One path: check, review, land, live (#7703)

Every change follows four steps, and a person only needs those four words:

  1. Check: types, lint and the tests the change reaches run on the pull's own code.
  2. Review: a reviewer approves the change or asks for changes.
  3. Land: an approved, green pull merges into the default branch.
  4. Live: a release builds the default branch and production serves it.

/pipeline says this to anyone. Each step's finer parts (the check queue, the merge queue's gate, the release build and boot check) sit one click down in a <details>. The live counts per stage stay on /boards, whose Pipeline panel links back to /pipeline.

Where a pull is

src/flow/path.ts decides it: pathStep is plain data in, plain data out, and stepOfPull reads one pull's facts. The rule is "the earliest step not yet passed":

(if no check is recorded, there is nothing to wait for);

deploy has served since the merge. A merge where nothing deploys ends at land, done.

For an open pull the sentence comes from nextActor (src/flow/next_actor.ts), so the step line and the "Next:" line never disagree. On the check step the check is the reason, even before anyone has reviewed.

Under every pull's title (both tabs) the line reads, for example, Check · Review · Land · Live — No review yet., linked to /pipeline. A finished path (live, or landed where nothing deploys) carries a "done" badge. The MCP pull tool answers step: { step, why, done }, and the terminal's pull screen says step: Review — No review yet. beside next: in its header line. The same word and sentence (from stepOfStored) sit on a pull's post in the feed, its row on the repository's pulls list and on /review (the word links to /pipeline), and under its row in the terminal's feed, pulls list and on-you list (#7783). List rows skip the merge-conflict probe (it costs a git run per pull), so a conflicted pull can read differently on its own page than on a list; the row links to the pull, whose page answers. The step is looked up per row (a few queries each); batching it is left for when the lists measure slow. changeHead computes "who acts next" once and hands it to stepOfPull.

The full suite as a report

A verification after landing (src/checks/verify.ts) runs every shard against the default branch's tip. When it is red, src/checks/report.ts files one issue per failing test file through fileIssueOnce (src/checks/flake.ts), titled Full suite red: <file> and labelled full-suite; a later red adds a comment to the open issue instead of filing again. Nothing that decides a pull, a batch or a release reads a verification, so the report holds none.

How it is scheduled: the checks service looks every minute (VERIFY_SWEEP_MS) for a repository whose default branch has moved since its last verification and verifies it, at most once per VERIFY_EVERY_MS (30 minutes, set in src/main.ts) and never two at once. So the report runs after a landing, not on a daily clock: a quiet main is not re-run, a busy one at most every half hour.

/pipeline links the label (in the forge's own repository, ownRepoPath()) in the card under "Where to look next". The terminal's Flow dialog ends with a line counting the open report issues; the issue list there filters by label with l.

Not here yet

The rest of #7703 is filed as its own issues: re-run a shard that crashed outside any test (#7781), reuse a verdict for the same tree (#7782), and measure each step (#7784).

↑↓ navigate ↵ open ⌘↵ its app esc close

Keyboard shortcuts

Docs