chrome-release-verify

End-to-end Chrome security backport for an Electron release branch. Given a Chrome Releases blog URL and a branch (e.g. 41-x-y), determines which CVE fixes are missing from the *actual synced source*, writes the cherry-pick patches locally, validates them with `e sync --3` + `lint --patches`, then pushes a single PR. Use when asked to backport a Chrome security release to N-x-y, "is CVE-X already in N-x-y?", or to produce/validate the cherry-pick set for a release branch.

Install
npx skills add 'https://github.com/electron/electron/tree/main/.claude/skills/chrome-release-verify'
Download bundle ↓
main · b1525cfScanned 2026-09-17

Contributors

GitHub-linked commit authors for this SKILL.md at the saved revision. Co-authors and history before file renames are not included.

File history ↗

SKILL.md

SKILL.mdBrowse 1 file
View on GitHub

Chrome Release → Validated Backport PR

Input: $ARGUMENTS = <release-branch> <chrome-releases-blog-url> (e.g. 41-x-y https://chromereleases.googleblog.com/2026/04/stable-channel-update-for-desktop_15.html). Ask if either is missing.

The flow is local-first: nothing is pushed until every patch applies via e sync --3 and passes lint --patches.

1. Map CVE → bug → fix CL

Run /chrome-release-cls <blog-url> (or its inline procedure) to produce /tmp/cve_bugs.txt (CVE|bug|severity|desc) and a per-bug canonical fix CL. For each CL also note repo (path under src/: ., v8, third_party/{skia,angle,pdfium,dawn}, third_party/libaom/source/libaom) and gerrit-host.

Prefer the target-milestone merge CL if one exists (e.g. on 41-x-y ≈ M146, prefer the [M146] cherry-pick over the main CL) — it's already rebased and far less likely to conflict. Find it via git log --all --grep on the Change-Id, or Gerrit ?q=bug:<n>. If Chrome did not merge a fix to the target milestone, that is not by itself a reason to skip — Chrome's merge window is shorter than Electron's support window. Fall back to the main CL (or a nearby-milestone rebase, see step 5) and let step 3 decide whether the vulnerable code exists on the branch.

2. Prepare a synced worktree

Reuse bp-<NN> from e show configs if present, else e worktree add bp-<NN> ~/src/electron-bp-<NN> --source <current> --no-sync.

cd <root>/src/electron
git fetch origin <branch>
git checkout -B security-backport/<branch>/<short-date> origin/<branch>
e use bp-<NN>
e sync 2>&1 | tee /tmp/bp_sync.log

If sync fails with NotADirectoryError: '<root>/src/.git/objects/info/alternates', remove GIT_CACHE_PATH from the bp config's env and retry.

3. Verify IN-TREE vs NEEDS-BACKPORT

For each bug, three checks against the synced repo:

  1. git -C "$repo" log HEAD --since='1 year ago' -E --grep="\b${bug}\b" --format='%h %s'
  2. Fetch Change-Id from Gerrit, then git log HEAD --grep="^Change-Id: ${cid}$"
  3. grep -rlE "(\b${bug}\b|${cid})" <root>/src/electron/patches/

Any hit ⇒ IN-TREE. All empty ⇒ NEEDS-BACKPORT.

Rows that step 1 (or an upstream triage report you were handed) left as RESTRICTED / no public CL are not inherited — re-resolve them here before classifying. Run git -C "$repo" fetch origin and git -C "$repo" log --all --remotes -E --grep="(Bug|Fixed):.*\b${bug}\b" --format='%h %s', the chrome-release-cls skill's gitiles_grep on refs/heads/main and the branch-heads refs of chromium/src, v8 and the component's repo, and a fresh Gerrit ?q=bug:<id> / ?q=message:<id> query. The Gerrit change record is often restricted while the fix commit is already public in git, and restrictions lift during a run, so an earlier RESTRICTED verdict is a hint, not a result. A row stays RESTRICTED only when all of those are empty; it is never IN-TREE by omission — list it separately in the report so a human can follow up.

For each NEEDS-BACKPORT CL, also fetch its file list (/changes/<proj>~<cl>/revisions/current/files) and skip if every file is under chrome/browser/, chrome/android/, ios/, or components/**/android/ — Electron doesn't compile those.

Report the table now (CVE | Sev | Bug | Component | Verdict | CL) and the proposed backport set; get user sign-off before continuing.

4. Write patches locally (no push yet)

For each backport CL, fetch the raw patch and write it into patches/<dir>/:

curl -s "https://${host}.googlesource.com/changes/${proj//\//%2F}~${cl}/revisions/current/patch" \
  | base64 -d > "patches/${dir}/cherry-pick-${short}.patch"
echo "cherry-pick-${short}.patch" >> "patches/${dir}/.patches"

For repos with no Gerrit host e cherry-pick supports (e.g. libaom on aomedia), instead git cherry-pick the upstream commits onto the synced sub-repo HEAD and git format-patch the result.

For any newly-created patches/<dir>/, append to patches/config.json preserving the compact one-line-per-entry style:

  { "patch_dir": "src/electron/patches/<dir>", "repo": "src/third_party/<dir-or-nested-path>" }

5. Validate with e sync --3

e sync --3 2>&1 | tee /tmp/bp_sync3.log

On Patch failed at NNNN <subject>:

  • cd into the failing repo, inspect git diff for conflict markers.
  • Test-only files (e.g. web_tests/VirtualTestSuites, *_unittest.cc context drift): take ours (git checkout --ours -- <file>) if the security-relevant hunks merged cleanly.
  • Substantive code conflicts: a conflict is not a reason to drop. Work through, in order:
    1. A merge of the same Change-Id on the target milestone (git log --all --grep on the Change-Id, or Gerrit ?q=change:<Change-Id>) — swap to it.
    2. The same Change-Id on a nearby milestone (the author's rebase, e.g. an [M152] CL when the target is M150 — status NEW is fine), and any prerequisite CL the fix was written against (the refactor that introduced the helpers or signatures the fix touches). Pick those so the fix applies verbatim; list prerequisites in the report.
    3. Otherwise hand-port the fix's semantics onto the branch's code, inside the pick's patch file: resolve the hunks in the target repo, fold them into the pick's commit (git commit --fixup + autosquash rebase), then re-export. Keep the upstream commit header verbatim and append an Electron adaptation notes (<branch>) paragraph to the patch header describing every adaptation (renamed member, changed signature, helper inlined, ...). Adapt mechanically and structurally only — never change what the fix checks or guards, and never invent semantics the upstream diff doesn't have. Mark the row BACKPORTED (adapted) and flag it for reviewer attention in the report and the PR body. An adapted pick is held to the same bar as any other: e sync --3, lint --patches, and whatever build gate the caller imposes.
  • DROPPED is reserved for two cases: (a) the vulnerable code is absent on the branch (file, function or feature doesn't exist), or (b) a subsystem rewrite where the fix has no counterpart to attach to. Every DROPPED row must say which, with the evidence. Remove a dropped patch fully (delete the file, remove from .patches and config.json).
  • After resolving: git add <files> && git -c commit.gpgsign=false am --continue, then e patches <repo> to export the resolved patch, then re-run e sync --3. Repeat until clean.

6. Export → lint → re-apply loop

e patches all
node script/lint.js --patches   # must exit 0

If lint reports findings (typically trailing whitespace on + content lines), fixing them changes the bytes the patch writes, which invalidates the index <old>..<new> blob hashes that e patches baked in. Hand-editing a .patch and pushing it as-is will pass lint locally but fail CI's Apply Patches re-export check with a one-line index hash diff.

So whenever lint (or you) modifies any .patch file after export, round-trip once more:

# fix the lint findings in patches/**/*.patch, then:
e sync            # re-apply the edited patches (no --3 needed; they applied cleanly last time)
e patches all     # re-export so index blob hashes match the edited content
node script/lint.js --patches      # must now exit 0
git diff --quiet -- patches/ || { echo "patches changed again — repeat the loop"; }

Repeat until lint --patches exits 0 and git diff -- patches/ is empty after the final e patches all. Only then is the patch set CI-stable.

7. Commit, push, PR

git add patches/
git commit -m "chore: cherry-pick <N> changes from <dirs>"
git push origin HEAD
gh pr create --repo electron/electron --base <branch> --head <this-branch> \
  --title "chore: cherry-pick <N> changes from <dirs>" \
  --label "<branch>" --label backport-check-skip --label semver/patch --label "security 🔒" \
  --body-file /tmp/pr_body.md

PR body format — follow .github/PULL_REQUEST_TEMPLATE.md (Description of Change / Checklist / Release Notes sections; PRs that don't follow it are auto-closed):

#### Description of Change

Backports the following changes:

* [`<shortCommit>`](<gerrit-CL-url>) from <patchDir> — <subject>
* [`<shortCommit>`](<gerrit-CL-url>) from <patchDir> — <subject> (adapted for <branch>; see patch header)
* ...

#### Checklist

- [x] I have filled out the PR description
- [x] [I have reviewed and verified the changes](https://github.com/electron/governance/blob/main/policy/ai.md)

#### Release Notes

Notes: Backported fixes from upstream Chromium and V8.

Keep the PR body low-key about security. The PR is public the moment it opens, while the fixes only protect users after a release ships — don't advertise the exploit map in between. Concretely: no CVE numbers, no crbug.com links, no "security"/"0-day"/severity wording anywhere in the title, body, or Notes: line. The security 🔒 label is fine (it's how releases track these). Upstream commit subjects stay verbatim even when they say "UAF" — rewriting them breaks traceability to the Gerrit CL. Adjust the Notes: wording to match what's actually in the set (e.g. drop "and V8" if no v8 patches).

Keep the CVE↔CL mapping out of the PR entirely: leave it in the local notes file (/tmp/cve_cls.txt) and report it to the user in chat, including any adapted, dropped or skipped fixes with reasons.

Restore e use <previous> when done.

Discovery context

Discovered by repository scan. No exact path reference found in the snapshot’s root CLAUDE.md.