AI library

Weekly summary

Week 2026-W44

1 November 2026

The week in AI

  • Coding agents as CI jobs: GitHub Agentic Workflows is GA. A Markdown file becomes an Actions job that runs Copilot, Claude, Codex, Gemini, or Pi in a read-only sandbox, and it can write only through declared "safe outputs". At the same GitHub Universe event, Copilot connectors added 26 verified MCP servers and Agent Merge fixes failing checks on a PR.
  • A hard write fence for Claude Code: 2.1.314 adds permissions.blockWritesOutsideWorkingDirectories, which holds in every mode, bypass included. Earlier in the week, 2.1.311 added ultrathink for a single xhigh turn and compaction past 200K tokens.
  • Multi-model routing goes mainstream: Copilot Fusion routes each prompt to a single model, a cascade, or a critique loop. GitHub claims it is close to Claude Code with Opus 5.5 at half the cost (vendor figures). VS Code 1.143 compares and merges results from several models.
  • Gemini from the shell: the gemini-api CLI (beta) lets any agent generate media, analyse files, or start Deep Research with one command, with -o json and --dry-run.
  • Know who is spending: Anthropic's Usage and Cost Analytics API (beta) breaks cost down by key, model, workspace, and user.
  • Models: OpenAI's GPT-6.1 Terra ($1 / $5) targets computer use. Kimi K3.1 Pro comes close to Opus 5.5 on Moonshot's own figures, with weights promised by 2026-11-04. Mistral Large 4 weights are now public, under a licence that restricts model-as-a-service use.

Rising on GitHub

  • ChromeDevTools/chrome-devtools-mcp (54,581 stars, +479 this week): lets a coding agent drive and inspect a live Chrome for debugging and performance traces. Apache-2.0; install with /plugin marketplace add ChromeDevTools/chrome-devtools-mcp. Usage statistics are on by default (--no-usage-statistics).
  • upstash/context7 (64,033 stars, +405): current, version-specific library docs for agents through an MCP server or the ctx7 CLI (npx ctx7 setup --claude). The docs are community-contributed, so they can be wrong.
  • thedotmack/claude-mem (108,183 stars, +5,290; in library): automatic memory across sessions for Claude Code, Codex, and others.
  • DietrichGebert/ponytail (181,666 stars, +5,998): a MIT skill and plugin that pushes agents to write the least code a task needs, with /ponytail-review and /ponytail-audit. Its benchmarks are self-reported and the star count is unusually high for a new repo.
  • Leonxlnx/taste-skill (107,230 stars, +5,107): a MIT front-end design skill that gives the agent an art direction, real photos, and a browser check of its own output (npx skills add https://github.com/Leonxlnx/taste-skill --skill "design-taste-frontend"). The README is promotional, and the star count could not be verified.

Library maintenance

  • Links: the weekly check found 0 dead and 0 moved out of 145. Two OpenAI pages are blocked (403), so no action was taken.
  • Consolidated 10 topic pages, folding durable facts from "Current state" into the body and dropping superseded entries: coding agents (14 entries to 7), cost and caching (8 to 4), open-weight models (8 to 5), security, MCP, Gemini, framework landscape, memory, skills, and GPT.
  • Resolved the open question about Mistral Large 4 weights: they are public on Hugging Face under the Mistral Ecosystem Open-Weight License (new source note). The Decisions API GA pricing check moved to Open questions on cost and caching.
  • Stale-page review: none was due, because no topic page was last reviewed more than 30 days ago (oldest: 2026-10-04).
  • Inbox was empty; no duplicate tags found.