Watch the first minute of this video as an introduction. This extension adds 1 tool for AI agent harness to call, named neuralyzer (no arguments). When the tool is called, all user and assistant messages in the session context are wiped, and a copy of the first message is sent again. Example:
USER: Hi, how are you?
ASSISTANT: Good. How can I help?
USER: Call neuralyzer tool
🕶️✨ Neuralyzer has flashed.
USER: Hi, how are you? [sent automatically]
ASSISTANT: Ready to help!
USER: Was neuralyzer tool used in this conversation?
ASSISTANT: No, never used.
Easier and more ergonomic loop engineering. A traditional Ralph loop is basically running this command in your shell: while :; do cat PROMPT.md | pi -p ; done, but then you have to save the prompt to a file, handle loop exit conditions, or adapt your workflow to whatever a third-party tool or extension demands. The loop controller lives outside the agent. This tool gives it back to the agent. You can just send the agent a message with control flow like this:
Check if @john has submitted a GitHub PR in this repo fixing authentication bug.
If yes -> add GitHub comment to that PR saying "Thank you".
If no -> wait 5 min and call neuralyzer.
/loop keeps adding to your session's context window, causing context rot and increased session continuation cost due to more tokens being in it, whereas neuralyzer gives the agent a fresh start and makes a loop setup super easy.
Pick your harness:
pi
pi install npm:@gintasz/pi-neuralyzerOpenCode
opencode plugin @gintasz/opencode-neuralyzerClaude Code
No. As of June 20, 2026, Claude Code exposes no extension surface (hooks, skills, MCP, agents) that can wipe, fork, rewind, or rewrite session context mid-session. Fork (--fork-session) and /rewind exist, but they are startup-flag / interactive-human only — an agent cannot neuralyze itself.
Don't see your harness? Adding one is the most welcome kind of PR — see CONTRIBUTING.