Coverage alerts: hook integrity, DLP posture, and silent devices

Enterprise admins can now see whether developer machines still have ReclaimLLM hooks and DLP, get alerted when that protection is removed or bypassed, and track org and team coverage percentage from a shared device check-in plane.

ReclaimLLM's controls live in local hooks. A developer can delete them, disable them, or install with --no-dlp. Until now that posture was mostly invisible to the org. Coverage closes that gap with device enrollment, integrity check-ins, deterministic findings (AD-010…AD-013), and a Coverage tab — detection without claiming OS-level prevention.

What you can do

• Open Enterprise → Coverage for org coverage %, a 30-day sparkline, per-team bars, and a device table (client chips, DLP, heartbeat, routing class, last seen, open findings).
• Filter Uncovered, DLP off, Silent, and Direct routing; edit device aliases; retire machines; export CSV.
• Turn on Require DLP and optionally Require heartbeat; grant time-boxed DLP exceptions.
• Investigate findings on Signals (same acknowledge / resolve / dismiss flow as other anomalies).
• As a developer, check Settings → Device health and run rclm-hooks-install --repair.

How it works

Devices enroll lazily on install, update, session start, repair, or heartbeat. There are no synthetic devices for existing members. The first observation of a client status is a baseline; AD-010 fires only on later transitions away from installed. A short grace window after an rclm version change holds upgrade/reinstall races.

Check-ins report per-client hook status, DLP on/off, heartbeat, and a gateway-vs-direct routing class (never URLs or keys). Hostnames are hashed. Integrity treats global or cwd-local hooks as healthy, so --local installs are not false-missing.

Detectors:

| Code | Cadence | Meaning | | --- | --- | --- | | AD-010 | On check-in | Hooks leave installed, bad policy signature, or uninstall ping | | AD-011 | On check-in | DLP off without an active exception (higher if Require DLP is on) | | AD-012 | Daily | Formerly active device goes silent; medium severity needs corroboration | | AD-013a/b | Daily / check-in | Gateway usage without hooks, or direct routing for a required provider | | AD-013c | Daily (Enterprise) | Provider admin usage vs gateway tokens cannot be reconciled |

rclm-heartbeat (install with --heartbeat, or when the org requires it) refreshes policy and check-in on an OS timer. Uninstall sends a best-effort check-in before config deletion.

Docs: Coverage Alerts and the Coverage tab.

Why this shape

We reused one device identity for MCP inventory and coverage instead of two enrollment systems. Lazy enroll plus first-observation baseline avoids an AD-010 storm when existing orgs turn this on. Event-driven detectors cover tamper and DLP; silence and gateway-without-hooks stay daily so they can corroborate.

Limits (intentional)

Coverage is detection, not MDM. Removing rclm and the heartbeat is mainly visible as silence, and only when something else corroborates activity (gateway, another device, login). Routing class unknown is never treated as bypass; AD-013a/b stay off if the org has no gateway hosts configured. Check-ins come from a user-writable client — signed policy helps with cache integrity, not patched binaries. Ephemeral environments are classified and do not drive silence email.

If you hit false positives on upgrade, Shared hosts, or IDE-hosted models that cannot use the gateway — or if Require DLP / exceptions feel too coarse — tell us. Criticism on the corroboration rules for AD-012 is especially useful.