The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill's TypeScript scripts (e.g., scripts/docs-probe.mts, scripts/architecture/check-page.mts, scripts/architecture/lint.mts) include an ensureBuilt function. This function automatically executes bun install to fetch Node.js dependencies from the skill's package.json if the built bundle is missing or determined to be stale based on file modification times.
[DYNAMIC_EXECUTION]: The same scripts utilize a self-bundling mechanism where they invoke bun build at runtime to compile their own source code into a .build directory before executing the resulting JavaScript bundle. This is used to ensure high performance and dependency resolution.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted local data including Markdown documents, YAML configuration files, and repository source code. While the skill uses structured AST parsers for analysis, this creates a potential surface for indirect prompt injection if the files being reviewed contain instructions intended to influence the agent's behavior during documentation reshaping or review tasks.
Ingestion points: Processes any Markdown file in a repository, the architecture.yml declaration, and source code files (TS/JS) for symbol extraction.
Boundary markers: The scripts use AST-based segmentation (mdast, babel) to isolate content, but the skill instructions do not specify explicit boundary markers for the agent when processing prose.
Capability inventory: The skill possesses file system read/write access (for updating architecture maps) and subprocess execution capabilities (specifically for its self-build/bundling logic).
Sanitization: Employs robust structured parsers (mdast-util-from-markdown, @babel/parser) to extract data from files, which mitigates simple injection attempts within the parsing logic itself.

working-text