The assistant now acts under its own name rather than mineThe assistant now acts under its own name rather than mine
August 3, 2026 · It now acts in Linear as itself rather than as me. Separately, typing a setup guide's placeholder literally locked me out of my own machine.The assistant now acts under its own name rather than mine. Separately, copying an example from my own guide too literally locked me out.
last updated
An identity of its own
Until today everything the system did in Linear was done as me, which makes a task list useless as a record of who did what. It now completes a proper OAuth handshake and holds its own credential, so its comments, status changes and sessions are attributable to it.
With an identity in place the work turned to what happens once it is handed something: classification, a completion protocol, and the table mapping tasks to issues picking up columns for a live session. A real bug got caught in the middle: the module caching “who am I” answers kept exactly one identity in a single shared slot, which was fine while there was exactly one identity and wrong the moment there were two.
The day's biggest change is the cutover itself: real traffic now goes through this path. Two gaps the cutover opened were closed the same day, and two documentation commits fixed real drift. The project's own status document had no mention of any of this, despite days of it running.
Notifications cut off mid-sentence
A background job's reply was being truncated to a cap sized for a spoken Alexa summary, whatever channel it went out on. A phone notification and a chat message are not a voice reply and do not need to be one sentence long. The cap now depends on where the message is going.
A placeholder, typed literally
I hit this one myself, first thing in the morning: working through this project's own restricted-login setup guide, I typed a placeholder that was meant to be replaced with a real value. It was accepted, and it locked SSH.
Worth recording rather than quietly fixing. A setup guide that can be followed exactly and still break is a guide with a bug in it, and the fault is the document's rather than the reader's.
Its own account
Until today, everything the assistant did in the task list was recorded as me doing it, which makes the history useless. It now has its own account and its own credentials, so its comments and changes are recorded under its own name.
Setting that up caught a real bug: the part that remembers who it is kept exactly one answer in one slot, which was fine when there was one identity and wrong the moment there were two.
Messages cut off mid-sentence
Replies from background jobs were being shortened to a length meant for speaking aloud, no matter where they were going. A phone notification does not need to be one sentence, so the limit now depends on the destination.
Locked out by my own guide
First thing in the morning I followed this project's own setup guide, typed an example value exactly as written, and locked myself out of the machine. Worth recording rather than quietly fixing: a guide you can follow exactly and still break is a guide with a bug in it.