The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data which could contain malicious instructions designed to influence the agent's behavior.
Ingestion points: The skill ingests 'Findings' objects (containing natural language fields like body and evidence) and 'Artifact paths' representing external source code (SKILL.md).
Boundary markers: The Finding schema provides structural boundaries, but text fields within the schema lack explicit delimiters or 'ignore' instructions for the sub-agents.
Capability inventory: The skill utilizes git add, allows file writes through 'fixer' sub-agents, and uses AskUserQuestion for interactive decision-making (SKILL.md).
Sanitization: The skill implements a 'Staging safety' protocol (Steps 1-5) that records snapshots of the Git index and validates files_changed against approved paths. Any unauthorized changes or scope expansions require human approval via AskUserQuestion (SKILL.md).
[COMMAND_EXECUTION]: The protocol relies on the execution of Git commands to manage the state of the fix-verify loop. Specifically, it uses git diff, git show, git add, and git diff --staged --stat to validate, stage, and report changes (SKILL.md). These commands are limited to local repository management and are guided by strict safety rules provided in the sub-agent briefs.

fix-verify-loop