The most useful part is the session recording. Instead of just seeing the final diff, you get the full trace: the initial prompt, the agent's internal reasoning process, and the resulting action. It makes debugging "hallucinated" architecture much easier.
To get this running with your Grok agent, use this command:
npx decispher@latest init --agents grok
My current AI workflow #
I've been testing a few different LLM agents for codebase management, and the biggest hurdle is always the "memory" of the project's rules. By plugging Grok into a decision-tracking layer, you stop repeating the same prompt instructions every time you start a new session.
This setup is essentially a practical tutorial in creating a persistent memory for your AI coding tools. If you are already using Cursor or Claude Code, adding this layer to Grok helps maintain a consistent technical direction across a team rather than letting each dev's agent drift in different directions.
Next Qwen 27B vs Claude: Testing Local Tool Use →