| name | pruner |
| description | Generate synthetic code context for LLM coding tasks. Automatically use this before making code changes, fixing bugs, refactoring, or answering questions about the codebase. Provides execution paths, relevant files, symbols, tests, and code snippets. |
| argument-hint | <ask> |
| allowed-tools | Bash(pruner *), Read, Grep, Glob |
| user-invocable | false |
Pruner — Code Context Engine
Pruner context is automatically injected into your conversation via a prompt-submit hook. You do NOT need to run pruner context manually — it has already run before you see the user's message.
How it works
A hook runs pruner context on every prompt submit and injects the output as additional context. Look for the "Pruner context (pre-computed codebase analysis)" section in the conversation — that's the pruner output.
Working with pruner context
- Use the injected context directly — it contains relevant files, symbols, execution paths, and code snippets.
- Read source files only if needed — if a snippet is truncated or you need surrounding context, use the file:line pointers from the output.
- Do NOT re-explore — pruner already searched the index. Do not grep or glob for the same keywords.
- Do NOT run
pruner context again — it already ran via the hook.
When to escalate
The hook output is a brief summary (~2K tokens) with file/symbol pointers. For deeper context:
- Read
.pruner/context.md — already written with full execution paths and code snippets. Zero cost, instant. Use this for understanding queries ("how does X work?"), debugging, or when you need to trace a flow.
- Run
pruner context --detail — regenerates focused output (~10-15K tokens). Use only if .pruner/context.md is stale or missing.
For code changes, refactoring, and simple lookups, brief pointers are sufficient — just read the specific files listed.
- Single symbol:
pruner show-symbol /path/to/repo "<name>" — signature, callers, callees.
- Single file:
pruner show-file /path/to/repo "<path>" — all symbols and imports.
Other modes
--detail — execution paths + code snippets (~10-15K tokens). Use when brief pointers aren't enough.--brief — metadata only, no snippets (~2K tokens). Use when you only need file/symbol pointers.--full — uncapped output with all snippets (~50-70K tokens). Use for deep analysis.
