The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill instructions specify the use of the codex tool with high-privilege flags (--sandbox danger-full-access) and explicitly disables user confirmation (--ask-for-approval never). This configuration enables the agent to implement code and execute shell commands without human oversight during the loop iterations.
[INDIRECT_PROMPT_INJECTION]: The multi-agent pipeline (Planner, Implementor, Reviewer, Judge) operates on external data from prd.json and progress.txt which directly influences agent behavior.
Ingestion points: The PLANNER reads prd.json to select implementation tasks, and the IMPLEMENTOR uses the project context to generate code changes (SKILL.md).
Boundary markers: There are no documented delimiters or safety instructions to prevent malicious content within the PRD or codebase from hijacking the agent's actions.
Capability inventory: The implementor utilizes codex with full sandbox access to write to the file system and commit changes, while the reviewer and judge execute generated tests and code.
Sanitization: The skill mentions structural schema validation for the PRD, but no content-level sanitization or instruction filtering is present.
[COMMAND_EXECUTION]: The skill provides instructions for executing local CLI tools and deployment scripts located in the vendor's repository (~/Code/joelhooks/joelclaw/), including publish-system-bus-worker.sh and several joelclaw subcommands for loop management.

agent-loop