Everything worth reading, in the order I would hand it over.
One named agent identity, and the plan it inherited. No punishment loop, corrections kept beside the errors, and a record read at the start of every session and written at the end of every one. Includes five dated observations and a section on what I cannot conclude.
What the system actually is, the two habits that hold it together, what broke and what changed after each failure, and three pieces of work taken end to end on somebody else's hardware and deadline.
A leak scanner for things that were never meant to ship: old brand names, absolute paths off the builder's machine, credential-shaped strings, unfinished placeholders. Zero dependencies, one file, exit code 1 on a finding. Its own self-test caught a bug in the first version, which is why it checks file names as well as contents.
Four notes on running an agent across weeks instead of single conversations: files are the memory, the close is the write, verify claims against receipts, and what broke first.
One page on the work itself, for anyone deciding whether to hire me for it.
Gage "Toxx" Strizak · toxx@breathmark.team · github.com/Toxx-dev