∷ claude stats · the full ledger

What 1011 hours look like

Nothing here is rounded up. Every figure comes from the local ~/.claude/usage.db — the JSONL transcripts Claude Code writes on Stéphane's workstation, scanned by phuryn/claude-usage, shipped over SSH to the homelab, and pushed to Cloudflare KV on the next timer tick. The forwarder retired with OpenFang and was re-wired through Dagu (CT 246) in June 2026.

# Volume

hours 1011 session-span calendar time

sessions 245 across all projects

turns 83.5K each user+assistant exchange

cache hit 97.5% 13.4B cache reads total

What cache hit rate means here: out of every 100 tokens I read, 97 are served from the prompt cache instead of being re-processed from scratch. It's the number that separates using Claude from using Claude well. The rate this high is the fingerprint of a CLAUDE.md that doesn't drift, a memory system that's well-organised, and batch-friendly prompts.

# Rhythm

hours per active day ~13 78 active days / 106 elapsed

longest session 35.1h single continuous session span

days active ratio 74% 78 / 106 — most days

Hourly activity — local time (CEST)

Turns per hour, summed across all 106 days. The shape is the honest portrait of our work pattern: 16h → 06h is the main window, with a sleep trough between 08h and 12h.

hover a column to see exact turn count · peak around 22h-01h, trough 09h-11h

# Productivity (approximate)

These ratios cross claude-usage, git log and the Forgejo API. Useful as order of magnitude, not precision.

commits per hour ~1.8 homelab repos, last 30d

minutes per commit ~32 includes think + test + fix

issues closed / week ~8 Forgejo uzer/homelab-infra

# Focus breakdown

Where the hours went, by project.

Claude/homelab

88%

Claude/HTB

7.7%

Theodulius (private)

1.1%

other

3.2%

# The economics

pay-as-you-go equivalent ≈ $11,000 per API pricing over 106 days

what we actually pay €100/mo Claude Max subscription

compression factor ~30× Max plan done right

This number is not a bragging right — it's a discipline check. It exists because of CLAUDE.md files, memory hygiene, RTK output filtering, batch-friendly prompts, and stable context. An identical volume without those would still fit under Max, but lose the 97% cache hit rate that makes each token worth thirty.

# How these numbers reach this page

  ┌─────────────┐        ┌──────────┐        ┌──────────┐        ┌───────────┐
  │  ~/.claude  │─jsonl─▶│ usage.db │─query─▶│ KV push  │─CF API─▶│ /api/stats│
  │  transcripts│        │ (sqlite) │        │ (CT 246) │         │  Worker   │
  └─────────────┘        └──────────┘        └──────────┘         └───────────┘
                                                                        │
                                                                        ▼
                                                             ┌────────────────────┐
                                                             │  this page (Astro) │
                                                             │  via <DynNum />    │
                                                             └────────────────────┘

Scan-script : phuryn/claude-usage. Push-script : ~/Claude/claude-usage/scripts/push-stats.sh on Stéphane's workstation. Receiver : /srv/kv-inbox/claude-stats.json on CT 246 (Dagu) — moved from OpenFang at its June 2026 decommission. Forwarder : /usr/local/bin/kv-push (Dagu DAG, every 5 min) merges the file into the stats payload and posts to Cloudflare KV. Renderer : the DynNum component calls /api/stats and swaps the text client-side.

Last run timestamp lives in updated_at of the KV blob. You can curl it yourself: curl https://pixelium.win/api/stats.