WHAT NEW IN X10.29 THU

AI collections from X · 22 collections · 58,925 posts

jevVIDEO

Claude Code /goal loop: Opus edits, Fable evaluates, Jev routes forks

An unattended /goal setup where Opus 5.5 plans and edits, Sonnet 5.5 dispatches worker/explorer/verifier subagents, Fable 5.1 judges each turn, and Jev resolves 1,700+ micro-forks in 16ms.

Grirk
@GrirkuArchived 3 days agoOriginally posted 2026.10.25
Grirk @GrirkuOriginally posted 2026.10.25
Open original on X

STOP BABYSITTING F*CKING EVERY TURN Anthropic just dropped the cleanest /goal loop in Claude Code. Opus 5.5 runs in auto mode and plans + edits completely unattended. Fable 5.1 sits as the evaluator and only wakes up after each turn to check if the goal is actually met. Sonnet 5.5 dispatches three subagents in the background: → worker reads the test harness and patches → explorer maps token call sites and AST → verifier runs pytest and lint JEV sits underneath and resolves 1,700+ micro-forks in under 16ms (which file, which tool, retry or pause) before anything reaches the main session. Fable never writes code. It just reads the full transcript after every turn and says “token signature expired” or “CORS header missing Origin.” Opus applies the guidance on the next turn. 18/18 tests pass. Lint clean. Goal achieved. You launch it like this: claude --auto-mode "/goal all tests in test/auth pass and lint is clean" The full tree > Opus 5.5 auto mode plans + edits unattended > JEV 16ms fork layer resolves deterministic micro-forks > Sonnet 5.5 subagent dispatcher runs worker / explorer / verifier > Fable 5.1 evaluator checks after every turn > proves the condition in the conversation transcript > zero tool calls made, Fable never edits code Paste the tree + this prompt into Claude Code ↓ Rebuild my Claude Code setup around this tree: Run the main session on Opus 5.5 in auto mode with effort high. Use /goal to remove per-turn prompts. Dispatch background work to Sonnet 5.5 subagents (worker, explorer, verifier). Route deterministic forks (which file, which tool, retry or pause) through JEV at 16ms. Keep Fable 5.1 as the evaluator. It checks after each turn, reads the session transcript, and gives the reason for the next turn. Fable never edits code. It only evaluates whether acceptance criteria hold. Stop when the goal is met or an unachievable goal is caught. Show me every change as a diff first. No edits until I say go. save it, it’ll come in handy.

This post quotes another post: x.com/mirku21/status/2098836008468992471

x.com/grirku/status/2114092693768876280Archived 2026.10.25 10:04