evals/results/workspaces/2026-08-26-04-newcomer/.agents/skills/akinator-resource-guard/SKILL.md

4.89 KB · Oct 3, 2026 · 06:33 UTC

---
name: akinator-resource-guard
description: Use before starting anything heavy - a build, a test suite, a container rebuild, a long-running process - and at the end of every task. Checks CPU and memory pressure, reduces your own load rather than the developer's, and leaves the machine as you found it.
---

# Akinator Resource Guard

Development happens on a machine someone else is also using - often the machine
of the person you are working for, with their editor, their browser, their
containers and their other work running on it.

Two rules: **reduce your own load, never theirs. Leave the machine as you found
it.**

## When to use

- Before starting any build, test suite, container rebuild or long-running
process.
- When a command times out, hangs or is unexpectedly slow - saturation is a more
common cause than a real bug, and it is the cheaper hypothesis to test.
- At the end of every task, always.

## When NOT to use

- Before quick, cheap commands. Checking pressure before `git status` is its own
kind of waste.

## Procedure

### 1. Check pressure before heavy work

Read CPU and memory before starting. Portable enough approximations:

```bash
# Linux / macOS
uptime # load average
free -m 2>/dev/null || vm_stat # memory
ps aux --sort=-%cpu | head -15 # top consumers
```

```powershell
# Windows
Get-CimInstance Win32_Processor | Select-Object LoadPercentage
Get-CimInstance Win32_OperatingSystem |
Select-Object FreePhysicalMemory, TotalVisibleMemorySize
Get-Process | Sort-Object CPU -Descending | Select-Object -First 15
```

If the machine is comfortable, proceed. If it is saturated, go to step 2.

### 2. If it is saturated, find out whose fault it is

Identify the top consumers and determine which of them **you** started in this
session: watchers you launched, dev servers, scratch containers, background test
runs, an earlier build that never exited.

- **Yours** - stop them. A watcher you started three tasks ago and forgot is the
most common cause of a saturated machine during agent work.
- **Not yours** - leave them alone. The developer's editor, their browser, their
containers, their other agent session and their unrelated builds are not yours
to kill. Never kill a process or container you did not start.

If the machine is saturated and none of it is yours, **wait or ask** - do not add
load on top, and do not silently proceed into a run that will time out and be
misread as a failure.

### 3. Serialize your own heavy work

- One long job at a time.
- Never run parallel test suites.
- Never race a build against a suite.
- Background long jobs and judge them by real exit codes, not by summaries.

Parallelism is for independent services during a rebuild, where the procedure
says so (`akinator-ops-map`) - not for stacking your own gates.

### 4. Prefer the cheap operation

- `restart` over rebuild for code-only changes.
- Scoped tests over the full suite (`akinator-gate-economy`).
- Incremental builds over clean ones, unless a dependency or schema change makes
the cache a liar.

### 5. Clean up - always

At the end of the task, the machine is as you found it:

- No orphaned watchers, dev servers or background jobs you started.
- No scratch containers, images or volumes you created for a one-off.
- No half-finished installs, no partially applied migrations on a shared
database, no leftover lock files.
- Temporary files in the scratch directory, not in the user's project tree.

Verify rather than assume: list your background jobs and containers and confirm
they are gone.

## Failure modes and pitfalls

- **Killing the developer's processes.** The single worst failure available at
this station. Their editor, browser, containers and other sessions are off
limits, however much CPU they are using.
- **Adding load to a saturated machine** and then debugging the resulting timeout
as if it were a code failure.
- **Forgetting your own watcher** from an earlier task, then blaming the machine.
- **Leaving containers running** at the end of a task. They survive the session
and quietly consume memory for days.
- **Parallelizing your own gates** to "save time" - on a loaded machine this is
slower, and flakier.
- **Cleaning up by assumption.** "It should have exited" is not verification.

## Definition of done

- [ ] Pressure was checked before heavy work.
- [ ] If saturated, only processes you started were stopped; none of the
developer's were touched.
- [ ] Your heavy work ran serially, judged by exit codes.
- [ ] No watchers, dev servers, containers, images or volumes you started are
still running.
- [ ] Temporary files are in the scratch directory, not the project tree.
- [ ] Cleanup was verified by listing, not assumed.

SHA-256: 34ac7fbd6300090d36fc256f8fea3536f75e83528df060c60e0ac5a745538528