Weekly summary

Week 2026-W47

22 November 2026

The week in AI

  • Claude Code Code Mode can now be tried. Start Claude Code with CLAUDE_CODE_EXPERIMENTAL_CODE_MODE=1 (interactive terminal only, 5.5 models or Fable 5.2) and Claude writes JavaScript that makes several tool calls at once. From 2.1.329, a PreToolUse hook that guards it needs "onFailure": "block" to stay fail-closed. (docs, 2.1.329)
  • Claude Code 2.1.325 to 2.1.329 tightened defaults and added background-task controls. Sandboxed Bash prompts again unless sandbox.autoAllowBashIfSandboxed is on, and scheduled tasks need a Skill(name) rule to run skills that declare tools or hooks. claude mcp add-json keeps tokens out of shell history, /status lists broken customizations, and /tasks stops scheduled tasks and /loop wakeups. (2.1.325, 2.1.327)
  • GitHub Copilot at Ignite: browser use in your signed-in Chrome, Copilot Evals (/eval) to test skills, agents, and MCP tools, on-demand secret scanning, adaptive code review, team plugin defaults (GA), and eight new MCP connectors. (Evals, browser use, team plugins)
  • External sandboxes for coding agents (Prime Radiant, experimental): run Claude Code's Bash, Codex, or Pi commands in a throwaway Linux VM that sees only the folders and hosts you allow, with API keys swapped in by a proxy. (source)
  • New models and prices: GPT-6.2 Astra (20% cheaper) and Sol, plus 30–40% off for 2–8 hour background windows; Gemini 4 Neon is GA at $3 / $15 with high thinking by default. Reminder: Gemini 4 Carbon and Argon double to $4 / $20 after 2026-11-30, the same day fast mode on Opus 4.8 stops working. (GPT-6.2, Neon)
  • Deferred tool loading is spreading: the Gemini API added tool retrieval (preview), including for remote MCP servers, and Pydantic AI added ToolBridge. (Gemini, Pydantic AI)

Rising on GitHub

  • alibaba/open-code-review (about 55,200 stars, +3,600 this week): free AI code review as a CLI or skills; its delegation mode lets your own coding agent do the review (npx skills add alibaba/open-code-review).
  • erkamyaman/rust-anything (about 7,600 stars in 6 days): MCP server and skill for porting code to Rust one unit at a time, each checked against recorded behaviour.
  • cathrynlavery/diagram-design (about 69,900 stars, +2,700 this week; star count unverified): skill that turns a request or an existing Mermaid, draw.io, or PlantUML file into a self-contained HTML/SVG diagram (/plugin marketplace add cathrynlavery/diagram-design). MIT.
  • apify/apify-mcp-server (about 17,300 stars, +1,000 this week): Apify's official MCP server for running ready-made scrapers (search, maps, social media) and getting structured results (claude mcp add apify "https://mcp.apify.com" -t http). MIT; Apify runs are paid.
  • ChromeDevTools/chrome-devtools-mcp (about 58,700 stars, +2,300 this week): lets an agent check pages with performance traces, console, network, and screenshots.

Library maintenance

  • Links: 244 checked, 1 dead. The Microsoft Agent Framework Python 1.23.0 release page (and the repo's releases list) returns 404; archived now points to a Wayback snapshot of 2026-10-24. One blocked link (Physical Intelligence, 429) left as is.
  • Consolidated (11 pages): coding agents, security, skills, MCP, tool use, evals, frameworks landscape, GPT, Gemini, Claude, and cost and caching. This week's Claude Code, Copilot, and Gemini facts moved from "Current state" into the body sections. "Current state" now keeps only open previews, betas, and coming deadlines. Unchecked OpenAI benchmark claims are grouped as one open question. Gemini 4 Neon added to the reference price table; Genie 4 and document slicing added to the Gemini page.
  • Reviewed: reasoning models (last reviewed 2026-10-18), checked against Anthropic's thinking docs. Fable 5.2 and the Mythos models are "Always on", Sonnet 5.5 is the only model that takes between_tools, and display: "omitted" hides thinking. A note on differing default thinking depth across vendors was added, plus an open question on Fable 5.2's default effort. No other page was more than 30 days old.
  • Other: the inbox was empty. No duplicate tags were found.