Researched and written with 312,930 tokens (in 10 · out 31,767 · cache-read 66,228 · cache-create 214,925) across 5 Claude calls

Coding agents & model behavior

• v2.1.319: Claude Code auto mode with telemetry or flags off (including behind LLM gateways) now asks or denies unverifiable calls instead of falling back to the built-in classifier; set CLAUDE_CODE_AUTO_MODE_SERVER=0 to keep the old behavior (Bedrock, Vertex, Foundry unchanged).
• RT @AdamHoltererer: Opus 5.5 fails twice as many T3 turns on max as on high: T3 Code telemetry shows max reasoning fails more in every turn-length bucket, mostly within the first 30 seconds, so it is not just longer turns.
• I gave Adam access to the T3 Code anonymized telemetry data...: Theo's takeaway from the same data is to avoid max reasoning.
• idk how theo and his team pulled this off...: One user's tests claim t3code now saves around 40% of usage on average, unverified but worth benchmarking against your own setup.
• A Jev that can ask Opus for help gives Opus accuracy at 4x lower cost: A hybrid where a cheap model escalates hard cases to Opus reaches frontier accuracy at more than 4x lower cost on identity-resolution grading, the same pattern as cheap-first routing with an Opus advisor.
• Why asking the same coding agent session to review its own branch misses bugs: Argues for clean-room review across Claude Code and Codex instead of self-review, with two CLI commands to do it.
• I made Claude Code and Codex message each other by name, no tmux: Named inter-agent messaging between Claude Code and Codex without tmux plumbing, relevant to cross-provider swarm setups.
• Tried Anthropic's webapp-testing skill on a tiny local app: Both checks passed and Claude flagged an unrequested bug, a small data point for the skill's usefulness.
• cloudflare/security-audit-skill: Multi-phase security-audit skill (recon, hunting, validation, independent record verification) that seeded Cloudflare's vulnerability harness, a good template for isolated-agent pipelines.

Context, memory & token tricks

• Stanford researchers develop a new AI memory system...: Recalls from full conversation histories using 16 memory tokens, cutting document-token usage up to 4,212x and running 27x faster, beating baselines on 5 of 8 benchmarks.
• Make0 Brain OSS: Open-source tool to sync context across multiple agents, in the same space as shared-memory efforts.

CLI / TUI

• Libghostty 1.0: libghostty-vt and its C ABI are now stable, giving a portable terminal-emulation core to embed; critics ask why the API is C rather than Zig, and Windows support is still missing.
• Show HN: Tmux-restore: Saves tmux windows every minute and before shutdown, then resumes Claude Code, Codex, Gemini CLI and opencode sessions after reboot; only worth it over tmux-resurrect if agent resume is reliable.
• crystl: macOS multitasking terminal built for running several AI coding agents.
• I built clibridge: Lets OpenAI and LangChain SDKs use your Claude Code login, so check the terms-of-service and auth implications before relying on it.
• nushell/nushell: Structured-data shell, usable as a daily driver but still carrying design-change and stability caveats.

Skills & vibe-coding repos

• sickn33/agentic-awesome-skills: Catalog of 2,926+ SKILL.md playbooks for Claude Code, Codex, Cursor and Gemini CLI, with local discovery, agent-driven selection and plan preview.
• Claude Code built my open-source AI voice studio, then made its 7-minute trailer too: End-to-end vibe-coding showcase covering both the app and its promo video.

GitHub Copilot & agentic SaaS

• @GitHub in Slack and Teams is generally available: Investigate bugs, create issues, implement changes and open PRs from chat, now GA.
• GitHub Copilot weekly releases, November 2: Copilot reaches JetBrains Air and Xcode's coding assistant, plus PR, settings and agent-session management updates.
• Copilot code review: API management and review overview updates: Code review settings are now configurable via REST and GraphQL from enterprise policy down to single repos, and the review summary comment is more compact.
• Copilot impact dashboard adds chat and VS Code agent usage: Agent adoption chart now covers Copilot Chat and the VS Code Agents window, spanning nine Copilot experiences.
• Introducing Claude Code for DaVinci Resolve: Motion drives a whole video edit from one prompt and opens it in Resolve with every layer as a real Fusion node.