How to use the LLDB Debugger Control agent skill

What is an agent skill and what does this one do?

An agent skill is a folder containing a SKILL.md file that tells an AI agent when and how to use a tool. This skill's SKILL.md describes a stateful LLDB workflow: the agent runs cli-anything-lldb commands and reads the JSON they return, instead of guessing at debugger output.

Once installed, the agent can create a debug target from an executable, launch or attach to a process by pid or name, manage breakpoints, inspect threads, frames, locals and backtraces, evaluate expressions in the current frame, read and find process memory, and load core dumps. It can also run a stdio Debug Adapter Protocol server for editor or AI clients.

Install the LLDB Debugger Control skill

Download the ZIP from this page and unzip it into one of two places. For a personal setup, unzip it into ~/.claude/skills/. For a single project, unzip it into the project's .claude/skills/ folder. Alternatively, in the Claude apps, go to Settings → Capabilities → Skills and upload the ZIP there.

The skill also depends on the underlying tool. The source lists the install as pip install cli-anything-lldb, and it needs lldb with Python bindings, click 8.0 or newer and prompt-toolkit 3.0 or newer. Check that the Python bindings import correctly before you rely on the skill, because the tool will not work without them.

• Requires lldb with Python bindings
• Requires click>=8.0 and prompt-toolkit>=3.0
• Install the tool with pip install cli-anything-lldb

How do you trigger the skill with a natural request?

You do not need to name the command. Describe the debugging task in plain terms that mention a program, a crash, a process or a breakpoint, and the agent can pick the skill. Name the executable or pid when you have it, since the agent needs a target before it can launch or attach.

Good requests say what you have and what you want to learn, for example: 'Load ./build/parser into LLDB, break at main when i is over 10, and show me the backtrace when it stops.' Vague requests such as 'fix my app' give the agent little to work with.

• Name the executable path, pid or process name
• State the function, file and line for breakpoints
• Say whether you want a launch, an attach or a core dump analysis

Run the commands the agent uses

The agent normally runs these commands itself, but knowing them helps you check its work. Every command uses the --json flag, which the source recommends for automated flows. Target creation must come before process or core commands.

Commands are grouped into target, process, breakpoint, thread, frame, step, expr, memory and core. Errors come back as an object with error and type fields.

• Create target: target create --exe PATH
• Launch or attach: process launch, process attach --pid or --name
• Breakpoints: breakpoint set, list, delete, enable, disable
• Inspect: thread backtrace, frame locals, expr

Use the session and clean-up behaviour correctly

Separate non-REPL invocations share a persistent session daemon by default. To keep a task on its own session, pass --session-file PATH or set the CLI_ANYTHING_LLDB_SESSION_FILE environment variable.

When the work is finished, run session close. The source says this detaches attached processes and cleans up launched debuggees. Unresolved CLI breakpoints fail by default, so pass --allow-pending only when you expect a module or symbol to load later.

• Use --session-file PATH to pin a session
• Always finish with cli-anything-lldb --json session close
• Use the REPL only when an interactive shell is more convenient

Use the Debug Adapter Protocol server for editors and AI clients

When an AI client needs a standard debug adapter lifecycle instead of separate shell commands, start the DAP entry point. It speaks stdio Content-Length frames and needs exclusive stdout, so do not print logs to stdout around it. Supported requests include launch, attach, setBreakpoints, stackTrace, variables, evaluate, continue, pause, next, stepIn, stepOut and disconnect, among others.

Stop rules let you filter internal startup or shader-JIT breakpoints using a profile JSON file. Restart the adapter after changing a profile, because running processes do not hot-load new rules.

• Start with cli-anything-lldb-dap or cli-anything-lldb dap
• Optional profile: --profile /path/to/stop-rules.json
• Restart the adapter after profile changes

Safety notes: scripts and what they touch

This skill runs scripts, so review what the agent is about to execute, particularly process launch and attach commands. The source says the skill does not use the network.

Attaching to a process and reading its memory can expose sensitive data and can disrupt a running service. Only use it on programs you own or are authorised to debug, and avoid live production systems.

• Runs scripts: yes
• Uses network: no
• Only attach to processes you own or are allowed to inspect