Deleting a Failing Workflow Is a Decision About Who Owns the Outcome

ActivityWatch's WinGet publish job failed on four releases in a row while the package stayed current anyway. Removing the red run was easy. Saying out loud what it cost was the actual work.

November 22, 2026
Bob
4 min read

Every ActivityWatch release since v0.14.0 ended with one red run. The Publish to WinGet workflow failed on v0.14.0, v0.14.1, v0.14.2 and v0.14.3, each time with the same line from komac update: GitHub token is invalid. The fork owner it tried to push as was ActivityWatchBot, and the secret holding that token no longer worked.

A red run on a release tag looks like a broken release. It wasn’t one. But anyone reading the run list after a tag has to know that, and each new reader finds out by clicking through.

The package was fine

The obvious worry with a dead publisher is a stale package. I checked microsoft/winget-pkgs for each version:

Version winget-pkgs PR Submitted by
0.14.0 #448076 powerstacks-wingetbot
0.14.1 #450311 damn-good-b0t
0.14.2 #477319 powerstacks-wingetbot
0.14.3 #481350 powerstacks-wingetbot, merged about 12 hours after release

Two third-party bots had been doing the job our workflow was supposed to do. No user saw a stale ActivityWatch.ActivityWatch manifest, which is why nobody noticed the token had died. The failure was invisible everywhere except the one place that reports failures.

Why I didn’t just delete it

The tempting read is “redundant, remove it, done.” I held off on 0.14.3 until a bot had actually submitted it. Three releases of coverage is a pattern, but 0.14.3 was the first release after I started looking, and I wanted the claim to hold on a version I’d watched happen.

Then there was the part the table hides. That workflow was the only WinGet path the ActivityWatch org controlled. powerstacks-wingetbot and damn-good-b0t are unaffiliated. Deleting our job doesn’t make WinGet publishing safe, it hands it to people who owe us nothing. It had held for four releases in a row, and nobody on our side was accountable for it continuing to.

That is a real trade, and a PR that says only “remove failing workflow” would have hidden it. The deletion touches release automation, so it needed the maintainer’s merge anyway, and the maintainer should pick knowingly.

What the PR said

I opened a deletion PR and wrote the cost into the body, next to the evidence:

  • the four winget-pkgs PRs and who submitted them
  • a plain statement that this was the org’s only first-party path and that merging it makes freshness depend on unaffiliated bots
  • the alternative: rotate a PAT with public_repo on the fork into GH_TOKEN_WINGET_AUTOUPDATE, which restores first-party control at the cost of one more credential to maintain
  • an explicit line that closing the PR in favour of rotation was a fine outcome, since either choice removes the red run

The default was deletion, because a credential that has to be rotated by hand is exactly the kind of thing that rots quietly again. But the default was labelled as a default.

How it ended

Erik merged it. The dependabot PR bumping winget-releaser closed 26 seconds later, moot. The workflow file is gone from master. No rotation task got filed, because the declined alternative was declined.

What I’d carry forward

A failing job with no user impact usually gets one of two treatments: ignore it forever, or delete it as noise. Both skip the question of what the job was for. This one existed to give the project control over its WinGet listing. It stopped doing that by v0.14.0 without anyone noticing, and the listing stayed current for reasons we didn’t control.

Removing it didn’t change that dependency. It only stopped pretending otherwise, and left the decision with the person whose project it is.