Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Open the follow-up "backmerge" PR that lands a release's version bump + changelog onto main. Use after (or alongside) creating a release PR for an SDK pod package.
.claude/skills/tetherto-qv-sdk-backmerge/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-07 | ✗→✓ | ▲ Improved | 145% | 0% |
| case-08 | ✗→✓ | ▲ Improved | 54% | 0% |
| case-09 | ✗→✓ | ▲ Improved | 36% | 0% |
| case-10 | ✗→✓ | ▲ Improved | 154% | 0% |
| case-11 | ✗→✓ | ▲ Improved | 199% | 0% |
Create the backmerge PR that keeps main aligned with what shipped on a release-<pkg>-<x.y.z> branch, per docs/gitflow.md "Keep main aligned" sections.
Applies to SDK pod packages as defined in .cursor/rules/sdk/sdk-pod-packages.mdc.
Use when:
release-<pkg>-<x.y.z>/qv-sdk-backmergesdk-pr-create chains into this flow automatically when the target is a release branch (see that skill's "Release Target Dual-PR Flow")The backmerge PR carries the version bump + changelog metadata from the release branch onto main so future development sees it. It is tagged [skiplog] to keep it out of subsequent changelogs.
sdk-pr-create): release PR number/URL, release branch, source head branch, ticket--package=<pkg> --version=<x.y.z> and one of --source=<head-branch> or --commit=<sha>release-<pkg>-<x.y.z>, derive package, version, and source from itSame policy as qv-sdk-pr-create: prefer pushing the backmerge head to the org remote (tetherto/qvac) and opening a same-repo PR. Personal-fork heads are a fallback and count as external for CI (fork-ci environment approval required per run on the current head SHA).
In command examples below, ORG_REMOTE / FORK_REMOTE are placeholders — substitute the resolved remote names from Step 1. Do not run those tokens literally.
gh is installed and authenticated (gh auth status)git status empty); if not, ASK whether to stashgit remote -v for the canonical org repo (tetherto/qvac.git); fall back to a remote literally named upstream. ASK if neither found.origin when org is upstream) — used only when org push is unavailable or the user explicitly chooses the fork path| Scenario | Source range | | ------------------------------------- | ---------------------------------------------------------------------------------------------------------------- | | Chained from sdk-pr-create | ORG_REMOTE/release-<pkg>-<x.y.z>..ORG_REMOTE/<source-head-branch> (fetch the head first). If the org remote tip is missing, use the release PR head SHA from gh pr view <num> --json headRefOid instead of an unqualified local branch name. | | Standalone with merged release PR | The PR's merge/squash commit on ORG_REMOTE/release-<pkg>-<x.y.z> (gh pr view <num> --json mergeCommit,headRefName) | | Standalone with --commit=<sha> | That single commit |
Sanity check: the source range must reflect release metadata only (version bump, changelog files, NOTICE, optional model registry/history). If git diff <range> --stat shows broad unrelated changes (suggesting the branch was based off main, not the release branch), STOP and ASK how to proceed — do not silently cherry-pick unrelated work onto main.
Before creating the backmerge branch, check whether main already contains the release content.
bashgit fetch ORG_REMOTE main git fetch ORG_REMOTE release-<pkg>-<x.y.z> git merge-tree --write-tree --merge-base=<src>^ ORG_REMOTE/main <src> git rev-parse ORG_REMOTE/main^{tree}
The first command prints the tree SHA produced by simulating the cherry-pick. The second prints main's current tree SHA. If they are identical, every change in the source range is already on main:
main so you can cite it:bash git log ORG_REMOTE/main --oneline -1 -- <pkg-dir>/changelog/<x.y.z>/ Resolve <pkg-dir> via node -p "require('./scripts/sdk/package-paths.cjs').getPackageDir('<pkg>')" (plugins use plugins/…, not packages/…).
No backmerge PR needed — main is already aligned with release-<pkg>-<x.y.z>. The release content landed on main via <commit-sha> (<commit-subject>).
This avoids pushing an empty branch and a gh pr create failure.
bashgit fetch ORG_REMOTE main git fetch ORG_REMOTE release-<pkg>-<x.y.z> git checkout -b backmerge/release-<pkg>-<x.y.z> ORG_REMOTE/main
If a local branch with that name already exists, ASK before overwriting.
bashgit cherry-pick -x <commit_or_range>
For a true merge commit (not squashed), add -m 1.
Resolve <pkg-dir> with scripts/sdk/package-paths.cjs (getPackageDir('<pkg>')).
Auto-resolvable (resolve, git add, then git cherry-pick --continue):
<pkg-dir>/package.json — version field conflict: take release-side.bash git checkout --theirs <pkg-dir>/package.json git add <pkg-dir>/package.json
<pkg-dir>/CHANGELOG.md (top-level aggregated): regenerate from the version folders just cherry-picked in.bash node scripts/sdk/generate-changelog-sdk-pod.cjs --package=<pkg> git add <pkg-dir>/CHANGELOG.md
Anything else → STOP. Hand control back to the user. Do not force-resolve, skip, or abort the cherry-pick on the user's behalf.
When stopping, print:
git status -sb # After resolving manually: git add <files> git cherry-pick --continue # Then re-run: /qv-sdk-backmerge --resume
Re-verify that the cherry-pick produced commits that actually change ORG_REMOTE/main. Defensive only — Step 3 catches the common no-op case; this guards against rarer paths (e.g. a --commit=<sha> arg that turned out to already be on main, or a -m 1 cherry-pick of a merge commit that resolved to nothing):
bashgit diff --stat ORG_REMOTE/main..HEAD
If the output is empty, treat it as a late no-op and STOP (same handling as Step 3 — report and exit). Otherwise push to the org remote when write access allows:
bash# Preferred — org-branch head (substitute real remote name for ORG_REMOTE): git push -u ORG_REMOTE backmerge/release-<pkg>-<x.y.z> # Fallback — personal fork (external CI path): git push -u FORK_REMOTE backmerge/release-<pkg>-<x.y.z>
Title (default — release PRs typically have a QVAC-#### ticket):
TICKET chore[skiplog]: backmerge release-<pkg>-<x.y.z> — <summary>Title (tickless fallback) — only when there is genuinely no ticket; combine the two tags inside a single bracket pair separated by | per .cursor/rules/sdk/commit-and-pr-format.mdc:
chore[skiplog|notask]: backmerge release-<pkg>-<x.y.z> — <summary><summary> lists what is being landed (e.g. version bump, changelog, NOTICE).Body (concise, copy-ready):
~~~markdown
Lands the release metadata for <pkg>@<x.y.z> on main, per gitflow.md "Keep main aligned". No functional changes — tagged [skiplog] so it does not appear in future changelogs.
<pkg-dir>/package.json — version <prev> → <x.y.z><pkg-dir>/changelog/<x.y.z>/ — generated changelog files<pkg-dir>/CHANGELOG.md — aggregated changelog<pkg-dir>/NOTICE — updated dependency attributions (if present)~~~
bash# Preferred — org-branch (same-repo) PR: gh pr create \ --repo tetherto/qvac \ --base main \ --head backmerge/release-<pkg>-<x.y.z> \ --title "<title>" \ --body "<body>" # Fallback — personal fork -> org PR: gh pr create \ --repo tetherto/qvac \ --base main \ --head <FORK_OWNER>:backmerge/release-<pkg>-<x.y.z> \ --title "<title>" \ --body "<body>"
Print the new PR URL as a clickable hyperlink. When chained from sdk-pr-create, the parent prints both URLs side by side. If the fork fallback was used, note that merge/release must approve the fork-ci environment for privileged CI on that head.
Before completing:
backmerge/release-<pkg>-<x.y.z>[skiplog] tag (combined as [skiplog|notask] if tickless)-x (so the original SHA is recorded in commit messages)gh pr view confirms target is tetherto/qvac:main and head is the expected org branch (or <fork>:backmerge/... if fallback).cursor/skills/qv-sdk-pr-create/SKILL.md — companion skill, auto-chains into this one for release targets.cursor/skills/qv-sdk-changelog/SKILL.md — changelog regeneration used during conflict resolution.cursor/rules/sdk/commit-and-pr-format.mdc — title format and [skiplog] semantics.cursor/rules/sdk/sdk-pod-packages.mdc — packages this skill applies todocs/gitflow.md — release flow and "Keep main aligned" rules (still documents fork-first contribution; prefer org-branch heads per this skill until DevOps updates gitflow)docs/ci/LABELS.md (fork-ci environment + fork-approval)| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-05 | pass→fail | 9,556 | 6,893 | -28% | 1 | 1 | 0% | 1,699 | 3,295 | +94% | 0 | 0 | — |
case-06 | pass→fail | 5,835 | 5,308 | -9% | 1 | 1 | 0% | 867 | 3,161 | +265% | 0 | 0 | — |
case-01 | fail→fail | 10,009 | 4,594 | -54% | 1 | 1 | 0% | 1,950 | 3,119 | +60% | 0 | 0 | — |
case-02 | fail→fail | 7,015 | 4,369 | -38% | 1 | 1 | 0% | 1,333 | 3,147 | +136% | 0 | 0 | — |
case-03 | fail→fail | 8,080 | 4,592 | -43% | 1 | 1 | 0% | 1,734 | 3,187 | +84% | 0 | 0 | — |
case-04 | pass→pass | 4,415 | 5,356 | +21% | 1 | 1 | 0% | 535 | 3,887 | +627% | 0 | 0 | — |
case-07 | fail→pass | 7,611 | 2,458 | -68% | 1 | 1 | 0% | 1,363 | 3,336 | +145% | 0 | 0 | — |
case-08 | fail→pass | 13,501 | 3,170 | -77% | 1 | 1 | 0% | 2,260 | 3,479 | +54% | 0 | 0 | — |
case-09 | fail→pass | 12,704 | 2,834 | -78% | 1 | 1 | 0% | 2,442 | 3,316 | +36% | 0 | 0 | — |
case-10 | fail→pass | 7,642 | 4,597 | -40% | 1 | 1 | 0% | 1,447 | 3,672 | +154% | 0 | 0 | — |
case-11 | fail→pass | 6,471 | 3,468 | -46% | 1 | 1 | 0% | 1,163 | 3,483 | +199% | 0 | 0 | — |
case-12 | fail→pass | 8,083 | 2,767 | -66% | 1 | 1 | 0% | 1,449 | 3,352 | +131% | 0 | 0 | — |
case-13 | fail→pass | 20,926 | 3,939 | -81% | 1 | 1 | 0% | 1,484 | 3,641 | +145% | 0 | 0 | — |
case-14 | fail→pass | 8,932 | 3,008 | -66% | 1 | 1 | 0% | 1,303 | 3,481 | +167% | 0 | 0 | — |
case-15 | fail→pass | 10,351 | 8,743 | -16% | 1 | 1 | 0% | 1,734 | 4,428 | +155% | 0 | 0 | — |
case-16 | fail→pass | 9,286 | 2,645 | -72% | 1 | 1 | 0% | 1,759 | 3,397 | +93% | 0 | 0 | — |
case-17 | pass→pass | 4,068 | 1,562 | -62% | 1 | 1 | 0% | 642 | 3,082 | +380% | 0 | 0 | — |
case-18 | fail→pass | 5,795 | 3,302 | -43% | 1 | 1 | 0% | 1,087 | 3,480 | +220% | 0 | 0 | — |
case-19 | pass→pass | 4,244 | 2,334 | -45% | 1 | 1 | 0% | 673 | 3,255 | +384% | 0 | 0 | — |
case-20 | fail→pass | 6,974 | 1,850 | -73% | 1 | 1 | 0% | 1,191 | 3,214 | +170% | 0 | 0 | — |
case-21 | fail→pass | 13,110 | 3,413 | -74% | 1 | 1 | 0% | 2,207 | 3,413 | +55% | 0 | 0 | — |
case-22 | pass→fail | 7,641 | 2,267 | -70% | 1 | 1 | 0% | 1,167 | 3,188 | +173% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted, and 16 counted toward the lift figure. The other 6 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +45 percentage points is the difference between those two pass rates over the 16 comparable cases. 4 cases got worse with the skill loaded, and they are included in that figure.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.