The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes the Bash tool to execute standard CLI utilities for configuration parsing and repository context discovery.
Evidence: Step 0.2 uses jq with the --arg flag to safely handle the TOPIC_SLUG variable, preventing command injection during JSON parsing. Step 0.3 uses git remote and gh repo view to determine the target GitHub repository.
[PROMPT_INJECTION]: The skill interpolates user-provided arguments into the prompt of a spawned sub-agent.
Evidence: Step 1.1 wraps the {repo} and {labels} variables in <user_input> tags within the instructions for the issue-architect agent.
Mitigation: Step 0.1 includes explicit sanitization logic that truncates arguments to 200 characters, strips backticks/angle brackets, and enforces a strict regex format [a-zA-Z0-9._-]+/[a-zA-Z0-9._-]+ for repository names, rejecting shell metacharacters and spaces.
[INDIRECT_PROMPT_INJECTION]: The skill processes research findings from local JSON files which could potentially contain instructions embedded in the analyzed data.
Ingestion points: reports/*/state.json specifically the topic and elicitation fields in SKILL.md (Step 0.2).
Boundary markers: Uses <user_input> tags in the sub-agent prompt to delimit variable content.
Capability inventory: The sub-agent (issue-architect) is tasked with creating GitHub issues based on the ingested data.
Sanitization: While metadata like repository names is sanitized, the research content itself is passed to the sub-agent without explicit filtering of potential instructions.

issues