The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructions direct the agent to execute several shell scripts for installation, verification, and runtime inspection.
Evidence: It uses plugins/sol-advisor/scripts/install-agents.sh to modify agent profiles and inspect-agent-runtime.sh to query metadata from other session thread IDs.
These scripts are located outside the skill's own directory, suggesting a dependency on an external repository structure.
[PROMPT_INJECTION]: The skill incorporates an indirect prompt injection surface by delegating tasks to sub-agents and processing their 'Implementation Reports' as authoritative claims.
Evidence: The workflow in role-contracts.md involves the primary session reading and acting upon structured reports from sub-agents, which could contain malicious instructions or misleading diffs.
Ingestion point: references/role-contracts.md (Implementation reports)
Boundary markers: Structured implementation report format (STATUS, OBJECTIVE, CHANGES, etc.)
Capability inventory: Shell script execution (sh ...), git commands, agent spawning (agent_type)
Sanitization: Parent session is instructed to inspect diffs and rerun verification, but the initial evaluation is based on sub-agent output.
[PROMPT_INJECTION]: The skill uses non-standard model identifiers and elaborate role-playing constraints to manage agent transitions.
Evidence: The instructions refer to 'GPT-5.6 Sol', 'GPT-5.6 Luna', and 'GPT-5.6 Terra', which are not standard model versions. This naming convention and the associated 'preflight' checks may conflict with the agent's actual system prompt or security guidelines.

orchestration