Storage

One file, one schema

core/db.py declares every table this process creates (SQLAlchemy metadata), and DATA_DIR/praxis.db (~/.praxis/praxis.db) holds all of them:

Tables	Accessors
chats, events — the transcript	core/transcript.py
results — full rows of a query result	core/transcript.py
dashboards, tiles, layouts, runs, tile_results — dashboards	core/boards.py
deployment — one row (CHECK id = 1): credentials, secrets, addresses, cached introspection	core/deployment.py
mcp — MCP enabled flag and token	core/access.py
checkpoints, writes — LangGraph	LangGraph's sqlite saver

core.db.engine is the one opener — one cached engine per file — so PRAGMA foreign_keys=ON cannot drift; the file runs in WAL mode because the checkpointer shares it. It is chmod 600 because deployment.credentials is plaintext JSON with the warehouse password.

Accessors live beside the schema, one module per group. A module belongs in core when it needs nothing from above, and a table's accessors never do.

No migrations, deliberately

create_all does not diff. A new table appears on an existing database; a new column silently does not, and changing one means deleting the file. Affordable for one machine, and the first thing to revisit if this ever runs shared. It is also why the flow chat's binding is flows/<flow>/chat.json, not a column on chats.

A file is not a row

An uploaded CSV, a script's chart and a file's profile live in the sandbox service's volume, one directory per chat, with no table — a row would only ever be a stale copy of ls.

Everything about a Flow — definition, runs, event logs, artifacts, Gate state — is files under FLOWS_DIR. See decision 0006.

Elsewhere in DATA_DIR

• jobs.lock — one model-build job at a time, machine-wide.
• drafts/<ref>/ — the model editor's working copy of each branch.