The Agent Skills Directory

[SAFE]: The skill defines a robust methodology for context reconstruction that prioritizes observable repository state (git, working tree, source code) over narrative summaries. This evidence-based approach promotes accuracy and reduces reliance on potentially manipulated or stale secondary sources.- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data including git history, PR evidence, and repository documentation (Ingestion points: SKILL.md, reality-check.md). While this represents a surface for indirect prompt injection, the skill explicitly mandates 'Reality Check' and 'Reconciliation' phases (Boundary markers/Sanitization: SKILL.md workflow step 1, reconcile.md) to verify claims against implementation ground truth. The agent uses its repository access tools to collect direct evidence (Capability inventory: evidence.md, reality-check.md), providing a significant defense-in-depth against malicious instructions embedded in project metadata or history.

continue