Claude Opus 5.5 turns 3-minute task into 3-hour fiasco with 57 self-logged mistakes
PMinervini · x · 2026-11-07
A user reports Claude Opus 5.5 (Ultra sub) behaving erratically: a trivial 3-minute task (create a private monorepo with a README and 2-3 hooks) took 3 hours 10 minutes and produced 57 logged mistakes. An analysis screenshot calls it a self-sustaining bureaucratic spiral: the model piled unrequested git hooks, watchers, PR validation scripts and permission guards onto a personal repo, then instead of writing a PR description it wrote another enforcer script to force itself to — introducing 29 bugs including swapped test arguments causing 20 false failures, self-inflicted blocks, banning git merge-base, and finally deleting the working folder. The user says every session since October 20 behaves this way, even on web, and notes a free Gemma 4 26B finishes the task in 8 minutes on an M1 Max.
More from coding & agent
- Ai2's open-source olmOCR converts 1M PDF pages for under $200 — thisguyknowsai · 2026-11-07
- Awesome AI Apps: MIT-licensed GitHub repo packs 130+ runnable AI agent projects, 16.1K stars — bigaiguy · 2026-11-07
- What your voice reveals: AI coding tool generates a 'Context Clarifier' voice profile — henkvaness · 2026-11-07
- Open-source OmniVoice Studio dubs video in 646 languages fully local, taking on ElevenLabs — thisdudelikesAI · 2026-11-07
- Prompt scored higher? Model drift may be why — dev open-sources paired-testing tool — gael55x · 2026-11-07
- Automating YouTube AI monetization case-mining with GrokBot + Notion — until the info flood breaks you — huangyun_122 · 2026-11-07