Claude Code 2.1.276 fixes a 400 crash that broke every proxy setup
The crash hit anyone routing through a gateway since yesterday's 2.1.275. Also new: a send-now key, smarter plugin installs, and a sub-1.58-bit quant trick.
Copy markdown
The 400 that broke your gateway
Claude Code 2.1.276 fixes a regression where every request returned `400 Input tag 'advisor_20260301'` when `ANTHROPIC_BASE_URL` points at a proxy or gateway — introduced one build earlier in 2.1.275. If you route Claude Code through any OpenAI-compatible proxy, a corporate egress gateway, or a self-hosted router, upgrade before your next session.
Send-now, minus the wait
2.1.275 adds a send-now key (Ctrl+Enter, or Ctrl+X Ctrl+S) that interrupts the current turn and fires all your queued messages at once. No more waiting for Claude to finish a long turn before your follow-ups register.
Install plugins from a fresh marketplace in one line
`/plugin install <plugin> --marketplace <source>` now adds the marketplace and installs in a single command. The same build stops credentials leaking in plugin and marketplace URLs, and un-buries @-mention file suggestions that were sitting below MCP resources.
The prompt-cache miss that was costing you tokens
2.1.274 and 2.1.275 fix memory-file age notes silently breaking your prompt cache — a needless cache miss on turns you never changed. Also patched: subagent messages dropping out of stream-json output, and large-file diffs freezing the UI.
Elsewhere: sub-1.58-bit weights are on the table
A new method, BITCOS, packs ternary LLM weights to 1.485 bits — past the 1.625-bit five-trit baseline — by exploiting that up to 51% of weights are zero. It reports up to 1.28x faster matrix-vector math and beats five-trit packing on 26 of 29 models tested, so local models get a little smaller on the RAM you already own.
Elsewhere: an open-source browser agent you can run tonight
browser-use shipped jev-ultrafast (MIT): a browser agent with a dynamic, indexed action space that ran a Zurich-to-London flight search in 7.1s and cut browser protocol calls from 1,092 to 101. Point it at a URL and a plain-English goal to watch it move.