@lucasmeijer This can be used for context isolation. E.g., you want more context/skills/agents in subfolders, that shouldn't leak to the parent - then you start pi in parent.
Why RTK broke 2/5 runs: it installs as a pre-Bash hook. The hook itself crashed, and Claude blocks Bash when the pre-Bash hook fails. The agent diagnosed it and tried to disable the hook in settings โ the bench refused the edit. Gave up: 8 bugs unfixed, tests 51/59.
I tested 8 tools that claim to cut Claude Code costs.
Two had to be excluded โ one modifies your config without asking, one routes prompts through a third party. And the most hyped tools? They didn't beat running Claude normally.
Here's what I found โ thread ๐
Bottom line.
โ Use these together (~24% savings):
- token-optimizer-mcp โ https://t.co/dBBmwcWooh
- code-review-graph โ https://t.co/fitU8Q6wVL
โ ๏ธ Skip the rest (see thread).
@davis7@davis7 Ralph loop is not what you said in this video. What you said is just parallel subagents or simple orchestration. Ralph loop is to run LLM on the same prompt multiple times until convergence to solve the issue of LLM not finishing the task, but calling out that it finished
Top-tier product marketing from @AnthropicAI:
1. Fork your current version.
2. Nerf your current version on main.
3. Release the fork as your next improved version.
4. Profit.