[nasser1941]

For #242. On 0.61.0's gate run, 2 of 3 sparql-university attempts wrote solution.sparql and said they were done without running it, and the third tried and gave up when it found no Python. All 3 failed. The check for code it never ran (#245) left query files out, so nothing sent them back.

What changes

• A new .sparql or .rq file counts like a new code file: if no command since its last change ran it, lcode sends the model back once, to run it or to say plainly that it's untested.
• The note adds "for a query, run it on its data". When uv is installed, it says how to run Python in a throwaway environment, and that the package for a SPARQL query is rdflib.

Left out on purpose

.sql files. One can be a migration, and asking the model to run it could change a real database. A SPARQL query rarely has that risk.

Testing

A new test covers the whole path with FakeOllama: the query is written, the model says it's done, the note sends it back once, and it runs. The full suite, ruff, basedpyright and mkdocs build --strict pass.

With a real model (qwen3.6-35b at 32K), two attempts at sparql-university on this branch: 1 of 2 passed, and it was 0/3 on each of the last two nights. Both attempts ran the query on their own, one by installing rdflib and the other from a script, so the new note never fired, and this doesn't show what the note itself does. Only the FakeOllama test covers that so far, and tonight's gate has three more attempts. Two attempts are also too few to credit this change for the pass.

This changes what the model is told, so it goes through the next gate before it ships.

🤖 Generated with Claude Code