feat: add Docker backup and restore workflow

.github/workflows/backup-contract.yml

@@ -0,0 +1,21 @@
+name: Backup Contract
+
+on:
+  push:
+    branches: [main]
+  pull_request:
+    branches: [main]
+
+permissions:
+  contents: read
+
+jobs:
+  backup-contract:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@34e114876b0b11c390a56381ad16ebd13914f8d5 # v4.3.1
+      - name: Run Docker-free backup contract test
+        run: |
+          bash -n deploy/backup.sh
+          bash -n deploy/test-backup.sh
+          bash deploy/test-backup.sh

README.md

@@ -59,6 +59,7 @@ We are especially interested in early bug reports and real-world feedback about
 - [Build instructions](BUILD.md)
 - [Contributing guide](CONTRIBUTING.md)
 - [Logging configuration](LOGGING.md)
+- [Docker backup and restore](deploy/backup-restore.md)
 - [Product documentation](https://windshift.sh/docs)
 
 ## License

deploy/backup-restore.md

@@ -0,0 +1,188 @@
+# Backing up and restoring Windshift
+
+`backup.sh` is a conservative host-side workflow for the stock Docker Compose deployments in this repository.
+It is not a general-purpose Docker or database recovery tool.
+It backs up the complete configured `/data` tree and, for the supported PostgreSQL shape, a custom-format logical dump.
+
+The helper image is pinned to the Alpine digest used by `deploy/coding-agent/Dockerfile`.
+It has no network, a read-only root filesystem, and no-new-privileges.
+The backup helper receives Windshift volumes read-only.
+The restore helper receives a writable data volume only after all preflight checks and a local rollback snapshot have completed.
+
+## Supported scope
+
+The workflow requires exactly one Compose `windshift` container using the stock `/windshift` entrypoint and exactly one persistent mount at the configured data path.
+Additional mounts below that path are rejected because deleting a mounted child can leave both restore and rollback incomplete.
+The backup directory and rollback root must be outside that mount's host source so an archive cannot include or delete itself.
+Replicas are rejected.
+It derives a stable, domain-separated hash from the Compose project and service labels and uses a system-wide mode-700 lock directory under `/var/tmp`.
+
+SQLite is supported only when `DB_PATH` is below `/data`.
+An empty `DB_PATH` is rejected because Windshift would otherwise fall back to a path outside the mounted data tree.
+The database path is recorded in the manifest and must exist as a nonempty regular archive entry.
+The script also checks `ATTACHMENT_PATH`, `PLUGIN_DIR`, `AI_PROMPTS_DIR`, `LLM_PROVIDERS_FILE`, and trimmed comma-separated `PLUGIN_DIRS` when configured.
+An empty `PLUGIN_DIR` is rejected because Windshift would otherwise fall back to a relative path outside the data mount.
+SSH host keys and SSH-enabled deployments are outside this workflow because the image-relative default path is not guaranteed to be under `/data`.
+Any extra host mounts or persisted paths outside `/data` are out of scope and cause a safe failure when they are configured through these variables.
+
+PostgreSQL is supported only when all of the following are true:
+
+- `DB_TYPE=postgres` and `POSTGRES_CONNECTION_STRING` is unset.
+- `POSTGRES_HOST` exactly equals the selected Compose service.
+- `POSTGRES_PORT` is unset or 5432.
+- `POSTGRES_USER` and `POSTGRES_DB` exactly match that service's container environment.
+- Exactly one PostgreSQL service container exists.
+
+The automatic service candidates are `postgres` and `db`.
+Use `--postgres-service NAME` for another stock-style service.
+On restore, that explicit option overrides the manifest value.
+External databases, connection strings, non-default ports, custom database selection, and replicas are deliberately unsupported.
+Any command-line use of `-postgres-connection-string`, `--postgres-connection-string`, `-pg-conn`, or `--pg-conn` is also out of scope and fails closed.
+The path override flags `-db`, `--db`, `-attachment-path`, `--attachment-path`, `-llm-providers`, `--llm-providers`, `-ai-prompts-dir`, and `--ai-prompts-dir` are out of scope too.
+Environment values containing CR or LF are rejected because this workflow cannot safely fingerprint or archive them.
+Use a database-specific procedure for those deployments.
+
+Use the same PostgreSQL major version for backup and restore.
+Restore only into the same Windshift image and schema generation that produced the backup unless a separately verified migration plan says otherwise.
+
+## Create a backup
+
+Run from the Compose directory:
+
+    cd deploy
+
+    ./backup.sh backup /srv/windshift-backups/windshift-$(date -u +%Y%m%dT%H%M%SZ)
+
+The backup path must be an absolute, normalized host path without a colon because it is passed to Docker's `-v` syntax.
+The created directory is mode 700.
+It contains `data.tar.gz`, `manifest.env`, and `checksums.sha256`.
+PostgreSQL backups also contain `database.dump`.
+
+The script takes an atomic lock before stopping the app or creating backup files.
+If a lock exists, do not run another operation.
+First confirm that no backup or restore is running, then remove only the reported stale lock directory.
+
+SHA-256 checksums detect accidental corruption and incomplete copies.
+They do not authenticate a backup.
+Use access-controlled, authenticated, encrypted storage and transport.
+
+The manifest stores a domain-separated SHA-256 fingerprint of the effective `SSO_SECRET`, or `SESSION_SECRET` when no `SSO_SECRET` is present.
+This is a compatibility check, not a password verifier.
+Weak secrets therefore still produce easily guessable fingerprints.
+
+Pass `--include-sso-secret` only for an explicitly encrypted secret-recovery copy.
+It writes the raw effective secret as mode-600 `sso-secret` inside the mode-700 backup and includes it in the checksum file.
+This artifact is highly sensitive and must never be logged or copied to ordinary backup media.
+
+## Restore
+
+Run from the same Compose directory:
+
+    cd deploy
+
+    ./backup.sh restore --force /srv/windshift-backups/windshift-20260829T120000Z
+
+Before stopping Windshift, restore rejects symlinked or non-regular input artifacts, copies approved artifacts into a private staging directory, and verifies only that staged snapshot thereafter.
+The helper checks the staged archive in the same image that performs extraction.
+It accepts only regular files and directories, rejects absolute and parent paths, and rejects symbolic links, hard links, devices, and FIFOs.
+
+Restore compares the live secret fingerprint before any state change.
+On mismatch it fails closed and never writes a secret into Compose.
+Configure the source secret in Compose or the secret manager first.
+
+After the app stops, the script creates a mode-700 rollback directory under `--rollback-root` or `WINDSHIFT_BACKUP_ROLLBACK_ROOT`.
+The default is disk-backed `/var/tmp`.
+That directory must already exist and be writable.
+Choose a local, private filesystem with space for a full compressed `/data` copy, staging copy, and, for PostgreSQL, a second logical dump.
+Do not put it on volatile `/tmp` unless its loss during reboot is acceptable.
+
+For PostgreSQL the rollback contains both `/data` and a pre-restore database dump.
+The script writes SHA-256 checksums for retained rollback artifacts and structurally validates custom-format PostgreSQL dumps with `pg_restore --list` before relying on them.
+The target database restore uses `pg_restore --single-transaction --exit-on-error`.
+It cleans and restores archive objects, not necessarily unrelated database objects.
+
+The rollback is retained until the started application is healthy.
+When Docker exposes a health check, the script waits for it for up to `WINDSHIFT_BACKUP_HEALTH_TIMEOUT` seconds, default 90.
+Without a Docker health status, it executes `/windshift healthcheck` through Compose until the configured deadline.
+Increase `WINDSHIFT_BACKUP_HEALTH_TIMEOUT` before restore when startup or migrations can legitimately take longer than 90 seconds.
+If the target is unhealthy, it restores both rollback data and PostgreSQL before trying the old state.
+If PostgreSQL state is uncertain because of an interruption or failed database command, Windshift remains stopped and the error retains both rollback and staged paths.
+The recovery lock also remains in place because a detached server-side `pg_restore` may still be running.
+Remove that lock only after confirming that no backup, restore, or PostgreSQL restore process is active and after resolving the database state.
+If Windshift was already stopped before restore, the script cannot perform a health check or automatic start.
+It reports success with the retained rollback path so the operator can validate a manual start before removing it.
+
+The command uses a 60-second Compose stop timeout.
+An interrupted command, host reboot, or Docker failure can still leave the service stopped.
+Read the printed recovery paths before restarting it.
+On a confirmed healthy restore, cleanup failures only warn and retain the private rollback or stage instead of changing the restored state.
+The ephemeral backup helper streams its tar archive to a host-created mode-600 file, so it does not need to chown host files.
+The Docker-free contract test does not prove behavior against a real Docker daemon, image, filesystem driver, or PostgreSQL server.
+
+## Manual emergency recovery
+
+This is a destructive emergency procedure.
+Leave Windshift stopped and first copy every reported stage and rollback directory to immutable, access-controlled storage.
+Use only the same Windshift image/schema generation and the same PostgreSQL major version.
+
+Before stopping or deleting anything, validate the retained checksums and data archive in a separate preflight.
+Use the SHA-256 tool available on the host and the exact pinned helper image shown below:
+
+    cd deploy
+    (cd "$ROLLBACK_DIR" && sha256sum -c checksums.sha256)
+    docker run --rm --network none --read-only --security-opt no-new-privileges -v "$ROLLBACK_DIR:/backup:ro" alpine:3.21@sha256:48b0309ca019d89d40f670aa1bc06e426dc0931948452e8491e3d65087abc07d sh -ec '
+      test -s /backup/data.tar.gz
+      tar -tzf /backup/data.tar.gz >/dev/null
+      tar -tvzf /backup/data.tar.gz >/dev/null
+      tar -tzf /backup/data.tar.gz | awk '\''{
+        path = $0
+        if (path == "./") key = "."
+        else {
+          if (substr(path, 1, 2) != "./") exit 1
+          path = substr(path, 3)
+          if (path == "" || path ~ /(^|\/)\.($|\/)/ || path ~ /(^|\/)\.\.($|\/)/ || path ~ /\/\//) exit 1
+          sub(/\/$/, "", path)
+          if (path == "") exit 1
+          key = path
+        }
+        if (seen[key]++) exit 1
+        count++
+      } END { if (!count) exit 1 }'\''
+      if tar -tvzf /backup/data.tar.gz | grep -Eqv "^[-d]"; then exit 1; fi
+      if tar -tvzf /backup/data.tar.gz | grep -E "^-.* -> "; then exit 1; fi
+    '
+
+If the host provides `shasum` instead, replace the checksum command with `shasum -a 256 -c checksums.sha256`.
+Do not continue unless every command succeeds.
+
+For PostgreSQL, validate the retained database dump before changing `/data`.
+Set `POSTGRES_SERVICE` to the confirmed service name and keep `remote_dump` for the restore step below:
+
+    POSTGRES_SERVICE=postgres
+    db_id=$(docker compose ps --all -q "$POSTGRES_SERVICE")
+    [ "$(printf '%s\n' "$db_id" | sed '/^$/d' | wc -l | tr -d ' ')" = 1 ] || exit 1
+    remote_dump=/tmp/windshift-manual-rollback-$$.dump
+    docker cp "$ROLLBACK_DIR/database-before.dump" "$db_id:$remote_dump"
+    docker compose exec -T -e REMOTE_DUMP="$remote_dump" "$POSTGRES_SERVICE" sh -ec 'pg_restore --list "$REMOTE_DUMP" >/dev/null'
+
+For `/data`, resolve exactly one Windshift container before stopping it:
+
+    cd deploy
+    windshift_id=$(docker compose ps --all -q windshift)
+    [ "$(printf '%s\n' "$windshift_id" | sed '/^$/d' | wc -l | tr -d ' ')" = 1 ] || exit 1
+    mount_destinations=$(docker inspect -f '{{range .Mounts}}{{println .Destination}}{{end}}' "$windshift_id")
+    [ "$(printf '%s\n' "$mount_destinations" | grep -Fxc /data)" = 1 ] || exit 1
+    if printf '%s\n' "$mount_destinations" | grep -Eq '^/data/'; then exit 1; fi
+    docker compose stop --timeout 60 windshift || exit 1
+    [ "$(docker compose ps --all -q windshift)" = "$windshift_id" ] || exit 1
+    [ "$(docker inspect -f '{{.State.Running}}' "$windshift_id")" = false ] || exit 1
+    docker run --rm --network none --read-only --security-opt no-new-privileges --volumes-from "$windshift_id" -v "$ROLLBACK_DIR:/backup:ro" alpine:3.21@sha256:48b0309ca019d89d40f670aa1bc06e426dc0931948452e8491e3d65087abc07d sh -ec 'find /data -mindepth 1 -maxdepth 1 -exec rm -rf -- {} +; tar -C /data -xzf /backup/data.tar.gz'
+
+For PostgreSQL, restore the already validated remote dump only after confirming the target database again:
+
+    docker compose exec -T -e REMOTE_DUMP="$remote_dump" "$POSTGRES_SERVICE" sh -ec 'pg_restore --clean --if-exists --no-owner --no-privileges --single-transaction --exit-on-error -U "$POSTGRES_USER" -d "$POSTGRES_DB" "$REMOTE_DUMP"'
+    docker compose exec -T "$POSTGRES_SERVICE" rm -f "$remote_dump"
+
+The PostgreSQL rollback cleans objects listed in its dump but may leave unrelated objects created by a newer schema.
+Remove any container path printed as `remote=` after the database state has been recovered and the retained dump has been preserved elsewhere.
+Only then start Windshift and run its health check.