pi-lsp-lite

Your agent can't see compiler errors. Now it can.

pi extension that runs language servers in the background and feeds diagnostics back inline after every edit. Errors appear on the same turn — no context switch, no separate command.

Go · Rust · TypeScript · Python · C/C++

Install

pi install npm:pi-lsp-lite

That's it. If you have gopls, rust-analyzer, typescript-language-server, pylsp, or clangd on PATH, diagnostics start flowing automatically.

What you see

  edit ─ src/main.go
  ✓ Edited src/main.go (replaced 2 lines)

  ⚠ LSP diagnostics for src/main.go (2 errors):
    error 12:5 [compiler] undefined: foo
    error 18:2 [compiler] too many arguments in call to bar
    + 1 diagnostic in 1 other file

The agent sees these too — they're appended to the tool result, so it can self-correct on the same turn.

Commands

Command	What it does
/lsp-status	Show running servers, PIDs, workspace roots, uptime
/lsp-diag	Show all current diagnostics (or /lsp-diag path/to/file for one file)
/lsp-add	Interactively add a new language server
/lsp-remove	Disable a configured server
/lsp-toggle	Flip a server on/off without removing config
/lsp-install	Install a missing server binary

Supported servers

Server	Language	Install
gopls	Go	go install golang.org/x/tools/gopls@latest
rust-analyzer	Rust	rustup component add rust-analyzer
typescript-language-server	TypeScript/JS	npm install -g typescript-language-server typescript
pylsp	Python	pip install python-lsp-server
clangd	C/C++	Xcode CLI tools / apt install clangd

Missing a server? /lsp-add lets you configure any LSP server that speaks stdio. Or add it to .pi-lsp-lite.json:

{
  "servers": {
    "haskell": {
      "extensions": [".hs"],
      "command": "haskell-language-server-wrapper",
      "args": ["--lsp"],
      "rootPatterns": ["cabal.project", "stack.yaml"]
    }
  }
}

Configuration

Works without config. For customisation, create .pi-lsp-lite.json (project) or ~/.pi-lsp-lite.json (global):

Field	Description	Default
servers.<id>.diagnosticTimeout	Per-attempt timeout (ms)	per-language
servers.<id>.maxRetries	Retry attempts on timeout (0-10)	3
servers.<id>.disabled	Disable this server	false
diagnosticTimeout	Global default timeout (ms)	5000
documentIdleTimeout	Close idle documents after (ms)	120000

Project config merges over global. Partial overrides work — only specify what you want to change.

How it works

1. Agent writes/edits a file
2. Extension detects the language, finds the workspace root
3. Spawns (or reuses) an LSP server for that language + root
4. Sends didChange, waits for publishDiagnostics
5. If timeout: retries with exponential backoff + jitter (up to maxRetries times)
6. Filters to errors + warnings, formats, appends to tool result + shows in TUI

Cross-file impact is detected via snapshot-diff: if editing lib.ts breaks caller.ts, you see "+ N diagnostics in M other files".

Servers are lazy (spawn on first edit), idle-shutdown after 240s, and clean up on session end.

Development

git clone https://github.com/mcphailtom/pi-lsp-lite
cd pi-lsp-lite && npm install
npm run check              # typecheck
npm test                   # unit tests (106, no servers needed)
npm run test:integration   # real server tests (needs servers on PATH)

See docs/ARCHITECTURE.md and docs/CONTRIBUTING.md.

License

MIT