The Agent Skills Directory

[DYNAMIC_EXECUTION]: The production_adapters function in scripts/adapters/__init__.py uses importlib.util to dynamically load and execute Python modules from file paths provided as strings. This enables the AI agent to execute arbitrary local Python code by specifying a path via the --adapter command-line argument used in the consult.py and panel.py scripts.
[COMMAND_EXECUTION]: The skill launches and manages external processes (specifically the claude and codex CLIs) using asyncio.create_subprocess_exec in scripts/adapters/base.py. It also allows for the execution of arbitrary verification commands (checks) provided in the panel roster or configuration.
[INDIRECT_PROMPT_INJECTION]: The skill functions as an orchestrator for multiple AI agents, ingesting their natural language outputs as evidence. This creates a vulnerability surface where a malicious or compromised participant could provide outputs designed to manipulate the host agent's reasoning or influence the behavior of other panel members.
Ingestion points: Model outputs are parsed from capture files in scripts/adapters/claude.py and scripts/adapters/codex.py and ingested as messages or reviews in scripts/engine.py.
Boundary markers: The system uses structural JSON envelopes and string delimiters (e.g., PANEL_INPUT) to separate instructions from participant data, but these do not provide robust protection against adversarial text.
Capability inventory: The skill can perform git operations, write to the filesystem, and execute shell commands and network requests (web search/fetch) via the participant harnesses.
Sanitization: There is no significant sanitization or filtering of the natural language content provided by participants before it is added to the shared context.
[DATA_EXFILTRATION]: The Claude and Codex adapters (scripts/adapters/claude.py and scripts/adapters/codex.py) explicitly grant read access to user-level skill installation directories such as ~/.claude/skills and ~/.agents/skills. This design intended for cross-skill collaboration increases the data exposure surface, allowing agents to read other installed instructions and metadata.

panel