The Agent Skills Directory

[SAFE]: The skill implements a strict Mutation Limits Contract that restricts file modifications to specific task-scoped paths and prohibits changes to git history or unrelated implementation files, ensuring high integrity of the development environment.
[SAFE]: Data handling instructions in contracts.md explicitly direct the agent to treat external content such as tracker payloads, command output, and fetched pages as data and never as instructions, providing a strong defense against indirect prompt injection.
[SAFE]: The skill demonstrates high security awareness by defining a "Category P" for sensitive information (credentials, tokens, secret-bearing logs), strictly prohibiting their persistence or inclusion in git history.
[SAFE]: A dedicated security-auditor subagent is included to perform automated security reviews of the implemented code, specifically checking for hardcoded secrets, unsafe input handling, and insecure dependency usage.
[SAFE]: All external references in external-sources.md point to well-known technology documentation (Anthropic, GitHub, Atlassian, OWASP) or established technical blogs, following secure documentation practices.
[SAFE]: Indirect Prompt Injection analysis: The skill ingests untrusted data from Jira tickets and GitHub issues (Ingestion points). It employs explicit instructions to treat this data as non-executable (Boundary markers). It possesses capabilities for file writing and shell execution through CLI tools (Capability inventory). While it relies on the LLM's adherence to the data/instruction boundary rather than hard sanitization (Sanitization), the design is defensive and safe.

executing-work-item-task