Python First Verification
The first Python verification should prove async execution through a real provider, not just import generated modules.
1. Static package checks
From the generated console directory, with the virtual environment and package installed as described in Python Quick Start:
python -m compileall app
Add application tests and declare their test dependencies before running
python -m pytest -q; the generated console does not itself supply this suite.
Use scoped Assist for exact terminal signatures instead of searching generated
library source. Execution must receive trusted context; a second data-service
argument must be rejected.
2. SQLite round trip
Run the shared School SQLite example
from a checkout of teaql-io-site:
npm run verify:school -- --language python
The complete Python application uses the same School
model as the other six runtimes. The runner generates and builds the workspace,
checks root/constant bootstrap twice, saves a school with an audit reason, then
starts a new process to read it back. An independent SQLite checker verifies
persisted fields and relationships after every phase. Logs, exit codes and row
snapshots are retained under artifacts/school/.
3. Loaded-state check
Run a partial projection containing a nullable field. Prove that a loaded null differs
from an unselected field through the generated E facade. This prevents application
logic from treating “not fetched” as a database null.
4. Governance failures
Add tests for missing comment, missing purpose, uninitialized context, missing
audit_as, stale version, unknown dynamic field/operator, invalid range/paging, and
attempts to supply tenant, actor, permissions, provider, or policy in JSON. Expected
behavior is an explicit error, never an empty or broader query.
with pytest.raises(TypeError):
await executable.execute_for_list(context, data_service)
The explicit arity test protects the context-only API from regression.
5. Native SQL check
Retain a masked trace showing bound parameters, stable ordering, native aggregate SQL, and exact per-parent limits across at least two parents. Depending on runtime policy, SQLite can use bounded probes instead of a window plan. Rows, count, sum, and facets must share the same active filter.
6. Regeneration check
Add a field, regenerate the whole package, reinstall it, compile, and rerun tests.
Confirm centralized plural/predicate naming remains correct. Never preserve a manual
edit under generated models/ or requests/.
Record Python/runtime/generator/provider versions, model hash, commands, exit codes, masked native response, and SQL shape as the baseline evidence.