Ada scanner

The experimental @groma/scanner-ada plugin reads selected .ads and .adb source files. It carries its parser as WebAssembly and needs no GNAT, Alire, GPRbuild, application dependencies, or network connection during a scan. The same package runs on Groma's supported Windows, Linux and macOS hosts. Default exclusions omit root alire/ and obj/ directories and test directories. Source packages such as src/alire/ and src/obj/ stay selected.

groma plugin add @groma/scanner-ada
groma scan

Source evidence

The scanner lists declarations from parsed source and reports named subprogram bodies, expression functions and null procedures as operations. Procedure-call statements belong to their nearest subprogram body, including nested local subprograms. Package elaboration and task/entry execution are not assigned to an enclosing subprogram. All procedure calls are unresolved: overloads, access-to- subprogram values and dispatch require semantic analysis beyond this parser. Function expressions are omitted because Ada uses the same syntax for calls, array indexing and type conversions. Duplicate-body comparison is not supported.

Exactly one selected library package specification and one selected body with the same case-insensitive declared unit name form a sourceUnits claim. The specification is primary. Filenames and directories do not establish this association. Duplicate declarations or bodies remain separate; no build variant is chosen. Subunits (separate) and standalone subprogram specifications/bodies are not associated in this first version.

There are no inferred execution entries: a library procedure is not necessarily a program's main procedure. GPR files, Alire manifests, dependency closure and launch configuration are not analyzed. Package hierarchy supplies source names, not C4 boundaries. Core owns component placement and OKF Markdown output.

Code outline

Package scopes act as namespaces. Their ordinary type declarations and subprogram declarations/bodies, including generic subprograms, appear in physical source order. Local subprograms and local types inside executable bodies are omitted from the outline. Subtypes, renamings, instantiations, generic formal parameters and protected/task types and entries are not listed. Ada primitive operations remain functions; the scanner does not invent methods on their parameter types. Separate subunit files (separate) have no Code outline in this version, even though their named subprogram operations remain in scan evidence.

Visible declarations are public. A package's private part and private child units are internal, because Ada permits access inside the package and its children. Package-body-only declarations are private. A subprogram body in a package inherits visibility only from one selected specification with the same syntactic profile, ignoring case, whitespace, comments and default expressions. When an identically named declaration exists but its profile cannot be matched, that body is omitted from the outline. A package body without its selected specification also omits subprogram outlines, while retaining operation evidence. The outline reads only requested Code files and supplied selected source context. Names and entry symbols use Ada's case-insensitive spelling. Lines refer to the physical name; evidence positions use zero-based UTF-16 offsets.

Parser limits and ownership

The MIT-licensed Tree-sitter Ada grammar is pinned to a45fd9c65a791cbe7d146e85a53fbe43a6d52a61 with a small local patch for GNAT array type indexes such as Positive range 1 .. <>. It is a partial Ada grammar, not a compiler or a complete Ada legality checker. A syntax error or a source form the grammar cannot parse fails with its file and physical position. GNAT preprocessor directives are rejected rather than treating both branches as active. The scanner's diagnostic states the supported evidence limits on every scan.

At HAC revision 8710a2c, the complete 99-file compiler implementation under src/ was verified. Its 396-file default selection correctly rejects four invalid examples under exm/not_working/: three use non-Ada colon formatting in procedure arguments, and one has an empty statement sequence instead of null;. These failures were checked against the Ada reference manual and separate minimal controls; the directory name was not used to classify them. The complete default selection is not counted as a successful scan.

src/index.ts receives selected paths, reads each source and constructs the shared observation. src/syntax.ts owns parser initialization, diagnostics and WASM memory. src/declarations.ts owns outlines and spec/body associations. src/evidence.ts owns subprogram boundaries and procedure-call facts. build.ts bundles JavaScript, both WASM assets, licenses and this documentation.

To regenerate the checked-in grammar asset, check out the pinned upstream commit and apply assets/gnat-fixed-lower-bound.patch with git apply. Run tree-sitter generate --abi 14, then tree-sitter build --wasm, using Tree-sitter CLI 0.26.3 and wasi-sdk 29. Copy its tree-sitter-ada.wasm into assets/. Normal package builds only need the repository's locked Bun dependencies and do not rebuild the grammar.