lsp-cli: daemon mode to keep language server alive between invocations

[c3d4r]

Problem

Each lsp-cli invocation spawns a fresh language server (e.g. gopls), waits for initialization, sends one request, and shuts down. The gopls startup cost is ~3 seconds per call, which adds up significantly when an agent makes 10+ navigation calls per task.

In our evaluation, Run B (lsp-cli) was 75% slower wall-clock than Run A (grep/read) despite using fewer tokens, primarily due to this overhead.

Proposal

Add a daemon mode that keeps the language server alive between invocations:

• lsp-cli daemon start — start a background language server, listening on a Unix socket
• lsp-cli daemon stop — shut it down
• One-shot commands transparently connect to the daemon if running, otherwise fall back to spawning a fresh server

This is the same model gopls uses and would reduce per-call latency from ~3s to near-zero.

Impact

This is the single highest-impact improvement identified in the evaluation. With daemon mode, lsp-cli would be strictly faster than grep for semantic queries while also being more precise.

Design notes

• Socket location: $XDG_RUNTIME_DIR/lsp-cli/<server>-<root-hash>.sock or similar
• Auto-shutdown after idle timeout (e.g. 10 minutes)
• One daemon per (server, workspace root) pair
• The original design doc discusses this in Phase 4: docs/lsp-cli-design.md