Researched and written with 319,578 tokens (in 10 · out 35,713 · cache-read 66,228 · cache-create 217,627) across 5 Claude calls

Security: Claude Code extension compromise

• Backdoor in Claude Code for VS Code extension on the Cursor Marketplace: Phorion found that the official extension bundle on the Cursor Marketplace drops a persistent SSH backdoor on macOS, Linux and Windows, and the write-up focuses on macOS persistence. Community notes say yanking the bad version doesn't clean machines already infected, and Cursor's own security notice corroborates the report, so treat any installed copy as compromised.
• One-line check for the Claude Code extension breach: Reddit thread with a quick command to test whether your machine was hit. Run it before anything else.

Coding agent releases

• Claude Code v2.1.314: adds permissions.blockWritesOutsideWorkingDirectories, which makes Write, Edit and NotebookEdit refuse paths outside the working directories in every permission mode. It also adds IdP groups routing for Claude apps gateway upstreams, including on failover.
• Claude Code v2.1.316: adds CLAUDE_CODE_DISABLE_ULTRATHINK to turn the ultrathink keyword off, plus pooled group budgets on the gateway via pooled_budget on the spend limits admin API.
• Claude Code desktop app update: faster session start, easier splitting of work across sessions, and a view of what each session is doing, aimed at running many sessions side by side.
• Codex 0.166.0: adds keyboard-driven transcript selection with word, line and page motions, integrated search and configurable shortcuts, which is a real TUI usability gain.
• Repository plugin defaults and enterprise plugin controls GA: in VS Code, Copilot CLI and the Copilot app, trusted repositories can now install and enable default plugins under enterprise-managed controls.

Agent workflow and context management

• Codex subagents inheriting parent context: Codex was biasing subagents toward forking the parent's context window, which breaks skills that need a clean window (code review, implement-spec). Matt Pocock shipped a skill fix, and the underlying default is worth knowing if you rely on isolated subagents.
• Anthropic builds in-house source control: merges went from about 2,500 per day in May to over 30,000 per day in September, and the resulting developer-infrastructure bottleneck led them to build a new VCS and migrate their monorepo off git. It's a strong signal that git is straining at agent-scale commit volume.
• Your claude -p script says a tool call was denied: three ways to fix headless permission denials and what each one lets Claude do. Useful when scripting Claude non-interactively.
• Claude Code Game Studios vs plain Claude Code: the multi-agent studio setup took 19 minutes and $5 to build a small Godot game, while plain Claude Code did it in 38 seconds. It's a data point against heavy orchestration on small tasks.
• Codebases as organic systems: opinion post arguing that agent-era codebases need cybernetics-style abstractions, and praising Anthropic's "SuperFans" blog post. Conceptual only.

Skills, memory and harness repos

• obra/superpowers: a software development methodology built from composable skills. The agent steps back and asks what you're really trying to do before it writes code.
• anthropics/skills: Anthropic's public skills repo, where skills are folders of instructions, scripts and resources that Claude loads dynamically. It's the reference implementation for the Agent Skills standard.
• thedotmack/claude-mem: captures what your agent does in a session, compresses it with AI and injects relevant context into future sessions. It works across Claude Code, Codex, Gemini, Copilot, OpenCode and others.
• affaan-m/ECC: a harness optimization system covering skills, instincts, memory and security for Claude Code, Codex, OpenCode and Cursor. Its README warns to install only from official channels, which matters given the extension compromise above.
• ruvnet/ruflo: an agent harness for multi-agent swarms with adaptive memory and federation, installed as Claude Code plugins (requires 2.1.287 or later). It's another entry in the orchestration-heavy camp.
• simonw/claude-system-prompts: turns Anthropic's published system prompts into files with a git history dated to each release, so git diff shows how the prompts changed.

Computer, browser and agentic SaaS

• agent-desktop: computer use for any coding agent on macOS, Windows and Linux. It runs in the background, with parallel agents each getting their own cursor.
• lightpanda-io/browser: a headless browser written from scratch in Zig for AI agents. Its benchmark shows 123MB peak memory over 100 pages versus 2GB for headless Chrome, so it could cut the cost of running browser agents at scale.
• CopilotKit/OpenBot: open-source AI coworkers that each get their own browser, files and granted tools, with every action decided before it happens and recorded after. It runs on your infrastructure and works with any AG-UI agent.
• Faylee: bug reports your coding agent can fix from a single link. Early Product Hunt launch.
• ilse: point at your running app and your AI agent changes the code. Early Product Hunt launch.