Files
narratio/docs/internal/command-restore.md

122 lines
5.1 KiB
Markdown

# Internal: Command Restore
## Purpose
Explain the implemented restore discovery, planning, installation, and
reporting flow in `internal/app`. User invocation belongs in
[CLI](../cli.md#session-restore), and the operator recovery procedure and
physical restore scope belong in
[Operations](../operations.md#restore-workflow).
Restore separates remote authority, local conflict policy, and filesystem
mutation so each remains testable independently.
## Discovery Contract
Discovery delegates current-state pointer and manifest loading to
`internal/artifacts`, then validates the result against the resolved request:
- campaign must match;
- session ID must match.
- run ID must match the pointer-selected committed run.
Restore treats any missing or invalid remote current state as a command error.
## Planning Contract
Restore planner action kinds:
- `download`;
- `skip_same`;
- `conflict`.
Planner behavior:
- a new-protocol restore uses only the selected commit's declared artifact set;
each action carries that artifact's immutable key, checksum, size, and
generation;
- remote-to-local mapping is traversal-safe;
- actions are sorted by local relative path and then remote key;
- planning opens the selected session root and inspects every target by its
normalized session-relative path. Missing ancestors produce a missing target;
an ancestor or leaf symlink is a non-regular conflict and is never opened for
hashing;
- force converts differing eligible regular files from conflicts to downloads;
directories, symlinks, and other non-regular targets remain conflicts.
For a non-dry-run restore, planning/classification happens only after acquiring
the session lock. Runner manifest/reuse checks acquire that same lock first.
Previous-cache readiness is resolved through `previouscache.Resolve` for restore,
prepare, status, and validation. Each requirement selects the unique
`published_output` for its canonical source in the pointer-selected commit,
then strips only that run's immutable `published/` prefix for its safe local
cache path. It never probes candidate objects or derives paths from manifests,
configuration, or basenames.
## Execution Contract
Execution order and safety:
- non-manifest downloads happen before manifest install;
- `manifest.json` installs last;
- downloads use sibling temp files, validate the complete object, then use an
atomic rename;
- manifest replacement is validated before rename;
- each committed object is verified against its declared checksum, size, and
generation before installation;
- installation reopens the selected session root, creates only validated
missing directory components with workspace modes, and rejects symlinked or
non-regular ancestors and leaves before replacement;
- a committed manifest already verified during discovery is retained for the
matching restore action and revalidated before installation, avoiding a
second body transfer;
- failed installs do not roll back files already written in the same execution.
- a durable `.restore-incomplete.json` marker is written before installation.
It blocks runners until a restore retry completes all verified installs and
the local manifest replacement, at which point it is removed.
- restored manifest local references are rebased beneath the selected local
session root. Unsafe relative references and producer-machine absolute paths
outside the manifest's producer session root are rejected; producer-local
spool paths and run-local diagnostic references are not restored as
authority. Cleanup remains ordinary restored stage state and is not reset or
synthesized by restore.
Audio restore path:
- uses `audio.MaterializeS3Audio`;
- integrates spool and S3 audio cache paths;
- reuses cached audio only when its opened regular file, content digest, and
identity sidecar all match the selected remote object version; otherwise it
refreshes through the validated download path.
## Reporting Contract
- dry-run mode prints a summary, performs no durable session writes, and may
read remote current-state or object-identity data to produce that summary;
- execution mode persists the canonical restore report described in
[Operations](../operations.md#restore-workflow);
- report includes plan counts, per-action status, and execution failures.
## Invariants
- restore uses committed remote current state as authority;
- one restore or status inspection observes the single pointer-selected commit
loaded at discovery; later pointer changes cannot add objects or substitute a
different run into its plan;
- a verified `current/commit-pointer.json` and its selected immutable commit
establish remote commitment; when absent, republish the session;
- restore does not execute pipeline stages.
## Implementation And Tests
- Discovery: `internal/app/restore_discovery.go`
- Planning: `internal/app/restore_plan.go`, `internal/previouscache`
- Execution: `internal/app/restore_execute.go`
- Reporting and command coordination: `internal/app/restore_report.go`,
`internal/app/restore.go`
- Tests: `internal/app/restore_discovery_test.go`,
`internal/app/restore_plan_test.go`,
`internal/app/restore_execution_test.go`,
`internal/app/restore_workflow_test.go`