// Handles all Rust-side work: control plane, native search engine (grep crates), native GPU engine (cudarc/CUDA), routing, index subsystem, rewrite substrate, and benchmark gates.
NOTE: Startup and cleanup are handled by worker-base. This skill defines the WORK PROCEDURE.
Use this worker for:
rust_core/src/main.rs, routing, JSON output)routing.rs)backend_ast.rs)index.rstest_routing.rs)$env:PYTHONPATH = '...\src' syntax.C:\Users\oimir\.cargo\bin\cargo.exe (may not be on PATH).uv pip install -e ".[dev,ast,nlp]" instead.benchmarks/rg.zip auto-extract or benchmarks/ripgrep-14.1.0-x86_64-pc-windows-msvc/rg.exe.Read feature description and preconditions carefully. Understand what assertions this feature fulfills.
Test-Driven Development FIRST: Write a failing Rust test (in rust_core/tests/ or inline #[cfg(test)]) that asserts the new behavior. Run it and verify it fails. Only then implement.
Implement the Rust logic to make the test pass. Follow these rules:
pyo3 auto-initialize is REMOVED. Never add it back.Sync — use thread_local! or per-task allocation.AstGrep::new(source, lang), Pattern::new(pattern, lang), root.find_all(pattern).For JSON output changes: Ensure all JSON outputs include the unified envelope: version (u32), routing_backend (string), routing_reason (string), sidecar_used (bool). Test by parsing output with serde_json.
For native search engine (grep crates): Use grep-searcher for search orchestration, grep-regex for pattern compilation, grep-matcher for the Matcher trait, grep-printer for output formatting, ignore crate for .gitignore-aware parallel directory walking. Reference the grep crate docs and the ripgrep source code for usage patterns. The grep crates handle binary detection, encoding, line counting, and output formatting — leverage them instead of reimplementing.
For CUDA/GPU work:
cfg(feature = "cuda").cudarc for CUDA runtime: CudaContext::new(device_id) for device init, compile_ptx() for NVRTC kernel compilation, stream.clone_htod() for host-to-device copy, stream.clone_dtoh() for device-to-host.cargo test --features cuda and cargo build --features cuda --release..factory/research/gpu-text-search.md for kernel design patterns.For AST search speed optimizations (parallel walk, pre-filter):
ignore::WalkParallel for directory traversal. The walker threads should do everything: walk, read, pre-filter, parse, match. Use mpsc::channel to collect results, then sort by (file, line) for deterministic output.source.contains(&literal) where literal is extracted from the AST pattern. If no literal is extractable (pure metavars), skip the pre-filter and scan all files.ignore::types::TypesBuilder to register language extensions at walk time — don't post-filter.min(cpus, 12) threads, read_to_string + contains check + parse + match per file..factory/research/sg_internals_report.md for sg's exact flow.For rewrite speed optimizations (fused I/O, no fsync):
plan_and_apply, pass the file content from the planning read to the apply phase — do NOT re-read the file.sync_all() with sync_data() or remove fsync entirely. Add a code comment citing NTFS rename atomicity guarantees. sg uses plain std::fs::write() with no fsync.ensure_file_not_stale inside apply_edits_to_file when ensure_files_not_stale already ran as a batch.For index changes: Preserve TGI\x00 magic + version byte. If format changes, bump FORMAT_VERSION and handle old-format migration (rebuild with warning, not crash). Verify query result parity with uncompressed index.
Run local Rust gates:
Set-Location "C:\dev\projects\tensor-grep\rust_core"
& "C:\Users\oimir\.cargo\bin\cargo.exe" test
& "C:\Users\oimir\.cargo\bin\cargo.exe" clippy -- -D warnings
For CUDA features:
& "C:\Users\oimir\.cargo\bin\cargo.exe" test --features cuda
& "C:\Users\oimir\.cargo\bin\cargo.exe" clippy --features cuda -- -D warnings
Run local Python gates:
uv run ruff check .
uv run mypy src/tensor_grep
uv run pytest -q
{
"salientSummary": "Implemented atomic writes in apply_edits_to_file: write-to-temp (.tg_tmp_<random>) + flush + rename. Added stale-file detection with mtime check before apply. Added 5 new tests: atomic_write_success, atomic_write_cleanup_on_failure, stale_file_rejected, no_temp_files_after_success, verify_still_works_with_atomic. cargo test: 100 passed. uv run pytest -q: 510 passed.",
"whatWasImplemented": "Changed apply_edits_to_file from std::fs::write to write-temp+rename. Added file_mtime_ns field to RewriteEdit for staleness tracking. Mtime captured during plan_file_rewrites, checked in apply_edits_to_file. On mtime mismatch: returns Err with 'file modified since plan' message. Temp files cleaned up in both success and error paths.",
"whatWasLeftUndone": "",
"verification": {
"commandsRun": [
{ "command": "cargo test", "exitCode": 0, "observation": "100 passed in 0.4s" },
{ "command": "cargo clippy -- -D warnings", "exitCode": 0, "observation": "no warnings" },
{ "command": "uv run pytest -q", "exitCode": 0, "observation": "510 passed, 14 skipped" }
],
"interactiveChecks": [
{ "action": "Built release binary, ran tg run --rewrite 'lambda $$$ARGS: $EXPR' 'def $F($$$ARGS): return $EXPR' bench_data", "observed": "Plan JSON includes routing_backend=AstBackend, routing_reason=ast-native" }
]
},
"tests": {
"added": [
{ "file": "rust_core/tests/test_ast_rewrite.rs", "cases": [
{ "name": "test_atomic_write_success", "verifies": "Write-to-temp + rename produces correct file" },
{ "name": "test_atomic_write_cleanup_on_failure", "verifies": "No temp files left on write failure" },
{ "name": "test_stale_file_rejected", "verifies": "Modified file between plan and apply is rejected" },
{ "name": "test_no_temp_files_after_success", "verifies": "No .tg_tmp_* files remain after successful apply" },
{ "name": "test_verify_with_atomic_write", "verifies": "Byte-level verify still works with new write path" }
]}
]
},
"discoveredIssues": []
}