tgrep

Trigram-indexed grep with a client/server architecture for fast regex search in large codebases.

tgrep is integrated into GitHub Copilot CLI to power fast grep searches across large repositories.

Install tgrep, then start a server:

tgrep serve .           # builds the index if needed; runs in the foreground

In another terminal:

tgrep -- "fn main" .    # automatically connects to the server
tgrep status .          # show indexing and refresh status

Without a server, run tgrep index . to build an on-disk index. Searches use the server when available, then the local index, or scan the filesystem if no usable index exists. Indexes are stored in .tgrep/ by default; add this directory to .gitignore. It is generated build output and must not be committed or distributed with a repository. Build indexes locally for repositories you search; index metadata is not a statement of trust or freshness.

Indexed file reads stay beneath the selected root and do not follow descendant symlinks or Windows reparse points. Invalid stored paths are rejected; indexes with an unavailable or unrelated recorded root fall back to a filesystem scan. Explicit file arguments and --no-index --follow retain their normal scan semantics.

Indexes can lag filesystem changes. The server updates asynchronously; without a server, rerun tgrep index . after changes. Use --no-index when a search must read the current files. Queries scan during an initial build until complete coverage is available.

For coding agents, see AGENTS.md. To install MCP search tools and startup hooks for Codex or pi, run bash install-agent.sh from this checkout (Linux/macOS, Python 3.11+). See the agent integration guide.

A trigram index narrows each query to files that could match, then verifies those files with the regex engine in parallel. Patterns without useful trigrams may still require searching every indexed file.

In the August 24, 2026 benchmark sweep, tgrep was faster than ripgrep in 17 of 18 repo/platform combinations, with speedups from 0.93x to 51.9x. These measure client/server search latency with the index already built, not indexing time. Results depend on the query, repository, storage, and match volume. See BENCHMARKS.md for measurements and methodology.

The shared-worktree baseline compares one ordinary server/index per worktree with one shared daemon across 1, 4, 16 and 32 synthetic Git worktrees. It covers LF, clean CRLF, private branch changes and bounded churn, with scan-checked results, fresh/restart startup, query percentiles, reconciliation counters, process resources and retained storage. It includes a small Windows/Linux smoke command and distinguishes logical sharing from OS page-cache and RSS effects. The measured baseline saves LF memory/storage but has higher startup/query latency; clean CRLF can cost more memory/storage too. Shared mode remains opt-in.

tgrep index .                        # index current directory
tgrep index /path/to/repo             # index a specific repo
tgrep index . --index-path /tmp/idx    # custom index location
tgrep index . --exclude vendor --exclude third_party  # skip directories

Each build reports elapsed time and, where available, peak memory:

Index built successfully at /tmp/idx
Indexed in 22.6s using external strategy (peak memory 160.1 MiB)

Windows reports peak private committed bytes; Linux samples anonymous resident memory. Both exclude file-backed mappings and report the working set separately when it is substantially larger. macOS reports resident memory instead.

Like ripgrep, tgrep applies .gitignore only inside a Git repository. It warns when a root .gitignore is present but not applied. Use --no-require-git on indexing, serving, and searching to apply those rules outside Git:

tgrep index . --no-require-git
tgrep serve . --no-require-git
tgrep --no-require-git -- "pattern" .

When Git's core.ignorecase is enabled, tgrep applies an additional case-insensitive exclusion pass for the repository's root .gitignore and .git/info/exclude. Git-tracked paths are exempt from this additional pass, but not from ordinary traversal, binary, or size filtering.

Nested .gitignore files and global ignore rules do not get this additional pass. --no-ignore disables it along with the normal ignore rules.

Use the same --no-require-git, --no-ignore, --max-filesize (or --no-max-filesize), and --exclude settings for index and serve. Startup reconciliation removes indexed files excluded by the server's current settings. For example, serving an uncapped index with --max-filesize 8M removes files larger than 8 MiB from that index.

Pass the same size policy, --no-require-git, and any custom --index-path to searches too. --exclude is only available on index and serve; --no-ignore on a search forces a filesystem scan.

tgrep index . --max-filesize 8M
tgrep serve . --max-filesize 8M
tgrep --max-filesize 8M -- "pattern" .

tgrep index defaults to an external merge sort with a 64 MiB posting buffer. When the buffer fills, sorted segments spill to disk and are merged into the index. This bounds posting-buffer memory, not total process memory: file tables, worker buffers, and decoded content also consume memory.

tgrep index .                          # external, 64 MiB posting buffer
tgrep index . --index-buffer 16        # smaller buffer, lower peak
tgrep index . --index-strategy=memory  # sort entirely in RAM

--index-buffer is in MiB and applies only to the external strategy. --index-strategy=memory avoids temporary spill files but holds all postings in RAM. Both strategies still need a writable index directory.

tgrep serve uses the external builder with the default buffer for new indexes; it does not accept --index-strategy or --index-buffer. Large files can be memory-mapped, but files needing decoding or UTF-8 repair require heap buffers. Use --max-filesize to limit admitted file sizes. See index-build benchmarks.

tgrep serve .                         # auto-build index if missing
tgrep serve . --index-path /tmp/idx   # custom index location
tgrep serve . --watch-mode poll       # poll without native subscriptions
tgrep serve . --poll-interval 60      # polling cadence after fallback
tgrep serve . --watch-budget 4096     # lower this process's native watch ceiling
tgrep serve . --no-watch              # disable all automatic refresh
tgrep serve . --exclude node_modules  # exclude directories from indexing

The server builds the index in the background if none exists and resumes incomplete builds. Clients scan the filesystem until the full corpus is ready; tgrep status reports indexing progress and hidden-file coverage. Legacy indexes are upgraded by startup reconciliation. Multiple clients can connect simultaneously.

Index builds and server refreshes exclude .git directory subtrees by default, including nested repositories' Git internals. --hidden does not override this; use --no-ignore consistently on index and serve to include them. Explicit filesystem scans (--no-index) retain ripgrep-style hidden and ignore behavior. Git metadata needed for ignore rules and tracked-file detection is still read.

On Windows, replaced index generations stay in .retired while readers have them memory-mapped. Cleanup retries after publication, every minute, and on startup; uncommitted backups are preserved for recovery. This does not reclaim storage already stranded in NTFS's $Deleted namespace by older versions.

These tuning options apply only to tgrep serve:

Fresh external builds use the default 64 MiB posting buffer and global Rayon pool, not --max-memory or --max-cpu. If external bootstrap fails, the server falls back to an in-memory build where these settings apply.

The server checks for pending saves once a minute. It saves at the mutation threshold, when filename-only membership changes, or when content changes remain unsaved for at least ten minutes since startup or the last successful save. Active builds and flushes defer this check.

Startup watcher traces count filesystem notifications, not confirmed content edits. Replay re-reads notified files and avoids content-index mutations when their decoded contents match the indexed fingerprint. Delta-save traces report the unique candidate paths plus separate scan and live-overlay counts; these inputs can overlap and do not represent a count of user edits.

Automatic refresh has two modes, configured on tgrep serve:

In auto mode, a watch-budget or native-registration failure releases this process's watches and switches it to polling until restart. Status reports the reason. Explicit poll mode creates no native subscriptions.

On Linux/Android, the budget counts admitted directories; ignored subtrees do not consume watches. The OS inotify quota is shared with other processes, so registration can fail before this budget is reached. Raising --watch-budget does not raise the OS quota. Windows and macOS use one recursive root subscription and filter ignored events after delivery.

Polling checks filesystem metadata for additions, changes, and deletions. The next poll waits --poll-interval seconds after the previous reconciliation finishes; scan time and ongoing builds/saves add to the delay. Queries do not defer polling. Queue overflows and native rescan notifications also request reconciliation; read-only access events are discarded.

Native mode reconciles about once an hour to recover missed notifications. It waits for a two-minute gap in queries, deferring no longer than four hours. --poll-interval does not change this native safety cadence.

No-change scans do not rewrite the index; changed merges may rewrite it. Metadata-preserving changes can be missed, and reconciliation is not an atomic filesystem snapshot. Failures appear in status. Use --no-index for current file contents.

--no-watch permits the initial build/startup reconciliation but disables later refresh. It conflicts with explicitly supplied --watch-mode, --poll-interval, and --watch-budget. --watch-mode poll also rejects an explicit --watch-budget.

tgrep -- "pattern" .                    # regex search
tgrep -- "pattern" file1.rs file2.rs     # multiple files/paths
tgrep -- "TODO|FIXME" .                 # alternation
tgrep -- '\w+(?!_test)' .               # backtracking-engine fallback
tgrep -i -- "error" .                   # case-insensitive
tgrep -F -- "Vec<T>" .                  # literal string
tgrep -l -- "MyStruct" .                # filenames only
tgrep -c -- "pattern" .                 # matching lines per file
tgrep -m 5 -- "pattern" .               # at most 5 matching lines per file
tgrep -g "*.rs" -g "*.toml" -- "pattern" .  # multiple globs (OR)
tgrep -t rust -C 3 -- "pattern" .       # Rust files, 3 lines of context
tgrep -e "pattern" -e "also_this" .     # multiple patterns (OR)
tgrep --json -- "pattern" .             # JSON stream
tgrep --vimgrep -- "pattern" .          # editor jump targets
tgrep --stats -- "pattern" .            # query plan and timing
tgrep --no-index -- "pattern" .         # read current files, bypass index
tgrep -U -- 'first\nsecond' .           # multiline match
tgrep -q -- "pattern" .                 # exit code only
tgrep --files .                        # list admitted paths, including binary files
tgrep --files -t rust .                # list Rust files
tgrep --type-list                      # show file types

With the default traversal rules, --files reads the live server or the local index instead of walking the repository. The local result is an index snapshot; use --no-index to inspect the filesystem as it exists now. Flags that change traversal membership or the file-size policy also fall back to a walk.

tgrep status .

Server status for /src/my-monorepo
  PID:        37980
  Port:       51043
  Files:      152
  Trigrams:   12265
  Cache:      2/50000
  Watcher:    active
  Watch mode: native (requested: auto)
  Watch budget: 8192
  Poll interval: 120s (after completion)
  Reconcile:  idle
  Last successful reconcile: 2m ago
  Reconcile pending: no
  Reconcile overdue: no
  Last reconcile duration: 42ms
  Indexing:   complete
  Hidden coverage: complete

Watcher refers to native notifications, so it is normally inactive in polling mode. Refresh fields report fallback reasons, errors, pending work, and elapsed deadlines. Older servers may omit these fields.

Indexing: complete and Hidden coverage: complete describe index readiness, not freshness. An existing index can be queried while startup reconciliation is still running. Without a server, status shows on-disk metadata.

tgrep count-files .              # count candidate text files (no server needed)
tgrep count-files /path/to/repo  # scan a specific repo

Counts files admitted by the walk's ignore, visibility, extension, and default size rules, without reading their contents. The reported "text files" can therefore include files containing NUL bytes. Prints the count to stdout and details to stderr:

284957
284957 text files (47516 binary skipped, 0 too large, 0 errors) in 1200ms

These tables describe search flags. Use tgrep index --help and tgrep serve --help for subcommand options.

Pattern matching

--engine auto (the default) uses Rust's regex crate, with fallback to fancy-regex for lookaround and backreferences. -P and --engine pcre2 select fancy-regex, not the PCRE2 library; they do not promise full PCRE2 syntax compatibility.

Output formatting

Encoding

By default, a UTF-8/UTF-16LE/UTF-16BE BOM selects decoding; otherwise files are treated as UTF-8. A BOM overrides a named encoding, but not -E none. Any non-auto encoding bypasses the index and server. Invalid UTF-8 is still repaired, including with -E none; this is not raw-byte matching.

File walking

Accepted for compatibility

--mmap/--no-mmap (tgrep chooses memory mapping automatically), --crlf/--no-crlf (a trailing \r is always stripped), --no-config (tgrep reads no config file), and --colors <SPEC> (colors are not yet configurable) are accepted and ignored so ripgrep command lines keep working. --debug/--trace imply --stats.

For indexed content searches, --stats reports Raw candidates before path, visibility, glob, and type filters, and Candidates after those filters but before size checks, reads, or matching. no index narrowing means the raw set covers the entire nonempty index. (via server) describes transport, not index effectiveness. Older servers may omit candidate statistics.

-z/--search-zip is not supported and exits with code 2 rather than silently reporting no matches in compressed files.

Note: -L means --follow (as in ripgrep). Use the long --files-without-match for the non-matching-files listing.

Without -e/-f, the first positional argument is the pattern and the rest are paths. As soon as -e or -f supplies a pattern, every positional becomes a path, matching ripgrep:

tgrep -e needle .            # searches for "needle" under .
tgrep -e needle -e other .   # both patterns, still just one path

Use -- before a positional pattern that starts with - or is a subcommand name, such as serve or index. Quoting alone does not prevent subcommand parsing. Put all flags before --:

tgrep -F -- serve .

tgrep matches ripgrep's context-dependent defaults rather than fixed ones:

• Line numbers are on only when stdout is a terminal. Piping to another command drops them, so tgrep needle . | cut -d: -f1 behaves the same as it does with ripgrep. --column, --vimgrep and -p turn them back on; -b and -A/-B/-C do not.
• Filenames are shown unless you named exactly one file. A directory argument always shows them. -H/-I override either way.
• Paths are printed by appending onto the argument you typed: the argument survives verbatim and only the appended part uses the platform separator. So tgrep needle src/ prints src/main.rs while tgrep needle src prints src\main.rs on Windows. --path-separator rewrites every separator.

-m/--max-count limits matching lines, not individual matches. All matches on an admitted line are reported. With -U, the unit is a contiguous block of lines covered by matches, so a multiline match is not cut short.

tgrep reads a whole-file buffer even with -m; its bytes_searched statistic can therefore exceed ripgrep's.

-U/--multiline allows matches across line boundaries and prints every covered line. --vimgrep reports one row per match, on its starting line. Unlike ripgrep, tgrep reports columns relative to each printed line rather than repeating the starting column on continuation lines.

A file is treated as binary if its decoded content contains a NUL byte:

• Binary files found by walking a directory are skipped silently — they appear in neither the output, -l, -c, nor --files-without-match.
• A binary file named explicitly on the command line reports a note: bin.dat: binary file matches (found "\0" byte around offset 7).
• --binary promotes traversal to the explicit behaviour, so binary files are searched and summarised with that note.
• -a/--text disables binary detection entirely and prints matches as text.
• --json has no note. As in ripgrep, the matching lines are emitted as ordinary match events and the file's end message carries binary_offset — the offset of the first NUL — so a consumer can still tell a binary hit apart from a text one. stats.bytes_searched stops at that offset rather than counting the whole file.

tgrep also skips known binary extensions during directory content searches and indexing, unlike ripgrep. For searches, --binary and -a lift this restriction and bypass the index; they do not change what index or serve indexes. --files lists binary paths too.

index and serve include hidden, non-ignored files by default. --hidden is accepted on either command but is redundant. Ordinary queries still hide hidden files and directories; -./--hidden includes them using the index or server, for content searches and --files. Visibility is relative to the requested search root and includes Windows hidden attributes. Ignore rules inside hidden directories remain active. The configured index directory, including its staging and retired generations, is always excluded from indexing and query filesystem walks.

Flags that widen or re-interpret the indexed corpus still walk the tree: non-auto -E/--encoding, -a/--text, --binary, and ignore-disabling flags such as --no-ignore. Explicit file arguments also bypass the index.

For content searches, --follow, --one-file-system, --ignore-file, and --ignore-file-case-insensitive do not trigger fallback: indexed searches ignore them. Add --no-index to apply them. --files falls back to walking for these options automatically. index and serve reject these traversal options, --max-depth, and individual --no-ignore-* discovery switches.

Both positive and negative --glob/--iglob patterns filter the indexed corpus when a compatible index or server is available, for content searches and --files. Positive globs such as --glob '*.ts' do not reinclude ignored files in indexed mode. Use --no-index when those matches are needed: filesystem scans retain ripgrep-style glob overrides, which can reinclude ignored files. Negative globs, such as --glob '!.git', exclude entire matching directory subtrees. --hidden never disables ignore rules.

Incomplete indexes and legacy indexes without hidden-file coverage fall back to scanning. Rebuild with tgrep index . or let a current server reconcile them. Older servers without coverage support also cause fallback.

Older binaries reject the new content-index format. To downgrade, rebuild with the older binary, preferably at a separate --index-path.

ripgrep can search raw bytes; tgrep searches decoded text, replacing invalid UTF-8 sequences with U+FFFD. A pattern can match those replacement characters. For UTF-8 repair, text-output columns and byte offsets are mapped back to the source bytes.

JSON output uses lines.text with replacements and submatch offsets into that text. ripgrep instead emits base64 lines.bytes for invalid UTF-8. Transcoded encodings use decoded-text offsets rather than original byte positions; offsets also exclude stripped BOM bytes.

Directory searches and indexing skip files larger than 64 MiB by default; ripgrep has no default limit. These files are absent from the index and its filename listing, so their matches will not appear.

Use --max-filesize to choose another bound, or --no-max-filesize to remove it. Rebuild or serve the index with the same setting; an uncapped query cannot recover paths that a capped index never recorded. Use --no-index --no-max-filesize to scan without either restriction.

A directly named file ignores the inherited cap, but respects an explicit --max-filesize: tgrep -- pattern ./huge.log searches that file regardless of its size.

Same as ripgrep:

A match plus an error yields 2, unless -q is set, which yields 0.

An ignore file that fails to parse is not one of these errors. Like ripgrep, tgrep reports it on stderr, skips the offending rule and carries on, leaving the exit code determined by the search alone. Suppress the message with --no-ignore-messages, or with --no-messages, which covers it as well.

1. Indexing walks the repository with ignore rules, including root-level p4ignore.ini, and filters by size and binary extension. It decodes files, checks the first 8 KiB for NUL bytes, and extracts overlapping three-byte trigrams in parallel. Sorted postings map trigrams to candidate files.
2. Querying decomposes regex literals into trigram lookups, intersects or unions posting lists, and verifies candidates with the full regex engine. Inline case flags participate in planning; unsupported constructs use conservative plans rather than excluding possible matches. Case-insensitive fixed strings use the same Unicode-aware planning after escaping regex syntax. Both paths require only trigrams safe for the index's ASCII folding; if none remain, all indexed files are candidates.
3. Serving combines a memory-mapped IndexReader with a mutable LiveIndex overlay in HybridIndex. Updates take precedence over disk entries. Clients use JSON-RPC 2.0 over newline-delimited TCP on loopback, with one thread per connection. Shared read locks allow concurrent queries but can contend with writers.

The server's content cache is limited to 50,000 entries and 1 GiB of decoded content, with a 64 MiB per-entry limit. --files combines content-index paths with a filename-only sidecar for admitted paths without searchable content.

An agent runtime can explicitly attach linked Git worktrees to one repository daemon. They share an immutable committed-tree index and keep private overlays. Ordinary serve/index and unattached queries retain their existing behavior. Do not point ordinary servers at a shared --index-path.

mkdir /trusted/external/tgrep-cache
tgrep serve --shared /repo --shared-storage /trusted/external/tgrep-cache
# In another terminal (the runtime owns the foreground daemon's supervision):
tgrep shared attach /worktrees/session-a --revision <starting-commit> --lease <session-a-token>
tgrep shared attach /worktrees/session-b --revision <starting-commit> --lease <session-b-token>
tgrep status /worktrees/session-a
tgrep -- 'pattern' /worktrees/session-a/src
tgrep --files /worktrees/session-b
tgrep shared refresh /worktrees/session-a --lease <lease-from-attach> --changed src/file.rs
tgrep shared refresh /worktrees/session-a --lease <lease-from-attach> --full
tgrep shared detach /worktrees/session-a --lease <lease-from-attach>

attach returns JSON with the caller's lease, view ID, exact generation key, requested commit and readiness. Wait for "ready": true in status; attaching does not expose the base before reconciliation. Use a unique caller-owned --lease token (1-128 ASCII letters, digits, hyphens or underscores) per logical attachment and retain it before sending the request. Retrying that token with the same root/revision/profile returns the original attachment without consuming another lease, even if symbolic HEAD moved or the lease budget is full. A lost response can also be recovered by detach --lease without knowing the view ID. Reusing a live token with another root or revision is rejected. Without --lease, the CLI generates a token and prints it on stderr before the request; capture it and pass it explicitly on retry. Independent compatible attachments use different tokens and share the view. The last detach removes only that instance/view's registration, not another worktree or the daemon. Leases do not expire automatically; the runtime must release them or restart the daemon. Marker cleanup failures do not prevent lease release: detach reports registration_warning and leaves unrecognized/corrupt metadata for safe scan fallback. Before removing a worktree, detach all its leases and let in-flight operations finish. On Windows, retained root handles prevent removal while the view is live. An attached view stays pinned across commits/checkouts. Selecting another tree for the same root requires releasing all its leases first; incompatible reattachment is rejected before building or publishing a generation.

Attachment is the opt-in: subsequent normal search, --files, and status discover a versioned registration in the worktree's Git metadata, including from subdirectories. Merely running the daemon does not attach other roots. Nested Git repositories/worktrees are discovery boundaries. --shared on a query explicitly requires shared discovery; missing, stale, not-ready, wrong-protocol, or incompatible registrations scan the filesystem, never a legacy index or bare base. status reports those errors instead of displaying legacy status. An explicit --index-path retains legacy intent unless --shared is also set (then queries scan). --no-index always reads the filesystem. Each attached CLI query validates the repository once (three Git processes); daemon lookup/search/files/status reuse the attached identity with filesystem boundary/gitfile/common-directory checks and launch no Git subprocesses. Queries open candidates through the view's retained root handle and verify that identity before and after snapshot callbacks, including empty and filename-only results. Replacing the root closes readiness rather than mixing old candidates with a new directory. Daemon readiness also stays closed until reconciliation and checkpoint publication have both finished.

The v1 profile is raw-Git-blob automatic decoding, tracked regular files and a 64 MiB size cap. Worktree reconciliation separately applies local ignores, hidden visibility, binary classification, and filename-only membership. Search/types/globs/output flags reuse the ordinary matching/output code. Positive indexed globs do not reinclude ignored paths; use --no-index for filesystem glob overrides. Explicit files, non-default encoding/size/traversal options, --no-ignore variants, --text, --binary, --follow, --one-file-system, --ignore-file, passthru, files-without-match and zero-count output scan with a diagnostic rather than claiming incompatible coverage.

Freshness: native notifications close readiness as they arrive; bounded hints are an optimization, not proof that all events arrived. Every automatic mode also fully reconciles after --poll-interval seconds (default 120, measured after completion). Native registration failures/budget exhaustion fall back to polling. --watch-mode poll skips native registration; --no-watch disables both automatic modes after initial reconciliation, leaving refresh to the runtime. refresh --changed returns processed_epoch; this acknowledges those changes, not an atomic filesystem snapshot. refresh --full also detects missed same-size/restored-mtime changes and Git/ignore configuration changes. Persistent reconciliation failures retain not-ready/last-error status and retry with completion-based exponential delays from 1 to 30 seconds. Explicit refresh bypasses and resets this backoff; a successful repair resets it too. Status includes attempt counts, consecutive failures and the current retry delay. Native registration and event filtering include searchable nested .tgrep directories; only the root .tgrep storage directory is excluded.

The daemon bounds views (default 32, --shared-max-views), leases (256, --shared-max-leases), native watches (8192, --watch-budget), and hint slots (16384, --watcher-queue-cap). Watch/hint budgets are partitioned equally by the configured maximum views; unused shares are not borrowed. There is one index/reconcile worker, two independent query workers, bounded connection and work queues, 1 MiB requests and 64 MiB responses. There is no content cache. The response limit is enforced while producing individual match/context rows, file paths and statistics, including JSON escaping and envelope overhead, and again during final serialization. Large queries fail explicitly and CLI queries scan; generation mappings and private posting memory still scale with admitted data.

Storage must already exist, be trusted and outside all registered worktrees, Git metadata and index snapshots. It contains separate bases and overlays trees. Each successful reconciliation saves only the private delta with its exact base/root key. Restart the same daemon, then reattach with the same starting revision: old leases/instance IDs are invalid and restored checkpoints undergo full validation before readiness. Corrupt checkpoints/generations are explicit errors, not base-only results. Published generations/checkpoints are retained: there is no online GC, automatic migration, or checkpoint eviction. Shared JSON RPC requires UTF-8 canonical worktree roots; unsupported roots, including symlink aliases to them, fail before attachment or generation publication.

Sharing avoids repeated trigram extraction, not all content reads. Tests measure three identical tracked files as three reads/decodes and zero private extractions. A linked worktree additionally reads/decodes and extracts its ordinary .git pointer file once, for totals of four reads and one private extraction; that file participates in --hidden searches and filename listings, unlike a real .git metadata directory. Four CRLF-transformed files need four private extractions, plus one for a linked worktree's gitfile (five total). Full no-op verification reuses those overlays with zero new extractions. Smudge/encoding transformations may similarly require full overlays. See the wire contract and design for generic runtime integration.

See the shared worktree index design for architecture diagrams, the agent runtime boundary, base-generation lifecycle, and the staged implementation plan.

tgrep-core::generations manages immutable committed-tree bases:

use std::path::Path;
use tgrep_core::generations::{GenerationManager, IndexingProfile, Repository};

let repository = Repository::discover(Path::new("."))?;
let manager = GenerationManager::new(repository)?;
let first = manager.ensure("HEAD", IndexingProfile::default(), None)?;
let next = manager.ensure("main", IndexingProfile::default(), Some(&first.generation))?;
// Keep this exact pin with the synchronized worktree registration below.
let pin = next.generation;

Repository identity is the canonical Git common directory, shared by linked worktrees but not independent clones. Git is required; commands use arguments, not shell interpolation, and ignore ambient GIT_* overrides. Symbolic revisions such as HEAD resolve in the discovered worktree. Keys contain repository identity, exact committed tree OID, indexing profile, and format/schema versions; commit OIDs are retained for diagnostics. new stores generations below the common directory's tgrep-bases-v1. with_storage accepts an existing trusted directory outside registered worktrees and Git metadata. Both that parent and its effective repository-identity subdirectory are checked; the effective directory cannot itself be a linked worktree or index snapshot.

The current profile indexes raw Git blobs with the existing automatic text decoder. It covers all tracked regular files, including hidden, ignored, and extension-filtered paths: a superset, not a worktree's searchable membership. entries() records each tracked path's mode, object OID, size and classification. Indexed { content_id } includes empty/short files; Binary, TooLarge, and NotRegular have no content-index entry. Symlinks and gitlinks are never followed. Non-UTF-8 tracked paths and paths the platform/index cannot represent return explicit unsupported errors; repository roots preserve their native identity. Clean Git status does not establish compatibility with CRLF, encoding, LFS or smudge transformations. entry.matches_worktree_bytes(bytes) compares the decoded identity only; it proves neither stable filesystem reads nor visibility.

A compatible predecessor reuses unchanged blob postings and masks, including copies/renames, while extracting only new content through one Git batch process. Destination mode and size eligibility are recomputed. Without a predecessor, a missing generation requires a full build. EnsureResult.stats reports actual blob reads/extractions, bytes, reused files/postings, and publication/reuse. predecessor_posting_lists_read counts actual predecessor traversal; it is zero when no indexed paths can reuse postings, including when only empty/short files are unchanged. Their paths and content identities still reuse normally. Posting presence is collected during existing strict snapshot validation; no persisted metadata or index-format change is required. An OS file lock deduplicates cooperating starters across processes; live generation pins share one in-process Arc<Generation> and Arc<SharedBase>. The spill sorter bounds posting accumulation; one raw/decoded blob and its extracted masks are held at a time. Removing the size cap can therefore require substantial memory. Publication still streams a complete new index.

open(key) reopens an exact generation; list() returns validated published keys. Complete staging directories are validated and atomically renamed before discovery. Corrupt/incomplete final generations error rather than being replaced; abandoned staging directories are not discoverable. File contents are synced, but parent-directory power-loss durability is not guaranteed. Retention is explicitly RetainAll: no published generation is deleted, even after pins drop. Persist the key with retained overlay checkpoints. Online GC, checkpoint eviction remain deferred; daemon request/view budgets are described above; offline cleanup requires stopping all users and discarding dependent checkpoints. Storage and its ancestors must not be externally renamed, modified or removed while in use. Do not run mutable index builders against generation directories.

tgrep_core::worktrees provides the core registration and synchronization layer:

use tgrep_core::query::build_query_plan;
use tgrep_core::worktrees::{WorktreeOptions, WorktreeView};

let view = WorktreeView::new(Path::new("."), pin, WorktreeOptions::default())?;
// An agent runtime subscribes to changes NOW, before the first refresh.
// Forward file/subtree events to invalidate_path; overflow/config changes
// and unknown events to invalidate_all. The core does not create a watcher.
let stats = view.refresh()?; // initially a full verification; now ready

let plan = build_query_plan("needle", false)?;
view.with_snapshot(|snapshot| -> std::io::Result<()> {
    for path in snapshot.candidates(&plan, "", false) {
        let file = snapshot.open_file(&path)?; // uses the view's retained root handle
        // Bounded-read `file`, auto-decode and run your final matcher here,
        // buffering results until with_snapshot succeeds; or use a private cache.
    }
    Ok(())
})??;
let files = view.with_snapshot(|snapshot| snapshot.files("", false))?;
// Candidates are root-relative paths, not final matches. Do not read a joined
// pathname with ordinary filesystem APIs: a raced link could leave the worktree.

The root must be the actual worktree root in the pin's repository, not a subdirectory, bare repository or independent clone. root(), repository() and generation() preserve its canonical identity and exact pin. Each view keeps private whole-file replacements and tombstones relative to that pin, not current HEAD; committing an edit does not clear its override. Missing, sparse, ignored and ineligible base paths are hidden. Membership and visibility use the existing hidden-inclusive walker and a frozen case-insensitive tracked exemption snapshot. files() includes eligible filename-only binary paths. Query prefix is empty or a root-relative directory ending in /.

invalidate_path(relative_path) immediately closes the query gate and queues a bounded file/subtree hint (both paths for renames). Case aliases trigger conservative verification; non-ASCII hints and queue overflow force a full pass. Accepted trailing/repeated separators and interior . components are normalized, so dir/ still invalidates the entire dir subtree. Absolute paths, parent components, and leading . remain explicit errors that force full repair. invalidate_all() handles native watcher overflow, polling uncertainty, Git or ignore configuration changes, and missed-event repair. refresh() rewalks membership/visibility and processes hints; with no hints it does full content verification. reconcile_full() always verifies all admitted bytes, including same-size/restored-mtime edits, assume-unchanged and skip-worktree files.

Construction and checkpoint restoration are not ready. A successful refresh atomically publishes overlay, filenames and visibility. status() exposes readiness, invalidation/published epochs and pending/full work. A concurrent invalidation causes ChangedDuringReconcile, not stale ready publication: retry or scan. There is one attempt per call, never an unbounded churn loop. Every reconciliation advances the epoch, including refreshes without hints; the published epoch acknowledges all earlier invalidation tokens. Discovery/read errors likewise leave the view unavailable and force full retry. Reconciliation pins a root directory handle and opens only regular files: Unix uses component-relative no-follow, nonblocking opens; Windows guards ancestor handles against replacement and checks resolved handle containment before reading. Detected root identity changes and read-path swaps leave the view not-ready. Windows retains the root guard until view drop, so an agent runtime should release registrations before removing or renaming a worktree. Ordinary servers also retain their root guard for their lifetime: stop tgrep serve before removing or renaming its served root on Windows. The shared helper is tgrep_core::rooted::RootedDir (open, open_file with a relative path, and verify_root); ordinary serving retains one per server and reuses it across build, watcher and verification reads. WorktreeSnapshot::open_file uses the view's existing reader, not a new root registration per candidate. Neither API freezes file contents. Metadata discovery rejects unrepresentable native names before conversion: non-Unicode paths and literal Unix backslashes cannot alias other indexed paths. Ordinary full scans retain native paths and remain available as the fallback. with_snapshot holds readiness and overlay guards through candidate-ID resolution, and verifies the pinned root before and after the callback, even for empty/file-only results. Verification failure closes readiness and queues full repair. Candidate-open failures are recorded too: with_snapshot returns an outer I/O error and invalidates readiness before releasing the guard, even if the callback swallowed the inner error or subsequently opened another file. No live IDs or mutable HybridIndex escape. A callback may return an owned read-only file handle for bounded matching outside the guard; before publishing buffered results, reenter with_snapshot and reject a changed epoch. Later reads through that handle remain caller-owned: report read errors and call invalidate_all() after leaving the guard rather than publishing partial results. Do not reenter the view from its closure. A refresh acknowledges processed hints, not an atomic filesystem snapshot. Periodic full reconciliation remains necessary; no-watch callers must explicitly refresh, and final reads can race subsequent edits.

Full verification reads, auto-decodes and hashes checkout bytes; clean Git status, blob identity and index stat data are never byte-equivalence proofs. Matching decoded content reuses base postings without extraction. Verified renames/copies can copy base masks in one streaming posting pass; unchanged private overlays also avoid extraction. A hinted refresh can avoid rereading unaffected, previously verified files under the event-driven freshness contract, but still performs a metadata/membership walk. ReconcileStats separates reads/bytes/decodes, actual extraction calls, base and overlay reuse, copied files/postings, and content reads avoided. hint_lookups counts ordered-set probes: each file checks its normalized path and ancestor prefixes, at most one probe per component rather than a scan of all queued hints. Each probe is logarithmic in the hint count. Even a hinted pass opens eligible file handles to verify regular-file metadata safely; avoided content reads do not mean zero filesystem I/O.

Raw LF bases versus CRLF/smudge checkouts can require an all-file overlay, even for equal committed trees and clean Git status. Position and next-byte masks make line-ending normalization unsafe. Regression fixtures measure zero extractions for three byte-identical tracked files, but four for four transformed tracked files. A linked worktree's plain .git file adds one private extraction in either case; subsequent unchanged reconciliation reuses those private postings. There is no cross-view transformed-content cache. Preparation retains per-path evidence and changed-file masks until atomic publication; removing the size cap increases per-file memory, and widespread transforms can make the private overlay large.

For persistence, configure an existing dedicated checkpoint_directory in WorktreeOptions. Construction, restoration and saving explicitly reject non-directory checkpoint paths. save_checkpoint() requires readiness and atomically writes only overlay.json: private postings/tombstones plus exact generation key, base fingerprint and canonical root, through the existing directory-bound publisher. The entire checkpoint directory, .tgrep, Git metadata directories and the generation store are excluded, even with no_ignore. A linked worktree's plain .git pointer file remains eligible under ordinary walker rules: --hidden exposes its filename and searchable text unless ignored or over the size cap. This does not expose the metadata directory it names. Custom storage directories belong in walk.exclude_paths. Checkpoint storage and ancestors must remain trusted and not externally renamed while in use. The directory cannot be an index snapshot, Git metadata, or a worktree ancestor. WorktreeView::restore(root, exact_pin, options) rejects missing, malformed, wrong-key and wrong-root checkpoints. Successful restoration still requires subscribing and full reconciliation; restored overlays have no trusted read evidence and may re-extract private postings on that first pass. Generation retention remains RetainAll.

tgrep-core::shared::SharedBase is the first building block for sharing one content index across worktrees. Open a complete, current-format index in an immutable snapshot directory once, then call create_worktree(root) for each worktree. The returned HybridIndex instances share the same Arc<IndexReader> (including the path table), but have independent roots, live postings, and deletion tombstones. Cloning SharedBase does not copy the base index.

Shared-base opening rejects mismatched empty lookup/posting sections and metadata counts inconsistent with the opened index. Legitimately empty indexes and files too short to produce trigrams remain supported. Shared snapshots also require aligned, contiguous posting ranges covering index.bin, valid trigrams, strictly increasing valid file IDs within each posting list, and nonzero location masks; ordinary readers retain their existing validation behavior.

The caller must populate each overlay before exposing it to searches: index whole changed/new files using view.live.upsert_file, and hide deleted or ineligible base paths using view.live.delete_file. Include committed branch differences, staged and unstaged edits, and eligible untracked files, not just git diff HEAD. Only reuse base postings when their decoded content matches the worktree's indexing semantics; Git checkout filters, encoding, and line-ending conversion can change that content. Clear a path's override with clear_reconciled_paths only after proving it matches the base again.

save_overlay(&view, checkpoint_path) atomically saves only the worktree's live postings, masks, and tombstones. It never merges changes into the base. restore_worktree(root, checkpoint_path) checks the checkpoint version, canonical root, and a fingerprint of the base's path table, lookup table, and postings. Wrong-base, wrong-root, missing, and malformed checkpoints return errors rather than silently revealing base entries. Opening the base computes its fingerprint once; attaching more worktrees does not rescan it. Checkpoint parent directories must already exist and be outside the base snapshot directory. Repeated saves atomically replace the existing checkpoint, including on Windows. Base-directory exclusion checks directory identities, not just pathname prefixes. On Unix, staging, replacement, and cleanup use the opened parent directory handle. On Windows, open handles prevent renaming or deleting the canonical parent and its ancestors until publication finishes. Replacing a parent pathname cannot redirect a save into the shared base. Paths with trailing separators or /. are rejected rather than normalized into filenames. Unicode worktree roots retain the existing JSON string representation; non-Unicode roots use tagged Unix bytes or Windows UTF-16 units to preserve their exact identity. Existing Unicode-root checkpoints remain readable; older readers cannot restore the new non-Unicode representation. Atomic replacement does not guarantee power-loss durability: file contents are synced before replacement, but the parent directory is not synced afterwards. A successful save may be lost after a system crash; callers must reconcile or rebuild stale/missing checkpoints.

The lower-level SharedBase API alone does not synchronize a checkout; use WorktreeView for the reconciliation/readiness contract above. Neither API creates native watchers or content caches. The opt-in shared daemon supplies worktree registration, native/poll refresh and CLI routing separately. The existing single-root RPC protocol, index format, and server are unchanged. Do not point existing servers at a common --index-path: their publication path still writes a complete index. Keep shared base files immutable for the lifetime of every view or query holding their reader.

cargo build --release --locked  # build optimized binary
cargo test --workspace          # run tests
make check                      # check formatting and run clippy
make install                    # install to ~/.cargo/bin (Unix)

git clone https://github.com/microsoft/tgrep.git
cd tgrep
cargo install --path tgrep-cli --locked

Keep vendor/ignore in the checkout: it contains the locked dependency with a small native Git-metadata path fix. Source installs and release builds use it directly; see its provenance and patch notes.

brew install tgrep

Download from GitHub Releases for Linux, macOS (Intel & Apple Silicon), and Windows.

Download the archive for your architecture, extract it to a temporary directory, then copy the executable into a directory on PATH.

# Linux (x86_64)
tmpdir="$(mktemp -d)"
gh release download --repo microsoft/tgrep -p '*x86_64-unknown-linux-musl.tar.gz' -D "$tmpdir"
tar xzf "$tmpdir"/tgrep-*-x86_64-unknown-linux-musl.tar.gz -C "$tmpdir"
install -Dm755 "$tmpdir/tgrep" "$HOME/.local/bin/tgrep"
rm -rf "$tmpdir"

# macOS (Apple Silicon)
# For Intel, replace aarch64 with x86_64.
tmpdir="$(mktemp -d)"
gh release download --repo microsoft/tgrep -p '*aarch64-apple-darwin.tar.gz' -D "$tmpdir"
tar xzf "$tmpdir"/tgrep-*-aarch64-apple-darwin.tar.gz -C "$tmpdir"
mkdir -p "$HOME/.local/bin"
install -m755 "$tmpdir/tgrep" "$HOME/.local/bin/tgrep"
rm -rf "$tmpdir"

# Windows x64 (PowerShell); for ARM64, replace x86_64 with aarch64.
$download = Join-Path $env:TEMP ("tgrep-dl-" + [guid]::NewGuid())
New-Item -ItemType Directory -Path $download | Out-Null
gh release download --repo microsoft/tgrep -p '*x86_64-pc-windows-msvc.zip' -D $download
Get-ChildItem $download -Filter '*.zip' | ForEach-Object {
    Expand-Archive -LiteralPath $_.FullName -DestinationPath "$HOME\.cargo\bin" -Force
}
Remove-Item -LiteralPath $download -Recurse

Ensure $HOME/.local/bin (Unix) or $HOME\.cargo\bin (Windows) is on PATH.

This project welcomes contributions and suggestions. Most contributions require you to agree to a Contributor License Agreement (CLA) declaring that you have the right to, and actually do, grant us the rights to use your contribution. For details, visit https://cla.microsoft.com.

When you submit a pull request, a CLA-bot will automatically determine whether you need to provide a CLA and decorate the PR appropriately (e.g., label, comment). Simply follow the instructions provided by the bot. You will only need to do this once across all repos using our CLA.

This project has adopted the Microsoft Open Source Code of Conduct. For more information see the Code of Conduct FAQ or contact opencode@microsoft.com with any additional questions or comments.

MIT