Session Analytics

What the agent actually did in a session: tokens, subagents, context pressure.

A desktop app over Claude Code's own session logs. It shows what happened inside one session: tokens and cache use, subagents, context growth, plan usage, and a turn-by-turn replay.

01 · Scale
1,800+
Claude Code sessions analyzed, as of July 2026
02 · Session list
16.2 s to 349 ms
session list, median load time after the rebuild
03 · Idle CPU
39.4% to 4.6%
idle CPU over 90 seconds, after the same rebuild
01 · Problem

Gigabytes of session logs that nobody reads

Claude Code writes every session to JSONL on disk. Mine grew to 2.4 GB. The built-in usage stats are totals, so when one run was slow or expensive, the reason sat somewhere in those files, unread.

02 · What it does

One session, opened up

  1. A1: Per-session tabs

    Overview, tokens and cache use, subagents, context-window growth turn by turn, plan usage, and a replay of the session.

  2. A2: Across sessions

    A dashboard, a journey view of how usage changed over time, and search across sessions.

  3. A3: Jump from elsewhere

    Other tools open a session directly by id, so a run report or the Cockpit can link straight to the session behind it.

03 · How it's built

A light index first, the heavy detail on demand

A Rust scanner walks the log folders, caches what it has parsed, and builds a light index of sessions; the full detail of one session loads only when I open it. That split came from a rebuild: the first version loaded everything, and on real volume the session list took 16.2 s. After the rebuild it takes 349 ms, and idle CPU dropped from 39.4% to 4.6%. One size target was missed, and the notes say so.

The code is written by AI agents under my specification, test contracts and review of the result.

04 · Key numbers

Measured, with dates

Sessions analyzed
1,800+

Claude Code sessions, as of July 2026.

Session list
46×

Faster after the rebuild (16.2 s to 349 ms).

05 · Screens

The context tab, on synthetic data

fig.1 · session analytics · context real capture · synthetic fixture
Session Analytics on a synthetic fixture: a list of eight invented sessions across three projects, and one session open on its Context tab with end context, growth, cache hit rate and turn count, a chart of context-window growth per turn, and a per-turn detail table. Click to zoom
FIG. 1 · A real capture of the app on an invented fixture: no session, path or project of mine appears in it.
06 · Status and next

Paused, on purpose

I set it aside in September 2026 to put my time into TarkOS. It works, but it isn’t in daily use, and I’d rather say that than dress it up. The plan is to restore and polish it, then publish it.

Talk to me about this →