fix(rust): match managed rg globs under ./-prefixed search paths - #229

Hi, just checking in. This still merges cleanly against main. Happy to rework it if you'd prefer a different approach.

Reviewed PR head bf4391810e against the fetched origin/main at 6e71bbb46f9.

The original bug is still present on main. build_walker still resolves search paths through root.join(path). On Unix, the pinned ignore 0.4.33 strips the workspace root as a byte prefix, leaving ./src/... for inputs such as . or ./src. Directory-qualified include/exclude globs and explicit ignore-file rules therefore still encounter the mismatch described here.

The focused fix at the walker boundary looks appropriate for the original baseline. However, two changes are needed before integrating it with current main:

1. [P1] Update the regression test for the current match API. The new test accesses item.relative_path.clone(), but perf(rust): streamline rg search and reduce match allocations #287 replaced that field with the relative_path() accessor. The test target will fail to compile after merging with current main. Please use item.relative_path().to_path_buf().

2. [P2] Preserve trailing-slash errors for file discovery. components().collect() also removes trailing / and /.. The newer file-discovery caller shares build_walker but deliberately bypasses check_paths, relying on the walker to report invalid inputs. If src/keep.rs is a regular file, zg --rg --files src/keep.rs/ (or src/keep.rs/.) should report a traversal error on Unix. This normalization instead turns the input into a valid file path; discovery then returns the original invalid spelling and exits successfully. This breaks the stated intent to preserve trailing-slash error behavior. Please retain that directory constraint, or validate the original discovery input while preserving its errors, and add a regression case.

Both findings concern integration with changes made on main after this PR's baseline. These conclusions are based on static inspection of the source and pinned dependencies; I did not compile or run tests.

Thanks for checking it against current main. Rebased onto 891497f and pushed c790f2a for both points:

• The test now uses item.relative_path().to_path_buf().
• Trailing slash: you were right. A zg built from the previous head printed src/keep.rs/ for zg --rg --files src/keep.rs/ and exited 0, and did the same for src/keep.rs/., where main exits 2 with Not a directory. resolve_match_path still drops the . segments, but now puts the trailing / or /. back when the path names something below the root. Both inputs exit 2 again with main's error text (with a ./ prefix, the error shows the path without it).

Two new tests in lexical/files.rs. file_inputs_spelled_as_directories_report_errors (Unix only) checks that src/keep.rs/, src/keep.rs/., ./src/keep.rs/ and ./src/keep.rs/. each give one error and no paths; it fails on the previous head and passes on main. path_globs_match_inputs_with_a_current_directory_prefix runs --files -g 'src/**' over ., ./ and ./src/, since file discovery uses the same walker and had the same glob bug (on main, zg --rg --files -g 'src/**' . prints nothing and exits 1); it fails on main.

I also added ./. to the search test. Without the "names something below the root" check it kept its trailing /. and the walker was back at <root>/./src/.... Removing any part of the new check fails a test.

cargo fmt --check, cargo clippy --workspace --all-targets -- -D warnings and cargo test -p zg-engine --lib (663 passed) pass locally.

Reviewed the updated head c790f2a8ac633f8f9009abb57385e5e0238eea1d against base 891497fa87d3e03d0c9273fed0ba3e75af62f179.

Both previous findings are addressed: the search test now uses relative_path(), and file discovery preserves errors for regular-file inputs ending in / or /., with regression coverage. Keeping the fix at the shared walker boundary is appropriate.

One remaining edge-case regression:

[P2] Preserve the workspace root's prefix when normalizing relative inputs. At resolve_match_path, line 889, root.join(path).components().collect() normalizes the absolute workspace root as well as the relative search path.

On Unix, std::path::absolute deliberately preserves exactly two leading slashes. An MCP/API root such as //tmp/workspace therefore reaches this helper unchanged. For a relative input src, the previous implementation produced candidates such as //tmp/workspace/src/keep.rs, which matched src/**. The new implementation produces /tmp/workspace/src/keep.rs, while OverrideBuilder::new(root) still uses //tmp/workspace. The pinned ignore implementation strips that prefix byte-for-byte on Unix, so stripping now fails: directory-qualified include globs miss files, and corresponding exclude globs no longer exclude them.

For example, given an existing Unix workspace with src/keep.rs, this MCP request exercises the affected path:

{"root":"//tmp/workspace","command":"rg --files -g 'src/**' src"}

Please normalize only the relative input while preserving the workspace root's prefix, and add coverage for a double-slash root with include and exclude rules. This keeps the change focused on removing redundant . segments without breaking previously working inputs.

This review is based on static inspection of the source and pinned dependencies; I did not compile or run tests.

ZG Retrieval Benchmark

Retrieval-only · PR base vs head

✅ Complete

Base: zvec-ai/zvec-grep@891497fa87d3e03d0c9273fed0ba3e75af62f179

Head: JayOfTheKeyboard/zvec-grep@bdfe27e1e57746bee2e930e0a9a79afa6b0dc7ed

Each cell is base / head / Δ. Δ = head − base; all five metrics use the 0–1 scale and higher is better. Both versions use the same frozen harness, queries, gold labels and local embedding models, with fresh indexes and five calls per query/mode. Quality is report-only; no regression threshold is applied.

SWE-QA nDCG uses repository macro averaging; the other suites use query means. Suites are reported separately. Full base/head reports and per-call evidence are available in the run artifacts.

View run and artifacts

Embedding: local. e2e: skipped. This comment is updated by automatic PR Retrieval runs.

Good catch, thanks. Pushed bdfe27e: resolve_match_path now drops the . segments from the relative input only and joins that onto the root as given, so //tmp/workspace keeps its prefix.

New test path_globs_match_inputs_under_a_double_slash_root in lexical/files.rs lists src and ./src under a // root with -g 'src/**' and -g '!src/**'. It fails on c790f2a and passes now. I also sent your MCP request through zvec_grep_rg with a // root: on c790f2a, rg --files -g 'src/**' src returned nothing and -g '!src/**' listed src/keep.rs; with the fix it's the other way round.