Generates structured session summaries and archives session state when a Claude Code session ends

session-end Hook

The session-end hook fires when a Claude Code session ends. It generates a structured summary from all observations captured during the session, captures conversation turns and tool call details from the transcript, marks the session as completed, and auto-generates/archives the CURRENT.md session state file.

Trigger Event

Fires once when the Claude Code session stops (user exits or session times out).

What It Does

Collects all observations for the current session from the memory database
Collects all user prompts captured by the user-prompt hook
Generates a structured summary including:

- Original request (first user prompt) - Discoveries and investigations - Decisions made - Work completed (features, bugfixes, refactors) - Failed attempts - Next steps (if work is incomplete) - Files created and modified - Verification results (VR pass/fail) - Plan progress tracking

Captures conversation data by parsing the JSONL transcript (incremental -- only new lines)
Marks session as completed in the database
Archives CURRENT.md and generates a fresh one

Example Input

json

{
  "session_id": "abc123-def456",
  "transcript_path": "/Users/dev/.claude/projects/my-app/transcript.jsonl",
  "cwd": "/Users/dev/my-app",
  "hook_event_name": "stop"
}

Summary Structure

The generated summary includes these fields:

Field	Source	Description
`request`	First user prompt	What the user asked for
`investigated`	Discovery observations	What was explored
`decisions`	Decision observations	Architectural choices made
`completed`	Feature/bugfix/refactor observations	Work finished
`failedAttempts`	Failed attempt observations	What did not work
`nextSteps`	Last 10% of observations	What to do next
`filesCreated`	File change observations	New files
`filesModified`	File change observations	Changed files
`verificationResults`	VR check observations	Pass/fail results
`planProgress`	Plan item observations	Plan completion tracking

Conversation Capture

The session-end hook also parses the JSONL transcript to capture:

Conversation turns: User prompts paired with assistant responses
Tool call details: Tool name, input summary, I/O sizes, success/failure, files touched
Token estimates: Approximate token counts per turn

This data powers the observability tools (massu_session_replay, massu_prompt_analysis, massu_tool_patterns).

Tips

The session-end hook uses incremental transcript parsing -- it tracks the last processed line to avoid re-processing
Assistant responses are capped at 10,000 characters to prevent database bloat
The hook runs best-effort -- it never blocks Claude Code from shutting down