The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes several common development tools, including jq, rg (ripgrep), git, and the GitHub CLI (gh). These are used for legitimate technical tasks:
Parsing dependency manifests (package.json, requirements.txt, Cargo.toml).
Searching the local codebase for exported symbols and architectural patterns.
Managing a private state repository (sidecar) via git commands.
[EXTERNAL_DOWNLOADS]: During Phase 4, the skill performs targeted web searches to retrieve official documentation for project dependencies. It also uses the GitHub CLI to list and read issues from upstream repositories to identify known bugs and breaking changes. These operations are performed against well-known services.
[INDIRECT_PROMPT_INJECTION]: The skill processes potentially untrusted data from external documentation and issue trackers.
Ingestion points: Technical documentation from the web and bug reports from GitHub are ingested in Phase 4.
Boundary markers: The skill employs a structured markdown template for its output, creating a logical separation between ingested content and the research summary.
Capability inventory: The agent maintains capabilities for shell execution (bash), file system searching (rg), and platform interaction (gh).
Sanitization: While explicit sanitization is not detailed, the skill provides instructions to verify all external claims against the local codebase and official documentation before committing to a design.
[SAFE]: The utility script references/sidecar.sh manages private state repositories. It performs local file operations and git commands within a dedicated directory (~/.claude/sidecars). The script follows best practices for isolation, such as unsetting sensitive environment variables (e.g., GIT_DIR) before performing sidecar operations, and does not exhibit any suspicious network activity.

research