AI digest: 2026-10-28

Highlights

Gemini API CLI (gemini-api) beta: a dependency-free binary that lets any coding agent generate images, speech, and video, analyse files, or start a Deep Research job with one shell command, with -o json and --dry-run built for agents.
Claude Code 2.1.312: Write no longer overwrites a large CLAUDE.md it has only partly read, paths: skills work after a cd, and prompt and agent hooks are capped at 24 concurrent per session.
Two skill packs to try: BeautifulRust (write, review, and port idiomatic Rust, installable as a Claude Code plugin) and Archify (interactive HTML diagrams of a plan or codebase).

Claude Code release 2.1.312: CLAUDE.md overwrite fix, path skills after cd, hook concurrency cap, routine effort.
Codex CLI 0.165.0: Vertex AI provider, Memory v2 on by default with import from other agents.
Gemini API CLI (gemini-api) beta: Gemini media, file analysis, and managed agents from the shell.
BeautifulRust: three Rust skills plus a clippy config; Claude Code plugin or Codex skills.
Archify: diagram skill producing self-contained interactive HTML.

Pydantic AI v2.63.0: DetectLoops, IPython code mode, shared spend budgets in PostgreSQL.

Ollama v0.40.12: LFM3 and d1 2B: Liquid AI's 1.7B, 256K-context small models, Apple only for now.
GLM-5.3 model card (Z.ai): 753B open-weight coder; MIT-style licence with a review for hyperscale API providers.

tooling/coding-agents: Claude Code 2.1.312 and Codex 0.165; hook cap and removal of safety_event corrected in the body.
tooling/skills: BeautifulRust and Archify.
frameworks/landscape: Pydantic AI 2.63; version table updated.
models/gemini: Gemini API CLI beta; Gemini CLI 0.66 notice about Gemini 4.
agents/memory: Codex Memory v2 and memory import; open question on import formats.
agents/tool-use: gemini-api as a CLI-first tool for agents.
models/open-weight: gap filled with a primary source for GLM (GLM-5.3); LFM3 in Ollama.