AI usage by day, source, and work driver.

• exact
  Codex — sum of observed last_token_usage.total_tokens from ~/.codex/sessions/**/rollout-*.jsonl, suppressing repeated cumulative snapshots and identical events copied into forked rollouts. This is processed-token usage; historic cache and output composition may be unavailable.
• exact
  Claude Code — sum of input + cache_creation + cache_read + output once per model response from ~/.claude/projects/**/*.jsonl, identified by message ID with UUID fallback. Repeated content blocks count once. CC calls = unique model responses with usage.
• exact
  Perplexity Sonar — API-reported prompt + completion tokens per live-search call, from the local sonar-search CLI’s usage log. Powers weekly market-signal diffs and ad-hoc research.
• exact
  OpenRouter — provider-reported prompt + completion tokens by model. Taken from OpenRouter’s own daily activity report for the account when that is available (every call, so scripts and evaluations count, not only the CLI); days are UTC and appear the day after. Otherwise, and for older days, from the local openrouter CLI’s usage log, which holds only what that command called. Calls to Jev made through OpenRouter are counted under Jev, not here.
• exact
  Jev — provider-reported input + output tokens per call to Jev, the semantic mail-state pre-check. From the Jev adapters’ metadata-only usage log (no message content) and, where OpenRouter’s account activity report is available, from that report too, since Jev is called through OpenRouter and ad-hoc calls appear there; the larger figure for each day is used. Counts start from the first logged call, plus one evaluation run recovered from its result files.
• est
  ChatGPT — timestamped visible assistant responses from the ChatGPT Mac app export, with visible user + assistant text estimated at characters ÷ 4. This excludes attachments, system prompts, hidden reasoning, tool traffic, and server-side context; it is not provider-reported usage.
• est
  Local · MLX — prompt tokens measured from the mlx_lm.server log (progress: x/Y ⇒ Y prompt tokens per request). Output tokens are not logged, so this is a conservative prompt-token floor; no output multiplier is assumed.
• est
  Local · Ollama — ≈ message characters ÷ 4 from the Ollama app database; Ollama does not record token counts. Calls made through the Ollama API (scripts, CLI) are counted separately from the server access log and shown as calls only, never as tokens.

The charts read scrubbed daily totals from data/daily-burn.json (committed). Day-detail (commit counts, project names) is read server-side from the git-ignored private/ file via /api/detail — never bundled into the page or committed, so a deployed build would show none of it. Re-run npm run extract to add a day.