/dejavu-rescue

The user is asking you to rescue a Weaviate class whose on-disk shard has a torn / truncated LSM segment that prevents the class from loading.

Use the bash command:

claude-dejavu rescue-shard $ARGUMENTS

Required:

• --class <C> — class name (e.g. ClaudeDejavuTurn)

Common:

• --dry-run — walk + report what WOULD happen. No DELETE, no batch insert, no gap-reindex. Safe to run anytime.
• --shard <id> — restrict to one shard id (default: all shards under the class).
• --yes — skip the typed-confirmation prompt before the destructive phase. The default flow requires typing yes interactively.
• --include-quarantined <path> — extra .db segment to walk alongside the live shard (e.g. a file you preserved before Weaviate stopped loading it).

Phases (destructive on apply, in this order):

1. Walk — Go binary dejavu-rescue scan-shard extracts records from every segment, with truncation tolerance for torn writes.
2. Reset class — DELETE /v1/schema/<C> then POST /v1/schema from the canonical class definition in code/weaviate-class.json.
3. Restore — batch-insert recovered records (with their original vectors) via /v1/batch/objects.
4. Gap-reindex — diff recovered UUIDs against the Postgres turns canonical set; re-embed any missing entries.

Recommended workflow:

1. Run with --dry-run first to confirm record counts look sane.
2. Make a claude-dejavu session-copy take --tag pre-rescue snapshot for rollback safety.
3. Run for real (omit --dry-run, add --yes when scripting).

Exit codes: 0 ok, 1 generic failure, 2 invalid args, 3 binary missing, 130 user abort.