Cache Playwright browsers in CI to avoid redundant downloads

[AndyTWF]

Summary

• Cache Playwright browser binaries using actions/cache@v4 keyed on Playwright version and OS
• On cache hit, skip browser download (~30-60s) and only install system-level apt dependencies via playwright install-deps
• On cache miss, do the full playwright install --with-deps as before
• Applied to both e2e-tests.yml (1 job) and e2e-web-cli-parallel.yml (5 jobs sharing the same cache key)

Why

Playwright browsers were downloaded fresh on every CI run. In the parallel web CLI workflow, this happened 5 times per run. Caching eliminates redundant downloads, speeds up CI, and reduces flakiness from download failures.

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
cli-web-cli	Ready	Preview, Comment	Apr 8, 2026 11:10pm

[claude-code-ably-assistant]

Walkthrough

This PR adds Playwright browser caching to the CI workflows to avoid redundant browser downloads on every run. It applies the same cache-then-conditional-install pattern to both e2e-tests.yml and e2e-web-cli-parallel.yml, using actions/cache@v4 keyed on Playwright version and OS. The parallel web CLI workflow benefits most — previously 5 jobs each downloaded browsers independently; now they share a single cache entry.

Changes

Area	Files	Summary
Config	.github/workflows/e2e-tests.yml	Added Playwright version detection, cache step, and conditional install/install-deps steps (1 job)
Config	.github/workflows/e2e-web-cli-parallel.yml	Same caching pattern applied to all 5 parallel test jobs

Review Notes

• Cache key design: Keyed on runner.os + Playwright version string (from pnpm exec playwright --version). This is correct — the cache will automatically invalidate when the Playwright version changes.
• On cache hit: Only playwright install-deps is run (installs system-level apt dependencies, not the browser binaries). This is the right approach since apt deps aren't cached.
• On cache miss: Falls back to playwright install --with-deps as before — no behavioral change for cold runs.
• e2e-tests.yml installs all browsers (playwright install --with-deps without a browser arg); e2e-web-cli-parallel.yml installs only chromium. Both are consistent before and after this change.
• No source code, test logic, or dependency changes — purely CI infrastructure.

🤖 Generated with Claude Code

[claude-code-ably-assistant]

Review: Cache Playwright browsers in CI

This PR only modifies GitHub Actions workflow files — no TypeScript source, commands, or tests were changed, so the CLI command review checklist doesn't apply.

Summary

File	Status	Notes
.github/workflows/e2e-tests.yml	OK	Caching added correctly to the single job
.github/workflows/e2e-web-cli-parallel.yml	OK	Caching added to all 5 jobs; setup seeds cache, test jobs restore it

Logic review

The pattern used is the canonical approach for Playwright browser caching:

1. Capture Playwright version → use as cache key
2. Restore ~/.cache/ms-playwright from cache
3. On miss: playwright install --with-deps (downloads browsers + installs system deps)
4. On hit: playwright install-deps (only installs system deps, browsers already present)

This is correct. The setup job in e2e-web-cli-parallel.yml runs first and writes the cache; the four downstream test jobs (auth-tests, session-tests, ui-tests, rate-limit-test) all restore from the same key — so only the setup job bears the download cost after the first run.

One observation (not a blocker)

pnpm exec playwright --version outputs Version 1.48.0 (with the Version prefix). The resulting cache key is playwright-Linux-Version 1.48.0 — note the space and capital V. GitHub Actions does allow spaces in cache keys (the docs only prohibit commas), so this won't break anything. If you want a cleaner key you could strip the prefix with pnpm exec playwright --version | sed 's/Version //', but it's not necessary.

Overall

Correct, clean, and well-structured. The setup job seeding the cache for all parallel test runners is the right architecture here. Good to merge.