Branch health

Honest limits:

• It has read only a recorded copy of six projects' GitHub data from 29 September 2026, through a stand-in for gh. It has never read GitHub live, and nothing polls on a schedule: live polling waits for the owner's commission.
• It only reads. It never reruns, pauses or re-enables a workflow, and it never asks gh to show a token.
• A signature is an observation, not a cause. "New-head red" means a workflow failed on a new commit after passing on an earlier one, a suspected regression. "Same-head red" means it failed on a commit where it had passed, which suggests something outside the code changed.
• An episode ends only when its failing jobs pass on the branch's latest commit, or when a person disables or removes the workflow (or removes a late schedule). A passing run that skipped those jobs ends nothing, and neither does a push while a schedule is late. Jobs are matched by their full names, so a job whose name merely looks like the failed one's, cut short or spaced differently, ends nothing. A failure an ended episode already covered never starts another one.
• If a person turns a workflow back on while its last run still shows the failure, the project stays unknown until the workflow runs again.
• Anything it could not read keeps a project unknown, never green: a listing cut short, a refused request (401, 403 or a rate limit), a timeout or an unreadable reply. Workflow files it did not read leave a missed schedule undecided. A schedule that has never run keeps the project unknown but is not called late, because nothing dates it.
• A check that was cancelled, went stale or waits for someone's approval is not a pass. A check from one workflow never hides a failed check of the same name from another.
• It reads at most five pages of a workflow's runs, so a failing spell that began earlier has an unknown start. When it stops early, it reads each other kind of trigger separately, so an older failure is not missed; such a page that fails or does not add up leaves the project unknown. GitHub does not date disabling a workflow, so that start is unknown too.
• Each poll reads the workflow's runs back to where the last poll stopped, and to any run that was still going then, so a gap cannot silently pass as complete history, even after an episode closes. A poll that cannot read that far back keeps the project unknown and prevents an episode from ending. It does not notice a rerun of a run older than what it reads.
• It reads the jobs of every failed run in an episode, 20 runs a poll. Until all are read, the episode cannot end, and a failure that began beyond what it can read never ends on its own.
• Evidence that names another commit or run, or a listing whose count does not match what it holds, is treated as unread.
• Its thresholds are proposals. Six failures started by a bot on one commit in a day mark a runaway. A run waiting an hour, or a schedule an hour late, gives no signal.