Skip to content

fix(ci): a met promotion criterion says so where somebody reads it - #795

Merged
stephrobert merged 1 commit into
mainfrom
fix/125-a-met-criterion-says-so-somewhere-read
Sep 25, 2026
Merged

stephrobert merged 1 commit into
mainfrom
fix/125-a-met-criterion-says-so-somewhere-read

Conversation

@stephrobert

Copy link
Copy Markdown
Owner

runtime-proof.yml counts consecutive green scheduled runs and, at fourteen,
writes into the step summary:

The criterion is met. Move this workflow onto pull_request and close #125.

It wrote exactly that on 2026-09-21, with fourteen greens running back to the
8th. Nobody read it. The streak broke the next night — #740 armed
guard.sh verification and a real broken claim has reddened every night since —
and the moment passed unnoticed. #125 is still open, and the number it asks for
was reached four days ago.

The failure is the one this workflow already names

Its own report job carries the sentence, about red nights:

left no trace outside job logs and the summary above — the two places nobody
opens without already knowing there is a problem

That lesson was applied to failure (#502: one issue, opened, updated, closed) and
never to the half that asks for action. The mechanism measuring the criterion had
no way to tell anyone it was met.

What changes

The criterion, once met, is announced on the issue. Once — a marker in the
comment body makes it idempotent, because a comment every night is how a
notification becomes noise somebody mutes, which is the shape #502 refuses for
red nights and would be no better here.

The decision lives in tools/ci/streak-report.sh rather than in a run: block,
for the reason night-report.sh states: a run: block cannot be executed outside
Actions, and this repository has paid for CI fixes described in comments and never
run.

What falsify found in this work

The workflow test asserted Contains(workflow, "streak-report.sh --apply") —
which echo streak-report.sh --apply satisfies while announcing nothing. It reads
the step's run: value now.

That is the third time here a Contains has been satisfied by a longer string
doing the opposite, after if: always() == false and a step commented out. The
helper reads the whole value rather than looking for a substring.

Proven

  • Seven tests driving the real script with a stubbed gh: below target, at
    target, past target, already announced, a maintainer's prose that must not
    silence it, a dry run, and the workflow actually calling it.
  • Six mutations, each compiling and each biting — including the two that matter
    most: the announcement reduced to a step summary again, and the job left without
    permission to write, a silence that reports success.
  • mise run prepush green.

What this does not do

It does not close #125. The streak is 0 and the criterion needs fourteen, so the
issue closes when the nights earn it — the rule its header states. What this
removes is the need for anyone to watch for the moment.

🤖 Generated with Claude Code

runtime-proof.yml counts consecutive green scheduled runs and, at fourteen,
writes into the step summary:

    The criterion is met. Move this workflow onto pull_request and close #125.

It wrote exactly that on 2026-09-21, with fourteen greens from the 8th. Nobody
read it. The streak broke the next night — #740 armed `guard.sh verification`
and a real broken claim has reddened every night since — and the moment passed
unnoticed. #125 is still open, and the number it asks for was reached four days
ago.

THE FAILURE IS THE ONE THIS WORKFLOW ALREADY NAMES

Its own report job carries the sentence, about red nights:

    left no trace outside job logs and the summary above — the two places
    nobody opens without already knowing there is a problem

That lesson was applied to failure (#502: one issue, opened, updated, closed)
and never to the half that asks for action. So the mechanism that measures the
criterion had no way to tell anyone it was met, and the streak job's verdict
had exactly the reach the red nights had before #502.

WHAT CHANGES

The criterion, once met, is announced on the issue. Once: a marker in the
comment body makes it idempotent, because a comment every night is how a
notification becomes noise somebody mutes — which is the shape #502 refuses for
red nights and would be no better here.

The decision lives in tools/ci/streak-report.sh rather than in a run: block,
for the reason night-report.sh states: a run: block cannot be executed outside
Actions, and this repository has paid for CI fixes described in comments and
never run.

WHAT FALSIFY FOUND IN THIS WORK

The workflow test asserted `Contains(workflow, "streak-report.sh --apply")`,
which `echo streak-report.sh --apply` satisfies while announcing nothing. It
reads the step's `run:` value now. That is the third time in this repository a
Contains has been satisfied by a longer string doing the opposite — after
`if: always() == false` and a step commented out — so the helper reads the whole
value rather than looking for a substring.

PROVEN

- Seven tests driving the real script with a stubbed gh: below target, at
  target, past target, already announced, a maintainer's prose that must not
  silence it, a dry run, and the workflow actually calling it.
- Six mutations, each compiling and each biting, including the two that matter
  most: the announcement reduced to a step summary again, and the job left
  without permission to write — a silence that reports success.
- `mise run prepush` green.

WHAT THIS DOES NOT DO

It does not close #125. The streak is 0 and the criterion needs fourteen, so the
issue closes when the nights earn it — which is the rule its header states. What
this removes is the need for anyone to watch for the moment.

Assisted-by: Claude Code (claude-opus-5)
@stephrobert
stephrobert merged commit 28be518 into main Sep 25, 2026
31 checks passed
@stephrobert
stephrobert deleted the fix/125-a-met-criterion-says-so-somewhere-read branch September 25, 2026 14:41
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

The runtime-backed proof earns its promotion: fourteen scheduled greens, then a PR gate

1 participant