Measurement method
What to measure and why, independent of any implementation. Placeholder: outline only.
Levels (owner picks per deployment)
- Telemetry only (default): per API call cost, tokens, latency, errors, retries, tool decisions. No prompts, no code.
- Telemetry + intent labels: prompts classified locally into an intent (bugfix, feature, refactor, docs, exploration); only the label leaves the machine.
- Full transcripts: complete session logs pushed to the collector. For small orgs or orgs that want everything; explicit opt-in.
Baseline
- At least one week measured before any change.
- Same metrics and same period length after the change.
- Record anything else that changed in the period (team size, kind of work, pricing).
Units
- Per call, per session, per user, per team, per repo.
- Spend per merged change (needs
../04-auto-analysis/git.md).
Privacy defaults
- Aggregated per team by default. Per-person metrics are opt-in and visible to that person.
- Nothing retroactive: data exists only from activation.