The Agent Skills Directory

[COMMAND_EXECUTION]: The skill templates incorporate shell functions and wrapper scripts that utilize eval to execute commands provided by the agent. Evidence includes templates/code-agent-workspace/debugging/wrapper-scripts/run-with-logging.sh which executes strings using the eval builtin.
[DATA_EXFILTRATION]: The skill documentation provides configuration templates suggesting broad file system access and includes scripts that handle file paths without strict traversal validation. Evidence includes references/permissions_setup_protocol.md which contains templates for .claude/settings.local.json that suggest allowing recursive read/write access to project and client directories. Additionally, templates/architect-workspace/scripts/tail-code-agent-logs.sh and templates/code-agent-workspace/.claude/hook-logger.py resolve file paths from user input or local configuration files without path validation.
[INDIRECT_PROMPT_INJECTION]: The skill's primary workflow involves a code agent executing instructions from a file controlled by an architect agent, creating an attack surface for cross-agent instruction injection. Ingestion point is debugging/instructions/current_instructions.md. Boundary markers and sanitization are absent, as the agent is instructed to read and follow the contents directly while possessing access to powerful tools like Bash and Write.
[EXTERNAL_DOWNLOADS]: The skill references external tools and dependencies associated with the author's organization. README.md instructs users to install the skilz universal installer via pip, and templates/code-agent-workspace/.opencode/package.json includes the @opencode-ai/plugin package.

architect-agent