The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes the Bash tool to conduct environmental probes and execute local verification tasks as defined in its operational guides. Evidence: Instructions in references/probes.md direct the agent to use commands like java -version and command -v to verify runtime installations. Evidence: When the --verify flag is enabled by the user, the agent is permitted to execute project-defined scripts such as scripts/verify.sh or Makefile commands to test functionality.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted project configuration files that could potentially contain instructions designed to manipulate the agent's logic or bypass auditing rules. 1. Ingestion points: Reads CLAUDE.md, .claude/rules/**/*.md, settings.json, and various build manifests (e.g., build.gradle, package.json) from the project being audited. 2. Boundary markers: Relies on natural language instructions stating that file contents are audit data and do not grant execution permissions, but lacks hard technical delimiters to isolate this data from the instruction context. 3. Capability inventory: The agent has access to Bash, Write, Read, and LSP tools, enabling it to execute local shell scripts and write report files. 4. Sanitization: The skill masks sensitive patterns like tokens and authentication values in its output reports but does not sanitize the content of ingested files before processing.
[EXTERNAL_DOWNLOADS]: The documentation suggests the use of standard package managers and official plugin installers to resolve missing project dependencies identified during the audit. Evidence: The references/probes.md file recommends the use of brew install and official Claude plugin installations (e.g., jdtls-lsp@claude-plugins-official) to install necessary development tools.

check-harness