The Agent Skills Directory

[PROMPT_INJECTION]: No evidence of instructions designed to bypass safety filters, extract system prompts, or override agent constraints. The 'Grill Me' mode is a requirement-drilling technique rather than a jailbreak attempt.
[DATA_EXFILTRATION]: The skill does not perform network operations or access sensitive credentials (e.g., SSH keys, AWS configs). It reads project files and documentation within the expected scope of a development tool.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from the codebase and user arguments. This risk is acknowledged and mitigated by:
Ingestion points: Codebase files (Stage 2) and user requirements ($ARGUMENTS).
Boundary markers: The process is governed by a rigid stage-based workflow.
Capability inventory: Uses file reading, agent spawning, and file writing tools.
Sanitization: Includes a dedicated 'Reviewer Agent' in Stage 5 specifically tasked with identifying injection vectors and security flaws.
[REMOTE_CODE_EXECUTION]: No remote scripts are downloaded or executed. The skill relies on local codebase exploration and internal agent prompts.
[CREDENTIALS_UNSAFE]: No hardcoded credentials were found. The skill proactively searches for exposed secrets in the modified code as a safety measure.
[COMMAND_EXECUTION]: Shell command execution is limited to legitimate development tasks (e.g., through the Agent tool prompts) and does not involve high-risk operations like 'sudo' or persistence modifications.

feature-dev