chadtmiller.com/knotless

Switch repository

Loading…

What changes

The feed leaves out a repository's bots. getFeed for everyone, and for whom one follows, no longer holds what a repository's bots said or pushed on it: the runners and triagers its record names, but its owner and collaborators, who are people a bot may act as. That holds for anyone's runners, read from their repositories' records, with nothing for an operator to set. A bot's own feed (actor=) still holds everything it did.

A merge the queue landed is shown as whoever queued it. A runner's merged state is kept, credited to the account whose queued record stands for the pull; with none, it stays the runner's.

Why

Of the last 60 events on knotless.dev, 17 were review.knotless.dev restating pulls already in the feed as opened, and the runner's batch landing added a push to main, a deleted branch and a merge under its own name. What a bot says belongs on the pull and in notifications. The filter is in the index's queries, so a page stays full.

An operator's list (KNOTLESS_BOT_DIDS) was tried first, but it covers only knotless.dev's bots, and on prod it names the owner, as the queue runner's account.

How it was checked

Index tests for recent and pushes quiet, with an owner and a collaborator named as runners kept; a unit test for who a merge is shown as. The rest is left to CI.

Shortcomings

A reviewer is named in no repository's record, so review.knotless.dev's reviews still show: a bot field an account sets on its dev.knotless.profile is the next pull. A runner that is a public member of the owner's organization, but no collaborator, counts as a bot.

This pull filters the accounts in KNOTLESS_BOT_DIDS out of getFeed (everyone's and followed) in the SQL for recent and pushes. Accounts that KNOTLESS_ADMIN_DIDS also names are not treated as bots. A bot's own actor= feed is unfiltered. A bot's merged state is shown as whoever's queued record stands for the pull. The deploy list gains the reviewer and the runner host, and the docs, lexicon description and changeset are updated.

I found no bugs. Filtering in the queries keeps pages full, and the cursor paging is unaffected. Because quiet is empty when by is set, a bot's own feed still holds everything. shown_by falls back to the bot when nothing is queued, and the unit tests cover that.

Two things to know, neither blocking:

• The recent filter keeps only dev.knotless.state records from a bot, so pulls and issues a bot opens are hidden too, not just comments. That fits the docs ("only the pulls they merged"), but the PR body says "comments and pushes".
• events now builds assembled(...) for every repository in the page, where it used to call forge(...). If assembled isn't cached, that adds cost per feed request. I didn't check how heavy it is.

I didn't run CI.

This round moves "bot" from the operator's KNOTLESS_BOT_DIDS list to what each repository's record names. A bot is a runner or triager that is neither the owner nor a collaborator. The SQL uses a new bot_on helper, and shown_by uses the repo record. I like the direction and found no logic bug in bot_on.

The change does not compile, though. The &[] → false rewrite was applied to calls where &[] was not the quiet argument:

• keep_profile(did, records: &[(String, String, Value)]) now gets false.
• heard(&index, "did:c", false) is passed where the third argument is the orgs slice. The same helper is called with &orgs a few lines above.
• said("2026-02-01T00:00:00Z", None, false) is a similar case.
• In tests/forge.rs, branch("g2", false) is passed where the closure takes files: &[&str].

Please revert those to &[] and change only the recent and pushes calls. CI had not finished when I looked, so a local cargo test would show any others.

Smaller points:

• The deploy README and cli.md no longer say that KNOTLESS_BOT_DIDS affects the feed, which is right. But the PR body still describes the old behaviour: the env var, the reviewer joining the list, and the admin exemption. The reviewer and the runner host will now leave the front page only if a repository names them as runners or triagers, and the body doesn't say so. Please update the body.
• The bot_on SQL runs json_each over the repo's JSON for every candidate row, and again for pushes. That should be fine at the LIMIT sizes in use, but it's worth a glance at the query plan.

This pull keeps a repository's runners and triagers out of the everyone and following feeds. It filters in the SQL for recent and pushes through the new bot_on helper, so pages stay full. Owners and collaborators are never treated as bots. A runner's merged state is shown as whoever queued the pull, through shown_by. A bot's own actor= feed is unfiltered, because quiet is off when by is set.

The compile errors from my last round are fixed. Only the recent and pushes calls take the new quiet argument, and calls such as keep_profile(..., &[]) keep their slices.

I read the SQL, the parameter binding, shown_by and the new tests and found no bugs. The tests cover an owner and a collaborator that are also named as runners. They also cover a merge shown as whoever queued it, and the cases where nothing was queued or the pull left the queue.

CI was still running when I looked: test, web and shots were queued and docs had passed. I did not see the Rust tests run, so wait for test before landing.