The Agent Skills Directory

[COMMAND_EXECUTION]: The script sfap_replay.py uses subprocess.run to invoke curl for API interactions. The command is constructed as a list of arguments and does not use shell=True, which is a secure practice to prevent shell injection. The resolve-token.sh script utilizes standard command-line tools such as sf, jq, and curl to manage authentication and interact with Salesforce infrastructure.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process historical conversation logs for model evaluation, creating a surface for indirect prompt injection. (1) Ingestion points: The skill reads JSON data containing user utterances and agent responses from /tmp/details_result.json and other temporary LLM step files. (2) Boundary markers: Replayed data is processed as structured JSON; no specific prompt delimiters or ignore instructions were observed for the replayed text. (3) Capability inventory: The skill has network access via curl to Salesforce AI Gateway endpoints and file-writing capabilities to generate CSV and HTML reports. (4) Sanitization: The skill uses json.dumps to sanitize input into the API payloads and uses a separate LLM-as-judge process to score the outcomes.
[SAFE]: Sensitive credentials are handled using secure patterns. The resolve-token.sh script writes session JWTs to temporary files with chmod 600 permissions, ensuring that access is restricted to the current user and not exposed to other system processes. The skill exclusively interacts with vendor-owned resources (Salesforce APIs).

agentforce-offline-experimentation-run