Atomic Transactions

repobrain maintains three stores: SQLite (metadata + metrics), LanceDB (embeddings + docs), and NetworkX (dependency graph). Writing to all three atomically is the key innovation that prevents the 5–15% silent consistency failures in repowise.

The Problem (repowise)

repowise writes to each store independently. If the process crashes between writes: - SQL has the file record, LanceDB doesn't - The graph has an edge to a file that has no embedding - Incremental updates skip files that are "already indexed" in SQL but missing from LanceDB

These inconsistencies are silent — no error is raised.

The Solution (AtomicStorageCoordinator)

async with coordinator.transaction() as txn:
    txn.pending_sql_calls.append(lambda: repo.upsert(...))
    txn.pending_vector_records.append({"id": file_id, ...})
    txn.pending_nodes.append(("src/main.py", {"type": "file", ...}))
    txn.pending_edges.append(("src/main.py", "src/utils.py", {}))

# On success: flush SQL → graph → vector (in order)
# On exception: SQL rollback + delete buffered vector IDs + remove buffered graph nodes

Rollback Behavior

Store	On Success	On Exception
SQLite	COMMIT	ROLLBACK
LanceDB	Insert records	Delete by buffered IDs
NetworkX	Add nodes + edges	Remove buffered nodes + edges

Health Check

health = await coordinator.health_check(repo_id)
# {
#   "sql_files": 500,
#   "graph_nodes": 500,
#   "vector_records": 500,
#   "consistent": true
# }

Use repobrain status to run a health check interactively.