How we build
pixtuoid is written, reviewed and maintained by AI coding agents working for one owner, across a handful of repositories. This page maps who does what; each part links the file that holds it, which wins wherever the two disagree.
The repositories
| Repository | What it holds |
|---|---|
| pixtuoid | The CLI, the office engine and the pixtuoid.dev site |
| pixtuoid/apple (private) | The App for iPhone and iPad |
| lofi | The procedural music the office plays |
| .github | What every repository shares, as its AGENTS.md lists |
Who does what
The owner (AGENTS.md)
Setting the bar and the Roadmap, reviewing every doc change, merging the release PR by hand
A session (Parallel sessions)
Working one branch in its own worktree, through the arc when the work is non-trivial and always through the merge gate
A subagent (local review)
Reviewing or verifying one scope for its session, then reporting
The CI review, an AI code review Claude Code runs
Posting a verdict and its findings on the PR
Merging once the queue’s conditions hold
A change, end to end
Non-trivial work runs as the arc loop, from picking a Roadmap item to the wrap; any PR, an arc’s or not, lands through the merge gate, each finding at a terminal state dispositions defines.
What keeps it honest
- One set of rules. The
org rules
are copied into each repository’s root agent guide (
AGENTS.md), and CI fails when that copy drifts. Each of those repositories adds aREVIEW.mdfor what a review checks. - Evidence, not recall. Every external claim follows the org rules’ Evidence bullet.
- Real CLIs, not mocks of them, as the hermetic tier runs them.
- A lesson becomes a mechanism, as the arc’s Wrap step says.