LSP Enforcement Kit for Claude Code

Stop burning tokens on Grep. Make Claude navigate code like an IDE.

The Problem

Claude Code defaults to Grep + Read for code navigation. This works, but it's wasteful:

"Where is handleSubmit defined?"

Grep approach:
  Grep("handleSubmit") → 23 matches, ~1500 tokens of output
  Read file1.tsx (wrong) → 2500 tokens
  Read file2.tsx (still wrong) → 2500 tokens
  Read file3.tsx (found it) → 2500 tokens
  ─────────────────────────────────────
  Total: ~9,000 tokens, 4 tool calls

LSP approach:
  find_definition("handleSubmit") → form-actions.ts:42, ~80 tokens
  Read form-actions.ts:35-55 → ~150 tokens
  ─────────────────────────────────────
  Total: ~230 tokens, 2 tool calls

~40x fewer tokens. Same answer.

A rule in CLAUDE.md saying "use LSP" helps ~60% of the time. Hooks make it 100%.

Token Savings: Grep vs LSP Per Operation

Real-World Data: 1 Week, 2 Projects

Aggregate from a week of development across 2 TypeScript projects:

How the estimate works:

• Each blocked Grep saves ~1200 tokens of noisy output
• Each avoided Read saves ~1500 tokens of file content loaded into context
• 39 LSP calls cost ~4k tokens total (precise, compact results)
• Without LSP: ~120 Greps + ~180 Reads = ~315k tokens for the same navigation work
• With LSP: 39 nav calls + 53 targeted Reads = ~84k tokens

Works with any LSP MCP server

v2.1 introduces provider-aware block messages. The kit detects which LSP MCP server(s) you have installed and tailors its suggestions accordingly:

• cclsp — standalone MCP server or bundled via the typescript-lsp Claude Code plugin. Suggestions use mcp__cclsp__find_definition, find_references, find_workspace_symbols, etc.
• Serena — high-level symbol MCP server (MIT, by Oraios AI). Multi-language support (Python, Go, Rust, Java, TypeScript, Vue, and more via its bundled solidlsp wrapper). Suggestions use mcp__serena__find_symbol, find_referencing_symbols, get_symbols_overview.
• Both installed — suggestions show entries for both providers.
• Neither installed — generic fallback with install hints for both.

Detection reads user-level Claude Code config (~/.claude.json, ~/.claude/settings.json) and matches known server names. The shared helper is in hooks/lib/detect-lsp-provider.js — adding a new provider means adding one entry to its PROVIDERS registry, with no changes to the individual hooks.

Architecture: 6 Hooks + 1 Tracker

                    PreToolUse                          PostToolUse
                    ──────────                          ───────────

 Grep call ──→ [lsp-first-guard.js] ──→ BLOCK
                  detects code symbols,
                  suggests LSP equivalent

 Glob call ──→ [lsp-first-glob-guard.js] ──→ BLOCK
                  blocks *UserService*, **/handleFoo*.ts;
                  allows *.ts, *subdomain*, src/**

 Bash(grep) ──→ [bash-grep-block.js] ──→ BLOCK
                  catches grep/rg/ag/ack
                  in shell commands

 Read(.tsx) ──→ [lsp-first-read-guard.js] ──→ GATE
                  5 progressive gates
                  (warmup → orient → nav → surgical)

 Agent(impl) ─→ [lsp-pre-delegation.js] ──→ BLOCK
                  subagents can't access MCP,
                  orchestrator must pre-resolve

 LSP call ─────────────────────────────────────→ [lsp-usage-tracker.js]
                                                   tracks nav_count,
                                                   read_count, state

                    SessionStart
                    ────────────

 New session ──→ [lsp-session-reset.js] ──→ WIPE
                    clears stale nav_count for current cwd,
                    forces fresh warmup + re-enforces gates

v2 note: versions before v2 had two silent bypass routes that let
Claude read code files without ever calling LSP:
(1) Glob("*SymbolName*") had no guard, and (2) nav_count persisted
for 24 h across sessions, so a new session inherited "surgical mode"
(unlimited reads) from yesterday's LSP work. Both are closed in v2 by
lsp-first-glob-guard.js and lsp-session-reset.js. If you installed
v1, re-run bash install.sh — it merges the new hooks without touching
your existing settings.

How Each Hook Works

1. lsp-first-guard.js — Grep Blocker

Hook type: PreToolUse | Matcher: Grep

Intercepts every Grep call. Detects code symbols in the pattern. Blocks with a suggestion to use the correct LSP tool.

Block message example (with both cclsp and Serena detected):

⛔ LSP-FIRST BLOCK: 1 code symbol(s) in Grep — use LSP instead
Symbols: handleSubmit
LSP tools:
  handleSubmit:
    mcp__cclsp__find_references("handleSubmit")  (cclsp)
    mcp__serena__find_referencing_symbols("handleSubmit")  (Serena)

If only one provider is installed, only that suggestion appears.

2. lsp-first-glob-guard.js — Glob Symbol Blocker

Hook type: PreToolUse | Matcher: Glob

Closes the gap where Claude searches for a symbol by filename pattern instead of content. Without this hook, Glob("*UserService*") silently returns the file, Claude reads it, and LSP enforcement never fires.

The guard parses the glob pattern, extracts alphabetic tokens, and blocks if any token looks like a code symbol (PascalCase, camelCase, or snake_case with 3+ parts). Lowercase-only tokens and short generic words are always allowed.

Allowed by design: lowercase concept searches (*auth*, *subdomain*) are legitimate file discovery by topic. Only symbol-shaped tokens (casing patterns) are blocked, because those should use find_workspace_symbols instead.

3. bash-grep-block.js — Shell Grep Blocker

Hook type: PreToolUse | Matcher: Bash

Same detection logic, but for Bash(grep "UserService" src/), Bash(rg handleSubmit), etc. Claude sometimes tries to bypass the Grep hook by shelling out.

Allows: git grep (history search), non-code paths, non-code file type filters.

4. lsp-first-read-guard.js — Progressive Read Gate

Hook type: PreToolUse | Matcher: Read

The most sophisticated hook. Forces a "navigate first, read targeted" workflow through 5 gates:

Gate 1 — Warmup Required
  No LSP state file → BLOCK
  Must call get_diagnostics(<any .ts file>) first

Gate 2 — Free Orientation (reads 1-2)
  ALLOW — explore freely, no restrictions

Gate 3 — Warning (read 3)
  WARN if no LSP nav calls yet
  "Next Read will be BLOCKED"

Gate 4 — Navigation Required (reads 4-5)
  BLOCK if nav_count < 1
  Must use at least 1 LSP navigation call

Gate 5 — Surgical Mode (reads 6+)
  BLOCK if nav_count < 2
  After 2 nav calls → unlimited reads forever

Session flow:

Session starts
  │
  ├─ Read(page.tsx) → Gate 1 BLOCKS → "warmup required"
  │
  ├─ get_diagnostics(file.ts) → tracker writes warmup_done=true
  │
  ├─ Read(page.tsx) → Gate 2 allows (1 of 2 free)
  ├─ Read(actions.ts) → Gate 2 allows (2 of 2 free)
  ├─ Read(types.ts) → Gate 3 WARNS
  ├─ Read(helpers.ts) → Gate 4 BLOCKS
  │
  ├─ find_workspace_symbols("MyFunc") → tracker: nav_count=1
  │
  ├─ Read(helpers.ts) → unlocked (reads 4-5)
  ├─ Read(utils.ts) → unlocked
  ├─ Read(service.ts) → Gate 5 BLOCKS
  │
  ├─ find_references("MyFunc") → tracker: nav_count=2
  │
  └─ SURGICAL MODE — all Reads unlimited

Always allowed (no gate):

• Non-code files: .md, .json, .yaml, .env, .sql, .css, .html
• Config files: tsconfig.json, next.config.ts, package.json
• Test files: *.test.ts, *.spec.tsx
• Non-code paths: .task/, .claude/, node_modules/, __tests__/

Dedup: Reading the same file at different line ranges counts as 1 Read.

5. lsp-pre-delegation.js — Agent Pre-Resolution

Hook type: PreToolUse | Matcher: Agent

Claude Code subagents cannot access MCP tools — this is an architectural limitation of the platform. Without this hook, every delegated agent falls back to Grep+Read, bypassing all LSP enforcement.

// BLOCKED — no LSP context
Agent({
  prompt: "Fix handleSubmit in the form component",
  isolation: "worktree"
})

// ALLOWED — pre-resolved LSP context
Agent({
  prompt: `Fix handleSubmit error handling.

    ## LSP CONTEXT (pre-resolved by orchestrator)
    - handleSubmit: defined at form-actions.ts:42, called from page.tsx:15
    - FormComponent: defined at form.tsx:8, used in page.tsx:120`,
  isolation: "worktree"
})

Three enforcement tiers:

6. lsp-session-reset.js — Stale State Wiper

Hook type: SessionStart | Matcher: true (runs on every session start)

The Read guard's state file (~/.claude/state/lsp-ready-<cwd-hash>) has a 24-hour expiry. Without this hook, a new session inherits yesterday's nav_count — and if that count was ≥ 2, the guard is permanently in surgical mode for today's session: unlimited Reads with zero LSP calls required. A full bypass of the enforcement chain.

This hook runs once on session start and deletes the state file for the current cwd. The next Read triggers Gate 1 (warmup required), forcing at least one get_diagnostics call before any code file can be opened. After warmup, the standard progression kicks in (Gate 2 → 3 → 4 → 5) requiring real LSP navigation calls before surgical mode unlocks.

Session lifecycle with reset:

Session start
  │
  ├─ lsp-session-reset.js → unlinks lsp-ready-<hash>
  │
  ├─ Read(page.tsx) → Gate 1 BLOCKS → "warmup required"
  │
  ├─ get_diagnostics(file.ts) → tracker writes warmup_done=true
  │
  ├─ Read × 2 (free) → Gate 3 warn → Gate 4 block → LSP nav → …
  │
  └─ (2 nav calls later) SURGICAL MODE unlocked

Safety: the hook only deletes the flag for the current cwd — other projects' state files are left alone. Failure is silent (never blocks session start).

7. lsp-usage-tracker.js — State Tracker

Hook type: PostToolUse | Matcher: all mcp__cclsp__* tools

Tracks successful LSP calls in a per-project state file. Other hooks read this state to make gate decisions.

State file: ~/.claude/state/lsp-ready-<md5-hash-of-cwd>

{
  "cwd": "/path/to/project",
  "warmup_done": true,
  "nav_count": 25,
  "read_count": 38,
  "read_files": ["src/page.tsx", "src/actions.ts"],
  "timestamp": 1775818285727,
  "last_tool": "mcp__cclsp__find_references"
}

Cold start handling: Detects the cclsp "No Project" error (upstream bug where find_workspace_symbols doesn't prime the TypeScript project). Emits a systemMessage with the correct fix — call a file-based tool first. It's an ordering bug, not a timing issue.

Installation

Option 1: Give the repo to Claude Code (recommended)

git clone https://github.com/nesaminua/claude-code-lsp-enforcement-kit.git
cd claude-code-lsp-enforcement-kit

Then tell Claude Code:

Run bash install.sh in this repo to set up LSP enforcement hooks.

The install script:

• Copies 7 hooks + shared lib/detect-lsp-provider.js helper to ~/.claude/hooks/
• Copies the LSP-first rule to ~/.claude/rules/
• Merges hook registrations into your existing ~/.claude/settings.json (won't overwrite your other hooks)
• Enables the built-in typescript-lsp plugin
• Creates ~/.claude/state/ for tracking
• Verifies everything at the end
• Safe to re-run: entries are deduped by command path, so upgrading from v1/v2.0 to v2.1 just adds what's missing without touching anything else

Option 2: Run the script yourself

macOS / Linux:

git clone https://github.com/nesaminua/claude-code-lsp-enforcement-kit.git
cd claude-code-lsp-enforcement-kit
bash install.sh

Windows (PowerShell):

git clone https://github.com/nesaminua/claude-code-lsp-enforcement-kit.git
cd claude-code-lsp-enforcement-kit
pwsh ./install.ps1
# or: powershell -ExecutionPolicy Bypass -File ./install.ps1

Output:

=== LSP Enforcement Kit — Install ===

[1/4] Directories ready
[2/4] Copied 7 hooks + lib + 1 rule
[3/4] settings.json updated (merged, not overwritten)
[4/4] Verifying...

  Hooks installed:  7/7
  Rule installed:   yes
  Plugin enabled:   yes
  State directory:  yes

Done. Restart Claude Code to activate.

Option 3: Manual setup

Prerequisites

• Claude Code (CLI, Desktop, or IDE extension)
• TypeScript/JavaScript project

Step 1: Copy files

mkdir -p ~/.claude/hooks ~/.claude/state ~/.claude/rules
cp hooks/*.js ~/.claude/hooks/
cp rules/lsp-first.md ~/.claude/rules/

Step 2: Enable the plugin

In ~/.claude/settings.json, add to enabledPlugins:

{
  "enabledPlugins": {
    "typescript-lsp@claude-plugins-official": true
  }
}

Step 3: Register hooks in settings.json

IMPORTANT: If you already have hooks, add these entries to your existing arrays — don't replace them.

Add to PreToolUse array:

{
  "matcher": "Grep",
  "hooks": [{ "type": "command", "command": "node ~/.claude/hooks/lsp-first-guard.js" }]
},
{
  "matcher": "Glob",
  "hooks": [{ "type": "command", "command": "node ~/.claude/hooks/lsp-first-glob-guard.js" }]
},
{
  "matcher": "Bash",
  "hooks": [{ "type": "command", "command": "node ~/.claude/hooks/bash-grep-block.js" }]
},
{
  "matcher": "Read",
  "hooks": [{ "type": "command", "command": "node ~/.claude/hooks/lsp-first-read-guard.js" }]
},
{
  "matcher": "Agent",
  "hooks": [{ "type": "command", "command": "node ~/.claude/hooks/lsp-pre-delegation.js" }]
}

Add to PostToolUse array:

{
  "matcher": "mcp__cclsp__find_definition|mcp__cclsp__find_references|mcp__cclsp__find_workspace_symbols|mcp__cclsp__find_implementation|mcp__cclsp__get_hover|mcp__cclsp__get_diagnostics|mcp__cclsp__get_incoming_calls|mcp__cclsp__get_outgoing_calls",
  "hooks": [{ "type": "command", "command": "node ~/.claude/hooks/lsp-usage-tracker.js" }]
}

Add to SessionStart array (create it if missing):

{
  "matcher": "true",
  "hooks": [{ "type": "command", "command": "node ~/.claude/hooks/lsp-session-reset.js" }]
}

Verify

Run the health-check script:

bash scripts/lsp-status.sh
# or from anywhere after install:
bash ~/.claude/scripts/lsp-status.sh

Expected output:

LSP Enforcement Kit — Status
============================

  Hook files:          ✓ 7/7
  Shared lib/helper:   ✓ yes
  Settings registered: ✓ PreToolUse(5) PostToolUse(1) SessionStart(1)
  Detected providers:  ✓ cclsp

State for current cwd (/path/to/project)
------------------------
  Warmup done:         yes
  nav_count:           5 (LSP navigation calls)
  read_count:          7 (unique code files read)
  Last tool:           mcp__cclsp__find_references (2min ago)

  ✓ Surgical mode active — all Reads unlimited for this session.

Diagnostic summary
------------------
  All checks passed. Enforcement is active.

Or restart Claude Code and ask "Where is handleSubmit defined?" — Claude should use find_definition, not Grep.

LSP Tool Reference

Optional: Python, Go, Rust Support

The built-in plugin only covers TypeScript/JavaScript. For other languages, install cclsp — a standalone MCP server that connects Claude Code to any Language Server:

npm install -g cclsp

Then install the language server for your language:

# Python
pip install python-lsp-server

# Go
go install golang.org/x/tools/gopls@latest

# Rust
rustup component add rust-analyzer

Create ~/.config/claude/cclsp.json:

{
  "servers": [
    {
      "extensions": ["py", "pyi"],
      "command": ["pylsp"]
    },
    {
      "extensions": ["go"],
      "command": ["gopls", "serve"]
    },
    {
      "extensions": ["rs"],
      "command": ["rust-analyzer"]
    }
  ]
}

Add to your Claude Code MCP config (~/.claude.json):

{
  "mcpServers": {
    "cclsp": {
      "type": "stdio",
      "command": "cclsp",
      "args": []
    }
  }
}

The hooks work identically — they detect code symbols by naming convention, not by language. Once cclsp is connected, find_definition, find_references, etc. work across all configured languages.

FAQ

Q: Does this work with Python/Go/Rust?
Out of the box — TypeScript/JavaScript only (built-in plugin). For other languages, install cclsp + the language server (see section above). The hooks themselves are language-agnostic.

Q: What if LSP gives wrong results?
The hooks don't eliminate Grep — they block Grep for code symbols. If LSP returns empty, Claude can still Grep with non-symbol patterns or search non-code files. The Read guard also gives 2 free reads before requiring navigation.

Q: Won't the Read gate slow down simple tasks?
After 2 LSP navigation calls, all gates open permanently (surgical mode). This happens within the first 30 seconds of a session. Non-code files (config, tests, docs) are never gated.

Q: Why block Agent delegation without LSP context?
Claude Code subagents cannot access MCP tools (architectural limitation). Without pre-resolved context, every delegated agent falls back to exploratory Grep+Read, burning thousands of tokens and bypassing all enforcement.

Q: Known issues?
find_workspace_symbols fails with "No Project" if called before any file-based LSP tool (cclsp upstream bug). The tracker detects this and tells Claude to call get_diagnostics first. Not a timing issue — ordering issue.

Q: I installed v1 and shared it with my team — should I upgrade?
Yes. v1 had two silent bypass routes (Glob symbol search and stale session state) that let Claude navigate code without ever calling LSP. Both are closed in v2. Just re-run bash install.sh — it's idempotent and only adds the missing hook entries to your settings.json. No existing configuration is touched.

Q: Does this work with Serena?
Yes. Since v2.1, the kit detects your LSP MCP provider and tailors its block-message suggestions. If you run Serena (the multi-language MCP symbol toolkit by Oraios AI — MIT), the hooks will point you at mcp__serena__find_symbol, find_referencing_symbols, and get_symbols_overview instead of cclsp tools. The enforcement logic (Grep/Glob/Read/Agent gates, session reset) is provider-agnostic — it works the same for both. You can also run cclsp and Serena side-by-side; suggestions then show both.

This is pure interop — the kit ships no Serena code, uses only their public tool names in suggestion strings, and reads only your own config to detect which provider is active.

Q: What about Python/Go/Rust? cclsp is TypeScript-only.
Two options:

1. Install a standalone cclsp MCP server with multi-language config (see the "Optional" section above), OR
2. Install Serena — it bundles solidlsp, a unified wrapper around language servers for Python, Go, Rust, Java, TypeScript, Vue, PHP, Ruby, Swift, Elixir, Clojure, Bash, PowerShell, and more. The kit will detect Serena automatically and adapt its suggestions.

The hook detection logic itself is language-agnostic — it works on naming conventions (PascalCase, camelCase, snake_case), not language-specific ASTs.

License

MIT