vis-lang-python

Python tools for Vis: ruff formatting and lint, pytest runs and a REPL that keeps its globals between calls.

Vis knows nothing about Python. This extension does, and it runs the tools your project already uses — its ruff, its pytest, its interpreter — so what you see here matches what your CI sees.

Install

vis-agent extension install 'blockether/vis-lang-python' --global --trust

Tools

py.format_code(["src"], is_written=True)     # ruff format
py.lint_code(["src"])                        # ruff check, with rule codes
py.run_tests(["tests"], keyword="repl")      # pytest, counted from its JUnit report
py.repl_start(cwd="~/app")                   # the project's own interpreter
py.repl_eval("model = load()", cwd="~/app")  # globals live between calls
py.repl_stop(cwd="~/app")

Which interpreter runs your tests and REPL is decided per project, first match winning: uv (uv.lock or [tool.uv]), Poetry (poetry.lock), a local .venv, then python3. Ruff is taken from the project's virtualenv, then PATH, then the copy installed with this extension.

The REPL stays alive when the Vis sandbox restarts, for example after /reload. The next call reaches the same interpreter, with its globals. Only py.repl_stop and the end of the session stop it.

Keep files parseable

Edit hooks repair structural mistakes and validate syntax locally in Python. They use ast.parse and compile the syntax tree without executing it. They do not start a project interpreter, write bytecode files or change REPL globals. Project lint and tests still use your project's tools and interpreter.

• Before a patch writes, the guard tries to repair invalid source within the changed lines. The host writes the validated result once and reports the corrections. If repair fails, a patch that breaks a parseable file writes nothing.
• Vis repairs a Python block itself before the block executes. This extension does not change blocks. Before each block, the guard records the state of the Python files for the check after the block.
• After a Python block, the guard can repair changed .py and .pyi files. These repairs happen after the original writes. The block is not transactional. Repairs include notes and diffs in session["python_syntax_repairs"]. Unresolved errors remain in session["python_syntax_errors"] until the files parse again.
• Before repl_eval runs code, the project's interpreter compiles all of it. If the code does not parse, none of it runs, and the guard's repair is tried on the code. When the interpreter accepts the repaired code, it runs, and the result lists each correction in repairs. If no safe repair exists, the error gives the line, the column and the parser's message.

The shared syntax guard selects matching files and prevents a repair from overwriting a detected concurrent change. Repairs require a Vis host with support for repair decisions. If the parser is unavailable, the guard allows the operation and logs the failure. It never accepts a repair without successful validation.

Requirements

pytest in the project you test. Everything else is installed with the extension.

Development

vis-agent python -m pytest tests -q

Results follow the contract in vis-lang-interface. The repair port and its regression corpus come from clj-parinferish under the MIT License. See NOTICE for the required license notice.