Your claude.ai skills follow you into Claude Code 2.1.275

Claude Code 2.1.275, released on September 17, brings the skills and plugins enabled on your claude.ai account into every terminal session signed in with that account. The same release adds a key that interrupts the current turn and sends everything you have queued.

The release notes on GitHub cover the CLI, the VS Code extension, Claude Code on the web, Claude Tag and Code Review. This article covers the changes that affect daily work, the settings you may want to change, and the fixes that change what a command returned.

Account skills in the terminal

Enabled skills and plugins now follow the account into the CLI. There are two opt-out settings: syncClaudeAiSkills: false for skills and syncClaudeAiPlugins: false for plugins. The sync is on by default, so keeping a terminal limited to what the repo defines takes both of those settings.

Each synced skill costs little context. At session start Claude Code reads only a skill’s name and description, roughly 30 to 100 tokens each. It loads the full instructions only when a request matches the skill. This approach is known as progressive disclosure, and it lets a large library of skills be registered without bloating the context window.

Two smaller changes support the sync. The ListPlugins tool description now tells Claude that it lists plugins enabled on your claude.ai account, not plugins installed locally with /plugin. When Claude writes to or edits a file in the synced account-skills folder, the result now says the change is not saved to your account and explains how to save it. Plugin installs also take one step less: /plugin install <plugin> --marketplace <source> offers to add the marketplace before installing the plugin.

Sending queued messages now

The send-now key is ctrl+enter, or ctrl+x ctrl+s. It interrupts the current turn and sends all queued messages at once. Sent and queued messages show in gray until the model receives them.

The VS Code extension also changes how messages sent mid-turn behave. A message sent while Claude is working now waits at the bottom of the conversation until Claude starts on it. Slash commands typed while Claude is responding now run once the response finishes. Before, they reached the model as plain text. A new claudeCode.scrollToBottomOnSend setting turns off the jump to the bottom on send. The conversation also no longer pulls you back down while a reply streams after you’ve scrolled up.

Prompt cache and usage reads

One fix affects compacted and resumed sessions. After a compaction or resume, a restored memory file’s age note could change between requests, which caused prompt cache misses. That no longer happens.

If you pass a string --system-prompt, the text above a __SYSTEM_PROMPT_DYNAMIC_BOUNDARY__ line is now cached globally. The SDK’s array form already worked this way. On one machine, editor windows and non-interactive sessions now share a plan-usage read made in the last minute. Previously each one called the usage endpoint separately.

VS Code, Artifacts and Claude Tag

• VS Code: accept and reject buttons under each change in the proposed-change diff tab; you can view, edit and delete saved memories in the Memory dialog; an attached image can be sent with no text; renaming a running session no longer reverts to the generated name (a regression in 2.1.269).
• Pasted and attached images are saved where Claude can open them as files without a permission prompt, including in Desktop and VS Code.
• Artifacts: the tool asks for a one-word tab icon on a first publish instead of an emoji favicon. Scheduled and Run now routines update an artifact you can edit without asking. Public artifacts, first publishes and deletes still ask.
• Claude in Chrome in auto mode skips the extension’s per-site check for classifier-approved calls, fixing browser_batch “Permission denied” after a redirect.
• Claude Tag: new presets for Amazon CloudWatch, CloudWatch Logs, Amazon SNS, Google Cloud Monitoring and Cloud Logging, and Datadog presets for the US3, AP1, AP2 and US1-FED sites. New Datadog connections are limited to Datadog’s read and query API routes.
• Code Review: pull requests with more than 100 Claude reviews now get the lighter merge-focused review on a clean merge from the base branch instead of a full re-review.

Fixes that change results

Some of the CLI fixes change what a command actually did. Users behind a network gateway that rewrites API error responses got a terminal API Error: 400 on every turn when a beta request header was rejected. /update-config wrote Write(path) permission rules, which file permission checks don’t match, instead of Edit(path) rules. Grep, Glob and @-file suggestions could hang or run out of memory on searches over the 20MB output cap. System ripgrep could also report “no matches” instead of an error after a flood of warnings.

Sandboxed Bash on Linux couldn’t write to project directories named hooks/ or config/, and with zsh as the shell it reported exit code 0 for failed commands. A failing otelHeadersHelper now produces a startup warning, so a session that exports no telemetry no longer goes unnoticed.

Gateway sign-in and plugin credentials

If you sign in through a Claude apps gateway, the gateway can now name the account being signed in. You confirm it before the credential is saved, and /status shows it afterwards. /logout now also ends the session on gateways that advertise token revocation. This is the second gateway change in two releases: 2.1.274 added a setting for the Postgres connect timeout of the gateway’s store.

Several fixes deal with secrets. Plugin and marketplace messages, logs and claude plugin marketplace list no longer show a password or token stored in a git, ssh or marketplace URL. The VS Code Manage plugins dialog gets the same fix for a token typed into a marketplace URL. claude plugin marketplace update no longer deletes a GitHub marketplace’s local copy when the fetch fails and the marketplace is named after its repository.

Plugins from an npm source are now fetched with npm pack --ignore-scripts and integrity-verified, so a package’s install scripts no longer run.