Skip to content

Code maps vertical slice for curated context #407

Description

@jamesrochabrun

Parent: #398

Depends on: Context Builder MVP and token estimates.

Goal

Add a first, narrow code-map implementation that lets users include signature-only API context for large files at much lower token cost.

Implementation Notes

Do not attempt the full multi-language RepoPrompt code-map system in one issue. Start with a vertical slice, preferably Swift first because AgentHub itself is Swift-heavy.

Suggested first slice:

  • signature/API model for one language
  • content fingerprint cache
  • per-file mode in Context Builder: full / code map / exclude
  • assembled <file_map> section for mapped files
  • token estimate uses code-map text instead of full file text
  • unsupported languages degrade to full/exclude without crashing

Tree-sitter may be used, but isolate dependencies so the rest of AgentHub does not pay unnecessary build complexity. If adding grammars, include license attribution for vendored grammar dependencies.

Acceptance Criteria

  • A large Swift file can be added as code map and contributes signatures without function bodies.
  • Code-map token estimate is materially lower than full-file estimate.
  • Re-running on unchanged content uses the cache.
  • Unsupported files do not crash context assembly.

Tests

Add tests for extraction output, cache invalidation by content fingerprint, fallback behavior, and assembly into the <file_map> section.

Metadata

Metadata

Assignees

No one assigned

    Labels

    enhancementNew feature or request

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions