fastfind

Find files by name, or search inside them, without waiting. Installs as ff.

ff cache.go                       find by name, under the current directory
ff -g 'func main' -name '*.go'    search file contents, like grep -r
ff -home '*.pem'                  the whole user profile

Measured against GNU find on the same tree, returning the same 165 results:

Documentation is in the wiki — a cookbook, how it works, what happens on filesystems other than NTFS, and how to make it the default search in Claude Code.

go install github.com/telebroad/fastfind/cmd/ff@latest

Windows, Linux and macOS. Tested on all three in CI.

Barely any of it is a cleverer search. It's three things find doesn't do.

It reads directories concurrently. A tree walk is latency-bound, not CPU-bound: a worker spends nearly all its life blocked inside a directory read, using no core at all. Sizing a worker pool to the core count — the reflex for compute work — leaves the disk idle waiting on the few workers that happen to be runnable. ff runs 8 workers per allowed core, because goroutines are cheap enough that a couple of hundred blocked ones cost a few hundred kilobytes of stack and nothing else.

It never follows a reparse point. Junctions, symlinks, and the virtual trees that Google Drive, OneDrive and Dropbox project are all reparse points. Windows leaves junctions all over a user profile — Application Data and My Documents point back at their own parents — so a walk that follows them never terminates. Worse, descending into a cloud provider's projection asks that service to materialise every file it is holding remotely, which is how a search for a source file ends up pulling gigabytes over the network.

It skips what holds most of the files and none of the answers. On a development machine node_modules alone is usually a clear majority of every file present.

The single biggest win was none of those three, though. The reparse check reads the attribute word that FindNextFile already returned — Go keeps it on the DirEntry, so entry.Info() costs nothing — rather than asking Windows again with GetFileAttributes. Deleting that one redundant syscall per directory took a user-profile scan from 36.6s to 4.2s.

-g switches from names to contents, and from there the flags mean what they mean in grep(1):

ff -g 'func main' -name '*.go'    # only inside Go files
ff -g TODO -i -n                  # ignore case, show line numbers
ff -g panic -l                    # just the files that panic
ff -g 'err != nil' -c             # count matching lines per file
ff -g Deprecated -C 2             # two lines of context either side
ff -g 'computed(' -F              # a literal, not a regex

The grep runs inside the walk, on the goroutine that found each file, so the content search finishes about when the walk does rather than starting then.

Two things it does that grep on Windows doesn't:

• It reads UTF-16. PowerShell redirection, Notepad and much of Windows still write UTF-16LE with a BOM. Every other byte is NUL, so a byte-oriented search calls the whole file binary — grep reports "binary file matches" and stops. This decodes it and searches the text.
• When your pattern isn't valid regex, it says what to do. Searching for computed( or foo[0] is the most common thing anyone types, and every one of those is a broken regex. The error names the fix rather than just the fault.

grep has no idea what JSON is, and jq prints values by default — so the normal way to find out what is in a config file is also the way to put its credentials on your screen. In a terminal that is bad. In a session with an AI assistant it is worse, because the transcript is stored: a secret that reaches the screen has been written down somewhere it will outlive the reason it was shown.

ff fails closed.

ff -json config.json     # every key path, no values

config.json
  api                 object(3 keys)
  api.debug           bool = true
  api.endpoints       array(2)
  api.endpoints[0]    string(17)
  api.token           string(20)  ** redacted: key looks like a credential **
  database.host       string(11)
  database.password   string(19)  ** redacted: key looks like a credential **
  database.port       number(4)
  logging.sinks       array(0)

That answers the questions actually being asked — is the key there, is it set, is it the right shape — without answering what is it. A value is printed only when it cannot carry a secret: a bool has two possible values, null is the absence of one. Everything else becomes a type and a length. Numbers are redacted too, which looks over-cautious until you remember that account numbers and PINs are numbers.

It composes with the finder, so ff -json with no argument describes every JSON file below you — 6,930 keys across a project in 72ms.

ff -get database.host config.json      # db.internal
ff -get 'api.endpoints[1]' config.json # array indexing works
ff -get database.password config.json  # prints it, and warns you it did

Separate from -json deliberately. Reading the shape of a config is casual and safe; reading a value out of one is neither, and the two should not end up one keystroke apart.

ff -g password over a config file would print the line, secret included. -mask keeps the keys and hides the values:

ff -g 'password|token' -mask config.json

config.json:  "database": {"host":<11 chars>,"port":<4 chars>,"password":<19 chars>},
.env:DB_PASSWORD= <19 chars>
.env:DEBUG=true

Each value is masked separately, so the structure and the keys survive — masking everything after the first colon would be safe but useless. DEBUG=true is left alone; there is no secret in a boolean. The length stays because it separates the setting is empty from the setting is filled in. Nothing else does — not a prefix, not a suffix. The first character of a password is a character of a password.

Where to look
  -in <dir>     start here (default: the current directory)
  -home         start at the user profile
  -all          do not skip node_modules, .git, Windows and the rest
  -hidden       descend into dot-directories too
  -cores <n>    cores to use (default: every core but one, minimum 1)

Finding by name (the default)
  -d            directories only
  -f            files only
  -max <n>      stop after this many hits

Searching contents
  -g <regex>    the pattern to search for inside files
  -name <glob>  only look inside files whose name matches
  -i -w -v -F   ignore case / whole words / invert / literal
  -l -c -n      files only / count per file / line numbers
  -A -B -C <n>  context after / before / either side

Output
  -abs          absolute paths rather than relative to the root
  -0            NUL-separated, for xargs -0
  -q            no summary line
  -color <when> always | never | auto (default: auto)

A name pattern matches anywhere in the name and ignores case; give it * or ? and it's a glob against the whole name. A content pattern is a regex and is case-sensitive by default — that's grep's rule, and muscle memory is worth more than internal consistency.

Skipped unless -all: node_modules, .git, .angular, .next, .nuxt, .gradle, .venv, venv, __pycache__, vendor, target, dist, build, Windows, WinSxS, $Recycle.Bin, System Volume Information.

Exit status follows grep — 0 found, 1 not found, 2 bad usage — so it composes:

ff -q -f '*.pem' && echo "keys present"
ff -0 -q '*.tmp' | xargs -0 rm

The default is every core but one. A file search is something you run while doing something else, and taking the whole machine to save a second is a bad trade — the editor stutters, the build crawls, and the search was going to finish either way. One core in reserve costs almost nothing here: the walk is latency-bound, so the cores sit idle waiting on the disk regardless. -cores n sets it explicitly, and it caps GOMAXPROCS rather than only the worker count, so -cores 2 really does stay out of the way.

internal/ntfs holds a direct reader for the NTFS Master File Table — the trick WizFile and Everything use. NTFS keeps one flat table with a ~1KB record per file on the volume, so a whole-drive answer is a single sequential read rather than millions of directory calls.

It's deliberately not the default. Raw volume access (\\.\C:) needs an elevated prompt, and it reads hundreds of megabytes of table to answer a question a scoped walk answers in milliseconds. Most of the time you don't want the whole system — you want your own directory.

Three details there produce plausible-looking numbers rather than errors when you get them wrong, so each has a test:

• Data-run offsets are signed and relative to the previous run. A later run can sit earlier on the disk.
• The boot sector's clustersPerRecord is a signed byte: −10 means 2¹⁰ = 1024 bytes. Read unsigned it's 246, and every offset after it is wrong.
• Every record needs its fixups applied — NTFS overwrites the last two bytes of each sector with a torn-write counter and keeps the originals in an array at the top of the record. Skip that and attribute lengths read as sequence numbers.

The walker builds and runs on Windows, Linux and macOS; the reparse-point handling is Windows-specific and falls back to Go's own symlink check elsewhere. The MFT reader is Windows-and-NTFS only, by nature. CI runs the suite on all three.

go test ./...
go test -race ./...

MIT — see LICENSE.