Claude Code 2.1.281 stops proxies from silently cutting replies
The point release also restores prompt caching after MCP disconnects, makes --max-turns halt runaway loops, and adds an auto-mode advisor to /insights.
Copy markdown
Proxy replies no longer masquerade as complete
2.1.281 fixes responses cut short by a proxy that were reported as finished with no warning. If you run Claude Code through a corporate or gateway proxy, earlier versions could hand you a truncated answer that looked whole — worth re-checking anything critical you shipped this week.
Cache survives MCP drops; --max-turns actually stops
Prompt cache is no longer lost when an MCP server disconnects mid-conversation, so a flaky server won't silently force full context re-reads (more latency, more tokens). The same release makes turns respect --max-turns instead of retrying forever, and fixes a crash during API request retries.
/insights now tells you if auto mode is worth it
A new recommendation in /insights estimates how many of your permission prompts auto mode could have handled — a concrete read on whether to turn it on. Gateway users also gain assume_role (STS) and Bedrock guardrails on upstreams.
Elsewhere: Strands Harness, an open Claude Code rival
AWS's Strands team shipped a prebuilt, Apache-2.0 agent you run locally or deploy anywhere — pip install strands-harness. It claims 28% lower token cost across six benchmarks at equal-or-better accuracy, and 77% cheaper than Claude Code when paired with Fable 5.