The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes data retrieved from external dLocal API endpoints, which creates a potential surface for instructions embedded in transaction records or merchant data.
Ingestion points: Untrusted external data is brought into the agent context via retrieval commands such as agent-dlocal payments get, orders get, and the investigate subcommands as described in SKILL.md and references/investigation/ files.
Boundary markers: The skill relies on the agent-dlocal CLI tool which uses NDJSON formatting and a default redaction policy to help maintain separation between data and instructions.
Capability inventory: The agent has the ability to execute agent-dlocal commands using the Bash tool and perform local file operations via Read, Grep, and Glob.
Sanitization: The tool automatically masks sensitive PII, including payer and beneficiary blocks (name, email, national document ID, bank accounts), as [REDACTED] by default to prevent data leakage and injection risks.
[COMMAND_EXECUTION]: The skill functions by wrapping a dedicated CLI tool, agent-dlocal, which is executed through the Bash tool. These commands are restricted to read-only operations for payment triage, ensuring the agent cannot perform unauthorized financial transfers or modifications.

agent-dlocal