MCP HubMCP Hub
SKILL·08B59D

lsp-understand

blackwell-systems
Actualizado 21 days ago
5 vistas
162
19
162
Ver en GitHub
Metadesign

Acerca de

Esta habilidad proporciona una exploración integral de código mediante la construcción de un mapa de dependencias navegable para cualquier símbolo o archivo. Sintetiza relaciones entre implementaciones, jerarquías de llamadas, referencias y código fuente para ayudar a los desarrolladores a comprender bases de código desconocidas. Úsela cuando necesite una visión estructural completa que vaya más allá de una simple búsqueda de símbolos.

Instalación rápida

Claude Code

Recomendado
Principal
npx skills add blackwell-systems/agent-lsp -a claude-code
Comando PluginAlternativo
/plugin add https://github.com/blackwell-systems/agent-lsp
Git CloneAlternativo
git clone https://github.com/blackwell-systems/agent-lsp.git ~/.claude/skills/lsp-understand

Copia y pega este comando en Claude Code para instalar esta habilidad

Documentación

Requires the agent-lsp MCP server.

lsp-understand

Deep-dive exploration of unfamiliar code — given a symbol or file, synthesizes hover info, implementations, call hierarchy (bounded to 2 levels), all references, and source into a structured Code Map.

Read-only — does not modify any files.


Differentiation from lsp-explore

/lsp-explore is a single-symbol pass: given one symbol name, it runs hover + implementations + call hierarchy + references and produces a per-symbol report. Use lsp-explore for quick "what is this one thing" questions.

/lsp-understand is broader in three ways:

  1. Accepts a file path as input — explores all exported symbols in that file as a group (Mode B), rather than requiring a single symbol name.
  2. Synthesizes cross-symbol relationships — produces a dependency map showing how entry points call each other, share callers, or implement the same interface, rather than isolated per-symbol reports.
  3. Enforces a 2-level call hierarchy depth limit — prevents infinite recursion in deeply connected code.

Use lsp-understand for "how does this module work as a whole."


Input — Two Modes

Mode A (symbol): User provides a symbol name in dot notation (e.g., "codec.Encode", "Handler.ServeHTTP").

Mode B (file): User provides an absolute file path. All exported symbols in the file become the entry points.


Prerequisites

Call mcp__lsp__get_server_capabilities before Step 2 to determine which capabilities are available. Skip steps that require missing capabilities:

  • go_to_implementation: skip Step 2b if implementationProvider: false
  • find_callers: skip Steps 2c and 2d if callHierarchyProvider: false; note in the Code Map output that call hierarchy was unavailable

Step 1 — Entry Point Resolution

Mode A: Single Symbol

Call mcp__lsp__go_to_symbol to locate the symbol definition:

mcp__lsp__go_to_symbol({
  "symbol_path": "<dot-notation name>",   // e.g. "codec.Encode"
  "workspace_root": "<root>"              // optional
})
→ returns: file_path, line, column (1-indexed)

Record file_path, line, and column. If go_to_symbol returns nothing, report:

Symbol not found: <name> Check the dot-notation path (e.g. "Package.Symbol") and ensure the workspace root covers the file.

Stop immediately — do not proceed to Step 2.

The single symbol becomes the sole entry point.

Mode B: File Path

Call mcp__lsp__open_document then mcp__lsp__list_symbols:

mcp__lsp__open_document({ "file_path": "<absolute path>" })

mcp__lsp__list_symbols({ "file_path": "<absolute path>" })
→ returns: list of symbols with kind, line, column

Filter to exported symbols:

  • Go: uppercase first letter
  • TypeScript/JavaScript: export keyword
  • Rust: pub visibility

Cap at 10 exported symbols maximum. If more than 10 are found, prioritize top-level functions and types; skip constants and variables.

Each filtered symbol becomes an entry point with its file_path, line, and column.


Step 2 — Per-Symbol Analysis

For each entry point, run the following sub-steps. Where possible, parallelize calls within each step.

2a — Type Info and Docs

Call mcp__lsp__inspect_symbol using position_pattern with the @@ marker (see references/patterns.md):

mcp__lsp__inspect_symbol({
  "file_path": "<file>",
  "position_pattern": "<symbol@@name>",
  "line_scope_start": <line - 5>,
  "line_scope_end": <line + 5>
})
→ returns: hover text with type signature and doc comment

Store result as hover_text. If the call fails or returns nothing, set hover_text to an empty string. Do not stop.

2b — Implementations (capability-gated)

If implementationProvider is available in server capabilities:

mcp__lsp__go_to_implementation({
  "file_path": "<file>",
  "line": <line>,
  "column": <column>
})
→ returns: list of concrete implementation locations

Skip if capability is absent. Record "not supported by this server" rather than stopping.

2c — Incoming Call Hierarchy (bounded to 2 levels)

If callHierarchyProvider is available:

Level 1 — Direct callers:

mcp__lsp__find_callers({
  "file_path": "<file>",
  "line": <line>,
  "column": <column>,
  "direction": "incoming"
})
→ returns: list of direct caller functions with file and line

Level 2 — Callers of callers:

For each Level 1 caller, call mcp__lsp__find_callers once more:

mcp__lsp__find_callers({
  "file_path": "<caller file>",
  "line": <caller line>,
  "column": <caller column>,
  "direction": "incoming"
})
→ returns: Level 2 callers

STOP at Level 2 — do not recurse further under any circumstances.

If Level 2 callers > 10: summarize by count and file, do not list individually.

2d — Outgoing Calls (Level 1 only)

If callHierarchyProvider is available:

mcp__lsp__find_callers({
  "file_path": "<file>",
  "line": <line>,
  "column": <column>,
  "direction": "outgoing"
})
→ returns: list of functions this symbol calls

Level 1 only — no recursion.

2e — All References

mcp__lsp__find_references({
  "file_path": "<file>",
  "line": <line>,
  "column": <column>,
  "include_declaration": false
})
→ returns: every usage site across the workspace

Group by file and count distinct files.

2f — Source

mcp__lsp__get_symbol_source({
  "file_path": "<file>",
  "line": <line>,
  "column": <column>
})
→ returns: implementation body

Step 3 — Synthesize Relationships

After analyzing all entry points, identify cross-symbol relationships:

  • Internal calls: Which entry points call each other? (from outgoing calls in Step 2d)
  • Shared callers: Which entry points are called by the same Level 1 callers?
  • Shared interface: Which entry points implement the same interface? (from Step 2b)

This synthesis step is what distinguishes /lsp-understand from running multiple /lsp-explore calls. The output is a dependency map, not isolated per-symbol reports.


Step 4 — Output: Code Map

Produce a structured Code Map with these sections:

## Code Map: <target>

### Summary
<2-3 sentence description of what this code does, synthesized from
hover docs and source reading>

### Symbols (<N> analyzed)

#### <SymbolName>
- **Type:** <type signature from hover>
- **Source:** <file:line>
- **Incoming callers (L1):** <list; count only if > 5>
- **Incoming callers (L2):** <summarized; e.g., "called by 3 HTTP handlers">
- **Outgoing calls:** <what this symbol calls>
- **Implements:** <interface name, if applicable>
- **References:** N sites across M files

### Dependency Relationships
<symbols that call each other, as a simple text diagram or list>
e.g.:
  HandlerA → Parse → Validate
  HandlerB → Parse

### Entry Points to This Code
<top-level callers that are NOT in this file — where does outside code
call in?>

### Depth-limit Note
Call hierarchy stopped at 2 levels. <N> additional callers exist beyond
Level 2 — use /lsp-explore on specific symbols to drill deeper.

Depth Control Rules

These limits are hard constraints — never exceed them:

  • Incoming call hierarchy recursion stops at Level 2
  • Outgoing calls: Level 1 only, no recursion
  • If Level 2 callers > 10: summarize by count and file, do not list individually
  • Do NOT follow call chains beyond these limits under any circumstances

Example

Goal: understand how the file pkg/codec/encoder.go works as a whole

Step 1 — Mode B (file path)
  open_document: pkg/codec/encoder.go
  list_symbols: pkg/codec/encoder.go
  → exported symbols: Encoder (type), Encode (func), Reset (func), NewEncoder (func)
  → 4 exported symbols (under 10 cap)

get_server_capabilities
  → go_to_implementation: supported
  → find_callers: supported

Step 2 — Per-symbol analysis (run in parallel across symbols)

  Symbol: NewEncoder (pkg/codec/encoder.go:12)
    inspect_symbol → "func NewEncoder(w io.Writer) *Encoder"
    go_to_implementation → 0 (concrete function)
    find_callers incoming L1 → 5 callers
    find_callers incoming L2 → 3 callers of those callers
    find_callers outgoing → calls: bufio.NewWriter
    find_references → 5 sites in 3 files
    get_symbol_source → implementation body

  Symbol: Encode (pkg/codec/encoder.go:28)
    inspect_symbol → "func (e *Encoder) Encode(v any) error"
    go_to_implementation → implements codec.Encoder interface
    find_callers incoming L1 → 8 callers (listed)
    find_callers incoming L2 → > 10: "12 additional callers across 5 files"
    find_callers outgoing → calls: NewEncoder, e.w.Flush
    find_references → 8 sites in 5 files
    get_symbol_source → implementation body

  (similar for Encoder type and Reset func...)

Step 3 — Synthesize relationships
  - Encode calls NewEncoder (internal dependency)
  - NewEncoder and Encode share callers in cmd/main.go
  - Encode implements codec.Encoder interface

## Code Map: pkg/codec/encoder.go

### Summary
This file implements a streaming JSON encoder backed by a buffered writer.
NewEncoder constructs an Encoder wrapping any io.Writer; Encode serializes
values and flushes. Reset allows reuse without allocation.

### Symbols (4 analyzed)

#### NewEncoder
- **Type:** func NewEncoder(w io.Writer) *Encoder
- **Source:** pkg/codec/encoder.go:12
- **Incoming callers (L1):** cmd.main, app.Start, loader.Load, test.Setup, bench.Run
- **Incoming callers (L2):** 3 callers across 2 files
- **Outgoing calls:** bufio.NewWriter
- **Implements:** n/a
- **References:** 5 sites across 3 files

#### Encode
- **Type:** func (e *Encoder) Encode(v any) error
- **Source:** pkg/codec/encoder.go:28
- **Incoming callers (L1):** 8 callers (cmd/main.go, internal/app.go, ...)
- **Incoming callers (L2):** 12 additional callers across 5 files (depth limit reached)
- **Outgoing calls:** NewEncoder, e.w.Flush
- **Implements:** codec.Encoder
- **References:** 8 sites across 5 files

...

### Dependency Relationships
  cmd.main → NewEncoder → bufio.NewWriter
  cmd.main → Encode → NewEncoder
  Encode → Reset

### Entry Points to This Code
- cmd.main (cmd/main.go:14)
- app.Start (internal/app.go:31)
- loader.Load (internal/loader.go:55)

### Depth-limit Note
Call hierarchy stopped at 2 levels. 12 additional callers exist beyond
Level 2 for Encode — use /lsp-explore on specific symbols to drill deeper.

Repositorio GitHub

blackwell-systems/agent-lsp
Ruta: skills/lsp-understand
0
agentskillsai-agentsai-toolingclaudeclaude-codecode-intelligence
FAQ

Preguntas frecuentes

¿Qué es el Skill lsp-understand?

lsp-understand es un Skill de Claude creado por blackwell-systems. Los Skills agrupan instrucciones y recursos que Claude carga cuando los necesita para realizar tareas relacionadas con lsp-understand sin indicaciones adicionales.

¿Cómo instalo lsp-understand?

Usa los comandos de instalación de esta página: añade lsp-understand a Claude Code como plugin o clona su repositorio en tu directorio de skills y reinicia Claude para cargarlo.

¿A qué categoría pertenece lsp-understand?

lsp-understand pertenece a la categoría Meta.

¿Se puede usar lsp-understand gratis?

Sí. lsp-understand aparece en AIMCP y se puede instalar gratis.

Habilidades relacionadas

content-collections
Meta

Esta habilidad proporciona una configuración probada en producción para Content Collections, una herramienta centrada en TypeScript que transforma archivos Markdown/MDX en colecciones de datos con tipado seguro mediante validación Zod. Úsala al construir blogs, sitios de documentación o aplicaciones Vite + React con mucho contenido para garantizar seguridad de tipos y validación automática de contenido. Abarca todo, desde la configuración del plugin de Vite y compilación MDX hasta la optimización de despliegue y validación de esquemas.

Ver habilidad
polymarket
Meta

Esta habilidad permite a los desarrolladores crear aplicaciones con la plataforma de mercados de predicción Polymarket, incluyendo la integración de API para operaciones y datos de mercado. También proporciona transmisión de datos en tiempo real a través de WebSocket para monitorear operaciones en vivo y actividad del mercado. Úsela para implementar estrategias de trading o crear herramientas que procesen actualizaciones de mercado en tiempo real.

Ver habilidad
creating-opencode-plugins
Meta

Esta habilidad ayuda a los desarrolladores a crear complementos de OpenCode que se conectan a más de 25 tipos de eventos, como comandos, archivos y operaciones LSP. Proporciona la estructura del complemento, las especificaciones de la API de eventos y los patrones de implementación para módulos en JavaScript/TypeScript. Úsala cuando necesites interceptar, monitorear o extender el ciclo de vida del asistente de IA de OpenCode con lógica personalizada basada en eventos.

Ver habilidad
sglang
Meta

SGLang es un framework de alto rendimiento para el servicio de LLM que se especializa en generación rápida y estructurada para JSON, expresiones regulares y flujos de trabajo de agentes utilizando su caché de prefijos RadixAttention. Ofrece una inferencia significativamente más rápida, especialmente para tareas con prefijos repetidos, lo que lo hace ideal para salidas complejas y estructuradas, y conversaciones multiturno. Elige SGLang sobre alternativas como vLLM cuando necesites decodificación restringida o estés construyendo aplicaciones con uso extensivo de prefijos compartidos.

Ver habilidad