Claude Code Daily Briefing - 2026-09-05

Release Summary

v2.1.261 landed just one day after v2.1.260 (9/3). It brings new features like /skill-doctor, which shows whether loaded skills are actually being used and their context cost, and bashOutputMaxChars/taskOutputMaxChars, which let command and background task output be received inline up to 128K characters, alongside more than 70 fixes and improvements.

Full release notes

New Features & Practical Usage

/skill-doctor — surfaces unused skills and their context cost (v2.1.261)

A new /skill-doctor command shows which loaded skills are actually going unused, along with how much context each one is costing. You can use this to prune skills you don’t need.

/skill-doctor
# Shows loaded skills, each one's context usage, and whether it's actually being invoked

When you accumulate plugins and skills over time, it’s easy for skills you never actually call to quietly eat into your context on every turn. /skill-doctor lets you find and remove those, freeing up both context headroom and prompt-cache efficiency. Full release notes

bashOutputMaxChars/taskOutputMaxChars — receive command and background task output inline up to 128K characters (v2.1.261)

Previously, output from command execution and background tasks that exceeded a certain length would get written to a file, with only a summary passed inline. New bashOutputMaxChars/taskOutputMaxChars settings let you raise that inline limit yourself, up to 128K characters.

// settings.json
{
  "bashOutputMaxChars": 131072,
  "taskOutputMaxChars": 131072
}

If you want to analyze long build logs or test output directly instead of being routed through a file, these settings let you raise the inline limit as high as you need. Full release notes

Developer Workflow Tips

--append-subagent-system-prompt-file — pass long subagent prompts without hitting command-line length limits (v2.1.261)

Until now, appending content to a subagent’s system prompt meant passing it as a command-line argument, which made it easy to hit the shell’s argument length limit. A new --append-subagent-system-prompt-file option reads the content from a file instead.

claude --append-subagent-system-prompt-file ./prompts/reviewer-guidelines.md

If you want to attach a large document, like your team’s code review standards or architecture guidelines, to a subagent in full, you can now manage it as a separate file and pass it in without worrying about command-line argument limits. Full release notes

How is MCP actually being used in production? — an Ask HN thread rounds up the answers (9/4)

In response to a question about why MCP generated so much buzz at launch yet real production use cases are hard to come by, practitioners shared what they’re actually doing. The consensus: MCP is being used for internal ops diagnostics, SaaS integrations, voice agents, and customer-facing chatbots, mainly because it cuts down on tool-discovery and integration overhead compared to wiring up plain APIs or CLIs directly.

If you’re designing a new MCP server for Claude Code, this is worth reading as evidence that narrowing your tool scope to a specific operational or customer-facing scenario tends to work better in practice than building a general-purpose API wrapper. GeekNews

Security & Limitations

Claude service status — no new incidents 9/4-9/5, all services operational

Checking status.claude.com directly, there have been no newly logged incidents as of 9/4-9/5 since the 9/3 incident covered in the 9/4 briefing (elevated error rates on Claude Sonnet 5, Mythos 5.1, Fable 5.1, and Opus 5, among others, all since resolved), and every service — including claude.ai, Claude Console, Claude API, Claude Code, Cowork, and Government — is Operational.

As of the StatusGator check (2026-09-05 02:08 UTC), there were 4 user reports in the past 24 hours, citing error messages, API overload (529s), and unresponsive requests from Brazil, France, and Oklahoma, US. These look like scattered reports concentrated in specific regions and time windows rather than a pattern pointing to a broader outage. Claude Status · StatusGator

Ecosystem & Plugins

wigolo — a local-first web search tool that bundles 18 search engines into a single MCP interface (9/5)

Provides search, page reading, crawling, data extraction, caching, similar-document search, and research capabilities for AI agents through a single MCP interface. Searches query all 18 engines at once, merge the results, re-rank them with a local model, and attach an excerpt plus source location and citation ID to each result.

If you’re wiring a research pipeline into Claude Code over MCP, this all-in-one local tool is worth considering as a gateway instead of stitching together multiple search engines yourself. GeekNews

CLIProxyAPI — a local proxy that exposes AI subscription accounts like Claude Code through standard APIs (9/4)

Connects OAuth-logged-in accounts from ChatGPT Codex, Claude Code, Gemini, Grok Build, Antigravity, and others to a local proxy, exposing them through several standard API formats. This lets you call your subscription account’s models from existing SDKs, API clients, or other coding agents without issuing a separate API key.

Reusing a subscription account’s OAuth through third-party tools is convenient, but as the Google Antigravity terms controversy below shows, vendors can change policy to explicitly ban this kind of reuse — worth keeping in mind. GeekNews

Community News

• Nvidia finalizes $12.89 billion acquisition of Hugging Face (9/3): Nvidia, long GPU-centric, has officially confirmed acquiring Hugging Face for roughly $12.9 billion, expanding into the AI platform space. Hugging Face is expected to remain an open platform post-acquisition, not locked into Nvidia compute. Teams integrating Hugging Face models or datasets into Claude Code workflows should keep an eye on how this ownership change affects API and licensing policy over the long run. GeekNews · NVIDIA Blog
• Anthropic publishes a computer-verified proof of Fermat’s Last Theorem, completed by Claude in 11 days (9/5): Anthropic released the first end-to-end, computer-verified proof of Fermat’s Last Theorem, completed largely autonomously by Claude over 11 days of work. This isn’t a newly discovered mathematical proof — it’s the existing proof formalized in Lean so it can be checked by computer from start to finish, making it the largest Lean proof to date at over 13 million lines, and along the way Claude also had to prove roughly 29,000 supporting lemmas. For teams pointing Claude Code at large-scale refactors or migrations that need long stretches of autonomous work, this is a useful data point for how far multi-day autonomous runs can actually go. Anthropic · GeekNews
• Google Antigravity draws controversy over terms banning third-party OAuth reuse and allowing account suspension (9/4): Antigravity’s terms now define using Antigravity’s OAuth login through third-party tools like OpenClaw as a violation, and state that Antigravity and Gemini CLI access can be suspended as a result. Because the original wording referred only to “your account” without further qualification, it raised concerns that an entire Google account could be suspended. If you’re using a third-party tool that reuses subscription-account OAuth, like CLIProxyAPI above, it’s worth checking in advance how each vendor’s terms treat this kind of reuse. GeekNews

Minor Changes

All items below are from v2.1.261 (9/4).

• /status and claude doctor now show an “Organization policy” line explaining why an organization policy failed to load — for example, when a proxy isn’t passing an endpoint through.
• The dangerous rm -rf safety prompt now also catches rm -rf on positional parameters and rm -rf inside double-quoted sh -c scripts.
• When the API sends no response headers at all, the retry wait time now extends up to API_TIMEOUT_MS (10 minutes by default) instead of capping at 3 minutes, with a message explaining what to change.
• Word-based editing shortcuts in the prompt input now match Bash’s behavior. Ctrl+W deletes through whitespace, Alt+F/Alt+D stop at word boundaries and treat punctuation as a word boundary, and the keybindingFlavor setting no longer has any effect.
• [VSCode] The MCP server dialog now has a form for adding servers and a way to remove them, so you can add or delete MCP servers without leaving the IDE.
• [VSCode] Permission and question prompts now have a collapse button, so you can scroll back through the conversation behind a prompt without dismissing it.

Recommended Reads

• “How do you survive code review in the age of AI?”: This piece argues that as colleagues lean harder on AI, the PRs they need reviewed have grown to an average of roughly 6,000 lines of diff, and the belief that a human reviewer must understand how the whole system works makes people hesitant to lean on AI for review help. The core problem: AI’s own explanations are often long-winded and don’t actually help reviewers understand the change. For developers who need to review large volumes of AI-generated code from Claude Code, this piece is less about diff size itself and more a prompt to rethink how review should work. GeekNews
• “DHH on the future of programming: from hands-on coding to running a team of AI agents”: DHH argues that, starting with Opus 4.5 and agent harnesses, AI has shifted from being an assistive tool to being the one actually solving problems end to end. He says his own project, Omarchy Quattro, has had nearly 100% of its code written by agents over the past three months, and that running multiple agents in parallel has changed his entire workflow. If you’re designing a workflow around multiple subagents or agent teams in Claude Code, this is worth reading as a well-known developer’s firsthand account of that shift. GeekNews
• “A simple program isn’t necessarily a small one”: This piece argues that a program’s simplicity isn’t about the size of its code or tools, but about whether distinct features and concepts stay untangled from each other. The key insight: Unix pipelines compose small tools, but once aggregation and sorting get combined, as in sort | uniq --count, even something as small as changing the output order ends up requiring temp files and re-sorting — simplicity breaks down. When designing scripts or pipelines with Claude Code, this is a good reminder to focus on separating concepts rather than just cutting line count. GeekNews

Interesting Projects & Tools

• Stave — a security verification tool that finds attack paths hidden across AWS configurations (9/5): Analyzes not just individual AWS resource settings but the permissions and connections between resources, surfacing compound security risks that individual checks miss. For example, even if S3 encryption and public access blocking both look fine on their own, it will flag a risky access path like Bedrock agent → Lambda → sensitive S3 data. For teams writing and reviewing AWS infrastructure code with Claude Code, this is worth trying as a path-level security check that goes beyond individual resource audits. GeekNews
• Show GN: FlowSuite — a Windows productivity toolset combining no-code screen automation and AI file organization (9/5): A collection of Windows utilities, built up over weekends out of frustration with daily chores like logging into web and system accounts, consolidating spreadsheets, and cleaning up download folders, now packaged as a website. It was built specifically to address how fragile existing PyAutoGUI/AutoHotkey scripts are when window size or position changes. If you’re planning a side project to automate repetitive local tasks, this is worth a look for how its feature set grew out of actual daily use. GeekNews