The Agent Skills Directory

[PROMPT_INJECTION]: Indirect prompt injection surface exists where untrusted data (user-provided problem and subagent-generated branch definitions) is interpolated into subsequent LLM prompts.
Ingestion points: tiling_tree.py accepts the problem string from command-line arguments and branches data from JSON responses generated by external subagents.
Boundary markers: Absent. Data is interpolated directly into prompts (e.g., _evaluator_prompt) without specific delimiters or instructions to ignore embedded commands.
Capability inventory: The skill uses invoke_claude and invoke_parallel for multi-agent orchestration and performs filesystem writes to /mnt/user-data/outputs/.
Sanitization: Absent. Content from subagent responses is parsed as JSON but the resulting strings are not sanitized for malicious instructions before being used in subsequent splitting or evaluation prompts.

tiling-tree