fix: stop forwarding Hub download options to the tool constructor

[TINGyu123644]

Fixes #2928

What changed

Tool.from_hub() documents that extra kwargs are split between Hub
download options and tool-constructor options. In practice the download
options (cache_dir, orce_download, proxies,
evision,
subfolder, local_files_only) were only read with kwargs.get(...),
so they stayed in kwargs and were forwarded to Tool.from_code(tool_code, **kwargs), which eventually calls ool_class(**kwargs).

A tool with a no-argument constructor therefore failed with:

ext TypeError: EchoTool.__init__() got an unexpected keyword argument 'cache_dir'

This change pops the Hub download options before the constructor is
invoked, so only tool-specific constructor arguments reach rom_code.

Why it matters

The documented split now actually happens: cache_dir,
local_files_only, etc. are used by the download layer and excluded from
the tool-constructor arguments, while tool-specific kwargs are still
forwarded unchanged.

How it was tested

• Added regression test est_from_hub_pops_hub_download_options that
  loads a no-argument-constructor tool through rom_hub with Hub
  download options plus a tool-specific kwarg, and asserts the download
  layer received the Hub options while the constructor only received its
  own kwarg.
• Verified the test fails against the old kwargs.get(...) behavior
  (TypeError) and passes with this fix.

[VANDRANKI]

Community review (not a maintainer, does not clear the merge gate). I read the full diff and compared it with its merge base (96f33fa, current main) byte for byte.

The first thing to know: the diff shows +2536 -2485 across src/smolagents/tools.py and tests/test_tools.py, but almost all of that is a line-ending change. In the merge base both files use LF (0 carriage returns in tools.py); on this PR head tools.py has 1422 CRLF lines out of 1422, and tests/test_tools.py is CRLF throughout as well. With git diff -w --ignore-cr-at-eol the real change is 6 lines in tools.py plus 51 added lines in the tests. As submitted, this would rewrite every line of two files, destroy git blame, and conflict with any other open PR touching them. I would redo the commit with LF endings (git config core.autocrlf input, or re-save the files as LF) so the diff is the 6 + 51 lines. The PR description also looks mangled (a BOM at the start and stray backslashes in the code spans), which is probably the same editor or shell problem.

The code change itself is right. Tool.from_hub now uses kwargs.pop(...) for cache_dir, force_download, proxies, revision, subfolder and local_files_only, so they reach hf_hub_download and no longer reach Tool.from_code(tool_code, **kwargs). What I ran, with PYTHONPATH=C:/sm/src and a mocked hf_hub_download (no network):

• The new test_from_hub_pops_hub_download_options passes on this PR head.
• With only tools.py put back to the merge base, the same test fails with TypeError: EchoTool.__init__() got an unexpected keyword argument 'cache_dir'. So the test reproduces #2928 and the fix removes it.
• I did not run the rest of the suite and did not call the real Hub.

Duplicates: #2931 (opened 06:59Z, 12 lines, one file) and #2936 (85 lines) fix the same issue, and this one is the third. Of the three, #2931 is the smallest, and this one's clean diff would be about the size of #2936. The maintainers will probably want only one, so it may be worth agreeing on which among you before more time goes into it. I only checked that the three target the same lines; I did not diff #2931 and #2936 against this one.

The tests use patch on smolagents.tools.hf_hub_download and assert the download kwargs, which is a good shape; a second assertion that token and repo_type are still forwarded would guard against a future regression, but that is optional.

[TINGyu123644]

@VANDRANKI Thanks for the byte-for-byte review — you were right about the line endings. Fixed on the branch (f736a64):

• src/smolagents/tools.py and tests/test_tools.py are now LF (the previous commit had written them with CRLF).
• The PR diff is now just the real change: tools.py +6/−6 (the kwargs.pop fix) and the new regression test (tests/test_tools.py +51).

On the duplication with #2931/#2936: the code here is functionally the same fix, and I'm happy to close this PR in favor of whichever one the maintainers prefer to keep — let me know.

[VANDRANKI]

Thanks for the LF fix. I re-checked the new head (f736a64): the diff is now tools.py +6/-6 and tests/test_tools.py +51, nothing else.

What I ran (Windows, Python 3.13, PYTHONPATH pointing at the PR checkout):

• test_from_hub_pops_hub_download_options passes on the PR.
• With src/smolagents/tools.py taken from origin/main it fails with TypeError: EchoTool.__init__() got an unexpected keyword argument 'cache_dir', so the test guards the fix.

One thing will fail CI. .github/workflows/quality.yml runs ruff format --check examples src tests. On this head ruff format --check tests/test_tools.py reports the file would be reformatted (ruff 0.15.7), while the same file from origin/main passes. The diff is three small hunks: one more blank line before def test_from_hub_pops_hub_download_options, ''' to """ around the dedent source, and the with patch(...) as mock_download: header collapsed onto one line. Running ruff format tests/test_tools.py should fix all three. ruff check is clean on both files.

On the overlap with #2931 and #2936: all three are still open and I cannot pick between them, that is for the maintainers. I did not run the other two.

[TINGyu123644]

@VANDRANKI Thanks for the precise report — fixed on the branch. The three hunks are applied to tests/test_tools.py (new head 2f255d415aeb12d50b5c381c7c2ccf3b9753838b):

1. one more blank line before def test_from_hub_pops_hub_download_options;
2. single-quote triple quotes → double-quote triple quotes around the dedent source;
3. with patch(...) as mock_download: header collapsed onto one line.

Verified with the repo's own config (line-length = 119): ruff format --check now reports the file as already formatted, and the PR diff is back to just tools.py +6/−6 plus the regression test (now +50 lines — the collapsed with header is one line shorter than the previous +51).

Thanks again for the byte-for-byte review.

[TINGyu123644]

@VANDRANKI One heads-up: the CI workflows on head 2f255d41 are showing action_required (waiting for approval) rather than running — typical for a first-time contributor. The Quality Check workflow is the one that runs ruff format --check tests; it should now pass since the three hunks are fixed. If you get a chance to approve the run, happy to address anything it flags. Thanks!