releng-next

name	releng-next
description	Switch all GLSP repositories from their release version to the next nightly (`-next`) version after a release — fresh-clones every repo, runs `releng prepare next` to open "Switch to nightly" PRs, and optionally polls CI and squash-merges each PR once green.
disable-model-invocation	true

Releng: Switch to Next

Switch every GLSP repository from its release version to the next nightly (-next) version by running releng prepare next in a fresh clone of each repo, opening one "Switch to nightly" PR per repo, and (optionally) auto-merging each PR after its CI passes.

This skill operates entirely on fresh clones in a temp directory — it never touches the user's existing working clones, so no git hygiene (branch/dirty/stale checks) is required.

Inputs

All inputs are taken from the user's prompt. Use the defaults below for anything not mentioned.

• target repos (optional): the user may name one or more specific repos to operate on (e.g. "just glsp-eclipse-integration", "only the npm repos"). Only those repos are cloned and processed; everything else is left untouched. This is the mechanism for a partial run — most importantly, re-running glsp-eclipse-integration on its own once glsp-server's nightly P2 is published (see Dependency ordering in Conventions). Default: all 8 repos.
• auto-merge (optional): after all PRs are created, poll each PR's checks and squash-merge it once green. With auto-merge the run also goes end-to-end — after glsp-server merges it waits for the server nightly P2 to publish, then retries and merges the deferred glsp-eclipse-integration (Phase 4b). Default: off.
• no-push (optional): if the user asks to not push / run locally / dry-run, pass --no-push to prepare next. The version bump and nightly-<version> commit are made in the temp clone but nothing is pushed and no PR is created. Default: push enabled.
• draft (optional): if the user asks for draft PRs, pass --draft to prepare next so the PRs are created as drafts. Only meaningful when pushing. Default: non-draft.
• Poll interval (default 60s) and overall timeout (default 60min): only relevant with auto-merge. The user may override these.

Input interactions

• no-push + auto-merge is contradictory — no PRs exist to merge. If both are requested, honor no-push and skip the auto-merge phase, telling the user why.
• draft + auto-merge is contradictory — a draft PR signals "not ready to merge". If both are requested, create the draft PRs but do not auto-merge them; leave them as drafts for manual review and say so.
• no-push + draft: draft has no effect (nothing is pushed). Just proceed with no-push.

Conventions

• Always use the local CLI built from this tooling repo: pnpm glsp … (never npx @eclipse-glsp/cli). The maintainer wants the behavior of the code as it currently stands.
• releng prepare next and releng version accept -r/--repoDir <dir> and cd into it internally, so stay in the tooling repo and point -r at each temp clone — do not run from inside each repo.
• Continue-on-error. Never abort the whole run because one repo failed. Collect outcomes and report them at the end in the four buckets.
• The 8 GLSP repos are: glsp, glsp-client, glsp-server-node, glsp-theia-integration, glsp-vscode-integration, glsp-eclipse-integration, glsp-server, glsp-playwright.
• Java repos (version lives in a pom.xml, not package.json): glsp-server and glsp-eclipse-integration. All others are npm repos.
• Dependency ordering — glsp-eclipse-integration depends on glsp-server already being on next. The Eclipse integration's server module builds (Tycho) against a P2 target platform that points at https://download.eclipse.org/glsp/server/p2/nightly/<major>.<minor>/ (e.g. …/nightly/2.8/). That P2 site only exists once glsp-server is on the matching next version and its nightly CI has deployed the new P2 repository — which happens only after glsp-server's "Switch to nightly" PR is merged. So running prepare next for glsp-eclipse-integration before that P2 is live fails at the build step with No repository found at …/p2/nightly/<major>.<minor>. Therefore: always process glsp-eclipse-integration last, and treat that specific error as ⏸️ deferred, not ❌ failed (Phase 2). With auto-merge, the skill unblocks it end-to-end (Phase 4b); without it, the user re-runs the skill for glsp-eclipse-integration alone once the P2 is up.

---

Phase 0: Preconditions

Run from the tooling repo root (glsp/glsp).

1. Verify the GitHub CLI is authenticated:

   gh auth status
   

   If this fails, stop and tell the user to authenticate gh first (they need push rights to the eclipse-glsp org).

2. Build the local CLI so dist/cli.js is current:

   pnpm install
   pnpm build
   

   Confirm pnpm glsp --help works before continuing.

---

Phase 1: Fresh checkout

3. Create a dedicated, timestamped temp directory and record its path as WORKDIR:

   WORKDIR="${TMPDIR:-/tmp}/glsp-release-next-$(date +%Y%m%d-%H%M%S)"
   mkdir -p "$WORKDIR"
   echo "$WORKDIR"
   

   Remember WORKDIR — it is printed again at cleanup time and referenced throughout.

4. Clone the target repos fresh into WORKDIR, from the real upstream (no fork, so origin = eclipse-glsp and pushes/PRs target upstream):

   # full run (default): all 8 repos
   pnpm glsp repo clone --preset all -d "$WORKDIR"
   
   # partial run: only the repos the user named (repo clone takes positional names)
   pnpm glsp repo clone <repo> [<repo> …] -d "$WORKDIR"
   

   Each repo lands at $WORKDIR/<repo> on its default branch. If any clone fails, record it as a failure for that repo but continue.

---

Phase 2: Per-repo prepare-next

Process each cloned repo independently. Track each repo's outcome in one of four buckets: ✅ PR created · ⏭️ skipped (already on next) · ⏸️ deferred (blocked on glsp-server nightly P2) · ❌ failed (reason).

Processing order: always process glsp-eclipse-integration last (after all other targeted repos), because its build depends on glsp-server's next P2 being published — see Dependency ordering in Conventions.

For each repo at $WORKDIR/<repo>:

Step 2.1: Precondition — must be on a release version, not already on next

Read the current version directly from the appropriate file:

• npm repos → root package.json → version
• glsp-server → pom.xml → first <version>
• glsp-eclipse-integration → server/pom.xml → first <version>

If the version ends with -next or .SNAPSHOT, the repo is already on next → ⏭️ skip it (do not run prepare, do not push anything) and record the reason. Otherwise continue.

Step 2.2: Run prepare next

Build the command from the resolved inputs:

pnpm glsp releng prepare next -r "$WORKDIR/<repo>" [--no-push] [--draft]

Add --no-push if no-push was requested, and --draft if draft was requested (draft is ignored by prepare when --no-push is also set).

This is atomic: version-bump → build → commit a nightly-<version> branch → (unless --no-push) push → open the "Switch to nightly" PR. The build step is the verification — there is no separate verify step.

• If the command throws (e.g. build failure, push rejected) → nothing was pushed:
  • Special case — glsp-eclipse-integration whose error contains No repository found at …/p2/nightly/<major>.<minor> → this is the expected upstream-ordering block (the glsp-server next P2 is not published yet), not a real failure → record ⏸️ deferred, capture the missing P2 URL and the <major>.<minor>. Then:
    • with auto-merge → it will be unblocked end-to-end in Phase 4b.
    • without auto-merge → remediation: merge the glsp-server "Switch to nightly" PR, wait for its nightly CI to deploy …/p2/nightly/<major>.<minor>/, then re-run this skill targeting only glsp-eclipse-integration.
  • Any other throw → record ❌ failed with the error reason.
  • Continue to the next repo either way.
• If it succeeds:
  • push enabled → record ✅ PR created and capture the PR URL (from the command output, or gh pr list --repo eclipse-glsp/<repo> --head nightly-<version> --json url). Note whether it is a draft.
  • no-push → record ✅ branch prepared (local, not pushed) with the nightly-<version> branch name and the temp-clone path.

---

Phase 3: Listing checkpoint

After all repos have been processed, print a clear summary with the four buckets:

✅ PR created                       (or, with no-push: "Branch prepared (local)")
   - <repo>: <pr-url>               (draft PRs marked "(draft)")
   ...                              (no-push: "<repo>: nightly-<version> @ <temp-path>")
⏭️ Skipped (already on next)
   - <repo>: version <x.y.z-next>
   ...
⏸️ Deferred (blocked on glsp-server nightly P2)
   - glsp-eclipse-integration: needs .../p2/nightly/<major>.<minor>/
   ...
❌ Failed
   - <repo>: <reason>
   ...

This listing always happens, regardless of auto-merge / no-push / draft.

If glsp-eclipse-integration landed in ⏸️ deferred and auto-merge is off, spell out the remediation in the report: merge the glsp-server "Switch to nightly" PR, wait for its nightly CI to publish …/p2/nightly/<major>.<minor>/, then re-run this skill targeting only glsp-eclipse-integration. (With auto-merge on, Phase 4b handles this automatically.)

---

Phase 4: Auto-merge (only if auto-merge)

Skip this phase entirely (and tell the user why) if no-push was used (no PRs exist) or if draft was used (drafts are left for manual review) — see Input interactions.

Operate per-PR and independently — merge each PR as soon as its own checks are green; do not wait for the others. Only ever act on the ✅ PR created bucket here; the ⏸️ deferred glsp-eclipse-integration is handled afterward in Phase 4b (which depends on the glsp-server PR merging first).

Loop every 60s (or the user-specified interval) until every ✅-PR has reached a terminal state or the 60min (or user-specified) overall timeout elapses. For each not-yet-resolved PR:

gh pr checks <pr-url>

• All checks green → squash-merge and delete the branch:

  gh pr merge <pr-url> --squash --delete-branch
  

  (No approval step — GLSP repos have no required-review rule.) Mark merged.
• No checks registered yet (just pushed) → treat as still pending; keep waiting. Never merge a PR before CI has attached.
• A check failed → stop watching it; mark CI failed (not merged).
• Timeout reached → leave any still-pending PRs open; mark them timed out (not merged).

Never force-merge and never bypass checks.

After the loop, print a final report listing every PR with its terminal state: merged / CI failed / timed out / pending.

Phase 4b: Unblock deferred glsp-eclipse-integration (auto-merge only)

Run this only when auto-merge is on and glsp-eclipse-integration is in the ⏸️ deferred bucket. It is what lets a full auto-merge run finish the entire switch end-to-end.

Preconditions inside this phase:

• The glsp-server PR must have merged in Phase 4. If it ended CI failed / timed out / pending, eclipse-integration stays deferred (its P2 will never appear) — report that and stop here.

Steps (share the same overall timeout as Phase 4):

1. Wait for the server nightly P2 to publish. After glsp-server merges, its nightly CI deploys …/p2/nightly/<major>.<minor>/. Poll that URL every poll-interval until it resolves to a real repository (not a 404):

   curl -sfI "https://download.eclipse.org/glsp/server/p2/nightly/<major>.<minor>/p2.index" >/dev/null && echo up
   

   The nightly deploy is not instant — it may take many minutes after merge. If the overall timeout elapses first, leave eclipse-integration deferred and report it.
2. Retry prepare next for eclipse-integration (same flags as the original run, push enabled):

   pnpm glsp releng prepare next -r "$WORKDIR/glsp-eclipse-integration"
   

   • Throws again with the same P2 error → P2 not actually ready; keep polling (step 1) until timeout.
   • Other throw → record ❌ failed.
   • Succeeds → record ✅ PR created, capture the PR URL.
3. Auto-merge the new PR using the same green-checks → gh pr merge --squash --delete-branch logic as Phase 4, within the remaining timeout.

Fold the eclipse-integration outcome into the final Phase 4 report.

---

Phase 5: Cleanup (always — final step)

Ask the user whether to delete WORKDIR. Timing:

• No auto-merge → ask right after the Phase 3 listing.
• auto-merge → ask after the Phase 4 merge loop (and Phase 4b, if it ran) finishes.

Always print the WORKDIR path so the user can inspect it. If the user declines, leave it in place.

If no-push was used, warn before deleting: the prepared nightly-<version> branches exist only in WORKDIR and have not been pushed anywhere — deleting it discards them.

rm -rf "$WORKDIR"   # only if the user confirms