The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill establishes an orchestration layer where a primary agent delegates tasks to sub-agents via the cursor-agent CLI. This architecture grants sub-agents the ability to execute shell commands and modify files autonomously (using the --force flag), creating a potential surface for indirect prompt injection if the agents process untrusted data from the codebase or external inputs.
Ingestion points: The cursor-worker-implement and cursor-worker-review sub-agent templates (found in assets/claude/skills/) ingest prompt instructions and task briefs generated by the orchestrating agent.
Boundary markers: The instructions recommend using "self-contained" briefs, but the skill does not implement strict technical delimiters or sanitization routines to isolate potentially malicious data within these briefs.
Capability inventory: Sub-agents are configured to use the cursor-agent tool, which possesses broad permissions for file system modification and shell command execution.
Sanitization: The architecture relies on the parent agent's reasoning and a separate "verifier" agent to detect anomalies, rather than using programmatic sanitization of the inputs passed to the CLI.

agent-orchestration