The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill instructions create a vulnerability surface by automatically integrating external data into the agent's primary reasoning context.
Ingestion points: In SKILL.md, the agent is directed to read the file foyer.json (if present) to automatically obtain the user's fiscal situation and context.
Boundary markers: Absent. There are no instructions for the agent to treat the contents of foyer.json as data only, nor are there delimiters to separate this data from the system prompt or instructions to ignore potential natural language commands within the JSON values.
Capability inventory: The skill possesses significant capabilities, including the ability to execute local Python scripts (calc_ir.py) to process the ingested data and a script (dgfip_oracle.py) that can perform network operations to an external government simulator.
Sanitization: The input is parsed using the standard Python json library without filtering or escaping content that might contain prompt injection payloads designed to manipulate the agent's behavior.
[EXTERNAL_DOWNLOADS]: The script scripts/dgfip_oracle.py interacts with an external official service to verify tax calculations.
Evidence: The script performs network requests to the official French tax simulator at simulateur-ir-ifi.impots.gouv.fr. This domain belongs to the French Ministry of Finance, a well-known public service. This finding is documented neutrally as a safe interaction with a well-known service.

fiscaliste