The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted local data (source code and documentation) to perform architectural mapping and audits (e.g., in agents/analyzer.md and agents/auditor.md). This creates a theoretical surface for indirect prompt injection.
Ingestion points: Source files and documentation are scanned using grep, find, and ls tools in agents/analyzer.md and agents/auditor.md.
Boundary markers: The skill includes explicit "Review boundary" instructions across its components (e.g., SKILL.md, agents/auditor.md) advising the agent to treat repository content as data, not as authority to override user rules.
Capability inventory: The skill uses shell commands for discovery and writes numerous template files to the local file system.
Sanitization: The skill uses structured JSON for analysis results (architecture.json, audit.json) but interpolates analyzed data into Markdown templates.
[DYNAMIC_EXECUTION]: The skill generates and writes various management scripts (Bash, PowerShell, Python, Node.js) such as scripts/harness-change.ps1 and scripts/lint-deps.mjs. These are intended for local execution to manage the development lifecycle.
[COMMAND_EXECUTION]: The skill instructions utilize shell commands (e.g., go list, grep, find) for architectural discovery and metadata extraction, which is a standard part of its design for developers.
[SAFE]: The skill references well-known and trusted external resources in its templates, such as official GitHub Actions (actions/checkout, actions/setup-go) and official Docker images from known entities (golang, node, python, postgres, redis). It also includes specific security checks, such as lint-encoding.ps1, to detect corrupted text markers in project files.

ecl-harness-engineer