Claude Code v2.1.243: Loops Breakdown, Model Picker, Cache TTLs, and Stability Fixes

Claude Code v2.1.243 is out, and it's a solid release for anyone running /loop tasks, managing models, or wrestling with prompt cache costs. Here's what changed and what it means for your daily workflow.

New in /usage: Loops breakdown

If you rely on /loop for iterative tasks, you can now see exactly what each loop costs. The /usage command now includes a Loops breakdown with per-loop run count, total tokens, tokens per run, and the last run time. This makes spotting runaway or chatty loops trivial.

Curate the /model picker with modelPicker

The new modelPicker setting lets you define an ordered, labeled list of models for the /model picker. You can use any model ID spelling, including Vertex/Bedrock IDs, and either append to or replace the built-in lineup. This is a game-changer for teams standardizing on specific models or regions.

Control prompt cache TTL

API-key and cloud-provider users can now set promptCacheTtl and subagentPromptCacheTtl to keep a 1-hour prompt cache on the main conversation while subagents stay at 5 minutes. This is a practical way to manage cache costs for long-running sessions.

Custom pricing for /cost and telemetry

Organizations with negotiated rates can now use the modelPricing managed setting to apply contracted per-model rates and discount multipliers to /cost, the status line, and telemetry. No more inaccurate list-price reports.

Keyless sign-in via Anthropic Console

For orgs that don't allow API keys, /login now offers "Sign in with your Console account" as a recommended alternative to creating an API key. This simplifies onboarding and aligns with enterprise security policies.

Skipped sources and managed markers

/status now shows a Skipped sources line listing managed settings sources (e.g., managed-settings.json) that were present but not applied because a higher-precedence source is active. Also, /mcp and /plugins now mark managed connectors with a managed label, so you know which auth is under your org's control.

GitHub connection status for web users

Claude Code on the web (Pro/Max) now shows a /status line indicating whether GitHub is connected, with a pointer to /web-setup if not. There's also a tip nudging users to connect GitHub.

Subagent model visibility

/tasks and agent detail dialogs now display the model (and effort level) each subagent ran on, giving you better insight into how your subagents are executing.

Key fixes and stability improvements

• Remote MCP servers: Now recover automatically or report failure after a dropped connection in non-interactive (-p) and SDK sessions.
• Auto mode: Fixed startup unavailability when a temporary server-side disable was cached, and fixed tool calls being denied after ~1 min when the API was overloaded.
• /model picker: Ultracode selection now applies correctly.
• /resume: Now loads more than 50 sessions as you scroll.
• API timeout: Sessions no longer go silent for 10+ minutes; requests time out after ~3 minutes, retry once, then show API Error: No response from API.
• Hook conditions: Fixed Bash(cat *) firing on unrelated commands with command substitution.
• Plugin dependencies: Fixed marketplace field resolution and /reload-plugins LSP handling.
• --agents: Now exits with a clear error for invalid JSON or agent definitions, instead of silently ignoring.

This release also includes fixes for workload identity federation in CI, cross-session messaging in rootless containers, and text rendering glitches in /login.

For developers running Claude Code in CI or with complex MCP setups, the reconnection and federation fixes are particularly valuable. The loops breakdown is a must-have for anyone using /loop in production.