The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructs the agent to use npx to download and run the likec4 package from the npm registry for generating architecture diagrams.
[REMOTE_CODE_EXECUTION]: Executes the likec4 CLI tool dynamically using npx, which allows for the execution of remote code at runtime.
[COMMAND_EXECUTION]: Spawns subprocesses to run commands such as npx likec4 --version and likec4 start. The latter initializes a local web server on port 5173.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection as it scans untrusted repository content.
Ingestion points: Manifests (package.json, pyproject.toml, etc.), README files, ADR directories, and LikeC4 files.
Boundary markers: Uses specific tagging such as [CONFIRMED YYYY-MM-DD] and [FROM CODE, UNCONFIRMED YYYY-MM-DD] to distinguish between human-verified and tool-extracted data.
Capability inventory: Performs file reads (globbing repo), file writes (creating discovery briefs and LikeC4 models), and subprocess execution (npx/likec4).
Sanitization: Implements a 'zero-hallucination' policy where no automated finding is treated as a fact until confirmed by the user.

adr-discovery