The Agent Skills Directory

[PRIVILEGE_ESCALATION]: The skill requires root privileges to function, as documented in SKILL.md and TASKS.md. It explicitly instructs the agent to use sudo for execution (sudo doca_telemetry_traceback) to access hardware-level telemetry data on BlueField DPUs or ConnectX NICs.
[DYNAMIC_EXECUTION]: The skill involves runtime compilation of source code. TASKS.md provides instructions for using meson and ninja to build the application from source located at /opt/mellanox/doca/applications/telemetry/traceback/ before execution.
[INDIRECT_PROMPT_INJECTION]: The skill has an ingestion surface for untrusted data via user-defined JSON configuration files (rules.json).
Ingestion points: The application reads parameters like data-ids and rules from a JSON file specified by the -r or --rules flag (documented in TASKS.md).
Boundary markers: The instructions in TASKS.md ## configure step 7 advise the agent to pre-validate inputs (e.g., data_id format, comparison_type values), but explicit LLM-level boundary markers for the data content are not defined.
Capability inventory: The skill executes shell commands with sudo and PCIe address parameters.
Sanitization: Validation steps are mentioned for data formats, but not for preventing malicious instructions embedded in friendly name strings or other JSON fields.

doca-telemetry-traceback-application