The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill uses an 'Evolutionary Prompt Optimization' (EPO) loop described in 'guides/ai/self-evolving-skills.md' to rewrite its own instructions. An LLM acts as an expert prompt engineer to generate new variants ('genes') which are then stored in a database. The system uses an epsilon-greedy policy to select these variants, where 80% of the time it chooses the variant with the highest 'success_rate'. Crucially, the reward signal for success is monotonic and only requires that the LLM call returned any response, meaning malicious or subverted prompt variants can accumulate fitness and be consistently used without human oversight.
[INDIRECT_PROMPT_INJECTION]: The 'cost-benefit-router' logic in 'SKILL.md' processes untrusted task descriptions and file lists. 1. Ingestion points: The 'task' and 'files' input fields provided by users. 2. Boundary markers: None explicitly implemented; the skill relies on conceptual regex or DLP scanners. 3. Capability inventory: The skill has the authority to route data to high-capability 'REMOTE SOTA' models and utilizes the 'gabbe route' CLI. 4. Sanitization: The logic depends on 'Context Analysis' rules that are described as conceptual rather than strictly enforced code, creating a surface for injection to influence routing decisions.
[DATA_EXFILTRATION]: The skill's 'Self-Correction' and escalation logic (Local -> Remote Mini -> Remote SOTA) creates a documented risk of data leakage. If a task involving sensitive or proprietary data fails to be processed by a local model, the router is designed to escalate the task to remote cloud providers. While the documentation acknowledges this risk and suggests re-running privacy checks, the underlying architecture provides a direct path for sensitive data to leave the local environment upon a processing failure.

cost-benefit-router