Human Token Tracker

A repo enters the tracker when someone commits a registration file to the tracker’s telemetry branch. The PR sync processes those files first, verifies that the committer is the person named in the request, and creates the repo record. There is no auto-enrol; an already-registered repo is a no-op, so a file that cannot be deleted is retried harmlessly every night. After that, the team just works.

The recorder plugin posts a Claude Code session the moment it ends, and the row is stored before the request returns. The Mac mini job covers the other paths: it pulls recorder telemetry from the tracker branch and backfills the Mac mini’s own local transcripts, deduplicating token counts by message id. Both paths apply the same rules: the repo must be registered, the client’s own identities are excluded, the Claude token row is keyed by session so a retry is an update, and the hours rule recomputes each day the session touched.

For every tracked repo with a GitHub repo, the job fetches pull requests updated since that repo’s last sync and upserts them by GitHub id. The author is resolved to a person from the author block in the PR body first, then from the GitHub login; an unresolved author is stored unattributed and counted. Session token rows on the same branch are linked to the PR afterwards, and the run writes its own sync row with repos synced, PRs upserted and errors. The hosted job returns as skipped until its GitHub token exists; the launchd job on the office Mac mini runs the same handler in the meantime.

Each tracked repo may carry a committed effort-log file. Every entry becomes up to two token rows: Claude tokens keyed on the session id, and human hours as centihours keyed on the session id plus a suffix. Only entries by the client’s own identities are kept, because Edge8 contributors are already captured by telemetry and would otherwise be counted twice. Persisting is delete-then-insert by session id, so the same file ingested twice yields the same rows.

Per repo, three things may be written and each has a rule for when the model is allowed to run. The executive summary is generated once from the repo’s status page and after that only when someone explicitly regenerates it. The digest of the latest ten PRs regenerates only when the set of PR numbers and states changed since last night. The FAST goal is suggested when the repo has none yet or its status page changed, and a goal set by hand is never replaced. The whole sweep shares a budget of forty model calls as a circuit breaker; steady-state nights spend a handful.