Weavatrix Scan
weavatrix-scan is a deterministic, read-only repository scanner for static
analysis, code intelligence, indexing, and AI tooling.
It does more than walk a directory. A scan produces a stable manifest with
normalized paths, file sizes, optional content hashes, an aggregate revision,
and explicit evidence explaining why files were skipped. Linux and macOS builds
have zero mandatory runtime dependencies; Windows uses only winapi-util for
native volume and file identities.
Why another repository walker?
walkdir and jwalk are excellent traversal libraries. ignore adds mature
Git-style filtering. Weavatrix Scan now exposes four deliberately separate
layers:
Walker: iterative, streaming, lossless low-level traversal;Scanner: ignore-aware deterministic manifest, hashes, revision, and typed evidence;RepositoryMatcher: cached path selection for incremental consumers;ParallelWalker: bounded collected or streaming traversal for wide trees.
| Capability | weavatrix-scan | ignore | walkdir | jwalk |
|---|---|---|---|---|
| Iterative traversal | Yes | Yes | Yes | Yes |
| Lossless native paths | Yes | Yes | Yes | Yes |
| Continue after local errors | Configurable | Yes | Yes | Yes |
max_depth / bounded handles |
Yes | Yes | Yes | Depth limit |
| Same-filesystem boundary | Yes | No | Yes | No |
.gitignore hierarchy |
Yes | Yes | No | No |
| Custom ignore files | Yes | Yes | No | No |
| Repository / Git-compatible ignore modes | Yes | Yes | No | No |
| Override globs / source switches | Yes | Yes | No | Directory callback |
| Reusable cached matcher | Yes | Yes | No | No |
| Stable normalized paths | Yes | No | No | Sorted traversal |
| File sizes and content hashes | Yes | No | No | No |
| Aggregate deterministic revision | Yes | No | No | No |
| Typed manifest delta / rename evidence | Yes | No | No | No |
| Binary and oversized-file policy | Yes | No | No | No |
| Typed skip reasons and warnings | Yes | No | No | No |
| Symlinks skipped by default / loop detection | Yes | Yes | Yes | Configurable |
| Parallel collected / streaming traversal | Yes / Yes | Yes / Yes | No | Yes / Yes |
| Cancellation and whole-scan budgets | Yes | Quit only | No | No |
| Minimum depth / hidden policy | Yes / Yes | Yes / Yes | Yes / No | Yes / Yes |
| Default runtime dependencies | 0 Unix / 1 Windows | Multiple | 2 platform helpers | Rayon stack |
Use Walker when you only need paths. Use Scanner when downstream results
must be reproducible and explainable.
Install
[]
= "0.2"
Enable serialization only when needed:
[]
= { = "0.2", = ["serde"] }
Quick start
use ;
let options = default
.with_extensions
.with_parallelism;
let report = new.options.scan?;
println!;
for file in &report.files
for skipped in &report.skipped
# Ok::
For the fastest path-only discovery, disable content reads:
use ;
let report = new
.options
.scan?;
assert!;
# Ok::
Low-level walkers
Walker is a streaming iterator. It keeps paths as native PathBuf/OsStr
values, uses iterative DFS, bounds open directory handles, and yields local
errors according to policy:
use ;
let options = default
.with_max_depth
.with_max_open
.with_same_file_system
.with_error_policy;
let mut walker = with_options?;
while let Some = walker.next
# Ok::
The root is depth zero. After receiving a directory, callers can invoke
skip_current_dir() before requesting the next item. Symbolic links are not
followed by default; enabling .with_follow_links(true) keeps traversal inside
the root and reports loops as typed skip reasons.
ParallelWalker is a separate collected mode for wide directory frontiers:
use ParallelWalker;
let report = new
.with_parallelism
.walk?;
println!;
# Ok::
For pipelines that should parse entries immediately instead of collecting
them, visit invokes a thread-safe callback directly on traversal workers:
use ;
let summary = new.visit?;
# Ok::
Scan modes
The same scanner supports three useful cost levels:
| Mode | Configuration | Reads content | Skip evidence | Hashes content |
|---|---|---|---|---|
| Rich manifest | ScanOptions::default() |
Yes | Complete | Yes |
| Safe discovery | hash_file_contents = false |
First 8 KiB | Complete | No |
| Metadata only | .metadata_only() |
No | Complete | No |
| Selected manifest | .metadata_only().selected_files_only() |
No | Omitted | No |
Content inspection uses available CPU parallelism by default. Set
.with_parallelism(1) for a serial run or pass a fixed worker count when a
host application owns the wider scheduling policy.
Output contract
ScanReport contains:
root: canonical absolute repository root;files: stable, lexicographically sortedScannedFilevalues;skipped: stable, sorted evidence for excluded entries;warnings: non-fatal ignore-file and local I/O diagnostics;ignore_sources: typed location and hash of every loaded selection input;revision: FNV-1a digest over ignore inputs, selected paths, optional content hashes, portability, and partial-termination state;complete: false when local errors made the evidence partial.termination: typed reason for a bounded or cancelled partial scan;portable: false when host-level Git configuration affected selection.
Each ScannedFile contains an absolute path, slash-normalized repository path,
byte size, and optional content hash. Default hashes are deterministic FNV-1a
digests intended for change detection, not cryptographic verification. Native
paths remain lossless in the walker and absolute PathBuf; invalid Unicode
units in normalized manifest names are escaped (%XX on Unix, %uXXXX on
Windows) instead of being replaced with the lossy Unicode replacement marker.
With the serde feature, invalid native path units use a tagged byte/wide-unit
representation and round-trip without loss; ordinary Unicode paths remain
plain JSON strings.
Incremental consumers
Two completed reports produce a stable changed-file set without filesystem access:
use ;
let previous = new.scan?;
// Apply repository changes, then scan again.
let current = new.scan?;
let delta = current.delta_from;
assert!;
println!;
# Ok::
Rename evidence is emitted only when the same content hash is unique in both
manifests; duplicate-content moves remain explicit add/remove pairs instead of
being guessed. Metadata-only scans classify same-size files as unchanged with
DeltaQuality::Metadata, so callers can decide whether to request content
hashes. Partial scans always produce DeltaQuality::Partial.
Long-lived file watchers can keep a RepositoryMatcher and call refresh()
after an ignore input changes. Refresh builds a replacement matcher first, so a
failure leaves the existing matcher usable.
SkipKind distinguishes:
BinaryExtensionFileSystemBoundaryHiddenIgnoredIoErrorMaxDepthOverrideOversizedPathEscapeScanLimitStandardDirectorySymlinkSymlinkLoop
This distinction matters to analyzers: "not selected by policy" is different from "unreadable" or "outside the repository."
Configuration
ScanOptions exposes:
| Option | Default | Purpose |
|---|---|---|
max_file_bytes |
1,500,000 | Reject oversized source candidates |
extensions |
Empty | Empty accepts every extension |
ignore_files |
.gitignore, .ignore, .weavatrixignore |
Hierarchical local ignore files |
ignore_policy |
Repository-only | Optional parents, .git/info/exclude, global Git and explicit files |
override_rules |
Empty | Request-level include/exclude globs above ignore sources |
ignore_case_insensitive |
false |
Optional ASCII case-insensitive ignore matching |
skip_hidden |
false |
Skip dot-prefixed and Windows-hidden entries unless included |
standard_skips |
Enabled | Skip generated/vendor directories |
hash_file_contents |
true |
Attach per-file hashes and content-sensitive revision |
detect_binary_files |
true |
Reject files containing a NUL byte |
evidence |
Complete |
Keep all typed exclusions, or only selected files |
parallelism |
0 |
Content workers; zero uses available parallelism |
limits.max_entries |
None | Bound examined filesystem entries |
limits.max_total_bytes |
None | Deterministically bound selected content bytes |
limits.timeout |
None | Stop traversal/content inspection after a duration |
cancellation |
None | Cooperative cross-thread cancellation token |
walk.max_depth |
None | Limit entry depth; root is zero |
walk.min_depth |
0 |
Suppress shallower results while still traversing them |
walk.max_open |
64 |
Bound live directory handles/workers |
walk.same_file_system |
false |
Stop at filesystem boundaries when enabled |
walk.follow_links |
false |
Follow only in-root links and detect cycles |
walk.error_policy |
Continue |
Continue with partial typed evidence or abort |
walk.collect_metadata |
true in ScanOptions |
Reuse directory-entry metadata without reopening selected paths |
The standard directory policy skips:
.git .hg .svn .venv __pycache__ build coverage dist
node_modules target vendor
Disable it when another layer owns generated-directory policy:
use ;
let mut options = default;
options.standard_skips = Disabled;
Ignore semantics
Ignore files are loaded hierarchically with source precedence
.weavatrixignore/custom > .ignore > .gitignore >
.git/info/exclude > global Git. Deeper files win within the same source
class. Supported Git-style constructs include:
- comments and escaped leading
#/!; - negation with
!; - root-anchored patterns;
- directory-only patterns;
*,**, and?;- character classes, negated classes, and ranges;
- brace alternatives such as
{foo,bar}; - escaped literals and escaped trailing spaces.
The default scanner intentionally does not read global Git configuration,
parent rules outside the scan root, or .git/info/exclude; repository-local
selection therefore stays portable. IgnorePolicy::git_compatible() enables
all three explicitly inside Git repositories, records their content hashes,
and marks host-dependent reports non-portable. Local .gitignore, .ignore,
and custom sources can be toggled independently. Request-level override globs
use ignore::Override semantics: ordinary patterns include and leading !
patterns exclude. Explicit includes can opt paths back into standard-directory
and extension filtering, but never bypass size or binary safety checks.
RepositoryMatcher::matched exposes the winning typed
decision without requiring a full walk. Differential tests compare
exact selected path sets against the
ignore crate for anchored, nested, negated, wildcard, and character-class
fixtures plus deterministic randomized rule sets. Stress cases cover deep
trees, permission errors, non-UTF8 names, and followed symlink loops. The
differential suite and competitor crates are dev-only.
Safety model
- never executes repository code;
- never starts subprocesses or accesses the network;
- canonicalizes and validates the root before traversal;
- does not follow symlink entries by default;
- rejects followed links outside the canonical root and detects cycles;
- can enforce a same-filesystem boundary;
- continues after independent local errors by default and marks the report partial;
- caps selected file size before content reads;
- rejects repository-local ignore-file symlinks and path traversal;
- supports entry, total-byte, timeout, and cooperative cancellation bounds;
- forbids unsafe Rust.
The scanner is read-only. Concurrent filesystem changes are surfaced as local
warnings/skips under Continue or as the first error under Abort.
Benchmarks
Run all included benchmarks:
Run the competitor comparison:
The Competitor benchmarks workflow runs the same output-equivalent comparison
on Ubuntu, Windows, and macOS for scanner or benchmark changes.
Run exact selected-path parity on a real repository:
$env:WEAVATRIX_BENCH_ROOT = "C:\path\to\repository"
cargo bench --locked --bench real_repository
The synthetic comparison uses 6,000 source files across Rust, Go, and TypeScript in 80 sibling directories. It runs two warmups and 11 interleaved measured samples, then reports the median. Raw walkers must produce the same fully sorted native relative-path set; the ignore-aware comparison additionally checks the same normalized path-and-size manifest.
Sample result on Windows 11, Rust 1.97.1, warm filesystem cache, measured
2026-07-24 against ignore 0.4.31, walkdir 2.5.0, and jwalk 0.8.1:
| Mode | Library | Files | Median |
|---|---|---|---|
| Raw paths | weavatrix Walker |
6,004 | 7.7 ms |
| Raw paths | weavatrix ParallelWalker |
6,004 | 5.0 ms |
| Raw paths | ignore | 6,004 | 10.1 ms |
| Raw paths | walkdir | 6,004 | 9.5 ms |
| Raw paths | jwalk | 6,004 | 7.8 ms |
| Ignore-aware manifest | weavatrix Scanner |
6,001 | 20.5 ms |
| Ignore-aware manifest | ignore | 6,001 | 24.1 ms |
| Rich manifest | weavatrix Scanner |
6,000 | 92.8 ms |
This is the median of five independent output-equivalent Windows benchmark
processes; each process itself reports the median of 11 interleaved samples
after two warmups. Walker and ParallelWalker beat walkdir on this corpus.
ParallelWalker is about 36% faster than jwalk, while the selected-manifest
Scanner is about 15% faster than ignore. The comparable scanner row omits
skip evidence on both sides. The rich row additionally records typed evidence,
reads content, detects binaries, hashes sources, and computes a deterministic
revision.
One cross-platform GitHub-hosted-runner sample on the same commit:
| Platform | ParallelWalker |
jwalk | walkdir | Scanner |
ignore |
|---|---|---|---|---|---|
| Ubuntu | 7.2 ms | 7.8 ms | 9.7 ms | 21.1 ms | 22.8 ms |
| Windows | 6.5 ms | 9.5 ms | 11.4 ms | 23.5 ms | 27.5 ms |
| macOS | 5.6 ms | 7.1 ms | 8.0 ms | 15.2 ms | 20.5 ms |
Absolute timings are not comparable between runner operating systems because their hardware differs. Within every row, Weavatrix produced identical output and led both the parallel-walker and ignore-aware comparisons.
Source review explains the remaining differences:
walkdirstreams unsorted directory entries and bounds open descriptors;jwalkschedulesread_dirwork through Rayon and restores ordered output;ignorecompiles patterns intoGlobSetmatchers and shares inherited matchers;- Weavatrix
Walkerstreams iterative DFS, bounds live handles and buffers the oldest remaining frame only whenmax_openis reached; - Weavatrix
ParallelWalkerkeeps round-robin directory balancing but restores discovery-task order independently of worker completion; - Weavatrix
Scannerreuses inherited rules, indexes exact literals, specializes prefix/suffix globs, prefilters complex patterns, and sorts only the final report.
Exact-path real-repository sample:
| Repository | Files | weavatrix-scan | ignore |
|---|---|---|---|
| radiochron | 86 | 18.7 ms | 22.9 ms |
| grpc-server | 30 | 5.4 ms | 5.6 ms |
| bgp-speaker | 30 | 4.7 ms | 5.6 ms |
| controller-rest-api | 1,085 | 38.3 ms | 38.4 ms |
| frontend | 1,689 | 39.8 ms | 46.3 ms |
| analytics | 361 | 40.6 ms | 40.4 ms |
| automation | 1,670 | 22.6 ms | 22.2 ms |
Every real row first asserts the exact same sorted (normalized path, bytes)
manifest. Weavatrix is faster on five repositories; the remaining two are
within 2%, below the observed run-to-run variance. Timing varies by filesystem,
cache, antivirus, and CPU, so treat the table as a reproducible sample rather
than a universal constant.
Correctness checks
The test suite covers:
- deterministic results and revisions;
- ignore-rule precedence and nested ignore files;
- repository-only, Git-exclude, parent, explicit and reusable-matcher policies;
- representative and randomized parity with
ignore; - raw entry parity with
walkdirandjwalk; - iterative deep trees, bounded handles, local error continuation, non-UTF8 paths, and symlink loops;
- binary, oversized, extension, generated-directory, and symlink policies;
- serial/parallel content-inspection equivalence;
- streaming parallel pruning and cancellation;
- manifest delta evidence and live matcher refresh;
- optional Serde support.
The real-repository benchmark compares the complete normalized selected-path
set against ignore. Its comparison policy disables Weavatrix's file-size cap
so an oversized file cannot masquerade as an ignore-rule mismatch.
Development
The MSRV is Rust 1.88. CI checks Rust 1.88 on Linux, Windows, and macOS, with stable test coverage on all three platforms.
Relationship to Weavatrix
weavatrix-scan owns repository discovery. It does not parse languages or
build graphs. weavatrix-graph
owns typed graph primitives. Higher-level Weavatrix crates can compose both
without coupling either library to MCP, a CLI, or language-specific parsers.
License
MIT © 2026 Sergii Ziborov.