Researched and written with 315,749 tokens (in 10 · out 34,320 · cache-read 63,690 · cache-create 217,729) across 5 Claude calls

Coding agents

• Gemini 4 Hydrogen in GitHub Copilot: Google's latest Flash model is now GA in Copilot, pitched at fast agentic coding, tool use and multistep workflows.
• Claude Code v2.1.325: Patches a security issue in zstd-compressed response decoding, plus small VS Code notebook and org-policy /rewind hints; update.
• google/elephant-goldfish-model: Plugin splitting design (an "elephant" holding full project context drafts the spec) from review (context-free "goldfish" reviewers see only the draft and sources), which exposes unstated assumptions.
• ChromeDevTools/chrome-devtools-mcp: Official MCP server and CLI that lets Claude, Cursor or Copilot control and inspect a live Chrome for debugging, automation and perf analysis.
• Expo Simulators: Cloud simulators built for agents, so mobile app testing can run without a local device setup.
• @copilot-extensions: GitHub's toolkit org for Copilot integrations (preview SDK, debugging tool); low-signal catalogue page, skim only.

LLM mechanics

• Read Less, Solve More: Token-Efficient Sparse Reading of External Content for AI Agents: Paper on cutting agent token spend by reading only the relevant slices of external content instead of whole documents.

Agentic SaaS and agent UX

• backnotprop/plannotator: Review layer to annotate plans, specs and diffs and send feedback to agents in one click, closing the human-in-the-loop gap.
• Agent Desk: Free open-source single screen for monitoring all your Claude Code agents.
• Burnstate: Menu-bar pacing for Claude Code, Codex and Cursor usage limits.
• T3 Code in-app UI customization: T3 Code now lets you ask the agent to rewrite its own UI live; a notable self-modifying-harness pattern.
• ADE tier list v3: Community ranking of agentic dev environments; T3 Code is the daily driver, with Capy, Zeron and bb as contenders.
• lobehub/lobehub: Team-chat "super-agents" that take on work autonomously; broad product, limited coding-agent relevance.

CLI / TUI

• Terminal borders that don't cost a cell: Ratatui maintainer on drawing borders without consuming character cells; community points to overstrike mode (DECSET 20, only mintty implements it) over a new protocol.
• tailscale/tailcat: netcat over Tailscale's WireGuard data plane (magicsock, DERP relay) without the control plane, handy for ad-hoc encrypted pipes between machines.

Vibe-coding and repos

• alibaba/open-code-review: Alibaba's internal AI review CLI, open-sourced, combining deterministic pipelines with an LLM agent for line-level comments (NPE, thread-safety, XSS, SQLi rules); works with OpenAI and Anthropic APIs.
• simonw/tools: Simon Willison's collection of LLM-built HTML+JS tools, with commit messages and transcripts for each, a good reference for prompt-driven development.
• Six weeks building an open-source video editor with Claude Code: Long-horizon build retrospective on what worked and what bit back.
• Windows system utility built with Claude Code (1089 tests): NitroSense replacement (fan control, battery limits) where a heavy test suite was used to keep hardware-touching code correct.