Researched and written with 309,399 tokens (in 10 · out 29,565 · cache-read 66,228 · cache-create 213,596) across 5 Claude calls

Claude Code internals and context

• Things that quietly break your prompt cache in Claude Code (and a few that don't) is a community list of the actions that invalidate the cached prefix, which matters because a cache break turns a cheap cached read into a full-price re-read of the context.
• Claude Code now gzips the conversation it sends on every turn, so requests reach the model faster, and cloud sessions now start spinning up as soon as you begin typing.
• The 3 kinds of "Claude forgot" argues that only one kind of forgetting is a context-window limit, so the other two need different fixes than just a bigger window.
• PSA: ultracode isn't locked to max effort anymore means you can now run ultracode below max effort, which gives you a way to trade reasoning depth against cost.
• How do you decide which rules to delete from CLAUDE.md? is a discussion of pruning instruction bloat, which is useful for keeping the static prefix small and the rules that remain effective.
• Hellhound 1.3.0 adds rule and memory cards, and the author's write-up of what their first test got wrong is a useful reference for anyone building agent memory tooling.
• Code Review is now GA for Claude Team and Enterprise, with faster reviews and admin-selectable Quick, Standard or Deep levels per repo.

Agent harness landscape

• You can now use your Claude subscription in OpenCode means a third-party open-source harness can run on subscription billing instead of API keys.
• Theo on the OpenCode news reads it as Anthropic loosening its restrictions on subscription use outside its own clients.
• GitHub Copilot Fusion is GitHub's coding agent that picks and combines models per task, which is a vendor-side take on model routing.
• Day 30 update from Theo makes the thread detail panel customizable and adds a section for monitoring usage limits and context usage, a pattern worth copying in any agent UI.

Multi-agent orchestration

• Memdoor team-solve runs five Qwen3.8-27B roles from fresh contexts that communicate only through files, with a shell script rather than a model deciding success; it reportedly beats Fable 5 on LiveCodeBench Hard. HN commenters note it reuses the GVS5H method posted a week earlier.
• Orkas is an MIT-licensed local-first desktop app where a Commander agent plans tasks and coordinates 16 built-in specialists plus Claude Code, Codex, OpenCode, OpenClaw and Hermes.
• Claude-Code-Game-Studios packages about 49 agents and 70+ skills into a studio-style hierarchy with design review and QA gates, a reference for structuring role-based agent teams.
• Running 2-3 Claude Code / Codex agents on one repo asks how to stop parallel agents from stepping on each other, which is the core worktree and coordination problem.

Vibe-coding patterns

• Patrick Collison's "improvements" box is a freeform text input in his personal apps wired to an agent that implements whatever is submitted, and a user of one app has already used it. It shows how low the barrier to fixing rough edges gets, and the risk of running it unsandboxed.
• frontend-design skill installs with npx skills add anthropics/skills --skill frontend-design and has the agent choose fonts, colors, layout and motion before writing code, to avoid generic AI-looking sites.
• Backenly pitches itself as an autonomous backend for AI coding agents.

Terminal and agentic SaaS

• Pluto Door is an SSH client with a built-in AI agent and a desktop view per server.
• Agents Prompt People is a Latent Space interview with Arca's lead AI engineer on long-horizon financial work, where agents do the background work but human advisors make the final decisions.