extract-api-surface

extract-api-surface

When this skill is the preferred path

The user wants to get oriented in a codebase quickly — they need to know what's there, not how each function is implemented. Reading every file just to answer "what does this library do" wastes context and misses the forest for the trees.

How to use it

1. Identify the target: a local directory (./src), a repo root, or a GitHub URL the user linked to.
2. Call read_skeleton (free-tier MCP tool) on the target path. You'll get function/class signatures, type definitions, module-level docstrings, and import maps — bodies stripped. Compression is adaptive: smaller files keep most of their structure; larger files compress harder. Sections not anchored come back as short summaries you can expand on demand with modulate_region. For cross-file understanding across a larger tree, compress_codebase (Pro+) returns the same AST-aware digest for every file in the corpus at once.
3. If the skeleton is still too large (rare, large monorepos), pipe it through ingest_context at fidelity=balanced.
4. Present a structured overview:
   • Top-level exports grouped by module
   • Key types / interfaces
   • Entry points (CLI, main, server)
   • One-sentence purpose per module
   • Total signature count (so the user knows the coverage)

When the raw read is actually fine

• You're actively editing a specific file — you need the body.
• You're debugging a specific function — skeleton hides it.
• Configs, data fixtures, generated code — skeleton isn't meaningful; either skip or run ingest_context on them.

Language support

AST-native: Python, JavaScript, TypeScript. Line-based fallback: Java, Go, Rust, C++. Other languages fall back to regex slicing — surface that caveat to the user if you hit a non-native language.

See references/overview-template.md for the reporting format.