deckhearth/.convoys/harden-visual-diff-gate.md

---
slug: harden-visual-diff-gate
status: shipping
opened: 2026-06-12
owner: rstillw
prerequisites:
  - Stable `main` post-`unify-glass-panel-surfaces` /
    `cleanup-card-item-list-and-share-modal-palette` /
    `migrate-button-input-mobilenav-to-glass-primitive` (all merged)
  - CT 111 self-hosted runner online (resolved by PR #132, 2026-06-07)
related:
  - PR #58 (`83a358b`) — initial Linux baseline seeded
  - PR #18 (`7b6f751`) — `adopt-playwright-smoke` Decision 4: `continue-on-error` end state
  - `tests/visual/homepage.spec.ts` module docblock
  - `.github/workflows/visual-diff.yml` line ~126
---

# harden-visual-diff-gate

## Problem

`.github/workflows/visual-diff.yml` still carries `continue-on-error: true`
on the "Capture screenshots (PR)" step. This was the documented end state
of `adopt-playwright-smoke` (Decision 4) because no Linux baseline existed
yet. PR #58 seeded the baseline on 2026-06-02, but the gate was never
flipped — visual drift continues to ship as artifacts + PR comments only,
not as a blocking check.

Additionally, the committed baseline (`tests/visual/__screenshots__/home.png`)
was generated against `main` at `83a358b`, which is several convoys behind
the current homepage rendering (glass redesign briefs, corner-light tone
adjustments, card-class retirement, etc.). Flipping `continue-on-error`
today would fail every PR touching `pages/**` / `components/**` / `styles/**`
against this stale reference.

## Two-step shape

This is one convoy with a strict ordering constraint:

### Step 1 — Re-seed the baseline against current `main`  *(SHIPPED)*

**Resolved via Option B** (see D1 below). New workflow
`.github/workflows/seed-visual-baselines.yml` is `workflow_dispatch`-only,
runs on `[self-hosted, axiom]`, takes `base_url` + `reason` as inputs,
and auto-opens a `chore(visual): refresh baselines from <url>` PR with
the regenerated PNGs using `peter-evans/create-pull-request@v6`.

The workflow is byte-equivalent to `visual-diff.yml` (same Chromium
version via shared `package-lock.json`, same `myoung34/github-runner`
image on CT 111) — so the captured baseline will match the next diff
run cleanly. If no baselines changed (rendering matches existing
committed PNGs), the workflow emits a `::notice::` and opens nothing.

**Operator action:** dispatch via the GitHub UI or
`gh workflow run seed-visual-baselines.yml -f base_url=<url> -f reason="..."`,
then review + merge the resulting `chore(visual):` PR. The "Known
staleness" callout in `AGENTS.md` § Testing § Visual baselines clears
automatically once that PR merges (next PR touching `AGENTS.md` should
sweep the line).

### Step 2 — Flip the gate  *(SHIPPED via PR #140)*

PR #139 (`54495fe`) landed the fresh baseline. PR #140 then:

- Removed `continue-on-error: true` from
  `.github/workflows/visual-diff.yml`'s `Capture screenshots (PR)` step.
- Added a 9th `forbidden-patterns` check in `.github/workflows/ci.yml`
  that greps `visual-diff.yml` for `^\s*continue-on-error:\s*true` and
  fails the build if it returns (Risk #3 mitigation made concrete).
  Scoped narrowly to that one file — other workflows (e.g.
  `seed-visual-baselines.yml`'s PR-open step, see D4 below)
  legitimately use the flag.
- Rewrote `AGENTS.md` § Testing § Screenshot diff to lead with "hard
  merge gate", document the intentional-change runbook (dispatch seed
  workflow → manually open PR → merge → re-run), and reference the
  new ci.yml check.
- Rewrote the module docblock in `tests/visual/homepage.spec.ts` to
  match the AGENTS.md runbook and drop the "advisory, not gating"
  language.

`maxDiffPixelRatio` (D2 below) intentionally not touched — left at the
Playwright default of 0 (any pixel diff fails). If subpixel jitter from
a future runner-image bump becomes a problem, that becomes the trigger
for a new convoy (`tune-visual-diff-tolerance`) rather than a quiet
config bump.

## Decisions to ratify

- **D1.** *(ratified)* Option B (new `seed-visual-baselines.yml`
  workflow) chosen over Option A (ad-hoc SSH). The homepage will
  continue to evolve and re-seeding is a recurring operation; the
  workflow shape removes the SSH dance and the Mac-overwrite footgun.
  Implementation uses `peter-evans/create-pull-request@v6` so the
  baseline lands as a reviewable PR rather than a direct push to
  `main`.
- **D2.** `maxDiffPixelRatio` value. Defer to operator preference; 0 is
  the strictest and what we currently use implicitly via Playwright
  defaults. 0.001-0.01 is a reasonable cushion.
- **D3.** Do we also harden `preview-smoke.yml`? It already runs as a
  blocking gate (no `continue-on-error`); no change needed. This convoy
  is scoped to `visual-diff.yml` only.

- **D4.** *(emerged during Step 1 dispatch, 2026-06-13)* How to handle
  the `peter-evans/create-pull-request@v6` PR-open failure caused by
  the stwl-labs org-level "Allow GitHub Actions to create and approve
  pull requests" setting being OFF. Three options:
  - **A. Flip the org setting on** — fastest path; but a real surface
    increase (any future PR-creating workflow could push untrusted
    PRs). The org default is OFF for a reason.
  - **B. Use a fine-grained PAT scoped to `tcg-vault` +
    `pull-requests:write`** seeded as `secrets.HOMELAB_CI_PAT` —
    sidesteps the org setting but adds another rotating secret.
  - **C. Accept the limitation; document the manual `gh pr create`
    step as the runbook.** Operator dispatches the workflow, the
    branch pushes cleanly, the PR-open step fails-soft with a loud
    notice telling the operator the exact `gh pr create` command.

  **Chosen: C.** Baseline regen is a low-frequency operation (~once
  per major UI change), the manual step adds ~30s of operator time,
  and there's no new attack surface or secret to rotate. PR #140
  updated the workflow to (a) flag the create-PR step with
  `continue-on-error: true` (narrowly scoped, with an inline rationale
  callout distinguishing it from the just-removed `visual-diff.yml`
  flag), (b) disambiguate the three possible outcomes (no-changes /
  pr-opened / branch-pushed-pr-blocked) via a `git ls-remote` check on
  the bot branch, and (c) exit non-zero on outcome C so the workflow
  run shows red and the operator can't miss the followup.

## Risks

| # | Risk | Mitigation |
|---|---|---|
| 1 | Re-seeded baseline drifts on next runner-image update | Watchtower's weekly update cycle (CT 111) could shift Chromium subpixel rendering. Mitigate via D2's `maxDiffPixelRatio` cushion |
| 2 | Step 1's baseline PR fails its own `Screenshot diff` | Expected — the new baseline IS the visual change. Use `pipeline: skip visual` directive in the PR body OR open the baseline PR with the `skip-metrics` label equivalent for visual-diff (currently none — would need a new bypass mechanism). Cleaner: land the baseline PR via the `tests/visual/**` path filter, which DOES trigger visual-diff but the new screenshot vs new baseline should match by construction |
| 3 | Step 2 lands but a third party reverts `continue-on-error` later | Add a 9th check to `forbidden-patterns` in `ci.yml`: `grep -n 'continue-on-error' .github/workflows/visual-diff.yml` should return zero |

## Non-goals

- Adding more visual baselines (login page, dashboard, etc.) — out of
  scope. The single homepage baseline is the smoke test of the
  visual-diff pipeline; deeper coverage is per-feature work.
- Switching to a hosted visual-regression service (Percy, Chromatic,
  Argos) — handles the platform problem cleanly but adds a paid
  dependency. Local + axiom is free and works.

## Acceptance

- `visual-diff.yml` gates merge (failed diff = red required check).
- Baseline regenerable via a reviewable Git-native workflow (no SSH
  required, no Mac-overwrite-Linux footgun).
- `AGENTS.md` + `tests/visual/homepage.spec.ts` docblock no longer
  describe the gate as advisory.
- Optional 9th forbidden-patterns check locks the gate in place.