Claude Code Daily Briefing - 2026-10-28

Release Summary

Just a day after v2.1.311 landed on 10/26, v2.1.312 shipped on 10/27. Beyond the new OIDC claim setting and the gateway-specific system prompt change, this release packs in several stability fixes with real-world impact: a data-loss risk where Claude would read only the start of a very long CLAUDE.md or rules file and then overwrite the whole thing, a bug where skills scoped with paths: went undiscovered after cd-ing into a subdirectory via --add-dir, and an issue where a single misnamed command could block all command loading, leaving claude -p producing no output at all.

Full release notes

New Features & Practical Usage

oidc.user_id_claim added to the Claude apps gateway (v2.1.312)

The Claude apps gateway can now identify users using whichever id_token claim your organization actually relies on - such as Microsoft Entra ID’s oid - instead of being locked into the default sub claim. Previously, only the sub claim could be used for user identification, forcing organizations with their own identity schemes to maintain a separate mapping layer.

# Example Claude apps gateway configuration
oidc:
  user_id_claim: oid  # for Microsoft Entra ID

If you’re connecting the Claude apps gateway to an OIDC provider like Entra ID that identifies users by a claim other than sub, this setting lets you tear out whatever custom user-mapping logic you’ve been maintaining. Full release notes

Shorter system prompts now apply to unrecognized model names on Bedrock, Vertex, Foundry, and Mantle (v2.1.312)

Even if you’re running Sonnet or Haiku under a name Claude Code doesn’t recognize - a custom deployment name or an inference profile ARN - Bedrock, Vertex, Foundry, and Mantle will now apply the shorter system prompt. If you’d rather keep the full prompt, you can turn this off with an environment variable.

# To keep the original full system prompt
export CLAUDE_CODE_SIMPLE_SYSTEM_PROMPT=0

If you’re running Claude Code through Bedrock, Vertex, or Foundry with a custom deployment name or ARN, it’s worth checking whether your token usage drops after v2.1.312 - and keeping this environment variable in mind if you need the old behavior back. Full release notes

Developer Workflow Tips

The Write tool no longer overwrites large CLAUDE.md and rules files after reading only part of them (v2.1.312)

A bug where Claude would read only the beginning of a very long CLAUDE.md, rules file, or @-included file in a subdirectory and then overwrite it outright has been fixed - Write now requires Claude to either read the entire file or edit it in place. Previously, there was a real risk that the latter portion of a file would simply vanish.

If you maintain long CLAUDE.md or rules files across multiple subpackages, you can now trust that this fix prevents Claude from carelessly overwriting content it hasn’t seen. Full release notes

Skills scoped with paths: are rediscovered after a Bash cd into a subdirectory (v2.1.312)

A bug where directory-scoped or paths:-conditioned skills went undiscovered after cd-ing into a subdirectory of a git repo via Bash has been fixed. Previously, moving to a deep path in the shell could cause skills designed for that exact path to suddenly disappear from view.

If you run skills with paths: conditions per package in a monorepo, this fix will matter more the more often your work involves moving across subdirectories via Bash. Full release notes

Prompt and agent hooks are now capped at 24 concurrent executions per session (v2.1.312)

Concurrent prompt and agent hooks within a single session - including those from subagents - are now capped at 24, with up to 12 additional hooks allowed to queue past their individual timeout, and any timeout set below 1 second is now treated as 1 second. Previously, there was no limit on how many hooks could fire concurrently when multiple subagents triggered them at once.

If you’re running automation that spins up many subagents or forks concurrently and triggers hooks on every turn, it’s worth reviewing your hook design with this concurrency ceiling in mind, since hooks can now get queued once you exceed it. Full release notes

Security & Limitations

Claude service status - both 10/27 incidents resolved, all services operating normally as of 10/28

Checking status.claude.com directly: an investigation into an increase in errors across Claude services began at 09:14 UTC on 10/27 and was resolved by 09:56 UTC; separately, an increase in Claude Sonnet 5.5 errors starting at 14:42 UTC affected claude.ai, the Claude API, Claude Code, and Claude Cowork, with impact ending at 15:17 UTC and the incident marked resolved at 15:50 UTC. As of 10/28, no new incidents have been reported, and claude.ai, the Claude API, Claude Code, Claude Console, Claude Cowork, and Claude for Government are all showing Operational status.

If you hit errors sending requests to Claude Sonnet 5.5 on the afternoon of 10/27, it was most likely this already-resolved incident - and since nothing new has been reported today, it should be safe to retry. Claude Status

A ruling has found that Anthropic trained Claude on illegally obtained books by Stanford professors (10/28)

It’s been confirmed that Anthropic illegally obtained works by more than 10 Stanford professors and used them to train Claude, and federal judge William Alsup has issued a ruling barring the practice of building a permanent central library out of pirated books. The ruling draws a line between the piracy itself and fair use in AI training, treating the use of legitimately acquired works for training as a separate matter.

If you’re reviewing training-data copyright issues while adopting Claude Code or the Claude API in an enterprise setting, it’s worth watching how this ruling draws the line between pirated-book collection and legitimate training use. GeekNews

A Tailscale SSH privilege escalation vulnerability is under investigation - disabling unneeded SSH forwarding is recommended (10/28)

Most of the issues identified in Tailscale SSH are privilege escalation problems where a user who already has SSH access can, under certain conditions, act beyond the scope of a specified local account. If you don’t need SSH forwarding, it’s recommended to set TS_SSH_DISABLE_FORWARDING=1 in the tailscaled service environment to reduce the attack surface.

If you’re connecting to a Claude Code self-hosted runner or a remote dev environment over Tailscale SSH, it’s safer to disable unneeded SSH forwarding until the investigation concludes. GeekNews

Ecosystem & Plugins

Show GN: Claude Code Desk Look (mod) - turns the Claude Code CLI into a desktop-app-like experience (10/27)

Built on the premise that people who prefer the terminal-based Claude Code CLI still find the Electron-based desktop app too heavy for running multiple sessions at once, this mod wraps an already-logged-in claude -p (and similar) to make it look like a desktop app. It keeps the lightness of the CLI while adding the window separation and visual layout typical of a desktop app.

If you’ve been running several Claude Code sessions in parallel in the terminal but found window-switching cumbersome, this mod is worth a try - it keeps the CLI’s lightness without needing the Electron app. GeekNews

Show GN: Translate Book: arXiv - a Claude Code skill that translates arXiv papers while preserving the original structure (10/27)

To solve the problem of feeding English papers into a generic translator and ending up with mangled equations, figures, and tables, this Claude Code skill doesn’t translate the PDF directly - instead it converts the arXiv source into a translated version that preserves the original document structure.

If you read arXiv papers often and have found yourself going back to the original because a translator kept breaking the equations and tables, adding this skill to Claude Code is a useful way to get a translation that preserves the source structure. GeekNews

Community News

• After repeated setbacks, Amazon is sharply scaling back its AI division (10/28): Amazon is scaling down development goals for most of its in-house Nova models and reorganizing its AI development organization to focus on top priorities, while its text and image/video generation models won’t be discontinued outright but will shift to receiving only the minimum resources needed to keep running. GeekNews
• Apple unveiled a new MacBook Pro 14 and iMac powered by M6 (10/28): With a 12-core CPU, 12-core GPU, a Neural Accelerator per GPU core, and a dual 16-core Neural Engine, the new chip boosts on-device AI workload performance - like running local LLMs and image generation - by up to 25-30% over the previous generation. GeekNews
• Is my data used to train the model? (10/28): Anthropic does not use inputs and outputs from commercial products like Claude for Work or the Anthropic API for model training by default, and chat or coding sessions are only used for training when a user submits feedback, explicitly reports a bug, or directly opts in to data usage. GeekNews

Minor Changes

All of the items below are from v2.1.312 (10/27).

• Clicking the “send now” hint beneath a queued message in full-screen mode sends it immediately, just like pressing the key.
• Usage limit messaging has been unified from “session limit” to “5-hour limit.”
• Automatic model switching after a refusal message now extends to unrecognized model names, including custom deployment names; it can be disabled with CLAUDE_CODE_DISABLE_REFUSAL_FALLBACK=1.
• The name and branch name of a cloud session started from the CLI are now fetched from the server.
• The OpenTelemetry mcp_server.name for pinned-URL remote MCP servers declared by official Anthropic marketplace plugins now shows the actual name instead of custom.
• Claude Code has stopped sending the claude_code.safety_event OpenTelemetry event.
• [Code Review] Fixed an issue where a delayed GitHub push event could cause an outdated commit to be reviewed, or no review to run at all - Code Review now reviews the PR’s current commit.

Recommended Reads

• AI promised to reduce my workload. I’m working more than ever (10/27): An account from someone using AI agents across all their engineering work, describing how lower implementation costs led not to finishing existing work faster, but to taking on a larger workload overall. The observation that multiple agents intermittently interrupting with questions, plan approvals, and code reviews adds up to more context-switching will resonate with anyone running several tasks at once through Claude Code. GeekNews
• Fixing vLLM’s tool-call accuracy problem (10/27): When tool calls violate their schema, the resulting errors and retries pile up and hurt both agent quality and cost - and in Ubicloud’s testing with Kimi K2.6, vLLM 0.28’s schema accuracy came in at just 92%. The piece’s approach of using constrained decoding to filter out schema-violating tokens at each generation step offers food for thought on how to improve reliability for agents like Claude Code that depend heavily on tool calls. GeekNews
• Software may be getting worse, but it used to be terrible (10/28): Even granting that modern software has real quality problems, this piece pushes back on the nostalgia that older software was better, arguing it forgets the frequent crashes and lost work that were routine after 1990. Revisiting the pain of the past - buggy drivers and operating systems, weak security, manual disk defragmentation - helps keep the quality of AI-generated code from being judged too harshly, or too generously, against a rose-tinted baseline. GeekNews

Interesting Projects & Tools

• Survival LLM - an AI assistant and emergency-response library that works without internet (10/28): A free, open-source app that stores an emergency-response library and an AI model locally on your computer, so once the initial download is done, you can ask questions and search the material entirely offline. It answers questions like food storage during a power outage, water purification, or first aid by searching the library, and you can open the cited reference document to verify the source directly. GeekNews
• Show GN: notepipe - a lecture-note and meeting-minutes pipeline you assemble yourself (10/27): While plenty of services already turn recordings into lecture notes or meeting minutes, this project is for people who dislike the canned summary format or specifically need local processing - it lets you assemble each stage of the pipeline yourself to build something that matches your own preferences. GeekNews