Everything I have described in this series, memory, skills, agents, session history, shares one failure mode. It decays if a human has to remember to maintain it. Skills break when a dependency changes and nobody notices until the next time one is needed. Context goes stale. The machine fills up with the residue of last week's work. So the workspace runs a set of background tasks, some on folder-open and some on a throttled cadence, and the difference between a tool I use and an environment I live in turns out to be who does the upkeep.
The Four Tasks
| Task | When it runs | What it prevents |
|---|---|---|
| Resource guard | Continuously | Watchers and agents piling up until the machine crawls; it watches memory and steps in before that |
| Skill watcher | When any skill file changes | A broken skill discovered at the moment it is needed; it runs smoke tests on change so the break is caught immediately |
| Scheduled maintenance | Nightly | Stale context; it runs quality checks, refreshes context from mail and chat, and pulls repositories so work starts from the latest source of truth |
| Throttled context capture | A few times a day | Transient data that is retained only a few days expiring unread; it pulls it before it expires, extracts action items, and classifies them |
None of these is clever. Each one is a chore that used to be mine, written down once with a trigger and a pass or fail check, which is exactly the definition of a tool from the operating principles: do something twice and it becomes one.
The Watcher Matters Most
A skill is code you do not run until you need it. That is its strength, since a library of dozens costs almost nothing to carry, and its weakness, since a broken skill is invisible until the worst moment. The skill watcher removes the gap. The instant a skill file changes, a smoke test runs against it: is the front matter intact, are the triggers still there, are the steps still numbered, are the checks still present, did a placeholder sneak back in. If it fails, I know now, not at the moment I am relying on it in front of someone.
The same test runs nightly over the whole library, because a skill can break without changing, when something it depends on moves underneath it.
I Owe and Owed to Me
The context-capture task deserves its own sentence, because it is the one people ask about. It pulls transient material, the kind of thing that a chat tool or a mail system keeps for a few days and then discards, extracts the action items from it, and sorts them into two lists: what I owe, and what is owed to me. Everyone keeps the first list. Almost nobody keeps the second, which is why commitments made to you quietly evaporate. A machine that reads the expiring material before it expires keeps both.
Take It
The watcher's test is the piece that transfers cleanest, so that is the download: skill-smoke.js. Point it at a folder of skills and it checks each one against the shape from the skill template.
| Check | Why it is there |
|---|---|
| Front matter with name, description, and use_when triggers | Without triggers the skill can never be selected; without a description the dispatcher cannot match it |
| Name matches the folder | A renamed folder with a stale name is a skill that loads under the wrong identity |
| Required inputs, numbered steps, pass or fail checks all present | The three sections that make a skill a procedure instead of a note |
| No template placeholders, no TBD or TODO | A half-finished skill is worse than none, because it is trusted |
| Warns on "as appropriate," "as needed," "consider" | Each one is a decision pushed into the model's judgment instead of the skill |
It exits zero when every skill passes and one otherwise, so it drops into any file watcher and any nightly job without ceremony. That is the whole point of this layer: the environment maintains itself, and I stop babysitting it. The principles that tie the five layers together get their own post in a few weeks.
This is part of a track on agent tooling and spec-driven development that goes deeper than a field report, with something you can download and run in each one. Free for now while the track grows; the deepest installments become a paid tier later.