The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes report summaries and titles that originate from external integrations (GitHub, Linear, Zendesk, etc.), which could potentially contain malicious instructions aimed at the agent.
Ingestion points: Data enters the agent's context through tools like inbox-reports-list and inbox-reports-retrieve in SKILL.md.
Boundary markers: The instructions explicitly warn the agent that reports are 'LLM diagnoses' and 'not ground truth,' requiring mandatory verification against source code in references/implementation.md before any fix is implemented.
Capability inventory: The skill enables high-privilege operations such as claiming reports, attaching pull requests, and triggering background agent actions (e.g., 'Get it merged') as seen in references/report-actions.md.
Sanitization: There is a robust automated mitigation where a 'safety judge' flags manipulative content. The agent is strictly instructed in SKILL.md to 'skip the report' and 'do not retrieve it' if the suppression_source is safety_judge.
[PROMPT_INJECTION]: A heuristic detection flagged potential concealment of actions in the instructions regarding held-back reports. Analysis confirms this is a benign UX guideline intended to ensure the agent correctly distinguishes between user-initiated dismissals and system-initiated safety suppression, preventing the agent from misleading the user about the origin of report state changes.

inbox-exploration