The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill implements a fan-out and aggregate pattern where an orchestrator agent processes results from multiple parallel workers, creating a surface for indirect prompt injection.
Ingestion points: In Phase C (Aggregate), the orchestrator reads terminal results produced by workers via the delegate_task tool.
Boundary markers: The instructions advise the agent to "not paste raw worker dumps" and to create a "consolidated in-chat report," but there are no explicit prompt delimiters or "ignore embedded instructions" warnings for the data being aggregated.
Capability inventory: The orchestrator and its workers possess significant capabilities, including file system access, browser interaction, and access to local authentication credentials.
Sanitization: No explicit sanitization, escaping, or validation of the worker-provided content is required before the aggregator processes it.
[COMMAND_EXECUTION]: The skill uses local Python scripts to manage roles and validate tasks.
Evidence: The skill executes python3 <pstack-runtime>/scripts/roles.py in Phase A and Phase B to resolve agent roles and check task briefs.

swarm