mirror of
https://github.com/Comfy-Org/ComfyUI_frontend.git
synced 2026-01-26 19:09:52 +00:00
87f5480462
## Problem
Our CI tests were experiencing non-reproducible results where:
- Tests would pass on a PR initially
- The same PR would fail later when main HEAD changed
- Screenshot comparisons showed excessive differences between expected
vs actual
- Blake identified: *"tests are not reproducible inside a branch - they
change every time main HEAD changes"*
## Root Cause
The issue was caused by **explicit `repository` parameters** in our
`actions/checkout` steps:
```yaml
- name: Checkout ComfyUI_frontend
uses: actions/checkout@v5
with:
repository: 'Comfy-Org/ComfyUI_frontend' # ← This was the problem!
path: 'ComfyUI_frontend'
```
According to GitHub Actions documentation:
> **When checking out the repository that triggered a workflow, `ref`
defaults to the reference or SHA for that event. Otherwise, uses the
default branch.**
When you specify an explicit `repository` parameter (even if it's the
same repo), GitHub Actions treats it as "otherwise" and defaults to the
**main branch** instead of using the **PR context**.
## The Fix
Remove the explicit `repository` parameter when checking out the same
repository:
```yaml
- name: Checkout ComfyUI_frontend
uses: actions/checkout@v5
with:
path: 'ComfyUI_frontend' # No repository parameter = uses PR context
```
## Changes Made
- ✅ Removed `repository: 'Comfy-Org/ComfyUI_frontend'` from setup job
checkout
- ✅ Removed `repository: 'Comfy-Org/ComfyUI_frontend'` from
merge-reports job checkout
- ✅ Updated setup-frontend action to use `actions/checkout@v5` for
consistency
- ✅ Simplified workflow by removing unnecessary `ref` and `fetch-depth`
parameters
## How This Fixes the Problem
**Before:**
- Setup job checked out main branch (due to explicit repository)
- Tests ran PR code against main branch snapshots
- Results varied based on what was in main at the time
**After:**
- Setup job checks out PR merge commit (natural PR context)
- Tests run PR code against PR snapshots
- Results are consistent and reproducible
## Why It Worked Before (Sometimes)
The explicit `repository` parameter has been there for a long time, but
the issue became more apparent recently due to:
1. GitHub Actions behavior changes over time
2. Increased frequency of main branch updates
3. More sensitive screenshot comparison tests
4. Complex cache/restore workflow where timing mattered
The fix ensures deterministic behavior regardless of GitHub's internal
changes.
## Testing
This change makes the CI behavior explicit and predictable:
- ✅ PR tests will always use PR context
- ✅ Push tests will always use pushed commit
- ✅ No dependency on GitHub's default behavior interpretation
- ✅ Simplified workflow with fewer moving parts
Resolves the issues described in `.github/workflows/problem.log`.
┆Issue is synchronized with this [Notion
page](https://www.notion.so/PR-5950-Fix-CI-Remove-explicit-repository-parameter-causing-non-reproducible-test-results-2846d73d36508159a848c4a2e14a0fb1)
by [Unito](https://www.unito.io)
Co-authored-by: Alexander Brown <drjkl@comfy.org>