Index not auto-loaded from vector DB on MCP server restart

[Retengart]

Problem

Snapshot file (~/.context/mcp-codebase-snapshot.json) is loaded on startup, but actual Milvus/Zilliz collections are never verified against local state.

Current Behavior

1. Index codebase → snapshot says "indexed", vector DB has data
2. Restart MCP server
3. loadCodebaseSnapshot() loads from JSON file ✓
4. But vector DB collection existence never checked
5. If cloud data was deleted (manually, TTL, etc.), local says "indexed" but search returns nothing

Expected Behavior

On startup, verify that collections listed in snapshot actually exist in vector DB. Remove from snapshot if collection missing.

Location

• packages/mcp/src/index.ts:79-80 - loads snapshot but no DB verification
• packages/mcp/src/snapshot.ts:440-467 - loadCodebaseSnapshot()

Suggested Fix

After loadCodebaseSnapshot(), iterate indexed codebases and call context.hasIndex() to verify collections exist.

[zc277584121]

Hi @Retengart, thanks for this detailed analysis — you nailed the root cause.

We've shipped a fix in v0.1.6 (PR #283) that addresses exactly this problem. Here's what changed:

The core issue was that the snapshot was treated as the sole source of truth, with no verification against the actual VectorDB state. We've now added VectorDB checks at multiple levels:

1. Bidirectional cloud sync on startup: syncIndexedCodebasesFromCloud() now not only removes local entries missing from the cloud, but also recovers cloud entries missing from the local snapshot. So if the snapshot lost an entry but the VectorDB collection still exists, it gets automatically recovered.

2. Search-time VectorDB fallback: Before returning "not indexed" error, search_code now checks the VectorDB directly. If the collection exists there, the snapshot is recovered and the search proceeds normally.

3. Index-time mismatch auto-repair: When index_codebase is called, it now detects snapshot/VectorDB inconsistencies in both directions and auto-fixes them — exactly the behavior you described in "Expected Behavior".

This also handles your scenario of cloud data being deleted: if the VectorDB collection was removed but the snapshot still says "indexed", the system will now detect the mismatch and clear the stale snapshot entry, prompting a proper re-index.

To update, simply reconnect your MCP server (/mcp reset or restart Claude) — it will pull the latest version automatically.

Would appreciate it if you could test with v0.1.6 and let us know if it resolves the issue. We'll keep this open for feedback. Thanks again! 🙏