The Agent Skills Directory

Indirect Prompt Injection Processing: The skill is designed to accept text data from various sources (such as files, globs, or structured tasks) and interpolate those values directly into prompts executed by subagents. When processing data from external or untrusted origins, this pattern presents an indirect prompt injection surface where adversarial text within the inputs could attempt to influence subagent behavior.
Ingestion points: External data enters the pipeline in scripts/table.ts through the createTable function, which processes glob patterns, file lists, or task objects.
Boundary markers: Prompts are constructed structurally in scripts/batching.ts using markdown headers (# Task, # Items) and structural brackets around item IDs, which helps separate data components but does not explicitly instruct the model to disregard potential embedded commands.
Capability inventory: The skill dispatches instructions to subagents using the ambient platform tool tools.swarmTask for parallel execution loops.
Sanitization: Values are formatted or JSON-serialized via formatValue in scripts/batching.ts without semantic sanitization for prompt override sequences.

swarm