Claude Code v2.1.247 adds feedback tool, API cost profiler, and admin API coverage

Anthropic released version 2.1.247 of Claude Code, which adds a SendFeedback tool to help users draft and send feedback reports from the /feedback command. The setting feedbackDrafts can be used to disable drafts.

The update introduces a /claude-api cost-optimize command to profile an existing project's Claude API spend and work through cost levers such as caching, token hygiene, batching, effort, and model choice one measured change at a time.

Administrative API coverage has been added to the /claude-api skill, including support for organization members, invites, workspaces, API keys, rate limit reports, workload identity federation, and CMEK.

The release fixes multiple issues, including fast arrow-key and Enter sequences acting on the wrong row in history search, /config, /mcp, /skills, background tasks, and /model; sub-agents dying on a first-call model 404 by falling back to the session's model chain; and overflow from hooks or background agents printing excessive error output that could wedge sessions on 'Prompt is too long'.

Additional fixes address terminal behavior under non-Latin keyboard layouts, mouse report parsing inserting unintended text into prompts, and the Bash sandbox's cleanup deleting a dotfile-managed ~/.claude/settings.json symlink in nix/home-manager or stow setups.

Terminal-setup commands no longer overwrite the entire Zed keymap.json and /rename now reports when the session registry cannot be updated. Sessions started with --agent now summarize under the conversation's own system prompt for /compact and 'Summarize from here'.

Background sessions no longer show 'opening…' indefinitely after their terminal host process dies, and memory growth is bounded when output files cannot be written. The /install-github-app command over SSH now shows how the sign-in URL was copied and displays the URL immediately when no browser is available.

Marketplace plugin caching issues were fixed for version-less plugins, and Remote Control sessions now report working-tree diffs to connected clients. Self-hosted runner sessions no longer report running before Claude Code has started, preventing premature notifications.

First-run setup no longer exits with 'Unable to connect to Anthropic services' when managed settings configure gateway sign-in and endpoints are unreachable. Cloud sessions now correctly show the current permission mode when switching modes, and resume sessions after container restarts while background agents or monitors are running.

Plugin marketplace hardening prevents names with control or invisible characters, and marketplace-supplied text in /plugin and claude plugin output is now escape-safe. Bedrock, Vertex, and Foundry sessions with telemetry disabled now receive connection failure notices for MCP servers.