The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill and its included scripts (e.g., collect-evidence.sh, validate-refine.sh) utilize a remote installation script via curl -fsSL https://sruja.ai/install.sh | bash. This pattern executes code directly from the vendor's domain (sruja.ai) to install the required CLI tool.
[COMMAND_EXECUTION]: The instructions and prompt templates in SKILL.md and references/PROMPTS.md guide the agent to execute shell commands such as sruja sync, sruja discover, sruja lint, and sruja drift. These are used to gather evidence from the local filesystem and validate the architecture DSL.
[EXTERNAL_DOWNLOADS]: The skill references external downloads from sruja.ai for tool installation. It also suggests using npx @sourcegraph/scip-typescript --infer for semantic indexing, which downloads packages from the npm registry.
[PROMPT_INJECTION]: The skill possesses an attack surface for indirect prompt injection as it ingests untrusted data from the local codebase (source code, imports, and comments) to synthesize architectural boundaries and elements.
Ingestion points: Files processed by sruja discover and sruja sync across the repository.
Boundary markers: Prompt templates do not define explicit delimiters or 'ignore' instructions for the data being analyzed during synthesis.
Capability inventory: Subprocess execution for CLI commands, file creation for architecture proposals and DSL files.
Sanitization: No specific sanitization or validation of the ingested code content is performed before processing.

sruja-architecture