What problem does it solve? LLM answers often hallucinate facts, ignore project-specific ground truth, and rely on single-pass generation. This Skill enforces code-level intelligence amplification by wiring shell hooks into every LLM call so responses are grounded in DAG entities and verified facts rather than prompt-level suggestions the model can ignore. ## Core Features & Use Cases - Hook-based enforcement: Four shell hooks (pre_llm_call, transform_llm_output, post_tool_call, on_session_end) automatically inject DAG entities and stored facts into context, then critique and refine significant responses with real API calls. - Real reasoning pipeline: self-refine.py (generate, critique, DAG-grounded refine) and multi-perspective.py (3 parallel expert calls plus judge synthesis) make actual API calls instead of emitting prompt templates. - Persistent fact memory: fact-store.py saves verified facts to disk and recalls them by keyword across sessions, while context-injector.py queries project .dag files for relevant entities before generation. - Use Case: When answering architecture questions about a codebase, the pipeline loads matching DAG entities and stored facts, generates an answer, then runs a critique-and-refine pass that can only cite entities present in the DAG, flagging anything else as unverified. ## Quick Start Enable the four intelligence hooks in your config.yaml with hooks_auto_accept set to true, then ask the agent a hard technical question so the pipeline grounds and refines the answer automatically.
