fix(core): tell the model what file tools skipped and accept numbers written as strings

[Max17190]

Why

The file tools hid what they left out, so the model drew the wrong conclusions:

• grep silently skipped files over 1.5 MB, files that are not UTF-8 text, and files it could not read. A "no matches" result could therefore mean "never searched", and the model would decide that code did not exist.
• read_file refused every file over 1.5 MB, even when the model asked for a small range of lines. The refusal told it to use grep, but grep also skips those files, so the advice led nowhere.
• read_file on a directory, or a failed read for any other reason, reported "not a UTF-8 text file". A call with no path failed with a confusing error about an empty path.
• Outside a git repository, glob and grep ignored the project's .gitignore, so node_modules/ and build output flooded results. Turning that on naively would let a .gitignore above the project (for example a home directory kept in git that ignores *), or the user's global git excludes, hide the whole project.
• Models often send numbers and booleans as strings ("offset": "200", "replace_all": "true"). Those values were silently dropped and the defaults used instead, so the model read the wrong lines or edited only the first match.

Summary

• grep counts the files it skipped and adds one line at the end, for example … not searched: 1 file over 1.5 MB, 2 files not UTF-8 text. Unreadable files are counted too. The size check measures what the read actually gets, which is a symlink's target.
• read_file with offset/limit on a file over 1.5 MB reads the file one line at a time, keeps only the requested lines (and the first 500 bytes of each), and formats them through the same code as small files, so the output is identical. The read stops as soon as the call is cancelled instead of running on to the end of the file. A whole-file read of a large file is still refused, now with advice that works: "read a range of lines with offset and limit".
• read_file on a directory says <path> is a directory; use list_dir to see its entries. A missing or blank path to read_file, write_file or edit_file says missing required argument: path. Other read errors report the real OS error.
• glob and grep share one walker builder:
  • Inside a repository, behavior is unchanged.
  • Outside a repository, the project's .gitignore files apply from the project root down. A .gitignore above the project and the user's global git excludes do not apply, as git would not apply them. .ignore files keep applying from every ancestor.
  • Ignore rules rank the same way in every walk: the nearest file that names a path decides, and a .ignore (from any ancestor) outranks every .gitignore. A search scoped to a subdirectory therefore finds the same files as a whole-project search, including files a .ignore above the walk or above the project whitelists.
  • Each directory's ignore files are read once per walk, not once per file they apply to.
  • .git is still excluded by name.
• offset, limit, timeout_secs and replace_all accept either their JSON type or a string. Strings are trimmed and matched without regard to case, and unusable values still fall back to the default.
• Tool schemas are unchanged, so the model-visible prompt bytes are unchanged.

Test Plan

New tests in crates/core/src/tools.rs. Each regression test was seen failing before the fix; the line reader test pins the new helper:

• grep_counts_the_files_it_could_not_search: the exact skipped-files line for oversized and non-UTF-8 files, with and without matches, and no line for a scoped search that skipped nothing.
• read_file_reads_a_window_of_a_file_too_large_to_load_whole: a range from a large file is byte-identical to the same range from a small file, and the whole-file refusal text is exact.
• a_cancelled_read_of_a_large_file_stops_reading: a range read of a large file returns as cancelled once its call is cancelled, instead of reading the file to its end.
• streamed_lines_split_like_str_lines: the streaming line reader splits lines exactly as str::lines does, across buffer boundaries and with CRLF endings.
• file_tools_name_a_directory_and_a_missing_path: the directory message and the missing or blank path message for all three file tools.
• numeric_and_boolean_arguments_written_as_strings_are_honored: string forms of offset, limit, timeout_secs and replace_all take effect.
• grep_and_glob_honor_gitignore_without_a_git_repository: ignored directories are left out of whole and scoped walks, a path named directly is still searched, and a nested repository is not hidden by an outer .gitignore.
• a_gitignore_above_a_project_without_git_cannot_hide_it: a .gitignore of * above the project hides nothing, while an ancestor .ignore still applies.
• a_scoped_walk_ranks_ignore_files_as_a_whole_project_walk_does: a nearer whitelist beats a farther ignore rule in a scoped walk.
• an_ignore_file_above_the_walk_outranks_a_gitignore_inside_it: a .ignore whitelist at the project root, and one above the project, keep a file a .gitignore ignores, in whole and scoped glob and grep.
• global_git_excludes_apply_only_inside_a_repository: global excludes hide files inside a repository and not outside one.

Results:

• cargo test --workspace --locked: passes (1089 passed, 0 failed), including frozen_prompt_fits_token_budget. The tools tests also pass with a global excludes file containing *.log, .env, .idea/ and node_modules/.
• cargo +1.97.0 clippy --workspace --all-targets --locked -- -D warnings: passes.

Confidence Score: 5/5

No outstanding findings block merging.

Summary

The PR clarifies file-tool results, supports ranged reads of large files, accepts string-form tool arguments, and applies project .gitignore rules outside repositories. The current code addresses the three previously reported search and cancellation issues.

_{Reviews (2) · Last reviewed commit: "fix(core): stop a cancelled read of a la..."}