I spent 2 years in prompt hell. Then I found DSPy.
Now I write Python, not prompts. I have tests. I can swap models. Optimizers do the tuning.
I then wrote a guide on what I learned: https://t.co/vS14MGR9hu
i shipped a much needed fix to @Av1dlive's agentic-stack for the pi agent this was something that was really bothering me https://t.co/U5VL8WNbYf now agentic-stack works seamlessly with pi.
@Av1dlive while working with agentic-stack and pi coding agent i found a gap. I've documented it in an issue here https://t.co/7jceyaJac1 I will see if I can submit a PR this week
I am loving the philosophy behind @badlogicgames’s pi coding harness. I just made it build itself an extension to interact with linear and then help me groom some issues for an audio generation project I am working on. Beyond impressed.
@Av1dlive i noticed a few more gaps between the repo and the article. The ACTIVE_PLAN.md file in working memory, plus three skills (api-scaffold, test-writer, code-reviewer) that the article describes but the repo doesn't include. Is this on purpose?
found a bug in the claude code adapter. the PostToolUse hook calls memory_reflect.py claude-code post-tool ok with hardcoded strings, so every episodic entry has an empty reflection and the dream cycle never stages any candidates regardless of how many sessions you run. wrote a proper hook that reads the JSON stdin claude code actually sends, fixed the salience math too. PR coming shortly
Starting Monday with @badlogicgames “slow the fuck down” and totally get where he is coming from. We need to start using smarter harnesses and stop churning out slop.