The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill features an 'Outer Loop' designed for cross-session learning. It utilizes a harness-updater tool to modify the agent's primary instruction files (e.g., CLAUDE.md, AGENTS.md) by promoting patterns identified during task execution. While the skill documentation states these writes are human-gated, the capability to modify the agent's governing rules based on dynamic data represents a significant risk vector.
[INDIRECT_PROMPT_INJECTION]: The orchestrator is designed to ingest and process arbitrary task descriptions, issue lists, and specifications from external sources, which constitutes a large attack surface for indirect prompt injection.
Ingestion points: Task classification logic in SKILL.md and references/classification-rules.md processes untrusted descriptions to determine pipeline routing.
Boundary markers: The skill uses 'Intent Frames' produced by intent-framed-agent to restate and confirm the task plan, which provides a layer of human-in-the-loop validation to mitigate unintended command execution.
Capability inventory: The pipeline dispatches tasks to skills with high-impact capabilities, including automated code patching (self-healing), test execution (verify-gate), and instruction file updates (harness-updater).
Sanitization: Sanitization is delegated to the simplify-and-harden skill, which is explicitly tasked with scanning for injection vectors and other security vulnerabilities in the code produced or modified.

skill-pipeline