Find code without grepping

AlRunner/ is ~139,000 lines across 341 tracked .cs files, the largest Program.cs at 6,952 (git ls-files | xargs wc -l, 2026-09-10). #3676 split the two files that used to dominate this — RecordPatches.cs and MockTestPage.cs — into partials by surface. Grep gives you line numbers you then have to read windows around; measured on one implementation agent, that loop was 63% of all its tool calls. The language server answers the same questions exactly, in one call.

The built-in LSP tool does not work in subagents on this Claude Code build — it returns No such tool available: LSP. This script gives you the same answers through Bash, which you do have.

Commands

tools/lsp-query.py callers <SymbolName>   # what calls it, and where it is defined
tools/lsp-query.py symbol  <SymbolName>   # where it is defined
tools/lsp-query.py refs <file> <line> <col>   # references at a position (1-based)
tools/lsp-query.py def  <file> <line> <col>   # definition of the symbol here

callers is the one you usually want, and it needs no line or column:

$ tools/lsp-query.py callers GetDataAccessForTableCore
object RecordPatches.GetDataAccessForTableCore(object self, NCLMetaTable table, bool isTemporary)  [defined AlRunner/Patches/RecordPatches.DataAccessDispatch.cs:87]
    AlRunner/Patches/RecordPatches.InstallBaselineDisk.cs:68:20
    AlRunner/Patches/RecordPatches.DataAccessDispatch.cs:71:26

That is complete — including the call site in a different partial-class file, which a grep for the name in one file would have missed.

Read the exit code. The three outcomes are NOT the same

exit	meaning	what to do
0	answered, results printed	use them
1	answered, genuinely nothing found	a real negative — you may rely on it
2	the server failed, timed out, or is not installed	NOT a negative. Say so and fall back to grep

Never treat exit 2 as "nothing calls this". That mistake reverses the meaning of your result and any conclusion built on it. If it says csharp-ls is not installed, tell the user to install it (README, tooling section) rather than silently grepping for the rest of the session.

Cost

Measured on this repo, one process per query, cold: ~8.5s for a hit, ~10s for a genuine miss. There is no daemon and none is needed. That is far cheaper than the grep-then-read-several-windows loop it replaces.

What it cannot tell you

It is static analysis. It cannot tell you whether a Hook(...) registration or a Cecil rewrite actually fires at runtime — an orphaned hook and a live one look identical to it. Use AL_RUNNER_HOOK_AUDIT=1 for that question. For orientation in an unfamiliar area ("what is near this"), the knowledge graph is better; see CLAUDE.md.

If you are the main session briefing a subagent

Resolve the symbols first and paste the answers into the brief, so the subagent does not have to go looking:

# LSP CONTEXT (pre-resolved)
GetDataAccessForTableCore — AlRunner/Patches/RecordPatches.DataAccessDispatch.cs:87
callers: RecordPatches.DataAccessDispatch.cs:71, RecordPatches.InstallBaselineDisk.cs:68