Working on knotless

For every agent working here: CLAUDE.md is this file, imported.

Use rustup run stable cargo (the Justfile's cargo), not a system cargo. Node is the version .node-version names, CI's: the harness's PDS loads a better-sqlite3 built for it, and refuses another major. Tests run against the npm reference PDS through harness/, never pds.js.

Running tests

Build and test on the runner, not on this machine: several agents building here at once have brought it down. Open the pull early, push to its branch as you go, and read what its checks say with knotless run list <repo> --pull <uri> --wait <secs> --log. Each push runs warm from main's caches (see What CI checks): check (fmt and clippy) in a minute or so, test in a few, running only the test binaries the push changed. just fmt (rustfmt, and oxfmt for the pages, the harness and the docs site) costs nothing, so run it before pushing; Claude Code runs it on each file it writes (.claude/settings.json).

Build here only for a loop the runner is too slow for, such as one test being fixed, and then in one checkout at a time, with CARGO_BUILD_JOBS=2, and just pds stopped once done. Never run just cov, just lint's clippy over the workspace, shots-linux, shots-check or shots-live here: their checks run them. What follows is for that loop.

• just pds in the background first: test binaries then share that PDS instead of each starting one.
• just changed names the files changed since tests last passed and the tests that cover them; just t-changed runs those and marks the pass. just t <cargo test args> runs any selection on the shared PDS.
• Working on the web pages (web/): web/AGENTS.md, loaded with them, says how they are built, shot and checked, and to load the ui skill before changing one.
• Done: all seven of the pull's checks pass. When one fails in a way its log and pictures do not explain, run that one thing here, alone (COV_CLEAN=1 rebuilds everything first, for a cov miss that makes no sense).

cargo fmt can put a ? on a line of its own, which coverage then counts as a line never run: bind the value first, or use a named function for a row mapper, rather than lowering the gate.

Commands and tools

A command is knotless <group> <action> [<url>] ...: the group a singular noun (repo, pull, branch), then the action. Only login, logout, upgrade and status stand alone. An opposite is --undo on the action (repo star --undo), never an un- command. A verb that works on any pull or issue by at:// URI exists under both pull and issue, on one handler.

A command and its MCP twin land together, in one pull. The tool is <group>_<action> (- as _), and takes the command's arguments: a parameter is the flag's name (--keep-branch, keep_branch), positionals as the usage line names them, --undo as undo. One tool per action: a tool that folds several commands is split to match. Every command has a tool and every tool a command, but those that act on this machine, which are CLI-only: login, logout, upgrade, repo mirror, workflow pin and release download.

Renaming a command keeps its old name working, hidden from the usage, for one release: the deploy host and shared steps run an installed binary. docs/reference/cli.md and docs/reference/mcp.md are laid out by group, one section each, and docs/reference/actions.md lines the two up: a new command or tool goes in its group's section.

atproto's primitives

Keys, JWTs, base32, TIDs, service auth, DID documents' keys and services, handle syntax, and the spaces alpha's set hash, MACs, STAR archives and tokens come from the atproto crate, in its own repository (knotless.dev/chadtmiller.com/atproto), published to that repository's Cargo registry, which .cargo/config.toml names. A primitive knotless needs that the crate lacks, or a fix to one, goes there, as a change to that repository and a new version, before knotless moves to it.

Lexicons

A new or changed schema under lexicons/ follows lexicons/AGENTS.md and passes just lex.

Queries

A query of the index (crates/knotless-serve/src/index.rs) is planned at prod's size as the tests run it, and test fails on one that scans or sorts a table that grows with the network: crates/knotless-serve/AGENTS.md says how to fix or mark one.

Docs

docs/ is a small wiki: design pages explain why, the guide shows how, the reference lists what. A change that alters what a page says updates that page in the same pull:

• storage records, push, claims, compaction: docs/design/storage.md
• gateway, delegation, pushRefs, createCommit: docs/design/gateway.md
• pulls, issues, states, merging, moderation: docs/design/collaboration.md
• the merge queue: docs/design/merge-queue.md
• knotless-run: docs/design/runner.md, docs/reference/runner-internals.md
• knotless-triage and triage.rs: docs/design/collaboration.md (Labels), docs/reference/cli.md
• knotless-serve, its cache, appview and pages: docs/design/serve.md, and its methods in docs/reference/api.md
• names, redirects, transfers, commit authors: docs/design/identity.md
• spaces and private repositories: docs/design/access.md
• releases, publish, versioning: docs/design/packages.md
• a flag, environment variable or command's output: docs/reference/cli.md
• an MCP tool: docs/reference/mcp.md
• a workflow key: docs/reference/workflows.md
• who can do what: docs/design/security.md

To write or rewrite a page, or to check one against the code, use the docs skill (.claude/skills/docs/SKILL.md): facts from the code first, then the page from the reader's questions. just docs checks links, anchors and diagrams (npm ci in web/ first, for Pele).

The pages also serve the docs as a site at /docs: docs-site/, a Starlight site that reads docs/ as it is, each page's # heading its title, its links made the site's, and its sidebar in astro.config.mjs. A page added to docs/ goes in that sidebar too. just docs-site shows it as you write.

Versioning

A pull that changes what those who use knotless see (a command, a flag, what the client does) adds a change file under .changeset/, as knope document-change writes one (https://knope.tech): default: patch, minor or major in its front matter, then a heading and a line for them. knope release turns them into the version and CHANGELOG.md. Commit titles stay prose: no conventional commits. The check check fails a pull that changes crates/knotless/src or crates/knotless-mcp/src and adds no change file, unless its body says, on a line of its own, Changeset: none: for a change no one using knotless sees.

What CI checks

The workflows in .knotless/workflows/ run on pushes to main and on pulls into it or into a feature/ branch (see Landing a change), each pull at its head merged with its base as it is now, and again whenever either moves; a push to any other branch runs nothing. On a pull, web, lexicons, shots, docs and e2e run only when it changes what they read (each file's paths); otherwise each passes at once, noted as skipped. A push to main runs all seven, and check and test run on every pull. main deploys only once they all pass:

• check: cargo fmt --check and clippy, warnings denied; on a pull, a change file for a change to the client (harness/changes.sh).
• test: just cov's 100% gate, over the spaces alpha, and the stock PDS for the few test binaries that reach what a PDS without spaces does (COV_STOCK in harness/cov.mjs), and that every XRPC answer the serve tests received, kept with a stored run or fresh, is as its lexicon says (harness/answers.mjs over what harness/cov.mjs gathers), and that no query the index ran scans a table that grows with the network, as SQLite would plan it at prod's size, or sorts many of its rows (harness/plans.mjs; see Queries). A test binary that passed before is not run again: when it, the workspace binaries it ran and the harness are the same, its stored profile counts instead (COV_STORE in harness/cov.mjs), so a change runs only the binaries it changes. A pull's run, its pre-flight, goes against both nets and holds the gate, as a batch's does, so a line nothing reaches fails the pull rather than a batch. A binary that fails is run once more; one that then passes is logged FLAKY: <net> <binary> <test> with its first failure: fix it. One that fails again only in tests harness/quarantine.json names, until the day it gives, passes too, logged QUARANTINED: <net> <binary> <test>, and is run alone until it passes, three times at most, so the lines only it reaches still count: a test that fails now and then is put there while it is fixed, and taken out with the fix.
• web: oxfmt's format, the component layers and svelte-check.
• lexicons: just lex, against main's schemas where the branch left main.
• shots: just shots-check in Playwright's Linux image: each page whose HTML differs from the base's, shot as the base draws it and as the change does. Its pictures (before, after, and where they differ) of each that moved, or is new, are the check's outputs, and one the pull's body does not name on a Shots-moved: line fails it.
• docs: just docs: every relative link and anchor in the docs lands, and every Mermaid block draws with Pele; and the site at /docs builds.
• e2e: just e2e: the pages used in Chromium over a world seeded for the run as the shots' is, served as a gateway behind the harness's https front: logging in through the PDS's approval page, then every write the pages make, from starring and commenting to merging, settings, organizations, transfers and deleting an account's data, as two accounts, each read back on the page and from the PDS (harness/e2e/flows.spec.mjs); and beside it a private repository's, over the spaces alpha, each read back from its space and as not there to anyone else (private.spec.mjs, just e2e-private). A flow that fails leaves its page among the check's outputs. A new control that writes gets a flow there, and in the private flows too when a private repository reads it back differently.

A push to main at the same tree as a pull's run that passed, as knotless pull merge leaves it, does not run these again: the runner files them as passed at once, each log naming the pull's run. deploy.yml still runs, and waits for them as ever.

ios.yml is not one of the seven either: on a pull that changes ios/, the lexicons or the recorded answers, it reads those answers with the iOS app's models, checks what its merge section says, and holds its writes to their lexicons. Xcode needs macOS, so the app itself is built by hand (ios/README.md), and no deploy waits for this. just ios-look <screen> shows a screen in any state, as just look does a page: run it before pushing a change to one.

scenarios.yml is not one of the seven: it runs the scenario tracks on main only when the owner asks for it, as knotless run rerun <repo> --ref refs/heads/main --workflow scenarios -e TRACKS="team", and costs model usage, so ask before starting one. Its report is the end of its log. look.yml is too, at any branch: just look for an agent with no checkout, its pictures the check's outputs (the ui skill says how). docs-audit.yml is the same kind: Claude Code sessions checking each docs page against the code (-e PAGES="design/storage.md" for some), whose report says what each page gets wrong or leaves out.

What just lint, just cov and just lex pass here passes there. A check's result and log: knotless run list <repo> --pull <pull uri> (--wait <seconds> waits for it; --log adds each log's end; --json gives each step), or the run_list tool over MCP. Its pictures: --images <dir> saves them to open, and over MCP images: true returns them to look at. A pull that moves pages on purpose names them in its body, on a line of their own, Shots-moved: home, pull-open (the names its check gives; all for every page), then runs that check again (knotless run rerun <repo> --pull <pull uri> --workflow shots): the body is read as the check runs, and editing it moves nothing. A batch reads the line from the squashed pulls' bodies. One moved by mistake is a fix to make. A change to a page in a state the check does not draw, as a menu or dialog open, an edit box, a filter that empties a list or a button busy, shows that state in the pull: just look's pictures of it, desktop light and phone dark, each given to knotless pull create (or knotless pull edit) as an --image. Only the states the change touches: the check already shows what moved of the pages it records.

Prod's health: knotless status https://knotless.dev --repo chadtmiller.com/knotless (--json for all of it): its feeds, index and a mirror, from /_status; deploy/README.md says what each means.

Landing a change

Open a pull rather than committing to main: push the work to a branch named for it (web/mermaid, serve/jetstream), then knotless pull create knotless://chadtmiller.com/knotless main knotless://chadtmiller.com/knotless <branch> "<title>" -m - < body.md, the title and body written from .knotless/pull.md (knotless pull template <repo>); -m - keeps the shell off the Markdown. The body opens with a short paragraph a person would write, what the pull is for and whether it needs a reader's attention, and folds the full write-up below it in <details>, under a <summary> naming the model and effort that wrote it, as the template lays out. It is merged into main and deployed from there. A body saying Fixes <issue> (or Closes, Resolves), the issue by its knotless.dev page or at:// URI, closes that issue as it merges. Once its own checks pass, land it with knotless pull merge <repo> <pull uri> --queue, which returns at once: the runner lands queued pulls in batches, main with each squashed onto it as a commit of its own, checked once at the batch's tip, so many landing together cost one run. It takes a pull queued in within half a minute, and checks the next batch on top of the one being checked, so a pull queued behind a batch does not wait for that one to land. knotless queue list <repo> lists the queue and what each is doing, and --leave <pull uri> takes one out. A pull that conflicts with those ahead of it, or breaks the batch (found by halving it), is ejected, and knotless pull show says why: fix it and queue it again. Plain knotless pull merge merges at once, and only once all seven checks have passed at its head merged with main as it is now, and otherwise says which failed, are running, or are of an older head or base: --wait <secs> waits for them, and --force is the owner's alone. A pull that conflicts with main fails each check without running: merge main into it, or rebase it onto main. The runner checks a pull as it moves only when its author and its head repository's owner are the owner or a collaborator; anyone else's is not checked until the owner runs it with knotless run rerun <repo> --pull <pull uri>. Every run starts warm from the caches main's last passing run left, each over a copy-on-write layer of its own, so runs of one workflow go side by side, and only pushes keep what they wrote: a pull's writes go when it ends. A branch whose pushes and pulls both run a workflow, as a feature branch's can, keeps caches of its own, from its pushes, which the pulls into it start from (docs/design/runner.md). (Anyone else's pull of a workflow that asks for secrets starts from empty caches instead.) While the runner lands the queue, a pull not in it is checked again only as its head moves, not as main does, and after the queue's batches: the queue checks it against main once it is queued.

Queue a pull at once, once pushed: the queue holds it until its pre-flights, check, test and web (queue: first in their workflows), pass at its head, and ejects it if one fails, so one that does not build, fails a test, misses a line or is left unformatted never fails a batch. A pass against an older main counts. Pulls queued behind it that are ready go ahead meanwhile, and knotless queue list says pre-flight check,test or the like while it waits. Queueing turns away a pull that conflicts with main or whose checks failed, saying why, so merge main into the branch just before queueing. One ejected: read why with knotless pull show, and the failed check's log, then fix it and queue it again. Queue the same head again only when the log shows the failure is not the pull's, such as a test or a line it does not touch that fails other pulls too, and then fix or quarantine what failed rather than leave it for the next pull. A pull that conflicts with one queued ahead of it waits for that one to land or leave; ejected then, it conflicts with main, so merge main into it. Ejected three times for the same reason, stop and say so rather than queue it again. A branch whose rule names a gate, as main may name review.knotless.dev (knotless branch rules), waits for its review of the pull's head: knotless pull show says waiting for its review meanwhile. Ejected with changes asked by @review.knotless.dev, read the review with knotless pull show, push the fix, and queue it again: the agent reads the new head. Ejected for not reviewed as main asks when it approved an earlier head and has stopped reviewing (it reviews a pull --rounds times), and nothing since is more than main merged in or its own notes answered, dismiss its review, saying so (knotless pull dismiss-review <repo> <pull uri> review.knotless.dev <reason>), and queue it again: a dismissal stands in for a gate's review, on the owner's own pull too, where an approval never counts. Where it asked for changes and is wrong, say why in a comment and ask the owner whether to dismiss it. Never --force past it: that skips the checks at the head too. An approving review can still carry notes it does not hold the pull for: once the pull lands, read its review with knotless pull show and judge each note, checking first that it is right, as a reviewer can be wrong. Fix it in a follow-up pull, open an issue for one worth doing but not now, or drop it, and say which, and why, in your report to the user: the reviewer reads no reply. When several agents land pulls at once, one of them lands them all, and the rest hand it their pulls' URIs rather than each watching its own. A restart of the runner host's runner, as its deploy does, pauses the queue. A batch whose checks pass while it stops still lands; one it cuts off is built and checked again, with nothing to do.

Most conflicts are in the docs pages nearly every change updates, docs/design/serve.md and docs/reference/api.md above all: add to them where the change belongs, and leave the lines around it as they are. Agents working at once take different parts of the code, and of those pages, where they can.

Work that should not reach main yet goes on a feature branch, feature/<name>, cut from main. Its parts are pulls into it, made as pulls into main are but with feature/<name> as the base: the seven checks run on them at their heads merged with the feature branch, and knotless pull merge lands them there, through the queue only if that branch's rule turns it on (knotless branch protect <url> <branch> --queue). Nothing deploys from a feature branch, and a push to one runs nothing. Keep one pull open from it into main, unqueued, as the check that the whole of it still passes on main as it is now; merge main into the feature branch as main moves, since a pull into main that conflicts fails without running. When it is ready, that pull is queued as any other.