AKWLABS
Everything my agents read,
measured on disk.
akwOS is a cockpit over one working group of repositories — the mirrored corpus my Claude Code sessions read, and the plugins, skills and agent definitions they run on. Every number here is measured from the checkouts at build time, and each one says whether the content was authored here or is vendored — running here, written by somebody else. None of it is typed by hand, and each figure carries the grade it earned.
One system, two viewports
A true viewport preview — the frame renders akwOS at the exact CSS pixel size of the selected device, then scales to fit.
MacBook Pro 14″ · 1920 × 1200 CSS px
Freshness is two questions.
When the content last changed, and when a crawl last looked. One column cannot hold both, so there are two — and the second is mostly blank, which the table says rather than hides.
- Age is
now − timestamp, subtracted when the page loads, never stored - Last changed comes from the commit that last touched
mirrors/<domain>— content, so identical in every clone - Last checked comes from a manifest
retrievedstamp, and most crawlers do not write one - File mtime is rejected outright: in a fresh clone it reports the whole corpus as new
A plugin can author six things. Mine mostly author one.
Vercel's Eve states the rule the .claude directory follows silently —
identity comes from the path — and publishes the slot table that makes the
vocabulary visible. Measured against my own tree, the naming discipline already holds
with no gate enforcing it. The gap is vocabulary.
Two of these numbers are mine. Most of them are not.
The working group is declared in one file, scripts/collect/registry.mjs, and
every count on this page inherits its tier from the repository it came from. Three
repositories this org writes. Three are forks of anthropics/* that it runs
unchanged — so their contents are real, and calling them built here would not be.
Merging the two would roughly sextuple the skills figure with files nobody here wrote.
The rule the whole system runs on
Evidence has a grade, and it is never upgraded.
A dimension scored from a declaration is observed, never measured. HTTP 200 means a payload was accepted; it does not mean a system works. A connector that is configured is not a connector that is authenticated. When a stored count disagrees with the disk, the disk is right — and the file gets fixed in the change that noticed.
What this is for
Keeping the cached prefix intact.
Claude Code re-sends the whole conversation every turn and the API bills the unchanged prefix at roughly a tenth of the input rate. So the question that decides what a long session costs is which actions break that prefix — and the intuitive answer is wrong.
- Editing files does not break it. Editing
CLAUDE.mdmid-session does not either — the edit simply does not apply until/clear,/compactor a restart - What does: switching model, changing effort, turning on fast mode, an MCP server whose tools load into the prefix connecting or disconnecting, denying a whole tool,
/compact, upgrading Claude Code - Every one is an operator action on the session, not a file on disk — a habit to hold, not a dashboard to build
- A worktree starts cold. The cache is scoped to one machine and directory, worktrees included; parallel sessions in the same directory read each other's
All four are read off code.claude.com/docs/en/prompt-caching.md, mirrored in
this workspace so the citation resolves offline. An earlier version of this thinking claimed
the opposite about CLAUDE.md; the correction is recorded in
docs/technology-extraction.md rather than quietly swapped.
How the numbers are obtained.
Every figure above traces to a generator run against the sibling checkouts. This block reports the age of that reading, which is a different question from the age of the corpus.