WHAT NEW IN X10.30 FRI

AI collections from X · 22 collections · 60,637 posts

jevVIDEO

Claude Code pre-merge review pipeline and Jev-style decision forking

Describes an /ultrareview multi-model review flow and how Jev resolves no-thinker forks in code so frontier models only see decisions that split.

mirku
@mirku21Archived 8 days agoOriginally posted 2026.10.21
mirku @mirku21Originally posted 2026.10.21
Open original on X

THIS IS AN ABSOLUTE F*CKING CHEATCODE Anthropic just documented the ultimate pre-merge engine for Claude Code The idea: run /ultrareview before pushing to main, and let the new Haiku 5.5 swarm the diff with 16 parallel agents while Opus 5.5 signs off Zero blind spots, zero style nits, and zero Opus context burned on routine file scans You just launch it in Claude Code like this: /ultrareview --post here's how it splits: → Opus 5.5 orchestrates the review workflow: drafts the inspection contract, filters false positives, and signs off on findings → Sonnet 5.5 proves critical bugs: writes isolated reproduction tests in a sandbox and drafts precision patches → Haiku 5.5 swarms the entire repository diff in parallel: audits 16 files simultaneously at 340 tok/s with native 1M context → you stop burning Opus context on routine files and keep your local session completely free The cloud fleet only reports findings through three proven gates: → cross-module inspection: scans taint flows and broken invariants before PR creation → reproduction proof: an independent verifier writes a failing test proving the bug exists → pre-merge signoff: zero style nits, every verified bug reproduced in bugs.json Opus orchestrates, Haiku swarms, Sonnet proves JEV engineering is the same move one layer down: the forks that need no thinker (which file, which tool, retry or stop) resolve in code in 15ms, and the frontier models only see the ones that split the full tree > Opus 5.5 orchestrates the review workflow (cloud sandbox master) > Haiku 5.5 swarms 16 parallel scan agents across repo diffs (340 tok/s, 1M context) > Sonnet 5.5 writes isolated reproduction tests to verify edge cases > Sonnet 5.5 drafts precision patches for confirmed bugs > bugs.json automatically stamps verified PR comments via --post > Opus 5.5 signs off on the unified report with zero false positives > JEV routes micro-forks in milliseconds paste the tree and this prompt into Claude Code ↓ "Reconfigure my Claude Code pre-merge workflow around this tree: 1. Set up /ultrareview on Anthropic cloud infrastructure for pre-PR audits 2. Delegate parallel diff scanning to the new Haiku 5.5 across 16 isolated agents at 340 tok/s 3. Route deep semantic verification and reproduction tests to Sonnet 5.5 4. Reserve Opus 5.5 for orchestration and final sign-off with zero false positives 5. Add one rule to CLAUDE.md: never merge to main until /ultrareview proves 0 regressions and reproduction passes Show me every change as a diff first. No edits until I say go" ↳ http://code.claude.com/docs/en/ultrareview save it, it'll come in handy

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

x.com/mirku21/status/2112650129195061496Archived 2026.10.21 10:04