This repo contains agent skills and scripts that can influence how coding agents read, write, verify, and sync local workspaces. Treat prompt-injection, unsafe shell behavior, and supply-chain drift as security issues.

Please report:

• prompt-injection directives inside bundled skills,
• descriptor-declared MCP/tools/hooks/commands/agents/auth/install surfaces that are hidden, misleading, or insufficiently reviewed,
• fetch-and-execute patterns,
• dynamic PowerShell execution,
• credential-shaped secrets in skills, docs, fixtures, or logs,
• converter path traversal or writes outside the requested output root,
• sync behavior that overwrites unexpected locations,
• false-success exits for blocked or unsupported conversions,
• platform adapter claims backed by unofficial or unsafe sources.

Run:

& .\scripts\scan_skills.ps1
& .\scripts\harness_test.ps1
& .\skills\stack-setup-audit\scripts\self_test.ps1 -Path (Get-Location)
& .\skills\stack-setup-audit\scripts\fixture_test.ps1
& .\scripts\sync_skills.ps1

The scanner rejects direct prompt-injection phrases, dynamic PowerShell execution, fetch-and-execute pipelines, and credential-shaped secret literals in bundled skills, descriptors, and docs. It also emits warnings for structured descriptors that declare tool, permission, auth, install, telemetry, or runtime surfaces requiring source review.

There is no bounty program. If GitHub private security advisories are available for this repository, use them. Otherwise, open a minimal public issue that does not include live credentials, private paths, or exploit details beyond what is needed to reproduce safely.