Skip to content

Project.validate() can't detect content changes served by a custom fscommands #925

Open

Description

@mcepl

A custom fscommands.read() lets an editor or a language server give Rope the contents of unsaved buffers. pylsp-rope's WorkspaceFileCommands does this, and so will rope-language-server. Rope caches parsed modules, though, and Project.validate() decides whether a file changed only from ChangeIndicator, which calls os.path.getmtime() and os.path.getsize() on the real file. An edit to a buffer changes neither, so Rope keeps analysing the old module even after validate().

To Reproduce

import pathlib
import tempfile

from rope.base import fscommands
from rope.base.project import Project


class Overlay(fscommands.FileSystemCommands):
    """Serve unsaved editor buffers, fall back to the disk."""

    def __init__(self):
        self.buffers = {}

    def read(self, path):
        if path in self.buffers:
            return self.buffers[path].encode()
        return super().read(path)


root = pathlib.Path(tempfile.mkdtemp())
(root / "m.py").write_text("def f():\n    return 1\n")
overlay = Overlay()
project = Project(str(root), ropefolder=None, fscommands=overlay)
resource = project.get_resource("m.py")
print(sorted(project.get_pymodule(resource).get_attributes()))

overlay.buffers[resource.real_path] = "def g():\n    return 2\n"
project.validate(resource)
print(sorted(project.get_pymodule(resource).get_attributes()))
print(resource.read())

Output on master (46ebf37), Python 3.13:

['f']
['f']
def g():
    return 2

resource.read() returns the new text, but the cached module still has f and not g. Refactorings then compute offsets against text the user no longer has.

Workarounds today

  • pylsp-rope calls the private project.pycore._invalidate_resource_cache(resource) for every open document on every request.
  • libutils.report_change() does invalidate, but the documentation describes it as the call to make after saving, and with the default automatic_soa=True it also runs static object analysis on every call. That's too slow to run on every keystroke.

Proposal

Either of these would be enough, and neither changes the existing fscommands contract:

  1. A public way to invalidate one resource, for example project.invalidate(resource) or project.validate(resource, force=True). This only needs to notify the resource observers, without the analysis that report_change() adds.
  2. An optional hook, for example fscommands.get_change_indicator(path), which ChangeIndicator uses when the fscommands object provides it, and falls back to os.stat() otherwise. An overlay could then return the buffer version.

The second option is close to part of #911, but kept optional, which matches the concern raised in the review of #911 that new fscommands methods must not break existing implementations such as pylsp-rope's.

Editor information

Activity

  1. lieryan commented on Nov 10, 2026

    @lieryan
    Member

    I think report_change() might be the correct place here. If you have an fscommand redirecting file accesses to the editor's unsaved buffer, I think for as far as rope is concerned, the state of the file in the editor buffer is what's relevant for any file operations. It's essentially the same as auto-saving editor.

    The issue is just performance when report_change() might be doing SOA on every keystroke and that's expensive. Maybe what we needed is a way to tell report_change() to skip SOA or we can implement a debouncer for report_changes() so that SOA only runs when the user stopped typing.

    Automatic SOA should probably run as a background thread in the first place.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions