The assistant had never once been reachable, then it did a day's workThe assistant had never once been reachable, then it did a day's work

August 4, 2026 · Yesterday's cutover looked live and had never delivered a single event. Capturing the raw payload settled it, and by night it was picking up its own next ticket.Yesterday's launch looked fine and had never actually worked. Finding out why took capturing what was really arriving, and by night it was working on its own.
last updated

It had never received a single event

Yesterday made the bot a real Linear Agent: something you delegate an issue to, rather than a listener reacting to comments after the fact. It looked live. It had never received one agent event.

Guessing at the shape of the payload is what had gone wrong, and capturing the raw delivery is what settled it. Two more things in the same stack were quietly wrong at the same time, which is why reasoning about it from the outside kept producing plausible answers that were not the answer.

With the right event category subscribed, a fresh signing secret loaded and the token reinstalled, a ticket went the whole way: comment to acknowledgement in under a second. The old path then came out entirely: two handlers, two routes, four test files, a mode flag and every gate around it.

Then it shipped fourteen tickets

The rest of the day went through the pipeline that had just started working. Twelve tickets, mostly this site's own design: a light/dark switch across every page, kept deliberately separate from the technical/simple content switch because they answer different questions; two blog-specific changes; a nav bar.

One ticket it refused to guess at was the home page, which is the right instinct. Taste is not a spec. And the re-scoring of ten backlog projects was the one that failed first time, killed at the ten-minute wall, then landed on the re-run. Re-deriving all ten from the project details rather than inheriting the old scores moved five of them.

Every way a delegated ticket could die quietly

Thirty seconds after the first successful run I restarted the backend to strip out diagnostic logging, and killed a live session mid-flight. That exposed three separate silent deaths: a reaper that only saw tasks dispatched from an issue, a reason a reaped task still said nothing, and a session killed by the restart it had itself been told to perform being recorded as a failure, which reads as “nothing happened”.

Two other liabilities surfaced the same day. Access tokens expired every 24 hours with nothing renewing them, so staying installed meant re-running a browser flow daily. And every request to the skill endpoint was being logged in full, which meant a signed token and device identifiers were sitting in a world-readable log.

Seventeen tickets filed into nowhere

The day's last find was on the ingestion side rather than the delegation one: seventeen issues had been filed with no project, for two independent causes either of which was sufficient alone. The prompt now requires a project whenever the team has one, and the answer is re-checked against the real project list rather than trusted.

By the end of the night it was picking up its own next ticket. A successor is taken only if every blocker is closed, nothing is assigned to it, and it is not already delegated. The first thing it built on its own was the chess archive pull: 1,836 games in about 22 seconds.

It had never actually been reachable

Yesterday's launch made the assistant something I could hand a task to directly. It looked like it was working. It had never once received anything.

The problem was that the format of the messages had been guessed at rather than checked. Capturing what was really being sent settled it in minutes, after a morning of reasonable-sounding theories that were all wrong.

Then it did a day's work

Once it worked, it completed fourteen tasks, mostly this site's own design: a light and dark mode, a navigation bar, and some blog changes. It refused to guess at the home page, which was the right call, since that is a matter of taste rather than instruction.

Every way a task could die silently

Half a minute after the first success I restarted the server and killed a job mid-run. That exposed three separate ways work could vanish without anyone being told, including a job killed by the restart it had been instructed to perform being recorded as a plain failure, which reads as though nothing happened at all.

By the end of the night it was picking up its own next task without being asked, and the first thing it built alone was the chess game downloader.