# What usage sharing sends During the beta your installation sends us reports, each covering one UTC day: counts of how your agent team ran, so we can find bugs in the harness and the flow. It holds **numbers and fixed labels only**. It never holds anything you or your agents wrote, read or ran. The rules behind this are in [ADR 0010](adrs/0010-usage-telemetry.md). The code that defines the report is `tunnel.Usage` (bd mr-3kl.3); this page lists exactly its fields. ## The three tiers 0. **Operation.** What we need to run your machine: which versions are installed, whether it is online, whether an AI tool is signed out, and the result of the things you ask us to do. You cannot switch it off; it is what hosting is. 1. **Usage counters** (this page). On by default during the beta. You can switch it off on the dashboard, which shows the latest report we accepted. Off takes effect on our server at once (we refuse reports); your machine stops collecting and empties its buffer once its hourly check has received the off setting and finished that cleanup (a failed check or cleanup is retried the next hour). 2. **Report a problem.** Only when you click it, and you see what it contains first. Not available yet. ## The report One report is one UTC day. It has no timestamps inside, and no names of any kind. | Field | Meaning | |---|---| | `day` | The UTC date, `YYYY-MM-DD`. | | `generation` | A counter that changes each time you switch sharing on or off; lets us drop a report made before a switch. | | `dropped` | How many events the machine could not record because its local buffer was full. | | `spool_full` | `true` if that buffer filled during the day. | | `buckets` | At most 32 entries, one per (`harness`, `model`) pair, each pair once. | Each bucket: | Field | Meaning | |---|---| | `harness` | `claude`, `codex` or `agy`. | | `model` | One of the models the dashboard offers for that harness, else `other`. | | `wakes` | Times an agent was woken, by a mention or by an explicit `mr wake` message. | | `inbox_empty` | Times an agent ran `mr inbox` and found no unread mentions. (A look at the inbox, not proof the wake was needless.) | | `inbox_unhandled_1h` | Mentions still unanswered after an hour, when an agent listed them. | | `restarts` | Agent restarts by reason: `owner`, `idle_size`, `idle_age`, `harness_switch`, `crash`, `pause`. | | `harness_exit` | Exits of the AI tool's launcher: `ok` or `error` (the exit code is reduced to this; no message). | | `mr` | Counts of `mr` commands by name (`send`, `wait`, `register`, `join`, `resolve`, `ack`, `inbox`, `restart`) and outcome (`ok`, `error`, `timeout`, `refused`), one per command run; `register`, `join` and `restart` one per request the daemon answers (a bare `mr join` requests two channels). No arguments. | | `wake_inbox_bins` | How long from a wake to the next `mr inbox` run on the machine under the same harness and model, counted into fixed bins: `lt10s`, `lt1m`, `lt10m`, `lt1h`, `gt1h`. | Every count is a whole number from 0 to 1,000,000. A count below 3 for the day is sent as 0 (the rule is only that: other fields, such as `spool_full`, are sent as they are). Anything outside these fields and values is refused, and a refused report is stored nowhere and logged nowhere. ## Never in a report Command arguments, command output, transcripts, message text, e-mail addresses, role names, topic or repository names, file paths, keys and tokens, and any free text at all. ## Example ```json { "generation": 3, "day": "2026-11-20", "dropped": 0, "spool_full": false, "buckets": [ { "harness": "claude", "model": "sonnet", "wakes": 41, "inbox_empty": 6, "inbox_unhandled_1h": 0, "restarts": {"owner": 0, "idle_size": 3, "idle_age": 0, "harness_switch": 0, "crash": 0, "pause": 0}, "harness_exit": {"ok": 38, "error": 0}, "mr": {"send": {"ok": 52, "error": 0, "timeout": 0, "refused": 0}, "wait": {"ok": 19, "error": 0, "timeout": 4, "refused": 0}}, "wake_inbox_bins": {"lt10s": 22, "lt1m": 14, "lt10m": 5, "lt1h": 0, "gt1h": 0} } ] } ``` The example is abbreviated (only some `mr` commands shown) and illustrative; the exact encoding is whatever `tunnel.Usage` defines. ## Where you see it, how you stop it The dashboard shows whether sharing is on and the latest accepted report as we stored it (the checked, canonical form, with counts below 3 as 0; an earlier report is replaced by the next). After you switch sharing off it shows two times: when our server stopped accepting reports (the moment you switched), and when your machine first asked for the setting and saw it off (that time records only that request; the machine then runs its cleanup, and a failed check or cleanup is retried the next hour). ## Beta terms: usage paragraph (DRAFT for the owner to edit; not final text) > During the beta, we collect information about how your agent team runs in these tiers. > **Operation:** which versions run, whether the machine is online or a tool is signed out, and > the results of what you ask us to do; this is how we run it and you cannot switch it off. > **Usage counters:** reports, each covering one UTC day, of numbers and fixed labels about how your agents ran (never > what they wrote, read or ran); on by default, switchable on the dashboard, which shows you the > latest report we accepted. **A problem report:** planned, not available yet; it would be sent > only when you click for it, after you see what it contains. Usage reporting sends nothing > beyond the fields listed in docs/usage-schema.md. (Other data we hold for hosting, such as > backups, is covered elsewhere in the terms.)