The Agent Skills Directory

[SAFE]: The skill's architecture is built on a clear policy that separates untrusted data from executable logic. No malicious code, persistence mechanisms, or obfuscation techniques were identified within the skill's bundled markdown or references.\n- [PROMPT_INJECTION]: The skill facilitates a surface for indirect prompt injection by processing external work items. However, it implements comprehensive defenses described in its evidence chain:\n
Ingestion points: Tracker content enters the context via ITEM_URL or ITEM_CONTEXT as defined in SKILL.md and subagents/refinement-reviewer.md.\n
Boundary markers: The skill relies on an explicit 'Untrusted Content Rule' in references/reviewer-policy.md which mandates treating tracker content strictly as data and recording any embedded imperative instructions as 'Injection notes' rather than executing them.\n
Capability inventory: The skill uses resolved tooling for read/write operations (MCP, CLI, API, fetch) but restricts mutations to a single posted comment, prohibiting metadata or hierarchy edits.\n
Sanitization: A human review preview gate is mandatory for all postings, and a validation checklist (references/review-quality-checklist.md) ensures compliance with safety policies before output production.\n- [EXTERNAL_DOWNLOADS]: The references/external-sources.md file contains links to official documentation for Jira, GitHub, OWASP, Anthropic, and AWS. These URLs target well-known services and organizations, and are used solely for grounding technical analysis and verifying implementation claims.\n- [COMMAND_EXECUTION]: The skill resolves local tooling for interaction with Jira and GitHub (such as platform CLIs). This usage is restricted to the intended purpose of the skill and follows a strict tooling resolution order specified in the coordinator's state machine.

refine-task