Third, independent recovery layer for keep-mode upgrades (Part of #791). PR1's
manifest keeps the sync out of operator paths and the ephemeral /tmp snapshot
rolls back a mid-write crash; this adds the case neither covers — a "successful"
upgrade that a manifest/logic bug silently let touch an operator file.
- install.sh make_durable_snapshot(): before any mutation, copy the operator-owned
surface that exists to $XDG_STATE_HOME/mosaic/backups/pre-update-<UTC-ts>/
(0700 dirs / 0600 files, umask 077 scoped + restored). Fail-open; retention
MOSAIC_BACKUP_RETENTION (default 5).
- install.sh verify_operator_surface(): post-sync, restore any operator file the
upgrade diverged/removed and warn loudly (a divergence = manifest bug). Honors
intentional migration removals; refuses to restore through a symlink (CWE-59);
guards mkdir under set -e.
- mosaic restore (TS): --list (default dry-run), --from <ts> confirmation-gated
restore. Strict timestamp selector (CWE-22), O_NOFOLLOW 0600 writes (CWE-59),
containment + managed-dir checks via secure-file.ts. Counts/paths only — never
emits file contents (secrev).
- Gates: typecheck/lint/format ✓; mosaic vitest 1252; restore.spec 30;
bash gates durable-snapshot 41 / manifest-guard 193 / rollback 28 / migration 21.
Co-Authored-By: Claude Opus 4.8 <[email protected]>
Each step was re-running pnpm install independently, and all quality
steps (typecheck, lint, format, test) ran in parallel. On merge commits
with more accumulated code this pushed the CI runner over its memory
limit (exit code 254 = OOM kill).
Fix:
- install once, share node_modules via Woodpecker workspace volume
- sequential execution: install → typecheck → lint → format → test → build
- corepack enable in each step (fresh container) but no redundant install
jason.woltje
jarvis