v1.0.0 stays proven
A weekly maintenance duty, v1-proven (#12989), runs the production probes of scripts/v1-proven.ts (the one list is
src/core/v1_proof.ts) against the forge's own public origin (FORGE_URL), GET only, no other host. Without
FORGE_URL it skips and says so.
- What it probes. One probe per box of the v1.0.0 proof. Each ends
pass,failorunproven(the forge could not
tell). Only fail counts: unproven files nothing.
- Where the record shows. Each run is a
maintenance_runrow. One line per probe sits in its actions
(2026-11-15 fail #123 <item>), a failing probe is also a finding. Read them with
GET /api/v1/maintenance/runs (admin token) or in the terminal's "Maintenance duty runs" screen. The web
/maintenance page is the power-off preparation and does not list duty runs.
- A defect issue. A run with failing probes opens one issue in the forge's own repository (
FORGE_REPO, e.g.
owner/forge), not one per probe (#13265), titled Defect of the production proof: <n> lines fail. Its body lists each
failing line, each with the evidence block of what that probe read (URL, status, bytes, ms, time), carries the stable
line Probe key: v1-proven/run and says to fix the cause and close it. With no FORGE_REPO nothing is filed and the
run says "no issue filed".
- Cannot reach. Probes whose read got no HTTP answer at all fold into one line, `cannot reach <origin>: <k> of <n>
probes got no answer; when that is the only failure the title is Defect of the production proof: cannot reach
<origin>`. An unreachable forge files one line, not one per probe.
- A repeat failure. The open defect is found by its
Probe key:line, so renaming the title loses nothing. While it
is open, a run that fails again (a second run the same day, or next week's) adds one comment ("Still failing on
<date>") listing that run's failing lines, and no second issue. Once it is closed, a new failure opens a new one. A
per-probe defect filed before #13265 (Probe key: v1-proven/<probe issue> or the title Defect of #<probe issue>: )
still takes its own probe's comments while it is open.
- A pass. A run with no failing line while the defect is open adds one comment ("Passing again on <date>") and leaves
it open: a person closes it. Repeated passes add nothing; a failure after a pass comments again, and the next pass notes again. A pass with no open defect says nothing.
- Limits. Runs at most twice a day, weekly by default, and spends no money. The findings only land on the run record;
the runner files no notice of its own.
One read-only probe (GET only) for each item on the v1 list, one line each:
pass: a stranger's request read the evidence.fail: the evidence is missing or wrong; the line says why.unproven: no public read can show the claim, so it is not counted as a pass.
Today: #11140 (the free plan sleeps rather than is deleted), #11141 (CPU slowing, the first change landing, lapsed-payment sleep: only the gated page is seen), and #11595 only against a forge whose health answer predates #12512.
Two probes read the release the forge runs, from version on
GET /api/v1/health (never /api/v1/version, which is the API's own constant
and says 1.0.0 on every release). #10304 says it: release 0.7.604. #13042
fails until that release is 1.0.0 or later: the release is 0.7.604, not 1.0.0.
Only the v1.0.0 tag running on production passes it; the proof alone is not
the owner's word that v1 is done.
Exit 0 only when every probe passes; 1 when any fails; 3 when none fails but
some are unproven; 2 on bad usage. Owner-only pages are probed as "there and
gated". The schedule probe reads schedules on the public GET /api/v1/health:
the forge-wide count of unpaused schedules more than five minutes past due and
the age of the oldest, counts only, naming no app or repository. It passes at
zero and fails naming the oldest one's age. It reports and never tags; scripts/cut-release.ts does that.