I've had a rule for months that any real work has to be recorded in the Weave. The agents agreed with the rule every time. They just kept forgetting to follow it.

This stretch, August 9 through 22, was mostly about making that rule mechanical. It was also the fastest release pace so far, with about 25 numbered hub releases in two weeks.

Enforcement beats a reminder

An agent's to-do list now mirrors into the Weave automatically. If a session does real work and leaves no record, a reminder fires before it settles. I also changed what counts as work. It used to mean a code commit, and now it means any durable change.

Every session gets graded on a capture scorecard that answers a plain question. Did it record what it did, or did it only talk about it? Scheduled routines got the same treatment, so when one misses its run, I hear about it. That came from finding out the dispatcher had been silently down for about a month.

Several rounds of multi-model review went after the guardrails themselves and found ways around them, like reaching a protected folder through a relative path. Those got closed one at a time.

Faster, better recall

Each turn, Loomfield pulls in the few pieces of knowledge most likely to matter. That retrieval now blends three approaches: keyword matching, direct lookups, and semantic search. On a blind-judged test set, the best single approach found the right item 98% of the time and the blend found it every time.

I set the bar for shipping before I saw the results, so I couldn't talk myself into a number afterward.

Mid-session lookups also went from 22 to 32 seconds down to about 22 milliseconds. The fix was ordering. It now ranks first and only checks permissions on the top 50 results.

Reliability, the unglamorous part

Idle sessions were piling up as background processes until an 18 GB machine was sitting on 9 GB of swap. Sessions idle for 30 minutes with nothing waiting on me now close on their own. That one change made the system usable again.

I also found that automatic retry for failed sessions had never worked on the real path. The code existed and the tests passed, but the live path never called it. I'm seeing that pattern often enough now that I go looking for it on purpose.

Deploys now run detached so a killed caller can't stop one halfway, and the hub reports what it's doing while it starts up instead of just going quiet.

Retiring routines that depended on me

I culled 12 recurring routines this stretch. They were genuinely useful, and I'd run all of them by hand for a good while. The problem was that every one of them needed me to kick it off, and once I stopped, they quietly died.

The redesign measures whether anyone actually uses a routine's output, demotes a routine on a clock that matches how often it should run, and suggests new routines from patterns it sees in the work. A review before building found that the draft assumed a scheduling action that didn't exist yet, so that got built first.

Still open

Long conversations can now hand off to a fresh session without losing the thread, but the automatic trigger shipped off until I trust it.

Release tooling for installers and updates also landed. It was the first real step toward putting Loomfield on someone else's machine.