The Agent Skills Directory

[SAFE]: The skill implements a disciplined orchestration for Spec Compliance, Code Quality, and Architectural reviews. It includes robust convergence controls, such as a three-cycle limit to prevent infinite loops, and requires explicit human authorization for architectural modifications.
[INDIRECT_PROMPT_INJECTION]: The skill processes implementation packets provided by external sources, creating a potential surface for indirect prompt injection. This risk is effectively addressed through the following architectural safeguards:
Ingestion points: Untrusted implementation packets (plans, diffs, and evidence) enter the agent context as described in the 'Establish The Gate' section of SKILL.md.
Boundary markers: The 'Preserve Blind Review' protocol in SKILL.md isolates reviewers from previous findings and discussions, preventing instructions embedded in previous feedback from influencing new subagents.
Capability inventory: The coordinating agent's capabilities are limited to orchestration and 'subagent-dispatch'; it does not perform direct file system writes or network operations.
Sanitization: The 'Finding Lifecycle' protocol (references/finding-lifecycle.md) requires objective proof (e.g., test cases or runtime reproduction) to confirm a finding before any remediation is permitted, preventing the agent from blindly following instructions found in implementation data.

review-implementation