[MomoSolaris8]

This PR moves the database code from psycopg and handwritten SQL to SQLAlchemy,
and makes Alembic migrations the only thing that creates or changes tables.

What changed

• SQLAlchemy models for all 20 tables in db/models/: auth, memory, MCP, claims, RAG chunks, uploads
• Migrations:
  • baseline: creates every table
  • seed: adds the default MCP servers and tool policies
  • HNSW index on chunks.embedding (vector_cosine_ops)
• Every store module now uses SQLAlchemy sessions: uploads, auth, MCP, memory, claims, RAG
• Removed all ensure_*_schema functions, so the app no longer runs DDL at startup or on every request
• Fix: a user could take over another user's thread by sending its thread_id
• Requires Python 3.11

Testing

• 70 integration tests run against a real Postgres (pgvector) in CI
• Each module got its tests before the swap. They passed on the old code and on the new code.
• test_migrations checks that the migrations create every table and that the models and migrations are in sync
• HNSW benchmark on 100k vectors (1024 dims): 307 ms (seq scan) → 4.6 ms (index scan)

Deployment

• Existing database (Supabase): checked that its schema matched the models
  except for the HNSW index, stamped it at the baseline, then ran alembic upgrade head
• Every deploy now runs alembic upgrade head before updating the Container App
• New database: just run alembic upgrade head

Follow-ups (not in this PR)

• Move the database to Azure Postgres and the files to Azure Blob Storage
• Clean up repo-wide lint (ruff, import order, B008)
• Fix with tests:
  • MCP credentials are stored in plaintext
  • any user can register a stdio MCP server
  • claim audit events can have ordering ties
  • a claim review run can be started twice

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
support-agent-one-coyl	Ready	Preview	Oct 29, 2026 5:08pm UTC