

Top-Level Fields

Field	Source	Example
name	Auto-generated	"Claude Code session (3 turns)"
description	First user prompt	"Fix the login bug in auth.py"
duration	Session timing	45.2 (seconds)
input_summary	First prompt	"Fix the login bug"
output_summary	Last assistant message	"Tests added and passing"

Execution Steps

Each trace includes an ordered list of steps:

Step Type	Action	Fields
User prompt	user_prompt	description, input_preview
Tool call	tool_use:{name}	input_summary, output_summary, duration_ms
Tool error	tool_use:{name}	error.error_type, error.message, error.recoverable
Subagent	subagent:{type}	metadata.subagent_id, metadata.agent_type

Environment Context

Every trace includes environment metadata:

Tag	Value
model	Claude model used (e.g., claude-sonnet-4-5-20250514)
session_id	Claude Code session identifier
working_directory	Project directory path
permission_mode	User’s permission level
framework	"Claude Code"

Error Tracking

Tool failures are captured as structured errors:

Field	Description
error_type	"ToolFailure"
message	Error message from Claude Code
recoverable	true (tool errors are recoverable)

Data Truncation

To keep payloads manageable, the connector truncates:

• Prompt text: First 500 characters (configurable via CONNECTOR_MAX_INPUT_PREVIEW)
• Tool output: First 500 characters (configurable via CONNECTOR_MAX_OUTPUT_PREVIEW)

Full data is never sent — only summaries that give Maestro enough context to evaluate quality.