Module api_surface 

The knowledge_base workspace API-surface technique (#669), with a pluggable language-pack model.

A language pack is pure DATA (crate::config::LanguagePack): file extensions, entry-point file globs, and regex symbol-extraction rules with free-form kind labels. So adding a language is config, not code.

• Built-in packs (first-class): Rust, Python, Bash, C/C++, Go, Java — see builtin_packs. They double as the canonical examples.
• External packs: drop a <name>.toml into ~/.newt/language-packs/ (global) or .newt/language-packs/ (project-local), or inline under [[context.api_surface.language_packs]]. Packs merge by name, so a community pack (Ruby, Swift, Objective-C, …) adds or overrides without ever touching the binary. See docs/language-packs.md + examples/language-packs/.

The rendered surface rides the frozen system prompt — the compressor’s protected head ([super::agentic::compress] head_len) — so it is a stable base that is never summarized (#661 group E).

Extraction engine: regex is the BOOTSTRAP; AST (tree-sitter) is the target

Per-line regex rules are a deliberate bootstrap — fragile for real code (multi-line declarations, macros, generics, attributes). The right engine is an AST parser (tree-sitter): a pack becomes a grammar + a tree-sitter query, and the query — like the regex rules here — is pluggable DATA. The architecture is chosen so that swap is local: extensions / entry_points / merge-by-name / drop-in loading are all engine-agnostic; only the per-pack extraction rules change shape (regex → query). Tracked as a follow-up.

Structs

ApiSurfaceProvider

Injects the workspace’s public API surface into the system prompt, driven by a resolved set of language packs.

Functions

builtin_packs

The first-class built-in language packs (Rust, Python, Bash, C/C++, Go, Java). These also serve as the worked examples a contributor copies — newt language-pack template <lang> emits one to a drop-in file. Best-effort regexes (a surface, not a parser); override any of them by name with your own pack.

load_packs_from_dir

Load language packs from a drop-in directory (<dir>/*.toml, one pack per file). A malformed pack file is skipped with a warning, never fatal — the same “drop-in, tolerant” contract as the [backends] directory. A missing directory yields [].

merge_packs

Merge pack layers by name — later layers win (built-ins < global dir < project dir < inline config). A custom pack named rust replaces the built-in rust; a new name adds a language. Output order is stable (insertion order of first appearance, then last value wins).