Claude Code v2.1.275 Adds Send-Now Key, Skill Sync

Claude Code v2.1.275 is out, and it's mostly a quality-of-life release with a handful of genuinely useful interaction changes plus a long tail of crash and cache fixes. Here's what actually matters for daily use.

New: send-now key

You can now interrupt the current turn and flush everything you've queued in one shot with ctrl+enter (or ctrl+x ctrl+s). Sent and queued messages both render in gray until the model actually receives them, so you can tell what's been dispatched versus what's still sitting in the queue.

New: claude.ai skill and plugin sync

Terminal sessions signed in with a claude.ai account now pull down the skills and plugins you've enabled on that account. Two opt-outs exist if you don't want that behavior:

syncClaudeAiSkills: false
syncClaudeAiPlugins: false

New: gateway sign-in confirmation

Claude apps gateway sign-in now shows the signed-in account — if the gateway names one, you confirm it before the credential is saved, and /status displays it afterward. Separately, a startup warning now fires when a configured otelHeadersHelper fails, so sessions silently exporting zero telemetry get surfaced instead of just looking quiet.

New: marketplace-aware plugin install

/plugin install <plugin> --marketplace <source> offers to add the marketplace before installing, so you don't have to do the two-step dance manually.

Notable fixes

A restored memory file's age note changing between requests after compaction or resume was causing prompt cache misses — fixed.
--forward-subagent-text stream-json and SDK output were dropping messages from subagents spawned by context: fork skills (and from forked skills invoked by other forked skills).
Resume paths hardened against malformed transcripts: bad task-reminder entries, @-file attachments, malformed message entries, and malformed message content blocks were crashing --resume, the picker preview, background agents, and the transcript view.
Sandboxed Bash on Linux was reporting exit code 0 for failed commands under zsh.
Terminal API Error: 400 on every turn behind network gateways that rewrite error responses when a beta request header is rejected.
Grep, Glob, and @-file suggestions were hanging or OOMing on searches over the 20MB output cap; system ripgrep could also report "no matches" instead of an error after a flood of warnings.
/rewind in forked or background sessions could restore a zero-filled or truncated file when file-history backups weren't fully copied.
Fullscreen mode was placing background-task completion notices in the wrong spot and freezing for seconds when scrolling past large diffs.
Plugin and marketplace logs, messages, and claude plugin marketplace list were leaking passwords or tokens embedded in git, ssh, or marketplace URLs. Also fixed: claude plugin marketplace update deleting a GitHub marketplace's local copy when the fetch failed and the marketplace was named after its repo.
Vim mode cursor off-by-one after a dot-repeated !, and fast-typed i! flipping a non-empty prompt into shell mode.
Read tool hanging instead of erroring when part of a large file couldn't be decoded under memory pressure.
--resume / --continue dropping earlier thinking when a tool it started with has since been disabled by a server-side flag.
Mouse-selected text in the fullscreen claude --resume picker never reaching the clipboard.
A stray </ccmemory>-style closing tag occasionally appearing in responses.

If you're on a signed-in claude.ai account and don't want your skills/plugins mirrored into the terminal, set the two sync flags before upgrading. Otherwise it's a low-risk update — most of the diff is resilience work around resume and transcript parsing.

📖 Read the full source: GitHub Claude-Code