refactor(mcp): tree-sitter masked source for text scanners#391
Conversation
|
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 590818df32
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| b',' => { | ||
| if let Some(context) = stack.last_mut() { | ||
| context.top_level_commas += 1; | ||
| } |
There was a problem hiding this comment.
When a formatting macro argument before the message contains turbofish/type arguments with commas, e.g. assert_eq!(make::<A, B>(), "{HashMap}"), this comma is counted as a top-level macro separator because only ()[]{} nesting is tracked. That makes the second argument look like assert_eq's format-string slot, so restore_format_captures preserves HashMap even though it is just literal string data; tracedecay_unused_imports will then miss a truly unused HashMap import in otherwise valid Rust code.
Useful? React with 👍 / 👎.
ScriptedAlchemy
force-pushed
the
codex/treesitter-source-mask
branch
from
590818d to
999bbcb
Compare
Replace the hand-rolled Rust lexer in analysis.rs (MaskState machine, mask_rust_noise, try_open_raw_or_byte_string, is_char_literal_start, raw_string_closes) with a shared tree-sitter-driven masking facility in src/extraction/source_mask.rs. It blanks comment and string/char literal node ranges reported by the existing Rust grammar, preserving newlines and byte offsets, with opt-in node sets (mask_comments / mask_strings) so tracedecay_todos is never affected. All three text scanners now share it: handle_unused_imports uses the UNUSED_IMPORTS preset (captures preserved); has_bare_call (via cached_lines) and contains_unsafe_block_start (via handle_unsafe_patterns) use the CODE_SCAN preset, fixing the comment/string false-positive class that was previously fixed for imports only. unsafe_patterns detects on the masked copy but still emits the original line as the snippet; non-Rust files are scanned raw. Format-capture carry-over is honored via tree-sitter spans plus a tiny delimiter/comma walk over the remaining code bytes, matching the EXPANDED semantics of PR #383 (codex/format-capture-completeness): panic/todo/ unimplemented/unreachable at arg 0, assert/write/writeln at arg 1, assert_eq/ne at arg 2, position-aware first-literal selection. The old mask_rust_noise unit tests plus #383's expanded capture test are carried over as contract tests against the new facility; the hand-rolled lexer and its helpers are deleted. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
ScriptedAlchemy
force-pushed
the
codex/treesitter-source-mask
branch
from
999bbcb to
c827eea
Compare
1 participant
What
Replaces the hand-rolled Rust lexer in
src/mcp/tools/handlers/analysis.rs(theMaskStatemachine,mask_rust_noise,try_open_raw_or_byte_string,is_char_literal_start,raw_string_closes,is_direct_format_literal_start,format_capture_identifier_end) with a shared tree-sitter-driven masking facility:src/extraction/source_mask.rs.The facility parses with the existing Rust grammar and blanks the byte ranges of
line_comment/block_comment/string_literal/raw_string_literal/char_literalnodes, preserving newlines and total byte length so line/byte indexing stays valid.API
masked_rust_source(&str) -> String— comments + strings, format captures preserved (theunused_importsbehavior).masked_rust_source_with(&str, MaskOptions) -> String—MaskOptions { mask_comments, mask_strings, preserve_format_captures }, withMaskOptions::UNUSED_IMPORTSandMaskOptions::CODE_SCANpresets. Node sets are opt-in per caller, sotracedecay_todos(which deliberately scans comments) is untouched.Per-scanner migration
handle_unused_imports— swapsmask_rust_noiseformasked_rust_source(UNUSED_IMPORTS preset); per-file caching infile_cacheretained (each file parsed once per call).has_bare_call(recursion self-call probe) —cached_linesnow caches the CODE_SCAN-masked source, so aname(appearing only in a comment/string no longer counts as a self-call.contains_unsafe_block_start(viahandle_unsafe_patterns) — the handler masks each Rust source (CODE_SCAN) and runs detection on the masked copy while emitting the original line as the snippet; non-Rust files are scanned raw (a newpath_is_rusthelper gates masking).Format-capture decision
Honored via tree-sitter spans plus a tiny linear delimiter/comma walk over the remaining code bytes — no second lexer. Tree-sitter does the only hard job the old lexer did (finding comment/string boundaries); the delimiter walk reproduces the format-arg logic. This matches the expanded semantics of merged PR #383 (
codex/format-capture-completeness), not the old masker's:panic/todo/unimplemented/unreachableat arg 0;assert/debug_assert/write/writelnat arg 1;assert_eq/assert_ne/debug_assert_eq/debug_assert_neat arg 2; position-aware first-literal selection; comments between the macro name and(tolerated. The branch is rebased after #383 and retains its expanded contract (both converge on the sameformat_macro_argument_index/take_format_literal_contextlogic and the same expanded contract test).Tests
All original
mask_rust_noiseunit tests (comment/string/char/raw-string/byte-string/format-capture) are carried over as contract tests insource_mask::tests, along with #383'simplicit_captures_survive_supported_format_macro_positions. The hand-rolled lexer and its helpers are deleted.Line counts
analysis.rs: 2982 → 2635 (net −347; hand-rolled lexer + old test module removed).source_mask.rs: new, 500 lines (facility + docs + contract tests).Gates (CI-exact)
cargo fmt -- --check→ clean (EXIT 0)cargo clippy --workspace --all-targets --locked -- -D warnings→ clean (EXIT 0)cargo nextest run --test mcp_suite --no-fail-fast→ 364 passed, 0 failed (incl.unused_imports_*,unsafe_patterns_reports_unsafe_block_in_markdown_and_json)cargo nextest run --lib(touched modules) → 17 passed (10source_mask, 3unsafe_pattern_detection_tests, render/unused)🤖 Generated with Claude Code