The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill implements an "Evolutionary Prompt Optimization" (EPO) mechanism as detailed in guides/ai/self-evolving-skills.md. This loop uses an LLM to rewrite the agent's own system prompts, creating mutated variants ("genes") that are stored and subsequently executed based on an epsilon-greedy selection policy. This results in the autonomous modification of the skill's own operational instructions at runtime.
[INDIRECT_PROMPT_INJECTION]: During the "Context Loading" phase (SKILL.md, Phase B01), the orchestrator ingests data from episodic_log and knowledge-map. The skill's own documentation warns that this "Episodic Memory Poisoning" vector could allow prompt injection payloads from previous, potentially untrusted tasks to be re-introduced and executed with the skill's elevated SDLC permissions.
[DATA_EXFILTRATION]: The skill uses a cost-benefit-router to dynamically transition task execution between local models and "Remote SOTA" providers (such as Claude or GPT-4o). This process involves transmitting project-specific data and context across the network to external cloud services.
[COMMAND_EXECUTION]: The skill integrates with a local command-line interface (gabbe brain activate) to drive its autonomous cognitive loop and manage the prompt evolution lifecycle.
[PERSISTENCE]: The skill maintains a persistent state through a SQLite database containing the prompt "gene pool" and historical episode logs. This allows self-modified instructions and potentially poisoned memory states to survive across different sessions and project environments.

brain-mode