What we’ve shipped

Since the provider switch it had been paying full price to re-send the same large, unchanging setup with every turn — about 933,000 tokens in a single day where it used to be a few thousand. No behavior changes; it just stops paying twice for the same thing.

Ask it to read, summarize, or pull a quote from a note and it will. Until now it could write notes, edit them, and undo those edits, but had no way to open one and read it — so when you asked, it told you it couldn't, which was true and useless.

Saying "the note we made on Sunday" now works the same as "my note". Follow-ups keep working too: once you've pointed at a note, "read it", "pull it up", and "yes that one" stay on topic instead of losing the thread mid-request.

Both had been failing on every single attempt since yesterday's provider migration — silently, because neither shows an error when it fails. Summaries just never appeared and threads never got named.

When a database write failed, the error it produced could carry a piece of whatever you had written, and that error was sent to our error-monitoring tool. Nothing was public and nothing was shared outside the tools we already use to run Selah, but study notes have no business being in an error report. They are now stripped before anything leaves.

A tracking script we're required to load on every scripture page throws when you navigate away mid-chapter. Those were being recorded as if they were our bugs, drowning out real ones.

The job remains queued until an eligible route is available, including when the note is processed in chunks.

, and successful extraction clears stale failure details.

When Selah reads a note into your belief map and something goes wrong — an outage at the model provider, a bad minute on the database — that note used to be retired permanently and silently. Nothing ever went back for it, so its ideas simply never entered your map and nothing said so. Failed notes now rest a few hours and are picked up again.

Every failure counted the same, so three bad minutes could retire a good sermon as surely as an unreadable file. Failures the platform caused now get more patience than failures in the content.

A single error while recording a failure took down the whole run, so notes queued behind it waited, and other background repairs were skipped along with them.

One bad label used to throw away every concept and connection found alongside it. Selah would try twice more, then give up on that note for good — it stayed out of your map with nothing to show you why. Now the good parts are kept and only the mislabelled item is dropped.

The old direct-to-Anthropic path is gone. It was never a live backup — switching to it needed a redeploy either way — and keeping two paths meant the unused one hid a real outage in the used one.