Session Report

#!/usr/bin/env node
/* eslint-disable */
/**
 * analyze-sessions.js
 *
 * Scans ~/.claude/projects/**.jsonl transcript files and reports token usage,
 * message counts, runtime, cache breaks, subagent and skill activity.
 *
 * Output is human-readable text by default; pass --json for machine-readable.
 *
 * Usage:
 *   node scripts/analyze-sessions.js [--dir <projects-dir>] [--json] [--since <ISO|7d|24h>] [--top N]
 *
 * Notes on JSONL structure (discovered empirically):
 *  - One API response is split into MULTIPLE `type:"assistant"` entries (one per
 *    content block). They share the same `requestId` / `message.id`, and only the
 *    LAST one carries the final `output_tokens`. We dedupe by requestId and keep
 *    the max output_tokens to avoid 3-10x overcounting.
 *  - `type:"user"` entries include tool_result messages, interrupt markers,
 *    compact summaries and meta-injected text. A "human" message is one where
 *    isSidechain/isMeta/isCompactSummary are falsy and the content is a plain
 *    string (or text block) that isn't a tool_result or interrupt marker.
 *  - Subagent transcripts live in <project>/<sessionId>/subagents/*.jsonl with a
 *    sibling *.meta.json containing {agentType}. When meta is absent we fall back
 *    to the filename label (`agent-a<label>-<hash>.jsonl` → label) or "fork".
 *  - Resumed sessions can re-serialize prior entries into a new file; we dedupe
 *    globally by entry `uuid` so replayed history isn't double-counted.
 */

import fs from 'fs'
import os from 'os'
import path from 'path'
import readline from 'readline'

// ---------------------------------------------------------------------------
// CLI args
// ---------------------------------------------------------------------------
const argv = process.argv.slice(2)
function flag(name, dflt) {
  const i = argv.indexOf(name)
  if (i === -1) return dflt
  const v = argv[i + 1]
  return v === undefined || v.startsWith('--') ? true : v
}
const ROOT = flag('--dir', path.join(os.homedir(), '.claude', 'projects'))
const AS_JSON = argv.includes('--json')
const TOP_N = parseInt(flag('--top', '15'), 10)
const SINCE = parseSince(flag('--since', null))
const CACHE_BREAK_THRESHOLD = parseInt(flag('--cache-break', '100000'), 10)
const IDLE_GAP_MS = 5 * 60 * 1000 // gaps >5min don't count toward "active" time

function parseSince(s) {
  if (!s) return null
  const m = /^(\d+)([dh])$/.exec(s)
  if (m) {
    const ms = m[2] === 'd' ? 86400000 : 3600000
    return new Date(Date.now() - parseInt(m[1], 10) * ms)
  }
  const d = new Date(s)
  return isNaN(d) ? null : d
}

// ---------------------------------------------------------------------------
// Stats container
// ---------------------------------------------------------------------------
function newStats() {
  return {
    sessions: new Set(),
    apiCalls: 0,
    inputUncached: 0, // usage.input_tokens
    inputCacheCreate: 0, // usage.cache_creation_input_tokens
    inputCacheRead: 0, // usage.cache_read_input_tokens
    outputTokens: 0,
    humanMessages: 0,
    wallClockMs: 0,
    activeMs: 0,
    cacheBreaks: [], // [{ts, session, project, uncached, total}]
    subagentCalls: 0,
    subagentTokens: 0, // total (in+out) inside subagent transcripts
    skillInvocations: {}, // name -> count
    firstTs: null,
    lastTs: null,
  }
}

function addUsage(s, u) {
  s.apiCalls++
  s.inputUncached += u.input_tokens || 0
  s.inputCacheCreate += u.cache_creation_input_tokens || 0
  s.inputCacheRead += u.cache_read_input_tokens || 0
  s.outputTokens += u.output_tokens || 0
}

// ---------------------------------------------------------------------------
// File discovery
// ---------------------------------------------------------------------------
function* walk(dir) {
  let ents
  try {
    ents = fs.readdirSync(dir, { withFileTypes: true })
  } catch {
    return
  }
  for (const e of ents) {
    const p = path.join(dir, e.name)
    if (e.isDirectory()) yield* walk(p)
    else if (e.isFile() &