Agent Skills: LSP Navigator

LSP Navigator

PREREQUISITE: LSP tools are deferred. Load first: ToolSearch({query:"select:LSP"}) CRITICAL: LSP returns EMPTY for .cjs files. Use ripgrep for .cjs instead.

When to Use

Use lsp-navigator when you need compiler-verified precision rather than text-based heuristics:

• Tracing where a symbol is defined (following imports, re-exports, index files)
• Finding every callsite of a function before refactoring
• Verifying parameter types and return types without reading the whole file
• Building complete call hierarchy trees for architecture review
• Validating that imports resolve correctly after a move/rename
• Checking what methods an interface requires vs. what a class provides

Use the diagnostics runner (lsp-diagnostics-runner.cjs) when you need:

• Bulk dead-code detection across many files (exports with zero importers)
• Broken import scanning (require() paths that don't resolve)
• Codebase health checks as part of QA or proactive-audit workflows

Do NOT use for:

• Discovering code you don't have a position for yet (use pnpm search:code first)
• Searching text inside comments or string literals (use ripgrep)
• Finding code by pattern shape (use code-structural-search)

Operation Reference

| Operation | Purpose | Best For | | ---------------------- | --------------------------------------------------------------------------------------------------------------- | -------------------------------------------------- | | goToDefinition | Find where a symbol is defined | Navigating to source of imported functions/classes | | findReferences | Find all usages of a symbol | Impact analysis before refactoring | | hover | Get type info and docs | Verifying parameter types, checking API contracts | | documentSymbol | List all symbols in a file | File structure overview, finding entry points | | workspaceSymbol | Search symbols by name | Finding functions/classes across the workspace | | goToImplementation | Find implementations of interface | Tracing concrete behavior of abstract types | | prepareCallHierarchy | Get call hierarchy item ⚠️ Always call this BEFORE incomingCalls/outgoingCalls or you get silent empty results. | Setting up for incoming/outgoing call analysis | | incomingCalls | Find all callers | Understanding who depends on this function | | outgoingCalls | Find all callees | Understanding what a function depends on |

All operations require: filePath (absolute path), line (1-based), character (1-based).

Search Intelligence Decision Table

| Question | Tool | Why | | ----------------------------------- | ------------------------------------- | -------------------------------------------- | | "Where is foo defined?" | lsp-navigator (goToDefinition) | Compiler-precise, follows imports/re-exports | | "Who calls foo?" | lsp-navigator (incomingCalls) | Complete call graph, not just text matches | | "What type does foo return?" | lsp-navigator (hover) | Compiler-resolved types, including generics | | "Find all usages of foo" | lsp-navigator (findReferences) | Compiler-aware, excludes comments/strings | | "Find files about authentication" | code-semantic-search | Conceptual search, not tied to symbol names | | "Find foo in all .cjs files" | ripgrep (rg -F "foo") | Fastest for exact text, works without LSP | | "Find all async functions" | code-structural-search | AST pattern matching, language-agnostic | | "What functions are in this file?" | lsp-navigator (documentSymbol) | Structured symbol list with types | | "Find class X across workspace" | lsp-navigator (workspaceSymbol) | Name-based but compiler-indexed | | "Find code similar to this pattern" | code-semantic-search | Embedding-based similarity | | "Find all try-catch blocks" | code-structural-search | AST pattern: try { $$ } catch ($E) { $$ } | | "What does this function call?" | lsp-navigator (outgoingCalls) | Full dependency graph of a function | | "Find foo in comments too" | ripgrep | LSP excludes non-code; rg searches all text |

Decision Flowchart:

START: What do you need?
  |
  +--> Need compiler-verified type/definition/reference info?
  |     YES --> lsp-navigator
  |     NO
  |      |
  +--> Need conceptual/intent-based code discovery?
  |     YES --> code-semantic-search / pnpm search:code
  |     NO
  |      |
  +--> Need AST pattern matching (structural shape)?
  |     YES --> code-structural-search
  |     NO
  |      |
  +--> Need exact text/literal match?
        YES --> ripgrep / rg -F

Updated Search Hierarchy

The complete search preference order (see also code-standards.md):

1. pnpm search:code         -- hybrid BM25 + semantic (discovery, default)
2. lsp-navigator            -- compiler-level definitions, references, types
3. Skill({ skill: 'ripgrep' })          -- fast text/regex search
4. Skill({ skill: 'code-semantic-search' })  -- conceptual/intent search
5. Skill({ skill: 'code-structural-search' }) -- AST pattern matching
6. Grep                     -- FALLBACK ONLY (advanced regex, single-file)

Note: lsp-navigator sits at position 2 because once you know a symbol/location (found via position 1), LSP provides the most precise navigation. Discovery (position 1) is still needed to find initial entry points.

Workflow Patterns

Use Case 1: Developer Passive Self-Correction

After editing a file, verify types and references are correct:

// After editing src/auth/jwt.ts line 45
// Step 1: Verify the function signature is correct
lsp_hover({ filePath: '/abs/path/src/auth/jwt.ts', line: 45, character: 10 });
// Step 2: Check that imported symbol resolves
lsp_goToDefinition({ filePath: '/abs/path/src/auth/jwt.ts', line: 3, character: 15 });
// Step 3: Verify references haven't broken
lsp_findReferences({ filePath: '/abs/path/src/auth/jwt.ts', line: 45, character: 10 });

Pattern: Edit → hover to verify types → goToDefinition to verify imports resolve → findReferences to verify callers still work.

Use Case 2: Specialist Active Discovery

Creator/updater agents trace execution flows:

// Tracing how a hook is called
// Step 1: Find the function definition
lsp_goToDefinition({
  filePath: '/abs/path/.claude/hooks/routing/routing-guard.cjs',
  line: 15,
  character: 5,
});
// Step 2: Find all callers
lsp_prepareCallHierarchy({
  filePath: '/abs/path/.claude/hooks/routing/routing-guard.cjs',
  line: 15,
  character: 5,
});
lsp_incomingCalls({
  filePath: '/abs/path/.claude/hooks/routing/routing-guard.cjs',
  line: 15,
  character: 5,
});
// Step 3: Find what this function calls
lsp_outgoingCalls({
  filePath: '/abs/path/.claude/hooks/routing/routing-guard.cjs',
  line: 15,
  character: 5,
});

Pattern: goToDefinition (anchor) → prepareCallHierarchy → incomingCalls (who calls me?) → outgoingCalls (what do I call?).

Use Case 3: QA Instant Static Analysis

QA agent checks types without running full test suite:

// After a code change, verify types across affected files
// Step 1: Get all symbols in the changed file
lsp_documentSymbol({ filePath: '/abs/path/src/middleware/auth.ts', line: 1, character: 1 });
// Step 2: For each exported symbol, check references still resolve
lsp_findReferences({ filePath: '/abs/path/src/middleware/auth.ts', line: 10, character: 15 });
// Step 3: Hover over key interfaces to verify type contracts
lsp_hover({ filePath: '/abs/path/src/middleware/auth.ts', line: 5, character: 20 });

Pattern: documentSymbol (what changed?) → findReferences (who's affected?) → hover (do types still match?).

Use Case 4: Reflection Structural Post-Mortems

Reflection agent verifies agent claims:

// Agent claimed it called `validateToken(token: string)` but review suggests wrong signature
// Step 1: Check actual type
lsp_hover({ filePath: '/abs/path/src/auth/validate.ts', line: 12, character: 10 });
// If hover shows `validateToken(token: JWTPayload)` not `string`, agent hallucinated
// Step 2: Check if the parameter type was changed recently
lsp_findReferences({ filePath: '/abs/path/src/auth/validate.ts', line: 12, character: 25 });

Pattern: hover (ground-truth type check) → findReferences (impact of the actual signature).

Use Case 5: Automated Dead Code Detection

Use documentSymbol → findReferences as an automated sweep to detect exported symbols with zero external references. This pattern powers the lsp-diagnostics-runner.cjs script.

// Step 1: List all symbols in the file
lsp_documentSymbol({
  filePath: '/abs/path/.claude/lib/memory/memory-manager.cjs',
  line: 1,
  character: 1,
});

// Step 2: For each exported symbol, check for external references
lsp_findReferences({
  filePath: '/abs/path/.claude/lib/memory/memory-manager.cjs',
  line: 42,
  character: 15,
});
// If results contain only the defining file → potential dead export

Automated script: .claude/tools/cli/lsp-diagnostics-runner.cjs implements this pattern using ripgrep as a fallback (since LSP has limited CJS support — see Anti-Patterns table).

# Run dead-exports check on all lib files
node .claude/tools/cli/lsp-diagnostics-runner.cjs --check dead-exports --format markdown

# Exclude archived directories
node .claude/tools/cli/lsp-diagnostics-runner.cjs --check dead-exports --exclude-pattern "_archive"

Note on hook exports (false positives): Hook files (.claude/hooks/**) export functions for testability but are invoked via stdin/stdout protocol, not require(). Their exports appear as dead code because test suites use dynamic require; the diagnostics runner marks hook export findings as LOW severity for this reason.

Use Case 6: Hook Wiring Verification

Verify that hooks registered in .claude/settings.json exist on disk and their internal require() chains resolve without errors.

// Step 1: Locate the hook registration in settings.json
// (use ripgrep — LSP won't index JSON well for CJS references)
// rg "routing-guard" .claude/settings.json

// Step 2: Confirm the hook file exists and resolves its imports
lsp_goToDefinition({
  filePath: '/abs/path/.claude/hooks/routing/routing-guard.cjs',
  line: 1,
  character: 1,
});

// Step 3: Check the exports the hook provides (for testability audit)
lsp_documentSymbol({
  filePath: '/abs/path/.claude/hooks/routing/routing-guard.cjs',
  line: 1,
  character: 1,
});

Hybrid approach (more reliable for .cjs files): Combine ripgrep for settings.json registration lookup with require.resolve() for import chain validation:

# Find all registered hooks
node -e "const s = require('./.claude/settings.json'); console.log(JSON.stringify(s.hooks, null, 2))"

# Verify a hook's require chain resolves
node -e "require('./.claude/hooks/routing/routing-guard.cjs'); console.log('OK')"

Diagnostics Runner Tool (RECOMMENDED for .cjs codebases)

Script: .claude/tools/cli/lsp-diagnostics-runner.cjs

An automated scanner that finds dead code, broken imports, and unreferenced functions using ripgrep + require.resolve(). Use this instead of manual LSP calls for bulk analysis of .cjs CommonJS files (where native LSP has limited support).

Quick Start

# Find exported symbols that nobody imports (dead code)
node .claude/tools/cli/lsp-diagnostics-runner.cjs --check dead-exports

# Find require() calls that don't resolve (broken imports)
node .claude/tools/cli/lsp-diagnostics-runner.cjs --check broken-imports

# Run both checks, exclude archived code
node .claude/tools/cli/lsp-diagnostics-runner.cjs --check dead-exports --check broken-imports --exclude-pattern "_archive"

# Target specific directories
node .claude/tools/cli/lsp-diagnostics-runner.cjs --check dead-exports --glob ".claude/lib/routing/*.cjs"
node .claude/tools/cli/lsp-diagnostics-runner.cjs --check dead-exports --glob ".claude/hooks/**/*.cjs"

# Output as markdown (for reports)
node .claude/tools/cli/lsp-diagnostics-runner.cjs --check dead-exports --format markdown --output report.md

CLI Flags

| Flag | Default | Description | | ------------------------- | ---------------------- | -------------------------------------------------------------------------- | | --check <type> | (required) | dead-exports, broken-imports, or unreferenced-functions (repeatable) | | --glob <pattern> | .claude/lib/**/*.cjs | File glob pattern to scan | | --exclude-pattern <pat> | (none) | Exclude files matching pattern (e.g., _archive) | | --format <fmt> | table | Output format: table, json, or markdown | | --output <file> | (stdout) | Write results to file |

Severity Levels

| Severity | Meaning | | -------- | -------------------------------------------------------------- | | HIGH | Broken import (require doesn't resolve) — likely runtime error | | MEDIUM | Dead export in non-hook file — probably unused code | | LOW | Dead export in hook file — expected (hooks export for testing) |

When Agents Should Use This

• qa / proactive-audit: Run as part of codebase health checks
• code-reviewer: Check for dead exports in changed files before approving
• architect: Audit module dependencies and find disconnected subsystems
• code-simplifier: Identify dead code candidates for removal
• developer: Verify new exports are actually imported somewhere after implementation

Important: Hook Export False Positives

Hook files (.claude/hooks/**) export functions for testability but are invoked via stdin/stdout JSON protocol, not require(). Their exports always appear as "dead" because no production code imports them — only test files do. The runner marks these as LOW severity automatically. Do not treat hook dead-exports as bugs.

Standard Invocation Pattern

// Invoke as a skill for guidance, then use native LSP tool directly
Skill({ skill: 'lsp-navigator' });

// Then call LSP operations natively:
// lsp_goToDefinition({ filePath, line, character })
// lsp_findReferences({ filePath, line, character })
// lsp_hover({ filePath, line, character })
// lsp_documentSymbol({ filePath, line, character })
// lsp_workspaceSymbol({ filePath, line, character })
// lsp_goToImplementation({ filePath, line, character })
// lsp_prepareCallHierarchy({ filePath, line, character })
// lsp_incomingCalls({ filePath, line, character })
// lsp_outgoingCalls({ filePath, line, character })

Agent-Specific Contracts

developer (always)

Use lsp-navigator to self-correct after edits:

• After writing/editing code: hover to verify types, goToDefinition to verify imports
• Before completing a refactor: findReferences to verify no callsites were missed
• When unsure about an API: hover on the call site to see the actual signature

qa (always)

Use lsp-navigator for lightweight static analysis:

• documentSymbol on changed files to inventory what was modified
• findReferences on exported symbols to find affected test targets
• hover on type boundaries to verify type contracts match test assumptions

code-reviewer (always)

Use lsp-navigator to improve review accuracy:

• findReferences when a renamed function might have missed callsites
• goToDefinition to trace where imported symbols come from
• hover to verify reviewer understands the actual type being used

architect (always)

Use lsp-navigator for architectural analysis:

• incomingCalls + outgoingCalls to build dependency maps
• workspaceSymbol to find all implementations of a pattern
• goToImplementation on interfaces to trace concrete implementations

code-simplifier (always)

Use lsp-navigator for safe refactoring:

• findReferences before renaming — must be exhaustive, not text-based
• goToDefinition to understand the full scope of a symbol before changing it
• hover to verify types before extracting functions

advanced-debugging (always)

Use lsp-navigator for root cause analysis:

• prepareCallHierarchy + incomingCalls to trace how a buggy path is reached
• outgoingCalls to understand what a suspect function depends on
• hover to verify types at boundary points where errors occur

reflection-agent (contextual)

Use lsp-navigator to verify agent claims:

• hover to ground-truth type claims made by agents
• findReferences to verify "all callsites updated" claims are accurate

security-architect (contextual)

Use lsp-navigator to trace data flows:

• outgoingCalls from entry points to trace how user input propagates
• findReferences on security-critical functions to verify all usages are safe
• goToImplementation on interfaces to find all concrete implementations that handle sensitive data

Enforcement Hooks

Input validated against schemas/input.schema.json before execution. Output contract defined in schemas/output.schema.json.

Pre-execute hook: hooks/pre-execute.cjs validates filePath is absolute and line/character are positive integers before invoking any LSP operation.

Iron Laws

1. ALWAYS provide absolute file paths — LSP requires full paths; relative paths cause silent failures or wrong-file resolution.
2. ALWAYS use 1-based line and character numbers — LSP protocol is 1-based; 0-based offsets produce off-by-one navigation errors.
3. NEVER use LSP for text discovery — LSP requires a known position to start from; use ripgrep or search:code to find initial entry points, then LSP to navigate from there.
4. ALWAYS verify the LSP server is active for the file's language — not all languages have LSP support in the workspace; fall back to ripgrep/structural search if LSP returns empty results.
5. ALWAYS use prepareCallHierarchy before incomingCalls/outgoingCalls — the call hierarchy operations require a prepared item from prepareCallHierarchy at the target position.

Anti-Patterns

| Anti-Pattern | Why It Fails | Correct Approach | | ----------------------------------------------------------- | -------------------------------------------------------------------- | -------------------------------------------------------- | | Using LSP to "search for" a concept | LSP needs a position, not a query | Use search:code for discovery, then LSP for navigation | | Relative file paths | LSP resolves against workspace root; may hit wrong file | Always use absolute paths | | 0-based line numbers | Off-by-one errors; navigates to wrong symbol | LSP is 1-based for both line and character | | Skipping prepareCallHierarchy | incomingCalls/outgoingCalls need a prepared item | Always call prepareCallHierarchy first | | Using LSP instead of rg for text-in-comments | LSP only sees code symbols, not text in comments/strings | Use ripgrep for text that includes non-code content | | Trusting empty LSP results as "no references" | Language server may not be running or file may not be indexed | Verify LSP is active; fall back to ripgrep if empty | | Using LSP on .cjs CommonJS files expecting TypeScript types | CJS files may have limited LSP support depending on workspace config | Fall back to ripgrep or structural search for .cjs files |

CJS File Limitations (Reflection from LSP Deep Dive)

In practice, most LSP operations (goToDefinition, findReferences, hover) return empty results for .cjs CommonJS files in this workspace. The TypeScript language server does not fully index CJS modules without explicit jsconfig.json or tsconfig.json coverage.

What this means:

• For .cjs files: prefer ripgrep (rg -F) for reference counting and require.resolve() for import validation.
• LSP documentSymbol may work on .cjs files for listing top-level exports, but is unreliable.
• The hybrid approach (ripgrep + require.resolve()) proved more effective than LSP for the agent-studio codebase's .cjs hook and lib files.

Recommendation: Use LSP as the primary tool for .ts and .js (ESM) files. For .cjs files, treat LSP as a secondary option and fall back to ripgrep immediately if LSP returns empty results.

Windows Path Normalization (SE-01)

On Windows, path.relative() and other Node.js path utilities return backslash (\) separators instead of forward slashes. LSP operations and file path comparisons expect forward-slash paths.

Rules for Windows compatibility:

• Always normalize paths before passing them to LSP operations: filePath.replace(/\\/g, '/')
• When comparing LSP result paths to local paths, normalize both sides
• The diagnostics runner (lsp-diagnostics-runner.cjs) handles this via its normalizePath() utility
• Use [^/\\]* in regex patterns if path normalization is uncertain
• Do NOT use path.relative() output directly in regex or glob patterns without normalizing

Example:

// WRONG: path.relative() returns backslashes on Windows
const relPath = path.relative(projectRoot, filePath); // ".claude\\lib\\routing.cjs"

// CORRECT: normalize before use
const relPath = path.relative(projectRoot, filePath).replace(/\\/g, '/');
// ".claude/lib/routing.cjs"

This is Sharp Edge SE-01 in the codebase. See .claude/rules/sharp-edges.md for full details.

Memory Protocol (MANDATORY)

Before starting: Read .claude/context/memory/learnings.md

After completing:

• New pattern -> .claude/context/memory/learnings.md
• Issue found -> .claude/context/memory/issues.md
• Decision made -> .claude/context/memory/decisions.md

ASSUME INTERRUPTION: If it's not in memory, it didn't happen.

Language Server Quick Reference

| Language | Server | Install | | -------- | ---------------------------- | --------------------------------------------- | | C/C++ | clangd | apt install clangd / brew install llvm | | Go | gopls | go install golang.org/x/tools/gopls@latest | | Python | pyright | npm i -g pyright | | Rust | rust-analyzer | rustup component add rust-analyzer | | Java | jdtls | brew install jdtls / Eclipse JDT LS release | | TS/JS | typescript-language-server | npm i -g typescript-language-server | | Lua | lua-language-server | brew install lua-language-server | | Bash | bashls | npm i -g bash-language-server | | JSON | jsonls | npm i -g vscode-langservers-extracted | | YAML | yamlls | npm i -g yaml-language-server | | TOML | taplo | cargo install taplo-cli | | Markdown | marksman | brew install marksman / GitHub releases |

Note: Claude Code activates these automatically when installed — no manual start required. LSP for .cjs files remains unreliable; use ripgrep as primary for CommonJS files.

Hookify Pattern

Auto-enable LSP server selection via a PostToolUse(Read) hook:

// .claude/hooks/lsp-enable.cjs
'use strict';
const path = require('path');

const input = JSON.parse(require('fs').readFileSync('/dev/stdin', 'utf8'));
const filePath = input?.tool_input?.file_path || '';
const ext = path.extname(filePath);

const serverMap = {
  '.py': 'pyright',
  '.rs': 'rust-analyzer',
  '.go': 'gopls',
  '.ts': 'typescript-language-server',
  '.js': 'typescript-language-server',
  '.lua': 'lua-language-server',
  '.sh': 'bashls',
};

const server = serverMap[ext];
if (server) {
  process.stdout.write(JSON.stringify({ lspServer: server }));
}
process.exit(0);

Register in .claude/settings.json:

{ "event": "PostToolUse", "tool": "Read", "command": "node .claude/hooks/lsp-enable.cjs" }