Claude Code 2.1.265 landed on Tuesday, and 2.1.266 followed a few hours behind it, because a single environment variable had taken down every gateway setup. In order.
What 2.1.265 adds
The changelog is unusually long — a good forty items. The genuinely new bits first:
--plugin-dircan now point at a folder of plugins. Every child folder with a manifest loads, and folders added or removed while Claude Code is running get picked up without a restart.- Tool results written to disk get a 1 GB cap. The in-conversation preview tells you when a saved file was truncated.
- Claude Desktop and Cowork now send
user.emailanduser.groupsin telemetry through a Claude apps gateway, matching what terminal sessions already did. Useful for admins, worth a read for everyone else.
The quieter block underneath interests me more: several fixes to the prompt cache. Resuming a foreground-spawned subagent used to change its tool list and system prompt prefix; agent teammates and resumed subagents pushed hook context and preloaded skills out of that prefix on later turns. Both wrecked cache reuse, and cache reuse is money. These are the lines people skim past, and they’re often the reason a long session suddenly costs more than it should.
Then the small stuff that shows up in daily work. After a crash mid-tool-call, resuming no longer rewrites your last prompt — the interrupted call stays and is marked as interrupted. /model opusplan[1m] is no longer rejected with “Model not found”. MCP servers configured as http that only speak the legacy HTTP+SSE transport now fall back to SSE instead of never connecting. And in non-interactive sessions a cd finally persists across turns, which has cost me time before.
The rollback
CLAUDE_CODE_USE_GATEWAY is an undocumented variable that used to be ignored unless ANTHROPIC_BASE_URL and ANTHROPIC_AUTH_TOKEN were both set. In 2.1.265 it started forcing Cloud-gateway sign-in on its own. Anyone who had it sitting next to an API key, an apiKeyHelper or custom auth headers got “Not signed in to the Cloud gateway” on every single request.
2.1.266 reverts exactly that. No config change needed — just update.
Reading the tea leaves
Shipping a release this size and patching it the same day doesn’t look great on paper. I’d still take it over waiting three days for the fix. What stands out is where the work is going: gateway, Remote Control, plugin management. That’s team infrastructure, not solo-developer-at-a-terminal infrastructure.
Sources: Claude Code changelog