Sheet 03 · Tags / AI
#AI
13 posts · 0 projects
AWriting
- Modern Web Guidance 1.0: Chrome's agent skills go stable, add a modernize workflow, and pick up Firefox and EdgeChrome released Modern Web Guidance 1.0 on November 10, 2026. The skill set for coding agents adds a static-analysis modernization workflow, expert-reviewed guides with a public eval dashboard, and Firefox and Edge as contributors.
- Postman desktop 12.33.4 installed plugins and a SessionStart hook into Claude Code and Codex without asking, then revertedA Postman user reported on October 25, 2026 that the desktop app auto-installed its agent plugins into Claude Code and Codex, added a startup hook, and wrote a hook trust record into ~/.codex/config.toml. On November 5 a Postman team member said the company was reverting the change.
- LiteLLM 1.107.0 shipped a credential stealer to PyPI and Docker for two hours: what to delete and what to rotateAttackers published a malicious litellm 1.107.0 on October 21, 2026, through LiteLLM's own release workflow pointed at a lookalike fork. The wheels drop cook.pth, cook.py and sitecustomize.py into site-packages, which run on every Python start. The last clean version is 1.106.2.
- Anthropic replaced git for its monorepo with Scion: 30,000 merges a day and checkouts in secondsOn October 13, 2026 Anthropic described Scion, the source control system a seven-person team built in five months after merges grew from 2,500 to more than 30,000 a day. It moves commits, merges and history to the server and fetches the trees and blobs a client asks for.
- Gemini 4 Argon on the API: 1M output tokens, thinking levels only, and a price that doubles on December 1Google opened Gemini 4 Argon (gemini-4-argon) to the Gemini API worldwide on October 13, 2026 at $2 input and $10 output per million tokens until November 30. It drops thinking_budget, sampling parameters and explicit caching, adds long decode up to 1,048,576 output tokens, and spends more than twice GPT-6 Astra's output tokens per benchmark task.
- Claude Haiku 5.5: a tenth of Haiku 4.5's price under 100K tokens, and a tokenizer that counts 30% moreAnthropic released Claude Haiku 5.5 on October 7, 2026 at $0.10 input and $0.50 output per million tokens for prompts up to 100K. It adds effort levels, rejects thinking budgets, sampling parameters and prefill, and counts about 30% more tokens than Haiku 4.5 for the same text.
- Claude Sonnet 5.5: same price as Sonnet 5, and five request settings that now return a 400Anthropic released Claude Sonnet 5.5 on September 28, 2026 at $2 input and $10 output per million tokens, with 70.6% on Terminal-Bench 4.0. The model rejects forced tool choice and disabled thinking, so Sonnet 5 code needs changes before you swap the ID.
- Pi 1.0: the minimal coding agent goes stable, two days after it added MCPEarendil tagged Pi 1.0.0 on October 1, 2026. The terminal coding agent runs fullscreen by default, cuts codemode prompt tokens by about 40%, hardens MCP OAuth, and ships an experimental durable harness, pi-durable.
- A web page could run code through OpenCode's local server: GHSA-632h-h47v-g4x4Datadog Security Labs disclosed a remote code execution bug in OpenCode 1.14.30 to 1.18.21 on September 24, 2026. A text/plain form post to the local /global/upgrade endpoint made npm install an attacker's tarball. OpenCode 1.18.22 fixes it.
- Claude Opus 5.5 from a front-end desk: price, speed and the agent partsAnthropic's September model claims Fable-level work at 40% lower cost than Opus 5. Notes on long codebase migrations, prompt caching and agents that read untrusted pages.
- OpenAI's API this quarter: GPT-6 arrived, the Assistants API leftBetween July and September 2026 OpenAI shipped GPT-5.6, then GPT-6 Astra, Sol and Luna, retired the Assistants API and changed its error codes. Notes from the changelog for teams with OpenAI calls in production.
- Grok 4.6 and 4.7 on the API: two frontier models in two monthsxAI shipped Grok 4.6 in August and Grok 4.7 in September 2026, with the same 500k context window and the same price. The release notes for developers, with a cost check and a deprecation to schedule.
- Building Next.js for agents starts with debugging for peopleOn February 12, 2026, Jiachi Liu described how the Next.js team built and shut down an in-browser agent called Vector, then moved the useful parts into the framework: an MCP endpoint in next dev, browser logs in the terminal and version-matched docs.

