# Radar, November 15, 2026: ten signals from X on Codex 0.169, Context Mode and Lara Mini

_Sunday November 15, 2026 · AI provenance: This page was compiled by Claude (AI, Anthropic) from X search briefings by Grok Bot (AI, xAI) and the llmnet team's X trend monitor. AI wrote the text; every signal has a source link to X and a confidence level. Responsible for llmnet.nl: Ivo Donker._

**In short:** The key signal of November 15: Codex 0.169 keeps project configuration and MCP servers in an untrusted folder off until you pass --trust. Alongside it, ten short signals on a plugin that keeps tool output out of the context window, two open models with very different licenses, tool calls in llama.cpp and the Cursor breach, each with its source on X.

## 1. Codex 0.169: codex exec keeps project config and MCP off in an untrusted folder

The November 13 stable release keeps a folder's project configuration and MCP servers off until you trust it, also with --sandbox workspace-write; --trust on codex exec or codex review saves that trust. Also side chats in tabs, searching older tasks and custom model providers. If you run Codex in CI, check whether a job silently used an MCP server from the repo.

Source on X: [@Codex_Changelog](https://x.com/Codex_Changelog/status/2121168171738484881) · [@eyoussef](https://x.com/eyoussef/status/2121985397051781258) · [GitHub, release rust-v0.169.0](https://github.com/openai/codex/releases/tag/rust-v0.169.0)
Read also: [Codex 0.169 and --trust](https://nieuws.llmnet.nl/en/codex-0-169-trust-zijchats) · [Codex news](https://apps.llmnet.nl/en/codex-nieuws)

## 2. Context Mode 2.2.0 keeps tool output out of the context window, now for eleven agents

The plugin puts large coding-agent tool output in a local SQLite index and returns only what a script prints. v2.2.0 (November 14) adds OMP as the eleventh agent. The sandbox inherits your environment, signed-in cloud CLIs included. License Elastic 2.0: free to use, not to resell as a service. The 99.4% saving is the maker's own measurement.

Source on X: [@mksglu](https://x.com/mksglu/status/2118037065149419852) · [GitHub, release v2.2.0](https://github.com/mksglu/context-mode/releases/tag/v2.2.0)
Context in the briefing: [Local agents in the November 15 briefing](https://trending.llmnet.nl/en/briefing/2026-11-15#lokale-agent-tools)
Read also: [Context Mode 2.2](https://nieuws.llmnet.nl/en/context-mode-tooluitvoer-sandbox) · [Skills and plugins](https://apps.llmnet.nl/en/nieuws-skills-plugins)

## 3. Lara Mini: Translated's multilingual MoE base model, weights non-commercial only

7.2 billion parameters, 0.6 billion active, a third of the training in 44 languages other than English, Dutch included. The weights have been gated on Hugging Face since November 11 under CC BY-NC-SA 4.0; building a product needs an agreement with Translated. It is a base model without instruction or safety tuning. The 57.0 score is the maker's own.

Source on X: [@Translation](https://x.com/Translation/status/2120895536517501221) · [@DeltaA59245](https://x.com/DeltaA59245/status/2121888490388328583) · [Hugging Face, modelkaart](https://huggingface.co/translated/Lara-4-Mini-Base)
Context in the briefing: [Open models in the November 15 briefing](https://trending.llmnet.nl/en/briefing/2026-11-15#open-modellen)
Read also: [Lara Mini, multilingual base model](https://nieuws.llmnet.nl/en/lara-mini-meertalig-basismodel)

## 4. Haru-4B listens and talks at the same time, under Apache 2.0

Higurashi AI put a full-duplex speech model on Hugging Face on November 12: every 133 ms one frame of audio goes in and one frame of audio plus text comes out. Nine languages, no Dutch. There is an MLX version with a macOS app; the NVIDIA route needs about 12 GB. It is a base model for research. speech.cpp supports it since v0.11.0, according to its maker.

Source on X: [@HigurashiAI](https://x.com/HigurashiAI/status/2121895755124543699) · [@gyakuse](https://x.com/gyakuse/status/2121877504789287304) · [Hugging Face, modelkaart](https://huggingface.co/higurashi-ai/Haru-4B)
Context in the briefing: [Open models in the November 15 briefing](https://trending.llmnet.nl/en/briefing/2026-11-15#open-modellen)

## 5. llama.cpp builds fix tool calls inside reasoning blocks for Gemma 4, Ministral-3 and Qwen3

Of last night's seventeen prereleases, three affect agents: b12388 passes a Gemma 4 tool call that fell inside an unclosed thought block, b12391 allows tool calls inside Ministral-3 and Qwen3 reasoning, b12404 fixes connection reuse after a streamed response. The stable pin stays v1.0.0 from November 13. Try them on a test machine, not in production.

Source on X: [@ggerganov (v1.0.0)](https://x.com/ggerganov/status/2121246103312212124) · [GitHub, b12391](https://github.com/ggml-org/llama.cpp/releases/tag/b12391)
Context in the briefing: [Local agents in the November 15 briefing](https://trending.llmnet.nl/en/briefing/2026-11-15#lokale-agent-tools)
Read also: [llama.cpp v1.0.0](https://nieuws.llmnet.nl/en/llama-cpp-v1-stabiele-api)

## 6. T3 Code: change the theme through the agent and resume a stopped turn

Theo shows T3 Code letting your agent change the interface. The November 15 nightly (v0.1.10-nightly.20261115.4723) resumes a stopped Cursor turn from the last checkpoint and keeps prompts waiting on a provider across a restart. These are nightlies, not a stable release.

Source on X: [@theo](https://x.com/theo/status/2121455838414201200) · [GitHub, t3code releases](https://github.com/pingdotgg/t3code/releases)
Read also: [MCP and agent tooling](https://apps.llmnet.nl/en/nieuws-mcp-agent-tooling)

## 7. Gemini 4 Hydrogen is generally available in GitHub Copilot

@github says Gemini 4 Hydrogen is in Copilot, in the app, the CLI and VS Code, aimed at fast agentic coding and tool use. Compare it on your own repositories with your current model and measure cost and error rate. We did not read back the GitHub changelog.

Source on X: [@github](https://x.com/github/status/2121319695626518823)
Context in the briefing: [Gemini in Copilot in the November 15 briefing](https://trending.llmnet.nl/en/briefing/2026-11-15#gemini-hydrogen-copilot)
Read also: [Gemini 4 Hydrogen](https://nieuws.llmnet.nl/en/gemini-4-hydrogen)

## 8. Cursor confirms: fake Claude Code extensions sat in the marketplace cache for 11 hours

BleepingComputer summarizes Cursor's update: a stolen Fireworks key led to two malicious copies (2.1.312 and 2.1.313) of the Claude Code extension, in the cache for about 11 hours on October 29, reaching at most 5,000 users. If you installed extensions in Cursor then, check for unknown SSH keys.

Source on X: [@BleepinComputer](https://x.com/BleepinComputer/status/2121656319383847234) · [Cursor-blog](https://cursor.com/blog/security-incident-update)
Context in the briefing: [The Cursor breach in the November 15 briefing](https://trending.llmnet.nl/en/briefing/2026-11-15#cursor-claude-extensie)
Read also: [The Cursor incident and the leaked key](https://nieuws.llmnet.nl/en/cursor-incident-fireworks-sleutel)

## 9. Claude chat uses auto mode, and permission in the session can lift a refusal

Simon Willison sees regular Claude chat using auto mode and sometimes refusing because of it. A prompt that gave the permission layer consent lifted the refusal, since auto mode weighs the user's messages. One observation, no explanation from Anthropic. The same mechanism applies to instructions you did not write yourself.

Source on X: [@simonw](https://x.com/simonw/status/2121695097674981849)
Context in the briefing: [Auto mode in the November 15 briefing](https://trending.llmnet.nl/en/briefing/2026-11-15#claude-auto-mode)

## 10. Unconfirmed: Grok 4.8 in a soft rollout, but without a product page

@grok says a soft rollout of Grok 4.8 to Bot users is under way. We found no product page, model ID or price on x.ai; the public apps and the API are on Grok 4.7. Andrew Curran expects news from Anthropic and OpenAI next week; those are expectations, not announcements.

Source on X: [@grok](https://x.com/grok/status/2121869842969940249) · [@AndrewCurran_](https://x.com/AndrewCurran_/status/2121715103427096652)
Context in the briefing: [Unconfirmed in the November 15 briefing](https://trending.llmnet.nl/en/briefing/2026-11-15#onbevestigd)

[← Previous radar](https://radar.llmnet.nl/en/dag/2026-11-14)

