Self-host: toolkit MCP endpoints 500 on legacy connection rows (empty-string JSON columns, bigint in blob tools_synced_at) — boot migration suggestion

[tejasgadhia]

Summary

Self-hosted Executor (Docker, ghcr.io/usefulsoftwareco/executor-selfhost:1.6.10) returns HTTP 500 / JSON-RPC -32603 on POST /mcp/toolkits/<slug> initialize when the connection table contains rows carried over from pre-FumaDB versions. Boot only runs additive schema-ensure (CREATE TABLE IF NOT EXISTS / ALTER TABLE ADD COLUMN) and never migrates existing row payloads, so two legacy value shapes crash the ORM.

Environment

• Image: ghcr.io/usefulsoftwareco/executor-selfhost:1.6.10
• DB originally created in the 0.8.x era, carried forward through the FumaDB cutover (May 2026) and unified-provider migration (June 2026)
• 22 rows in connection

Crash 1: empty string in JSON-mode columns

SQLiteTextJson.mapFromDriverValue does JSON.parse(value). Legacy rows have '' (empty string, not NULL) in JSON-mode columns (credential_write in my case), and JSON.parse('') throws:

StorageError: FumaDB connection.findMany failed
  [cause]: SyntaxError: JSON Parse error: Unexpected EOF
      at mapRelationalRow (serve.js:290690)
      at executor.connections.list (serve.js:284883)
      at mcp.host.create_executor_server

The JSON-mode columns on connection are: item_ids, credential_write, last_health, provider_state. Empty strings in any of them crash every toolkit endpoint because create_executor_server → connections.list runs on every MCP session init.

Scan to find affected rows:

SELECT row_id FROM connection
WHERE credential_write = '' OR last_health = '' OR provider_state = '' OR item_ids = '';

Fix:

UPDATE connection SET credential_write = NULL WHERE credential_write = '';
UPDATE connection SET last_health = NULL WHERE last_health = '';
UPDATE connection SET provider_state = NULL WHERE provider_state = '';
UPDATE connection SET item_ids = NULL WHERE item_ids = '';

Crash 2: integer timestamp in blob column tools_synced_at

tools_synced_at is declared SQLite blob, but legacy rows contain ASCII-digit timestamps (e.g. 1790039911318). The Drizzle bigint decoder receives a JS number and throws:

TypeError: The first argument must be of type string, Buffer, ArrayBuffer, Array,
or Array-like Object. Received type number (1790039911318)
    at mapFromDriverValue (serve.js:291456)

(The existing runSqliteNdjsonOutputMigration already NULLs this column for NDJSON-affected rows — but rows unaffected by that migration keep the crashing values.)

Fix:

UPDATE connection SET tools_synced_at = NULL;  -- forces stale re-sync, safe

Suggestion

A boot-time data migration would save every long-running self-hoster from hand-repairing their DB. Minimal version:

-- Crash 1
UPDATE connection SET credential_write = NULL WHERE credential_write = '';
UPDATE connection SET last_health = NULL WHERE last_health = '';
UPDATE connection SET provider_state = NULL WHERE provider_state = '';
-- Crash 2
UPDATE connection SET tools_synced_at = NULL WHERE typeof(tools_synced_at) = 'blob';

(item_ids is NOT NULL so '' rows there would indicate worse damage; none existed in my DB.)

Alternatively, SQLiteTextJson.mapFromDriverValue could treat '' as NULL instead of throwing, which fixes Crash 1 generically for every JSON column in every table.

Notes

• This is distinct from Desktop app crashing? Start here #966 (desktop app crash-triage), though the underlying theme is the same: legacy data + new storage layer. The advice there ("reset data") works but throws away connections unnecessarily — the rows above are repairable in place.
• After applying the two fixes, all toolkit endpoints return 200 on initialize and tools/list with no log errors.