grep: optional tgrep backend for large workspaces #1086

Feature Description

Let the grep tool run against a tgrep trigram index when one exists for the session root, and fall back to the bundled ripgrep otherwise. Opt-in, not a default change — tgrep only pays off with an index plus a live server, and it is slower than ripgrep on patterns the index cannot narrow.

Today there is no setting to change the grep backend, and a mod cannot register a tool named grep (colliding names are skipped with mod_error), so the only seam is beforeToolCall interception.

Use Case

On a 124,465-file / 29 GB workspace, a selective search such as TODO takes ~3.1s through the grep tool. Agents issue dozens per session, serially. With a live tgrep server the same search returns in 31ms end to end — 100x. Measured on macOS arm64, tgrep 1.2.5, ripgrep 15.2.0, warm cache, best of 5:

For comparison, engine-to-engine with full output the broad case is only 1.2x (3.31s vs 2.83s); the 21x through a tool result comes from stopping at the page limit instead of materialising everything.

Additional Context

I built this as a mod and hit six things a native implementation would need to handle:

1. output_mode defaults differ — the built-in is files_with_matches, tgrep is content, so -l must be passed when the field is absent.
2. tgrep omits line numbers unless -n is given explicitly.
3. head_limit/offset need a stream that aborts at offset + limit + 1 lines; buffering a broad match set hit 23 MB and a 285k-line result.
4. An index without a running server is not safe. It silently misses files written since the build. Reproduced: a newly created file returned exit 1 (no matches) from tgrep while ripgrep found it. A native backend should require a live server, not just an index on disk.
5. A cheap tgrep status . probe gates it (0.00s live, 0.08s dead), plus a stderr check for falling back to local index to cover a server dying mid-query.
6. Patterns with no literal trigrams scan every indexed file — \d{9,} took 40.5s against ripgrep's 3.0s — so a timeout fallback or trigram heuristic is required. My mod caps at 30s and then defers to the built-in.

Prototype with methodology, raw output and limitations: https://github.com/Hashimlokasher/cmd-mod-tgrep-grep (BENCHMARKS.md). tgrep also powers GitHub Copilot CLI's grep, so it is actively maintained.

Proposal: a grep.backend: "tgrep" setting, or auto-detect when tgrep is installed and a live index is present, keeping ripgrep as the default and the fallback. Happy to test or contribute.

Priority: Important for my workflow