Device Health

pakem health inspects the device and removes what the owner approves. The judgment is yours, the deleting is the tool's: it re-verifies every finding at the moment it acts.

1. Check first

Run pakem health check (add --json to read it programmatically). Never hand-run a recursive du over the home folder or ~/Library, which takes about a minute, and never clean by hand: no docker volume prune or docker system prune (they delete named database volumes), no cargo clean (its lock stalls a live session's build), no rm -rf on a guess. If the check has no workspace roots, add the owner's workspace folders as roots in ~/.config/pakem/health.toml and run it again.

2. Weigh urgency

The check rates the data volume critical, watch, or healthy by free space. Weigh it against the rest of the report: memory free and swap, thermal state, battery, and the agent sessions and builds running. A critical disk with swap in use needs space now. A healthy device needs nothing unless the owner asks.

3. Present the findings

Follow tiers. For each tier, give the total and the largest findings with their ids, what each costs to lose (nothing, a rebuild, a pull), and what is in use. Name what stays protected (for example the named Docker volumes). Recommend:

• critical: the zero-impact tier, plus the warm findings that free the most for the least rebuild cost.
• watch: the zero-impact tier.
• healthy: nothing.

Cold candidates (whole idle repositories, old toolchains) never come from the tool. If you see them, suggest archiving them to another machine and leave the decision to the owner.

4. Clean only what the owner approved

Run pakem health clean <id>... --plan with exactly the ids the owner approved in this session, show it, then run it again with --yes. Report what was removed, what was skipped and why (in use, changed, refused), and run pakem health check again to show the new free space. Never pass an id the owner did not approve.

5. Memory

The check lists idle Gradle and Kotlin daemons with the memory they hold. Stopping one (./gradlew --stop in its project) is the owner's call. Never stop an agent session or a process you did not start.

6. Platform notes

• macOS with Docker Desktop: Docker's disk image shrinks after a cleanup only as the VM returns blocks. The check shows its allocated and reported sizes.
• WSL: deleting files frees space inside Linux, but the Windows-side ext4.vhdx does not shrink by itself. Compacting it (wsl --manage <distro> --optimize) is an owner action on the Windows side.