running-tend

name	running-tend
description	Project-specific guidance loaded by tend workflows alongside CLAUDE.md.

Running tend — leaf

Landing

Work from Tend lands as a pull request that a maintainer merges. The bot has write access, and the Merge access ruleset holds merging to admins.

Review threshold

Apply CLAUDE.md's Stage section to the verdict. Once a change moves Leaf toward a coherent architecture and its claimed path works, the review is done. Reserve findings for architectural seams, cross-runtime invariants, public surface traps, regressions on the claimed path, or a central claim or test that is false. Omit bounded edge cases, exhaustive same-pattern cleanup, minor simplification, and prose or test polish unless they expose one of those problems.

Filing issues in other repos

Standing exception granted: file directly in agent-equipped targets (per Filing Issues in Other Repos in the bundled running-in-ci skill) without asking permission here first. The default rule (open an issue here asking permission first) still applies when the target shows no agent signals.

Leave outage trackers for the drain

Leave the "Bot temporarily unavailable" tracker open until tend-review-runs drains its rows; the outage ending is not enough because each row names a stranded trigger. The tracker's own body invites the opposite — "Close it once the outage is resolved" — but that boilerplate does not account for the rows. Ignore it. A comment can record recovery, but the drain owns the close. This applies by title, not by the tend-outage label: ci-fix diagnosis trackers have no rows and are closed by ci-fix itself.

A red ci on main is live

Pull requests and main run the everyday and website-worker gates. The complete suite runs on the daily schedule. A red ci on main is already affecting whoever pulls next. Treat it as live.

Review test selection

Before approving a product change, choose and run the smallest test selection that exercises the failures the diff could introduce. Select from the product paths and contracts in the diff, not from the test files it happens to touch. The review selection supplements the everyday CI gate. A docs-only or generated-workflow change may need no additional test; a selected failure withholds approval.

For a change that can alter browser startup, apply CLAUDE.md's Working on the repository performance rule. Read the candidate profile from CI and compare it with the base. If requests or bytes rise before presentation, check that the PR names the user-visible benefit and why the work must happen then.

When a high-level browser test is slow or fails on timing or geometry outside its contract, repair its arrangement or move that contract to the lower boundary that can prove it. Keep a browser case only where it proves that the boundaries work together.

Reading a red suite

Nearly every test drives a real browser, so the traceback can name a symptom several boundaries after its cause. Find the first violated contract, reproduce at the lowest boundary that preserves the failure, then determine whether the product, test, or execution environment owns it. Failure movement, determinism, and clustering can guide that search; none decides who owns the fix.

Two test-owned failures recur here:

• A read or press that ran before the page said it was ready. The dominant class here, and the one a re-run hides. The test measures a point, presses it, or asserts on it while the page is still arriving at the state the gesture needs — a panel still widening the document, a scroll still settling, a response landed but not yet reconciled. It surfaces at a wait far from the read that caused it, so the traceback names the symptom rather than the cause. tests/CLAUDE.md already owns the fix under State races are arrangements, not probabilities and A state the page passes through is not a state to poll for: state the ordering, do not repeat the gesture until it happens to hold.
• A reading that has been widened before. The failure is a case some accepted set, allowlist, or tolerance does not cover, and the fix that presents itself is one more member. Read the line's history first (git log -L). A set that has already grown is describing the noise the suite makes rather than the behaviour the test names, so the next wording reddens main again. tests/CLAUDE.md owns the fix under A test cannot assert over noise it makes itself. The PR is against the test.

Weekly: interface sweep

Run /ui-sweep before dependency maintenance. This is the discovery pass for visual and interaction behavior the suite has no stated invariant for yet. Follow its Reconcile route: a reproduced defect becomes a tested repair, while a design judgment stays in the run report.

Weekly: vendored browser dependencies

.github/dependabot.yml watches the action refs and uv.lock. It cannot watch the browser dependencies, whose versions live in scripts/vendor.py's PINS table rather than a manifest. They drift silently, and this is the step that catches it.

scripts/vendor.py --pins

Each row is a package, its pin, the bundle to rebuild if it has moved, and the newest release that pin could take where that differs. elkjs and entities are beautiful-mermaid's imports rather than Leaf's own choices, so their rows read against the range it declares: a release outside it is not a pin to take, because npm would install the declared version nested and the bundle would carry that one. esbuild is the tool the builds share rather than payload, so it moves when a bundle needs it rather than on every release.

On drift, bump the entry in PINS and run scripts/vendor.py <bundle> — the rebuilt bundle is the commit, not the version string on its own. A copy's output tracks its version directly. highlight and pierre also read the language list out of the registry's $languages.names, so their output is a function of both the pin and the registry: rerunning them after an unrelated registry change is how the bundle and the lint stay unable to disagree. Pierre and Shiki must move together when their compatibility requires it.

Run the suite afterwards. The browser tests load the bundles, so a bad rebuild surfaces there rather than in review.