The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructs the agent to probe the local system's PATH for specific binaries including claude, codex, cursor-agent, gemini, grok, and opencode and execute them with shell commands.
Evidence: Found in SKILL.md under the 'Procedure' section: 'Probe available CLIs on PATH... Invoke all members concurrently in background processes.'
Evidence: Specific execution flags are provided to bypass security prompts, such as cursor --trust and gemini --skip-trust.
[DYNAMIC_EXECUTION]: The deliberation loop involves managing concurrent background processes and utilizing system prompt overrides to control the behavior of headless agent instances.
Evidence: SKILL.md mentions: 'Invoke all members concurrently in background processes with a ~300s timeout' and 'Pass system prompt overrides to headless models (e.g. Grok)'.
[INDIRECT_PROMPT_INJECTION]: The skill processes and synthesizes architectural plans based on outputs received from other agents, creating a multi-step chain where a malicious or compromised agent response could influence the final synthesized plan.
Ingestion points: Member agent responses received from CLI executions (SKILL.md).
Boundary markers: The instructions mention parsing and collecting responses but do not specify robust delimiters or instructions to ignore embedded malicious payloads within member critiques.
Capability inventory: The skill has the capability to execute arbitrary shell commands for tool invocation (SKILL.md).
Sanitization: There is no explicit requirement for sanitizing or escaping the output of council members before it is synthesized into the final counciled_plan.

fable-council