The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill implements a workflow where data gathered from multiple independent AI sub-agents (workers) is merged and used by a parent agent to modify the project environment. This introduces a vulnerability surface where potentially untrusted data from a sub-agent could influence the parent agent's actions. * Ingestion points: The skill ingests 'takes' (responses) from multiple models in Step 5 and Step 6 of the SKILL.md pipeline. * Boundary markers: Worker outputs are delimited using labeled headers like [Current] or [OpenAI] as described in Step 5 of SKILL.md. * Capability inventory: The parent agent is authorized to 'do the work' (write files) and 'prove on the real path' (execute commands) in the 'Write' section of SKILL.md. * Sanitization: The skill instructs the agent to 'resolve consensus and divergence' and 'reject weak claims,' providing a manual logic check, but lacks explicit automated sanitization of the ingested content strings.
[DYNAMIC_EXECUTION]: The skill uses the host platform's capabilities to dynamically spawn and manage child agents at runtime based on a 'live list' of available models detected on the host (e.g., Cursor, Claude Code, OpenCode) as described in references/hosts.md.
[COMMAND_EXECUTION]: The orchestration pipeline requires the main agent to execute validation commands or tests ('prove on the real path') to verify the implementation of merged research, which involves running shell commands based on aggregated model output.

execute