[Release] anthropics/claude-code v2.1.243: v2.1.243

Claude Code v2.1.243 introduces significant improvements to usage monitoring, model management, and authentication. Key additions include a Loops breakdown in `/usage` for tracking runaway tasks, a `modelPicker` setting for curating model selections, prompt cache TTL controls, organization-managed pricing, and keyless sign-in via Anthropic Console. The release also includes numerous bug fixes for MCP servers, auto mode, session management, and performance optimizations reducing binary size by ~75% and memory usage by 40-70 MB.

Key Points

• •Added Loops breakdown to `/usage` showing per-loop run count, total tokens, tokens per run, and last run to identify runaway or chatty tasks
• •Introduced `modelPicker` setting to curate `/model` picker with ordered, labeled list of models (supporting Vertex/Bedrock IDs)
• •Added `promptCacheTtl` and `subagentPromptCacheTtl` settings for organizations to maintain 1-hour main conversation cache while subagents use 5-minute cache
• •Implemented `modelPricing` managed setting so organizations use contracted per-model rates and discount multipliers instead of list price for `/cost` and telemetry
• •Added keyless sign-in via Anthropic Console for organizations that don't allow API keys
• •Fixed remote MCP servers auto-reconnecting after dropped connections and improved MCP server sign-in handling for client ID metadata documents
• •Fixed auto mode availability issues at startup and tool call denials after API overload retries
• •Optimized native binary size to ~75 MB (from 340 MB) using zstd compression and reduced memory usage by 40-70 MB through on-demand code loading
• •Fixed cross-session messaging in user namespaces/rootless containers and improved `/resume` to load more sessions on scroll
• •Enhanced startup performance by deferring sandbox/MCP bring-up and reducing workflow discovery/settings work

Workflow Diagram

Concepts