Free system resources with consent
Goal
Reclaim CPU, memory, or I/O by stopping resource-heavy processes - only with
explicit per-process user approval, never by pattern or guesswork.
Input
- a Linux machine whose
/proc is readable;
- a debug verdict (ideally from the
debug-system skill) naming the resource
under pressure.
Output
A ranked hog list, the user's approved PID set, the signal result per PID, and
before/after resource numbers. Exit status: 0 all targets stopped, 1
something survived or was refused, 2 unusable environment.
Workflow
Rank the suspects (read-only; --sort cpu|mem|io to match the verdict):
python3 skills/free-resources/scripts/hogs.py list --sort mem
KIND marks critical (session-critical, never a target), agent
(stopping it kills a live agent session), and normal.
Present the list to the user and ask which PIDs to stop. Every PID needs an
explicit user decision. For agent kind, say out loud that killing it ends
a live session. If the user is not reachable, stop and report instead of
acting.
After approval, signal gracefully first and verify:
python3 skills/free-resources/scripts/hogs.py kill --pid PID1 --pid PID2 --yes
SIGKILL (--signal KILL) only after a TERM attempt survived its timeout
and the user re-approves that specific PID.
Confirm the effect with fresh evidence: re-run the ranking or the
debug-system diagnostic and report freed memory/IO before declaring done.
Rules
- Consent is per PID, per action. A list approval from earlier does not cover a
new kill; re-check that each PID still runs the same comm before signaling.
- Never signal: PID 1, this agent's own process group,
critical kind
processes, or processes owned by other users (root escalation is out of
scope for this skill).
- The script refuses unapproved, critical, foreign, and vanished PIDs; do not
work around a refusal with raw
kill, pkill, or xkill.
- Prefer TERM over KILL. If nothing is safe to stop, say so and suggest
non-destructive options (restart a wedged watcher, close idle sessions).
1---2name: free-resources3description: memory, cpu, resources, kill, consent, cleanup4---56# Free system resources with consent78## Goal910Reclaim CPU, memory, or I/O by stopping resource-heavy processes - only with11explicit per-process user approval, never by pattern or guesswork.1213## Input1415- a Linux machine whose `/proc` is readable;16- a debug verdict (ideally from the `debug-system` skill) naming the resource17 under pressure.1819## Output2021A ranked hog list, the user's approved PID set, the signal result per PID, and22before/after resource numbers. Exit status: `0` all targets stopped, `1`23something survived or was refused, `2` unusable environment.2425## Workflow26271. Rank the suspects (read-only; `--sort cpu|mem|io` to match the verdict):2829 ```sh30 python3 skills/free-resources/scripts/hogs.py list --sort mem31 ```3233 `KIND` marks `critical` (session-critical, never a target), `agent`34 (stopping it kills a live agent session), and `normal`.35362. Present the list to the user and ask which PIDs to stop. Every PID needs an37 explicit user decision. For `agent` kind, say out loud that killing it ends38 a live session. If the user is not reachable, stop and report instead of39 acting.40413. After approval, signal gracefully first and verify:4243 ```sh44 python3 skills/free-resources/scripts/hogs.py kill --pid PID1 --pid PID2 --yes45 ```4647 `SIGKILL` (`--signal KILL`) only after a TERM attempt survived its timeout48 and the user re-approves that specific PID.49504. Confirm the effect with fresh evidence: re-run the ranking or the51 `debug-system` diagnostic and report freed memory/IO before declaring done.5253## Rules5455- Consent is per PID, per action. A list approval from earlier does not cover a56 new kill; re-check that each PID still runs the same comm before signaling.57- Never signal: PID 1, this agent's own process group, `critical` kind58 processes, or processes owned by other users (root escalation is out of59 scope for this skill).60- The script refuses unapproved, critical, foreign, and vanished PIDs; do not61 work around a refusal with raw `kill`, `pkill`, or `xkill`.62- Prefer TERM over KILL. If nothing is safe to stop, say so and suggest63 non-destructive options (restart a wedged watcher, close idle sessions).
lost-rob0t