The REST/Web-UI HTTPServer shares one long-lived PG connection across all
requests. Any statement that raised mid-request (e.g. a UniqueViolation from
a desynced SERIAL sequence) aborted the transaction; the global handler
returned 500 without rolling back, so every subsequent request failed with
InFailedSqlTransaction until restart — surfacing as "duplicate keys cause
500s" and "500 immediately after login".
- server.py: global handler now always rolls back the shared connection on
error and maps constraint violations to 409 (was 500/400). This is the one
funnel that guarantees the connection is never left aborted.
- db.py: add is_integrity_error() — dual-backend (psycopg + sqlite3)
constraint-violation classifier; replaces the fragile `"UNIQUE" in msg`
string match that never matched Postgres' error text.
- Remove make_write_mcp_server: a never-run duplicate of the MCP write tools
that had bit-rotted (wrong file_update arg order + FK-violating hardcoded
actor). Live writes go through oauth_mcp_app, which is correct.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
overseer
bdc984e5ff