Claude Code Daily Briefing - 2026-11-13

Release Summary

Just two days after v2.1.322 (11/10), v2.1.323 shipped on 11/12. It’s a substantial release: Code Mode, an experimental feature that lets Claude bundle tool calls into JavaScript and run them in one go, makes its debut, alongside an engines field that enforces version compatibility across the plugin ecosystem, and an unusually large batch of sandbox security fixes.

Full release notes

New Features & Practical Usage

Code Mode, an experimental feature that lets Claude bundle tool calls into JavaScript instead of invoking them one at a time, has arrived (v2.1.323)

Until now, Claude Code had to go back to the model after every single tool call to ask what to do next. With Code Mode turned on, Claude can instead write a piece of JavaScript that bundles several tool calls together and run the whole thing at once. It’s still an experimental opt-in feature, off by default, and can be toggled at the organization level. Given that a feature of the same name already exists in Docker’s agent runtime and in OpenCode, where it works by cutting down the number of round trips to the model during multi-step tasks, Claude Code’s Code Mode appears to be aiming for the same payoff.

If you’ve found the round-trip latency and token cost of going back to the model on every step frustrating for tasks that chain together many tool calls, it’s worth checking the official docs and trying this experimental feature. That said, since it’s still experimental, it’s safer to try it out on low-stakes work first rather than rolling it straight into a critical workflow. Full release notes

Developer Workflow Tips

If subagent progress summaries feel like overhead, you can turn them off with CLAUDE_CODE_DISABLE_AGENT_PROGRESS_SUMMARIES (v2.1.323)

This environment variable has actually been working since 2.1.318, but it’s only now been documented. Set it to 1 and both the subagent progress-summarizing behavior and the separate request used to produce those summaries get turned off entirely.

# Turn off subagent progress summaries and the request used to generate them
export CLAUDE_CODE_DISABLE_AGENT_PROGRESS_SUMMARIES=1

If you run several subagents at once and have found the extra requests and token cost from progress summaries wasteful, this environment variable is worth using to cut that cost. Full release notes

Context-window estimates for unfamiliar model IDs jumped from 200K to 1M, which can be a trap when using third-party models (v2.1.323)

When you open a session with a model ID that Claude Code doesn’t recognize, it now assumes a 1M-token context window instead of 200K. If you’re actually running a third-party model with a smaller context window behind an LLM gateway, Claude Code will mistakenly think it has far more headroom than it really does, miss the point where it should compact, and once the session fills up, every turn can keep failing until you run /clear.

# To keep the old fixed 200K assumption
export CLAUDE_CODE_DISABLE_1M_CONTEXT=1

If you’re running Claude Code against a non-Claude model through a gateway and that model’s actual context window is smaller than 1M, after upgrading to v2.1.323 it’s safer to either compact proactively with /autocompact before the session fills up, or set the environment variable above to keep the previous behavior. Full release notes

Claude Code’s internal git and ripgrep calls now run without login tokens, which may break some git hooks (v2.1.323)

The ripgrep calls and most git commands that Claude Code runs internally on its own now execute without most login tokens or API keys. As a result, any credential helper or git hook running underneath these internal git commands that previously relied on Claude Code’s login credentials will no longer be able to find them.

If you’ve customized a git hook or credential helper to build automation that depends on Claude Code’s session token, it’s worth re-checking that behavior starting with v2.1.323, since that token is no longer accessible there. Full release notes

Security & Limitations

Claude service status — a Microsoft 365 connector incident is still causing degraded performance on claude.ai and Claude Cowork as of 11/13 (ongoing)

Checking the official status.claude.com directly confirms that a Microsoft 365 connector error incident reported starting 11/12 is still unresolved. The most recent update (as of 11/12, 22:23 UTC) attributes it to an upstream vendor issue, noting that most requests to connector-based mail, calendar, and search tools are failing, and that Claude for Outlook is affected as well. As of today, 11/13, claude.ai and Claude Cowork are still marked “Degraded Performance,” while Claude API, Claude Code, Claude Console, and Claude for Government all show “Operational.” Around the same time, GeekNews also reported a broad Microsoft 365 outage spanning Outlook, Teams, Exchange, and Partner Center, which Microsoft attributed to a recent change in a health-check component that caused affected servers to fail automated health checks and get pulled from service. GeekNews

If a task using Microsoft 365-family connectors (mail, calendar search, etc.) on claude.ai or Claude Cowork fails today, it’s likely this ongoing incident rather than a local configuration problem — the Claude API and Claude Code CLI themselves are unaffected, so you can keep working from the terminal as usual. Claude Status

v2.1.323 bundles a large batch of sandbox and permission-check security fixes

This release fixes several sandbox flaws that limit the blast radius of commands run without approval. It closes a hole where a symlink inside a writable folder pointing elsewhere or to a nonexistent path would still be granted write access; fixes deny rules and protected-file prompts not applying to files written via cp/mv into a directory; stops the Write and Edit tools from writing to paths covered by a Read deny rule; prevents sandboxed commands from reading Claude Code’s own checkpoint copies; fixes Bash permission checks missing parts of a command; and fixes the sandbox being able to silently turn off mid-session.

If you run Claude Code in sandbox mode and have been relying on the sandbox to contain the blast radius of unapproved command execution, it’s worth upgrading to v2.1.323 to make sure you’re on the version where all of the above are fixed. Full release notes

Ecosystem & Plugins

Reports put Anthropic’s Nasdaq IPO pricing at 11/18, with trading starting 11/19

Following the official roadshow launch covered in the 11/12 briefing ($70–78 per share, 825 million shares), outlets citing CNBC are now reporting that pricing is set for 11/18, with trading on Nasdaq beginning 11/19. IPO calendar site IPOX also lists 11/19 as the first trading day, though Anthropic has not officially locked in the date, so it could still slip. The $70–78 price range, the targeted offering size, and co-lead underwriters Goldman Sachs and Morgan Stanley remain unchanged from before.

If your organization runs on Anthropic products including Claude Code, it’s worth noting that concrete pricing and trading-start dates have now emerged as the next step after the roadshow launch — since these aren’t officially confirmed yet, check next week’s announcements for the actual schedule. Source

The engines["claude-code"] field in plugin.json now enforces plugin version compatibility (v2.1.323)

If a plugin declares the Claude Code version range it supports via engines["claude-code"] in plugin.json, it will no longer install, update, or load on a Claude Code version outside that range, and installs/updates will pick the latest version that fits within it. Until now, even if a plugin was designed to work only with specific Claude Code versions, there was no way to enforce that, which led to cases where a plugin got installed on an incompatible version and then failed with confusing errors. The same release also adds type checking to the plugin’s TypeScript hooks module in claude plugin validate, so running it with --strict can catch these problems ahead of time.

If you build and distribute plugins for Claude Code, it’s worth declaring your supported version range with engines["claude-code"] in plugin.json to head off errors users would otherwise hit from installing the plugin on an incompatible version. Full release notes

Community News

• Google Antigravity unveiled a new coding model, Gemini 4 Hydrogen, along with Orchestrator, a model-switching mode (11/13): Hydrogen is a fast execution model for everyday coding and agentic work, while Orchestrator automatically switches between the higher-tier Carbon model and Hydrogen depending on task complexity. In Google’s Terminal-Bench 4.1 evaluation, Orchestrator kept a success rate close to Carbon’s while cutting average quota usage by 24% and completing tasks 19% faster. GeekNews
• AWS officially launched its own AWS Emulator (11/12): it’s a tool for testing software that calls AWS APIs locally, on dev servers, or in CI without a real AWS account, explicitly scoped to dev/test use and not recommended for production or security testing. Reaction has been largely positive, though some have raised concerns that security policies like IAM don’t behave identically to real AWS, and about the quality of ongoing maintenance. GeekNews

Minor Changes

All of the following are from v2.1.323 (11/12).

• Added .docx support for asset uploads in the Artifact tool.
• “Don’t ask again” allow rules that appear to contain secrets are now kept session-only instead of being written to disk when they’d otherwise be saved into a git-tracked settings.local.json.
• The number of messages sessions on the same machine can send each other without user confirmation jumped from about 20 to 100,000.
• Claude now asks for confirmation before running /security-review.
• Running which with flags other than lookup/display now prompts for permission.
• Updating an artifact that’s externally visible or whose access permissions can’t be verified now prompts for confirmation first, even if an allow rule exists.
• Fixed an issue where a headless (claude -p) session failing to save its config file due to low memory would wipe login credentials, causing Claude Code to fail to start at all.

Recommended Reads

• Jane Street is scaling back live coding in undergraduate on-site interviews for the AI era (11/13): starting summer 2026, it’s piloting new interview rounds that assess software design, code verification, and skill with LLMs, on the reasoning that as the cost of generating code drops, judgment about what to build and the ability to verify correctness matter more. Comments were split between support for having candidates explain LLM-generated code themselves as an effective approach, and skepticism that the new interview format’s effectiveness has actually been validated. If generating code quickly with Claude Code has become routine for you, this is a signal worth watching, since it suggests the skills the job market evaluates are shifting too. GeekNews
• 62% of data center expansion costs get passed on to residents, and low-income, rural households bear a disproportionate share (11/13): Cornell researchers analyzing four U.S. power grids in 2025 found that additional energy costs for the lowest-income households reached as much as 28% of their income. Comments were split between the view that the root cause of power shortages is policy that makes it hard to build new generation capacity, and the view that data centers should directly bear the transmission costs they themselves create. It’s a piece that surfaces something invisible to anyone just using an AI model: who ends up footing the bill for the infrastructure expansion behind it. GeekNews
• smolmaxxing — a writeup on running LLMs on aging hardware like TV boxes and a 2013 iPad Air (11/13): using a custom kernel, memory optimization, and weight streaming, the author got Qwen3.5-2B running at 7.7 tok/s on a TV box and Bonsai-1.7B running at 23.7–25.8 tok/s on the iPad Air. Measurements showed flash bandwidth becomes the bottleneck for models larger than available RAM, and commenters noted that modern AI agents can now do this kind of low-level optimization work quickly. GeekNews

Interesting Projects & Tools

• Show GN: TokenWatch — a phone dashboard app for tracking usage and reset times across AI services like Claude and Codex (11/12): an iOS/Android app that lets you connect your own AI accounts to check remaining usage and the next reset time from a single screen on your phone, with a demo screen showing sample data before you connect an account. It’s well suited to leaving open next to your monitor as a persistent display, and it’s free until November 16. App Store · Google Play · GeekNews
• Show GN: Ssamcut — a browser-based video editor that Claude and ChatGPT can operate directly via MCP (11/13): a free, install-free browser video editor that processes video, photos, and audio entirely in the browser without uploading them to a server, with support for auto-captioning, multi-track editing, and keyframe animation. It also exposes a remote MCP server, so once connected to Claude or ChatGPT as a custom connector, the AI can directly operate the open editing screen to handle cutting, captions, effects, and exporting. Ssamcut editor · GeekNews