What ready proves
The five checks a mission's feature pull request must pass before it is marked ready, the AC format of a spec, the label that starts the full checks, the settings, and the runner fallback for e2e.
In a mission, a feature pull request is marked ready only when five checks prove it. A check that is not proven leaves the pull request a draft, and the verdict names the check that holds it. Nothing here merges: only a person merges.
The five checks
- CI is green on the latest commit. Every check run of that commit finished
success. A skipped or neutral run is left out, since a pull request skips jobs on purpose; a run that failed, was cancelled, timed out or needs an action is red. If only skipped or neutral runs are there, the check is not finished. Theomni / readycheck itself is left out. - The full tests and lint ran. The CI jobs named by
ready.full(heretestandlint, neverquick, which runs only the changed files) succeeded on that commit. - Every acceptance criterion has its proof. Each
AC-nof the spec has a passing proof on that commit (see the AC format below). - No test was loosened. The diff against the default branch holds no deleted test file, no
removed test or
describe, no.skip,.todo,.only,xitorxdescribe, no weaker assertion, no added suppression (eslint-disable,@ts-ignore,@ts-expect-error), no lowered mutation floor, no raised timeout or retry, no grown baseline, and no changedAC-ntest whose spec criterion did not change. An outbox item a person answered can excuse a named path, and only through itsexcuses:field. - The change stayed in its area. Every file of the diff sits under a slice's territory, the PRD's folder, its outbox, or a generated entry's path, unless an answered outbox item names it.
Each check reads one of four states: pass, fail, not finished (its evidence is not there yet,
such as a run still going or a preview not up) or not judged · <why> (its evidence could not be
read). Only pass passes. The verdict is ready only when all five read pass. The Ready panel on
the PRD page shows Ready for review · 5 of 5, or 4 of 5 with the failing check named at the top,
and each check links to its evidence.
Who it applies to
A setting decides, ready.checks:
offgates no pull request. It is the kit's default.missiongates the feature pull request of a PRD that belongs to a running or paused mission. This repository runs withmission.allgates every feature pull request. A later switch, which changes no code.
Under mission, /omni:review's pass is the one that marks a feature pull request ready. It runs
omni ready <prd> before gh pr ready, and on a red verdict leaves the pull request a draft and
ends with changes, the failing check named as the finding.
The label that starts the full checks
A draft runs no full checks by itself. Before omni ready runs, the gated skill adds the label
omni:ready-check to the draft. That label starts, on the draft's latest commit, the full checks, the
e2e job and the preview. A later push to a labelled draft runs them again. Ready still comes last: a
draft is never marked ready first and then turned back.
The label is made by a person, once per repository, as the other loop labels are not. The name can be
changed with labels.readyCheck; left out, the kit reads omni:ready-check. Until the label exists,
the gate cannot start those checks, and the checks read not finished.
The AC format in a spec
A spec's acceptance criteria are numbered scenarios in Gherkin, each with its kind:
### AC-2 · e2e
```gherkin
Scenario: A loosened test keeps the pull request a draft
Given a mission's pull request whose diff skips a test
When the Reviewer passes it
Then it stays a draft and its check names the skipped test
```
test: proven by a vitest test titledAC-n: …, with its body in the same Given, When, Then order, that passed on the latest commit.e2e: proven by the scenario run with e2e against the pull request's preview./omni:validate-e2eturns each scenario into one test titledAC-n: …, taggedprd-<n>, with the scenario's lines kept verbatim (see Validate with e2e).
The kind is written on each criterion. acceptance.kind names the default a new spec takes.
omni check inbox refuses a spec whose AC- headings have a missing or repeated number, a missing
kind, a kind other than test or e2e, or a scenario without a Given, a When and a Then. A spec with
plain bullets still reads, but its PRD fails check 3 when it is gated.
The settings
In .omni-loop/config.yml:
ready:
checks: mission # off, mission or all
full: [test, lint] # the CI jobs that run the whole suite
acceptance:
kind: test # the default kind a new AC takes
e2e:
ci: true # run the e2e scenarios in CI, on the draft's preview
repo: null # owner/name, when the e2e tests live in another repository
labels:
readyCheck: omni:ready-check # optional: the label that starts the full checks
e2e.ci is false by default. With it off, or with no key configured, the e2e scenarios are run by a
runner of the workspace instead (below). e2e.repo is unset for this repository, where the tests live
in e2e.dir. A product whose e2e tests live elsewhere names <owner>/<name> there.
Where the proof comes from
CI writes the check run acceptance on the head commit. The AC results come from the test job's
report. The e2e results come from a job that waits for the pull request's Vercel preview and runs the
e2e scenarios there, keeping its HTML report (steps, screenshots, video) as the run's artifact. For
CI to run them here, the repository needs the secrets OPENROUTER_API_KEY and
VERCEL_AUTOMATION_BYPASS_SECRET.
The runner fallback
When e2e.ci is false, or no key is configured, and the PRD has an e2e criterion, a new commit
queues a runner job. The runner runs the scenarios against the preview, and the App writes the
acceptance check run from its result. Two things make it work:
- the preview's address. With
e2e.url: github-deploymentthe runner asks GitHub for the preview of the commit, and the Omni App is not allowed to read deployments, so that lookup fails and says why. A fixede2e.urlruns as set instead. - a runner. With no runner online, check 3 reads
not finished · no runner has run the e2e scenarios yet, and the verdict stays not ready, the pull request a draft.
Where the verdict lives
omni ready <prd> runs the rule on facts it reads itself, through gh: the head commit's runs, the
acceptance report, the diff, the plan and the ledger at the head. The Omni App runs the same rule on
each new commit of a gated pull request and posts the check omni / ready, each state linking its
evidence. When the App cannot be reached, omni ready still decides alone: the gate never waits on the
App.
The acceptance check run, the omni / ready check and the App's saved verdicts are the App's own
status records, not a person's words. They are never read as an answer that excuses a finding.