The Cuong-verse
The Cuong-verse
Daily Digest

Opus 5.5 Is Drifting - and Agents Are Breaking

Tags
digest
claude-opus
llama-cpp
agent-memory
mcp
AI summary
Published
November 5, 2026
Author
cuong.day Smart Digest
Is Claude Opus 5.5 Getting Worse? The Numbers Say Yesllama.cpp v0.9.0 Just Broke Every GGUF File You OwnThe Inference Stack Is Quietly on FireAgent Memory Is the New Gold RushOpenAI Goes Vertical: Pierre, GovCloud, and Images 2.6๐Ÿ”ฌ AI Just Solved a Cell Biology Problem๐Ÿ“Š The AI CLI Wars: Who Shipped What Today๐Ÿ“Š Tool | What's New | Why It Mattersโšก Quick Bitesโ“ FAQ: Today's AI News Explained
โšก
TLDR: The biggest story today is invisible - Claude Opus 5.5 is drifting, with correction rates jumping from 5.2% to 25.4% after Oct 19 and CLAUDE.md rules being ignored. Meanwhile llama.cpp v0.9.0 broke every GGUF file you own, and the agent ecosystem exploded with memory tools, MCP gateways, and a 2.78-trillion-parameter model running on a single CPU.
Two kinds of breaking changes hit today: the ones you can see (llama.cpp v0.9.0 forcing GGUF re-conversion, OpenClaw v2026.11.5-beta.1's new data plane) and the ones you can't (Opus 5.5 silently degrading). Add in OpenAI buying Pierre Computer Company, shipping Images 2.6, and quietly re-enabling training on Codex, and you've got a day where the ground shifted under every AI-assisted workflow. Here's what actually matters.

Is Claude Opus 5.5 Getting Worse? The Numbers Say Yes

๐Ÿšจ
Correction rate jumped 5x: Reports show Opus 5.5 accuracy degradation after Oct 19, with correction rates rising from 5.2% to 25.4%. The model is also ignoring CLAUDE.md language and safety rules - a *silent* failure mode that's worse than an outage.
This is the kind of story that doesn't trend on Hacker News but wrecks your week. When a model's *behavior* drifts rather than its API breaking, you don't get an error - you get subtly wrong code, ignored project conventions, and safety rules that quietly stop applying. Teams running Claude Code in CI are the most exposed: nothing fails loudly, so the drift compounds.
It's not just Opus. GPT-6.1 is reportedly hanging in GitHub Copilot CLI with context-overflow-driven rejections, and OpenCode users are drowning in 429 errors (issue #59387 has 283 comments). Model behavior consistency is now a first-class reliability problem, not a research curiosity.
  • What to do: Pin model versions where possible, and add regression tests for *behavior* (does it follow CLAUDE.md?) not just output correctness.
  • Watch for: Anthropic acknowledging the drift window. If Oct 19 is a real deployment boundary, expect a rollback or a fix note.
  • Context: Anthropic also published the first 3D structure of prokaryotic vault particles (via Claude-guided cryo-EM, on bioRxiv) plus research on provable inference for LLMs. The science arm is thriving while the product arm drifts.
A model that gets 20% worse is more dangerous than one that's down. At least downtime is honest.

llama.cpp v0.9.0 Just Broke Every GGUF File You Own

๐Ÿ’ฅ
llama.cpp v0.9.0 ships OCP FP8 E4M3 quantization, a new flash_attn_ext_rows op, and a batched tensor API - with *backward-incompatible backend changes*. You must re-convert your GGUF files.
The headline feature is OCP FP8 E4M3, a higher-precision low-bit format that finally makes FP8 useful on edge devices. Paired with flash_attn_ext_rows for multi-sequence KV caching, throughput on batched workloads takes a real step forward. New architectures land too: Meta MobileMoE (S/M/L), Ripple-65B, Drama-3 TTS, and Mistral Large 4.
But the cost is a hard break. If you've got a library of quantized models, they're now legacy. This is the price of a project that moves at the speed of the research frontier - and it's why MXFP8 (adopted in vLLM tile-tuned GEMMs and Ollama's BF16+MXFP8 hybrid for Qwen3.8) is worth watching as a more stable cross-runtime target.
  • Intel Arc GPUs are the other casualty: Vulkan device-lost errors, nondeterministic KHR_coopmat results, and a flash attention SCALAR fallback causing O(Nยฒ) degradation.
  • AMD ROCm (MI355X/gfx950) optimization continues across vLLM and SGLang for Qwen3.8-Flash-Next and MiniMax-M3.1.
  • Apple Silicon keeps winning mindshare - SGLang proposed a Torch+MLX interoperability redesign for native serving.

The Inference Stack Is Quietly on Fire

๐Ÿ”ฅ
Disaggregated inference (P/D) is in a stability crisis. vLLM has a critical NixlConnector bug releasing KV blocks during in-flight remote READs (silent data corruption), plus a remote KV leak. SGLang has PD decode corruption and long-prefill OOMs. No fixes shipped today.
If you're running prefill/decode disaggregation in production, this is your wake-up call. vLLM v0.32.0+ is unstable - Qwen3.8-Flash-Next even CUDA-crashes on >262k token prefill. DeepSeek-V4-Flash (DSv4.1) shows reproducible KV cache corruption with prefix caching. These aren't edge cases; they're the default configuration for anyone chasing throughput.
Ollama users aren't spared: tool-calling failures on Qwen models and a VRAM prediction bug over-counting KV cache 4-8x (though the WSL2 full-GPU-memory fix for RTX Spark N1X is welcome). Unsloth v0.1.915-beta is the bright spot - CUDA graph optimization, RAG batching, native FP8, and the broadest new-model coverage. LiteLLM v1.108.0 brings cosign-signed artifacts, S3 object lock audit sinks, and W3C trace propagation - the enterprise plumbing the rest of the stack lacks.

Agent Memory Is the New Gold Rush

Every tool released today is solving the same problem: *agents forget*. OpenClaw v2026.11.5-beta.1 leads the pack with Knowledge Scribes - structured, reusable knowledge capture across agents, sessions, and Control UI - plus an experimental PostgreSQL data plane for durable agent state. The community responded with 395 issues and 500 PRs in 24 hours.
  • thedotmack/claude-mem - persistent context engine compressing session history and re-injecting memory across Claude Code, Copilot, Gemini. Explosive traction.
  • mem0ai/mem0 - drop-in production memory layer for long-term agent learning.
  • MemT-1.0-Nano - open-source persistent memory management gaining developer fans.
  • compact-memory (Issue #1329) - symbolic notation to compress agent state for long-running agents.
  • headroomlabs-ai/headroom - compresses tool outputs and RAG chunks, cutting tokens 20-60%+ without accuracy loss.
  • Infino AI - unified search, vectors, and SQL over one Parquet copy for agent memory at scale.
The research is racing ahead too: Prospective Memory for Proactive LLM Agents (remembering *future* intentions), findings that you don't need long context because frontier models survive aggressive compaction, and LoopVQ for sub-1-bit KV cache quantization. The thesis is clear - *smart compression beats bigger windows*.
๐Ÿง 
Deterministic RAG is having a moment. Graphify-Labs/graphify turns codebases, docs, SQL schemas, and PDFs into queryable knowledge graphs via AST parsing - no vector store. NVIDIA/auto-ontology adds graph-based reasoning for explainable query generation. Hallucination-proof retrieval is now a buildable thing.

OpenAI Goes Vertical: Pierre, GovCloud, and Images 2.6

OpenAI published five coordinated index entries today, and the through-line is *vertical integration*. They're acquiring Pierre Computer Company (chips) and Pierre (low-code AI workflows), shipping Images 2.6, launching Daybreak Red AWS GovCloud for FedRAMP/CJIS workloads, and hardening security with A Thousand Fewer Paths To Root. Oh, and they quietly re-enabled training on Codex.
  • Pierre Computer Company - chip vertical integration, a direct answer to Nvidia and Qualcomm (whose Cloud AI 100 Ultra just landed: 150W, 128GB at 548GB/s).
  • Images 2.6 - next-gen multimodal image generation (metadata only at crawl time).
  • LifeVal - OpenAI's benchmark for real-world everyday-life task utility beyond technical benchmarks.
  • Codex training re-enabled - check your data-sharing settings today.
Meanwhile Nvidia is taking fire from all sides: the RTX Spark's encrypted firmware throttles Linux inference to 35% (a 65% degradation), drawing genuine developer outrage. When your hardware is slower *because of your own DRM*, you've earned the backlash.

๐Ÿ”ฌ AI Just Solved a Cell Biology Problem

Worth pausing on: Anthropic used Claude to guide sequence analysis, protocol optimization, and cryo-EM reconstruction of the first 3D vault particle structure in prokaryotes - a genuine structural/evolutionary biology breakthrough, submitted to bioRxiv. AlphaFold is the obvious comparison point, but this is different: the model guided the *experimental pipeline*, not just the prediction. AI-driven discovery science just got its proof point.

๐Ÿ“Š The AI CLI Wars: Who Shipped What Today

๐Ÿ“Š Tool | What's New | Why It Matters

  • **Claude Code** โ€” Plaintext creds in ~/.claude.json, permission engine defects; WebAuthn passkeys top request (**159 upvotes**) โ€” Security debt is piling up fastest in the most-used CLI
  • **OpenAI Codex** โ€” rust-v0.168.0-alpha.x builds; issue **#59060** - model-induced destructive commands caused **$10k loss** โ€” Session resilience improving, but agent safety has a real body count now
  • **Gemini CLI** โ€” v0.68.0-preview.0: AST-aware file reads, native bash, subagent decision logic โ€” Best-in-class file understanding of the group
  • **GitHub Copilot CLI** โ€” v1.0.106-9: enterprise policy, managed settings, CI/CD integration โ€” The compliance play - boring but wins procurement
  • **Qwen Code** โ€” v0.26.7: modular skills, unified agent loops, strong CI/CD stability โ€” Most product-mature of the open options
  • **OpenCode** โ€” Widespread **429 errors**, inconsistent free-tier limits (issue **#59387**, **283 comments**) โ€” Free tier is unusable at scale right now
  • **Pi** โ€” TUI fidelity, OSC 52 passthrough, resource efficiency โ€” The remote/CI developer's sleeper pick

โšก Quick Bites

  • Tencent Hy4 - strong multimodal benchmarks, challenging U.S. dominance in foundational models.
  • Qwen3.5 / Qwenputer - a deterministic RISC-V CPU running *inside* the model's activations. LLMs as internal computation engines is now a demonstrated thing.
  • Weight-Space Programming - LLMs reading and writing neural weights natively. A small weight edit let Qwen finish its reasoning. Self-modifying models are here.
  • Modality-Aware Expert Pruning - cuts MoE multimodal memory overhead up to 90% with no performance loss.
  • FareedKhan-dev/kimi-k3-in-c - 2.78-trillion-parameter Kimi K3 inference on a single CPU with 8.24 GB RAM in portable C99, no BLAS, no GPU. Wild.
  • Postman - caught silently auto-installing AI plugins and attempting to share API data. Delete the plugins.
  • Gemini - revealed to train on personal chat history. Privacy backlash incoming.
  • Codra - open-source AI code review engine to automate PR reviews.
  • AgentKit by Scalekit - production MCP gateway, led Product Hunt with 419 votes. MCP is now infrastructure.
  • vercel-labs/agent-desktop - Rust computer-use CLI for desktop control. Panniantong/Agent-Reach - zero-API-fee web browsing (Twitter, Reddit, GitHub, YouTube, Bilibili).
  • Unprivileged Rust Sandbox - zero-trust containment for agent bash. Please use this.
  • Hermes Agent (Nous Research) - checkpoint/undo file-edit rollback. ZeroClaw - RBAC, OIDC, append-only logs for regulated environments. QwenPaw (AgentScope) - 50 PRs on Windows support. IronClaw - just a zip crate 8.6.0 โ†’ 9.0.5 Dependabot bump.
  • Claude Code Skills ecosystem - security hardening (webapp-testing shell=True removal), skill impersonation defenses in anthropic/ namespace, plus proposed skills: proofcore-contract-auditor (Solidity/Rust + TON Blockchain proof anchoring), md2video-audio (Marp), notion-spec-to-implementation, merged document-typography.
  • Conduit - mobile control of local Codex agents. claude-codex-taskbar - Windows taskbar rate-limit display.
  • Anthropic Messages API - community request #192632 for a Claude web_search provider, likely v2026.11.6.
  • Nemotron 3.5 ASR Streaming - requested in vLLM. Drama-3 TTS - now in llama.cpp (mtmd) and Unsloth.
  • W4A8 MoE online quantization on Ascend NPU in SGLang, plus sampling-mask export for RL.
  • Kairos - open-source risk/fraud decisioning. Prism - 'Cursor for hardware engineering'. kwant - AI bots auditing each other's work. Wavefunction - local Mac coding agent. nana - one terminal per coding agent on Windows. One Roof - human-oversight agent framework. Teamily AI - agent network in one messenger. Hookdeck Inspect - free webhook tester. Synter Canvas - ad automation across 30+ platforms.
  • GitHub trending: MoneyPrinterTurbo, TradingAgents, career-ops, CowAgent, nanobot, daily_stock_analysis, ppt-master, ragflow, tiny-llm, opencompass (100+ datasets), braindead-local-intelligence.
  • HN's biggest thread: AI technical debt - 1090 score, 639 comments - framing AI as exposing 40 years of legacy software fragility.
  • Research worth reading: Suppressing Misaligned Behaviors Durably, DNO-Prct (beats RLHF via game-theoretic equilibrium), Born-Measured Attention (word-level interpretability), TandemBench (1,200 multi-agent cooking scenarios), Routes of Reasoning (kills MoE overthinking), RAGCompass, SlackBench, MedPhysBench, CARTBench, PAYWORTHY (1,494 freelance jobs), PyRUA-DataFarm, False Reassurance in LLM Severe-Weather Guidance, Executor-Relative Consultation Value.
  • The Little Book of Generative AI Foundations - genuinely great math primer on diffusion and transformers.

โ“ FAQ: Today's AI News Explained

  • Q: Is Claude Opus 5.5 actually getting worse? - Yes, by reported metrics. Correction rates rose from 5.2% to 25.4% after Oct 19, and the model is ignoring CLAUDE.md instructions and safety rules. Pin your model version and add behavioral regression tests.
  • Q: Do I need to re-convert my GGUF models for llama.cpp v0.9.0? - Yes. llama.cpp v0.9.0 introduced backward-incompatible backend changes alongside OCP FP8 E4M3 quantization and the flash_attn_ext_rows op. Existing GGUF files won't work without re-conversion.
  • Q: What is OpenClaw Knowledge Scribes? - A new feature in OpenClaw v2026.11.5-beta.1 enabling structured, reusable knowledge capture and retrieval across agents, sessions, and Control UI. It's paired with an experimental PostgreSQL data plane for durable agent state storage.
  • Q: Why is disaggregated inference (P/D) unstable? - Both vLLM and SGLang have critical bugs: vLLM's NixlConnector releases KV blocks during in-flight remote READs causing silent corruption, while SGLang has PD decode corruption and long-prefill OOMs. Avoid P/D in production until fixes land.
  • Q: What did OpenAI announce on November 5, 2026? - Five things: Images 2.6, acquisition of Pierre Computer Company (chips) and Pierre (workflows), Daybreak Red AWS GovCloud, a security hardening post, and the LifeVal benchmark. They also quietly re-enabled training on Codex.
  • Q: Can a 2.78-trillion-parameter model really run on a CPU? - Yes. FareedKhan-dev/kimi-k3-in-c runs Kimi K3 inference in portable C99 using 8.24 GB RAM on a single CPU with no BLAS or GPU. It's a landmark for edge LLM deployment.
๐Ÿ”ฎ Editor's Take: We spent 2026 scaling context windows and we're ending it by proving you don't need them - compaction, memory layers, and knowledge graphs beat brute force every time. Meanwhile the models themselves are drifting under us and nobody's sounding alarms. The next competitive edge isn't a bigger model; it's *knowing what your model did yesterday*.