[roomote-roomote]

​Opened on behalf of @daniel-lxs. Follow up by mentioning @roomote-roomote, in the web UI, or in Telegram.

Related issue

Internal maintainer-requested reliability work; no issue was opened.

Why this PR exists

• A maintainer explicitly invited this PR in the linked issue or discussion
• I am a maintainer / this is internal Roomote work

A failed worker release package should leave the last complete release usable. Previously, packaging removed its checksum before copying or archiving the replacement, and a failed tar left both the staging tree and partial archive behind.

What changed

• Give each packaging attempt its own temporary staging directory and clean unpublished files on exit.
• Generate the replacement archive and checksum before replacing the previous files.
• Preserve the archive layout, packaged skills, sourcemap stripping, release names and publishing arguments.

This keeps a previous release usable after copy, tar or checksum-generation failure, and avoids deleting an unrelated same-named directory in the checkout. Archive and checksum publication remain two filesystem renames; this does not claim transactional multi-file publication, concurrent compiler isolation or SIGKILL cleanup.

How it was tested

• Baseline c97173a7: ran the actual packaging script with native tar/gzip under a controlled file-size limit. It failed with exit 2, preserved the previous archive, deleted its checksum, and retained a partial archive/staging tree.
• Head f44fe16f: repeated the same stimulus and a native TERM interruption; both previous files survived and unpublished files were cleaned. Compilation is substituted by prebuilt synthetic assets only in these failure fixtures.
• Regression tests cover native tar failure, failure of both checksum implementations, preservation of unrelated same-named directories, successful checksum/embedded skill contents, and extracted COMMIT, NODE_PTY_VERSION, PM2_VERSION and VERSION metadata.
• At the same head, ran the real worker build and release packaging with no compilation stub; the generated 4.4 MB archive passed sha256sum -c.
• All 100 dev-package tests, 317 compute-provider tests, 83 release-script tests, six Judgement isolation tests, pnpm lint, pnpm check-types, pnpm check-types:fast and pnpm knip passed. Browser proof is not applicable to this shell packaging path. Hosted publication and macOS execution were not exercised.
• Two initial hosted failures identified old staging-path expectations in the Modal and release-workflow source contracts. Both anchored expectations were updated, with actual archive metadata assertions added. Final-head hosted Test succeeded through full Vitest, Judgement isolation and release-script tests; current-head review is clean. A local full-repository test attempt hit its ten-minute command limit; all directly affected suites passed locally and the final hosted suite completed successfully.

Related PRs

• Mock Telegram duplicate-replay state #3634 — independent developer-harness reliability change.
• Runtime investigation source map #3635 — independent developer skill correction.

Checklist

• The PR title follows the repo convention: [Fix], [Feat], [Improve], [Refactor], [Docs], or [Chore] followed by a user-facing description
• This PR is small and scoped to one change
• pnpm lint and pnpm check-types pass locally
• I added tests or included a clear manual validation note above
• I removed secrets, tokens, private keys, and customer data from code, logs, and screenshots
• If this change should appear in the changelog, I ran pnpm changeset

No changeset: development/release tooling failure cleanup only.

[roomote-community]

No new code issues found. Reviewed the deterministic GITHUB_SHA fixture, extracted COMMIT assertion, and updated release-workflow expectation against the packaging script. Current-head CI lint, type checking, and Knip passed; tests and the app Docker build were still running when inspected. See session

_{Reviewed f44fe16}