The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill's primary function involves monitoring and capturing output from other agents (sub-agents) using tools like rig capture, rig transcript, and rig queue. This ingestion of untrusted content from the fleet creates a surface for indirect prompt injection, where a compromised agent could influence the orchestrator's behavior.
Ingestion points: rig capture <session>, rig transcript, rig queue, and rig ps --nodes --json.
Boundary markers: None specified for isolating captured agent output or queue payloads from the orchestrator's instruction space.
Capability inventory: The orchestrator has access to high-impact commands such as npm publish, git push --force, rig down, and rig fork. These capabilities are powerful tools that could be abused if the orchestrator is successfully manipulated.
Sanitization: There is no evidence of sanitization, filtering, or validation of the captured transcripts or queue metadata before they are processed by the LLM.

orchestration-team