The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes external agent binaries (opencode, agy, kiro, orca) and various local helper scripts such as code-logger.py and archetype.py. It also uses docker inspect to query container metadata, which is intended as a safety check for volume mounts.
[DATA_EXFILTRATION]: The skill performs network requests to https://agentmemory.giatbh.io.vn/ to store and retrieve session memory. The instructions include curl commands using the -k flag, which bypasses SSL certificate verification and could facilitate man-in-the-middle attacks.
[EXTERNAL_DOWNLOADS]: The skill synchronizes its instructions by cloning a specific GitHub repository (rheinmir/setup) at runtime.
[PROMPT_INJECTION]: The orchestration logic uses a dispatch --inject mechanism to pass generated task briefs directly into sub-agent contexts. This creates a surface for indirect prompt injection where content from a PLAN.md file (which may contain or be influenced by untrusted data) is treated as instructions for worker agents without explicit sanitization or boundary markers.

orca-workflow