The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes shell scripts (setup-ralph.sh, stop-hook.sh, and cancel-ralph.sh) to manage its internal state and control a repetitive task loop. These scripts perform standard local operations such as file creation, directory management, pattern matching with grep, and JSON parsing with jq to maintain execution context across sessions.
[PROMPT_INJECTION]: A 'Stop Hook' mechanism in stop-hook.sh intercepts the agent's exit signal and re-injects the task prompt (using exit 2). This behavior overrides the default exit logic to maintain a continuous loop for multi-session tasks. It includes a max-iterations safety guardrail (ranging from 10 to 50 iterations) to prevent infinite loops.
[PROMPT_INJECTION]: The skill's architecture creates a surface for indirect prompt injection as defined by Category 8.
Ingestion points: The instructions in SKILL.md guide the agent to analyze local source code, read package.json, and perform web browsing or visual verification using Playwright tools.
Boundary markers: The re-injection script (stop-hook.sh) does not wrap the re-injected prompt in delimiters or include instructions to ignore commands found in the processed data.
Capability inventory: The agent has the capability to execute shell commands (git operations, setup scripts), perform file system modifications, and use Playwright browser tools.
Sanitization: No evidence of input validation or sanitization is present for the external content processed during the execution loop before it is reflected in the agent's context.

ralph-planner