The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external diagnostic data (evidence) which represents a potential attack surface for indirect prompt injection.
Ingestion points: The Troubleshooter role collects data from the system/environment to build evidence tables as described in references/troubleshooter.md.
Boundary markers: The framework uses a mandatory 'authority profile' to delimit what data can be accessed. The Interrogator role specifically audits these boundaries as defined in references/interrogator.md.
Capability inventory: The skill can execute 'probes' for evidence collection but contains strict 'Non-Negotiables' in SKILL.md that prevent remediation or execution of fixes.
Sanitization: The Interrogator context acts as a logic-based audit layer, validating that every claim traces to numbered evidence and checking for authority violations before passing a diagnosis.
[COMMAND_EXECUTION]: The framework allows for running 'probes' to collect evidence, which involves executing system commands or tool calls.
Execution is restricted to diagnostic purposes only and is governed by the supplied authority profile provided by the wrapper.
The skill includes an explicit instruction: 'It never implements or executes remediation without a separate explicit user request,' ensuring that the command execution capability is not used for unauthorized modifications.
[SAFE]: The skill configuration in agents/openai.yaml explicitly disables implicit invocation, ensuring that the tool only runs when intentionally called by the system or user.

dig-deeper-core