The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes data from external vendor 'twins' (sandboxes), including request/response traces, manuals, and stored state. This creates an indirect prompt injection surface where a malicious twin could return content designed to influence the agent's behavior.
Ingestion points: The agent is instructed to read veris sandbox trace output, veris sandbox data get results, and twin manual/schema data during investigation.
Boundary markers: Instructions specify filtering traces by 'tiers' (e.g., handler, fault) to isolate application traffic from control traffic.
Capability inventory: The skill allows for command execution (veris, npm, shell scripts), management of local files, and interaction with sandbox provider APIs.
Sanitization: scripts/ledger.sh includes a secret_scan function to detect and prevent the recording of credentials in audit logs.
[COMMAND_EXECUTION]: The skill includes shell scripts (record.sh, ledger.sh) and instructions to execute various commands for testing, setup, and auditing. These commands are part of the skill's primary purpose of providing an integration testing framework.
Evidence: scripts/record.sh executes user-supplied commands passed via arguments and captures their output to create a persistent record of the test flow.
[EXTERNAL_DOWNLOADS]: The documentation instructs the agent to install Node.js packages from the vendor's own scope (@veris-ai) and well-known sandbox services (@daytona, @e2b) via npm install to set up the test environment.
Evidence: daytona.md and e2b.md provide shell commands to install @veris-ai/daytona, @daytona/sdk, and @veris-ai/e2b for managing remote test boxes.

veris-reference