Summary

/tokenwise:summary — Multi-session trend report

Aggregate .tokenwise/log.ndjson over a time window.

Parse $ARGUMENTS

• --week (default) — last 7 days
• --month — last 30 days
• --all — entire log
• --days <N> — last N days
• --out <path> — write the report to a markdown file instead of stdout
• --json — output JSON instead of text

Steps

1. Read ./.tokenwise/log.ndjson (or the path the user provides)

2. Filter entries to the time window:

   • Compute cutoff: now - <days>*86400
   • Keep entries where ts > cutoff

3. Aggregate:

   • Total: sessions (unique session_id), tasks, total cost, baseline cost, savings
   • Per model: task count, cost, % of total
   • Per task_class (mechanical|execution|review|planning): task count, avg cost, dominant model
   • Trend (only for --week or --days <≤14>): per-day cost + savings bar chart in plain text

4. Print:

TokenWise Summary — last <N> days
==================================

Sessions:        <count>
Tasks routed:    <count>
Total spent:     $<total>
Baseline:        $<baseline>
Savings:         $<savings> (<pct>%)

Per model:
  Haiku    <count> tasks  $<cost>  (<pct>%)
  Sonnet   <count> tasks  $<cost>  (<pct>%)
  Opus     <count> tasks  $<cost>  (<pct>%)
  Fable    <count> tasks  $<cost>  (<pct>%)

Top task classes:
  <class>          <count> tasks   avg cost $<avg>   model: <dominant>
  ...

Daily trend (cost):
  Mon  ███████░░░ $4.21
  Tue  ███░░░░░░░ $1.82
  Wed  ████████░░ $5.04
  ...

5. If --out <path> was provided, write the markdown version of the report to that path. Use a proper markdown table for the trend section.

6. If --json, dump the aggregated data structure as pretty-printed JSON.

Notes on small log files

• If <2 sessions in the window: print the report but add Note: too few sessions for meaningful trend data.
• If log file is empty/missing: print the same "No TokenWise log found" message that /tokenwise:report uses.
• Omit any "Per model" row whose count is 0 for the window — Fable's row will legitimately be absent for most users most of the time.

Tools

Read, Bash (for jq aggregation if helpful), Write (for --out).