Matrix evaluation for jobs.*.runs-on appears to have regressed #2320

alexrp commented 2026年01月17日 03:54:22 +01:00

Comment

Since the v14 deployment, Zig's translate-c CI has been doing weird things: https://codeberg.org/ziglang/translate-c/actions/runs/66

If you click into each job, you can see that they're all picked up by zach and orange - both aarch64-macos-labeled machines! So something is funky with matrix evaluation, I suspect. What's interesting is that these matrix.host checks seem to be seeing the correct values, but if I schedule a run manually, all 4 jobs are waiting for a runner with the labels self-hosted and aarch64-macos.

cc @Gusted @mfenniak

### Comment Since the v14 deployment, Zig's translate-c CI has been doing weird things: https://codeberg.org/ziglang/translate-c/actions/runs/66 If you click into each job, you can see that they're all picked up by `zach` and `orange` - both `aarch64-macos`-labeled machines! So something is funky with matrix evaluation, I suspect. What's interesting is that [these](https://codeberg.org/ziglang/translate-c/src/commit/827524d4c4b7dac3c5ee584d311c928b0a1c70ac/.forgejo/workflows/ci.yml#L51-L68) `matrix.host` checks seem to be seeing the correct values, but if I schedule a run manually, all 4 jobs are waiting for a runner with the labels `self-hosted` and `aarch64-macos`. cc @Gusted @mfenniak

mfenniak commented 2026年01月17日 05:52:14 +01:00

I'll take a look into this, but I am stacked up with other Forgejo 14 issues as well. Can you provide the label configs for your runners so that I can set up a similar reproduction case? The workflow looks straight forward, but this additional info would provide the information I'd need on the runner side to make a clean reproduction.

I'll take a look into this, but I am stacked up with other Forgejo 14 issues as well. Can you provide the label configs for your runners so that I can set up a similar reproduction case? The workflow looks straight forward, but this additional info would provide the information I'd need on the runner side to make a clean reproduction.

alexrp commented 2026年01月17日 05:56:04 +01:00

FreeBSD:

 "labels": [
 "self-hosted:host",
 "x86_64-freebsd:host"
 ]

Linux:

 "labels": [
 "self-hosted:host",
 "x86_64-linux:host"
 ]

macOS:

 "labels": [
 "self-hosted:host",
 "aarch64-macos:host"
 ]

Windows:

 "labels": [
 "self-hosted:host",
 "x86_64-windows:host"
 ]

FreeBSD: ``` "labels": [ "self-hosted:host", "x86_64-freebsd:host" ] ``` Linux: ``` "labels": [ "self-hosted:host", "x86_64-linux:host" ] ``` macOS: ``` "labels": [ "self-hosted:host", "aarch64-macos:host" ] ``` Windows: ``` "labels": [ "self-hosted:host", "x86_64-windows:host" ] ```

mfenniak commented 2026年01月18日 04:11:33 +01:00

This is a regression in the workflow parser in Forgejo 14, specifically when runs-on is an array and references ${{ matrix.... }} in an expression. PR for the core fix is up at https://code.forgejo.org/forgejo/runner/pulls/1307, but as it is an update to a library used by Forgejo it will require a few additional steps to get into a Forgejo release after merge.

As a short-term workaround: runs-on: ${{ fromJSON(format('["self-hosted", "{0}"]', matrix.host)) }} avoids the bug and behaves the same.

This is a regression in the workflow parser in Forgejo 14, specifically when `runs-on` is an array and references `${{ matrix.... }}` in an expression. PR for the core fix is up at https://code.forgejo.org/forgejo/runner/pulls/1307, but as it is an update to a library used by Forgejo it will require a few additional steps to get into a Forgejo release after merge. As a short-term workaround: `runs-on: ${{ fromJSON(format('["self-hosted", "{0}"]', matrix.host)) }}` avoids the bug and behaves the same.