[eric] ci: parallelize + cache the e2e gate to cut push wall-clock under ~10 min

- split the single windows e2e job into parallel jobs: gate (pure-node
  selftests, no build), verify (build + verify-all), playwright (build +
  renderer suite). verify and playwright run concurrently so wall-clock is
  max(verify, playwright), not their sum
- build the UNPACKED app on the gate path (electron-builder --dir via a new
  -DirOnly switch in build-app-win.ps1), skipping the ~2min NSIS LZMA
  compression. verify-all drives win-unpacked\OpenSwarm.exe; afterPack +
  locale paks still run in the pack phase, and verify-update-feed skips
  cleanly when no installer feed is present
- cache the heavy build inputs (bundled Python env, uv, MCP bundles) + npm
  across runs, keyed on their source manifests; the build script already skips
  any input already on disk, so warm builds are fast
- move the destructive installer cycle to its own job gated off routine pushes
  (runs on PR-to-main / dispatch; release-windows.yml covers v* tags)
- apply the same caching to dogfood (shares cache keys, so the two warm each other)
This commit is contained in:
Eric
2026-05-28 23:55:32 -07:00
parent 99f1e56497
commit 2b4d90edaa
3 changed files with 182 additions and 91 deletions
+24 -1
View File
@@ -49,10 +49,33 @@ jobs:
steps:
- uses: actions/checkout@v4
- uses: actions/setup-node@v4
with: { node-version: '20.18.1' }
with:
node-version: '20.18.1'
cache: npm
cache-dependency-path: |
electron/package-lock.json
frontend/package-lock.json
- uses: actions/setup-python@v5
with: { python-version: '3.13' }
# Reuse the heavy build inputs (shares keys with e2e.yml so the two warm
# each other's caches); the build script skips any input already on disk.
- name: Cache bundled Python env
uses: actions/cache@v4
with:
path: electron/python-env
key: pyenv-win-${{ hashFiles('scripts/build-python-env-win.ps1', 'backend/requirements.txt') }}
- name: Cache uv binaries
uses: actions/cache@v4
with:
path: backend/uv-bin
key: uvbin-win-${{ hashFiles('scripts/build-app-win.ps1') }}
- name: Cache MCP bundles
uses: actions/cache@v4
with:
path: backend/mcp-bundles
key: mcpbundles-win-${{ hashFiles('scripts/build-app-win.ps1') }}
- name: Build packaged app (Windows)
shell: pwsh
run: pwsh -NoProfile -File scripts/build-app-win.ps1
+147 -89
View File
@@ -1,38 +1,26 @@
name: E2E (packaged app, Windows)
# Builds the UNSIGNED packaged app on each OS, then runs the FULL deterministic
# gate against the real artifact plus the Playwright smoke. Signing/notarization +
# release upload are handled separately by release-windows.yml / release-macos.yml.
# Fast-feedback packaged-app gate, Windows-only. Split into parallel jobs so the
# push wall-clock stays well under ~10 min:
# gate - cheap pure-node selftests (no build): mutation gate + preflight
# Layers 1-5. Fails the "tests that test the tests" fast.
# verify - build the UNPACKED app (electron-builder --dir, which skips the
# slow ~2min NSIS LZMA compression) and run the deterministic
# verify-all gate against win-unpacked\OpenSwarm.exe.
# playwright - build the unpacked app and run the renderer-level Playwright e2e.
# installer - full NSIS build + destructive install->verify->uninstall. The
# heaviest leg, so it runs only on PR-to-main / dispatch (NOT on
# routine pushes); release-windows.yml covers it on v* tags.
# verify + playwright run concurrently; both reuse cached heavy build inputs
# (bundled Python env, uv, MCP bundles, npm) so warm builds are fast - the build
# script skips any input already on disk.
#
# The gate, per leg:
# 1. selftest-gate.js - "test the tests": mutation-checks the gate's own logic
# so a weakened guard fails loudly (runs first, no build needed).
# 2. build the packaged app.
# 3. verify-all.js - boot/paint/serve/provenance + signature state +
# resilience (locked-port, double-launch) + network/auth/9router. Provenance
# goes GREEN here because the build is AT the checked-out commit.
# 4. mcp/selftest.js - proves the Playwright "GUI hand" can launch + drive +
# read the real app.
# 5. playwright npm test - the renderer-level smoke.
# 6. (windows) verify-installer --destructive - the runner is a CLEAN box with no
# prior install, so the real install -> verify -> uninstall cycle is safe here
# (it refuses to run where an install already exists).
#
# Matrix: windows-latest (Win Server 2022) only. The macOS legs were removed -
# GitHub's scarce mac runners left them perpetually queued/starved and surfaced
# mac-only failures we cannot triage from the Windows dev box, so the whole
# matrix read red. Re-add macos-14/macos-13 (matrix edit + workflow_dispatch)
# when a Mac maintainer can own them. GitHub retired windows-2019 and ships no
# Win10 client image, so true Win10 coverage needs the SELF-HOSTED e2e-win10 job
# (activates when a runner labeled [self-hosted, windows, win10] exists and repo
# var WIN10_SELF_HOSTED=true).
# macOS legs were removed (runner starvation + untriageable mac-only failures);
# re-add when a Mac maintainer can own them. Real Win10 coverage still needs the
# SELF-HOSTED e2e-win10 job (repo var WIN10_SELF_HOSTED=true + a
# [self-hosted, windows, win10] runner).
on:
# This is the EXPENSIVE leg (packaged-app build + e2e matrix). Gate it so it only
# runs when code that actually affects the built artifact changes - not docs,
# .gitignore, or unrelated CI meta - plus on PRs into main and on demand. The
# cheap gates (hermetic phase-tests, gitleaks) keep running on every push via
# their own workflows, so every push still gets fast signal.
push:
branches: [eric/lock]
paths: &build-paths
@@ -41,8 +29,8 @@ on:
- 'backend/**'
- 'e2e/**'
- 'scripts/build-app**'
- 'scripts/fetch-router**'
- 'scripts/build-python-env**'
- 'scripts/fetch-router**'
- 'scripts/ci/**'
- '.github/workflows/e2e.yml'
pull_request:
@@ -53,37 +41,24 @@ on:
permissions:
contents: read
env:
CSC_IDENTITY_AUTO_DISCOVERY: 'false'
GOOGLE_OAUTH_CLIENT_ID: 'e2e-placeholder.apps.googleusercontent.com'
GOOGLE_OAUTH_CLIENT_SECRET: 'e2e-placeholder-secret'
jobs:
e2e:
strategy:
fail-fast: false
matrix:
# Windows-only. macOS legs removed: runner starvation + untriageable
# mac-only failures kept the matrix red. Re-add macos-14/macos-13 here
# (with workflow_dispatch) when a Mac maintainer can own them.
os: [windows-latest]
runs-on: ${{ matrix.os }}
timeout-minutes: 90
env:
CSC_IDENTITY_AUTO_DISCOVERY: 'false'
GOOGLE_OAUTH_CLIENT_ID: 'e2e-placeholder.apps.googleusercontent.com'
GOOGLE_OAUTH_CLIENT_SECRET: 'e2e-placeholder-secret'
# Pure-node, no build: cheap enough to always run and fail fast.
gate:
runs-on: windows-latest
timeout-minutes: 15
steps:
- uses: actions/checkout@v4
- uses: actions/setup-node@v4
with:
node-version: '20.18.1'
- uses: actions/setup-python@v5
with:
python-version: '3.13'
# Test the tests first: cheap, pure, no build. If a guard stopped
# discriminating good from broken, fail before spending a build.
- name: Gate selftest (mutation)
shell: bash
run: node scripts/ci/selftest-gate.js
# Preflight Layer 1+2+3+4 (unit, hang fuzz, failure rigs, race) - pure node, ~5s.
- name: Preflight selftest (Layers 1+2)
shell: bash
run: node scripts/ci/selftest-preflight.js
@@ -96,77 +71,160 @@ jobs:
- name: Pairwise generator selftest (covering-array math)
shell: bash
run: node scripts/ci/selftest-pairwise.js
- name: Preflight matrix (normal + hostile-env scenarios)
shell: bash
run: |
node scripts/ci/verify-preflight.js
OPENSWARM_TEST_NETWORK=blocked node scripts/ci/verify-preflight.js
OPENSWARM_TEST_APPDATA=readonly node scripts/ci/verify-preflight.js
OPENSWARM_TEST_LANG=de-DE node scripts/ci/verify-preflight.js
# Preflight Layer 5: hostile-env scenarios. Each leg asserts the verdict
# matches the expected envelope for that scenario; network=blocked
# producing 'fail' instead of 'warn' is the false-positive bug class.
- name: Preflight matrix (normal)
shell: bash
run: node scripts/ci/verify-preflight.js
- name: Preflight matrix (network=blocked must -> warn)
shell: bash
env: { OPENSWARM_TEST_NETWORK: blocked }
run: node scripts/ci/verify-preflight.js
- name: Preflight matrix (appdata=readonly must -> fail)
shell: bash
env: { OPENSWARM_TEST_APPDATA: readonly }
run: node scripts/ci/verify-preflight.js
- name: Preflight matrix (lang=de-DE must stay -> ok)
shell: bash
env: { OPENSWARM_TEST_LANG: de-DE }
run: node scripts/ci/verify-preflight.js
- name: Build packaged app (Windows)
# Build the unpacked app + run the deterministic gate.
verify:
runs-on: windows-latest
timeout-minutes: 30
steps:
- uses: actions/checkout@v4
- uses: actions/setup-node@v4
with:
node-version: '20.18.1'
cache: npm
cache-dependency-path: |
electron/package-lock.json
frontend/package-lock.json
- uses: actions/setup-python@v5
with:
python-version: '3.13'
- name: Cache bundled Python env
uses: actions/cache@v4
with:
path: electron/python-env
key: pyenv-win-${{ hashFiles('scripts/build-python-env-win.ps1', 'backend/requirements.txt') }}
- name: Cache uv binaries
uses: actions/cache@v4
with:
path: backend/uv-bin
key: uvbin-win-${{ hashFiles('scripts/build-app-win.ps1') }}
- name: Cache MCP bundles
uses: actions/cache@v4
with:
path: backend/mcp-bundles
key: mcpbundles-win-${{ hashFiles('scripts/build-app-win.ps1') }}
- name: Build packaged app (unpacked, no installer)
shell: pwsh
run: pwsh -NoProfile -File scripts/build-app-win.ps1
run: pwsh -NoProfile -File scripts/build-app-win.ps1 -DirOnly
- name: Deterministic gate (verify-all)
shell: bash
run: node scripts/ci/verify-all.js
# Build the unpacked app + run the Playwright renderer suite, concurrently with verify.
playwright:
runs-on: windows-latest
timeout-minutes: 30
steps:
- uses: actions/checkout@v4
- uses: actions/setup-node@v4
with:
node-version: '20.18.1'
cache: npm
cache-dependency-path: |
electron/package-lock.json
frontend/package-lock.json
e2e/package-lock.json
- uses: actions/setup-python@v5
with:
python-version: '3.13'
- name: Cache bundled Python env
uses: actions/cache@v4
with:
path: electron/python-env
key: pyenv-win-${{ hashFiles('scripts/build-python-env-win.ps1', 'backend/requirements.txt') }}
- name: Cache uv binaries
uses: actions/cache@v4
with:
path: backend/uv-bin
key: uvbin-win-${{ hashFiles('scripts/build-app-win.ps1') }}
- name: Cache MCP bundles
uses: actions/cache@v4
with:
path: backend/mcp-bundles
key: mcpbundles-win-${{ hashFiles('scripts/build-app-win.ps1') }}
- name: Build packaged app (unpacked, no installer)
shell: pwsh
run: pwsh -NoProfile -File scripts/build-app-win.ps1 -DirOnly
- name: Install e2e deps
shell: bash
working-directory: e2e
env:
PLAYWRIGHT_SKIP_BROWSER_DOWNLOAD: '1'
run: npm ci
- name: GUI hand selftest (MCP)
shell: bash
working-directory: e2e
run: node mcp/selftest.js
- name: Run E2E (Playwright)
shell: bash
working-directory: e2e
run: npm test
- name: Installer cycle (Windows, clean runner)
if: startsWith(matrix.os, 'windows')
shell: pwsh
run: node scripts/ci/verify-installer.js --destructive
- name: Upload E2E results
if: always()
uses: actions/upload-artifact@v4
with:
name: e2e-results-${{ matrix.os }}
name: e2e-results
path: e2e/results.json
if-no-files-found: ignore
retention-days: 14
# Per-test traces: playwright-trace.zip (open via `npx playwright show-trace`),
# events.jsonl (unified timestamped stream), mousepath.jsonl, backend.log.tail.
# Always-upload so a failed run is debuggable without re-running.
# Per-test traces: playwright-trace.zip, events.jsonl, mousepath.jsonl,
# backend.log.tail - always uploaded so a failed run is debuggable.
- name: Upload E2E visibility traces
if: always()
uses: actions/upload-artifact@v4
with:
name: e2e-traces-${{ matrix.os }}
name: e2e-traces
path: e2e/traces/
if-no-files-found: ignore
retention-days: 14
# Heaviest leg: full NSIS installer + destructive install->verify->uninstall on
# a clean runner. Skip on routine pushes; run on PRs into main + manual dispatch.
installer:
if: github.event_name != 'push'
runs-on: windows-latest
timeout-minutes: 45
steps:
- uses: actions/checkout@v4
- uses: actions/setup-node@v4
with:
node-version: '20.18.1'
cache: npm
cache-dependency-path: |
electron/package-lock.json
frontend/package-lock.json
- uses: actions/setup-python@v5
with:
python-version: '3.13'
- name: Cache bundled Python env
uses: actions/cache@v4
with:
path: electron/python-env
key: pyenv-win-${{ hashFiles('scripts/build-python-env-win.ps1', 'backend/requirements.txt') }}
- name: Cache uv binaries
uses: actions/cache@v4
with:
path: backend/uv-bin
key: uvbin-win-${{ hashFiles('scripts/build-app-win.ps1') }}
- name: Cache MCP bundles
uses: actions/cache@v4
with:
path: backend/mcp-bundles
key: mcpbundles-win-${{ hashFiles('scripts/build-app-win.ps1') }}
- name: Build packaged app (full NSIS installer)
shell: pwsh
run: pwsh -NoProfile -File scripts/build-app-win.ps1
- name: Installer cycle (clean runner)
shell: pwsh
run: node scripts/ci/verify-installer.js --destructive
# Real Windows 10 coverage. Skipped unless a self-hosted Win10 runner exists
# and WIN10_SELF_HOSTED=true (repo variable). Mirrors the windows steps above.
e2e-win10: