Claude Code 2.1.247 Adds a Self-Filing Feedback Tool and a Cost Optimizer for Claude API

Claude Code 2.1.247 landed on August 26 with two notable new capabilities and the usual round of targeted fixes. The headline features are a feedback-drafting tool that lets Claude document its own failures, and a cost optimization command aimed at teams running heavy Claude API workloads.

Claude now files its own feedback reports

The new SendFeedback tool gives Claude a way to draft structured feedback when something goes wrong in a session. When a tool keeps failing, when Claude notices it made a mistake, or when you point out a problem, Claude can now generate a feedback report and queue it for your review. You can view, edit, and submit reports from a new /feedback command.

The feature is toggled via the feedbackDrafts setting, so teams that don’t want automated feedback drafting can leave it off. For those who do enable it, this represents a meaningful shift: rather than issues disappearing at session end, they get documented in a format that can actually reach Anthropic or be reviewed by the team.

/claude-api cost-optimize

The /claude-api skill picks up a new cost-optimize subcommand that profiles an existing project’s Claude API spend and works through a checklist of cost levers. It covers caching opportunities, token hygiene (trimming unnecessary context), batch processing eligibility, effort level adjustments, and model choice. Each lever gets evaluated against the project’s actual usage patterns rather than generic recommendations.

This sits alongside the Admin API coverage also added in this release. The /claude-api skill now covers organization members and invites, workspaces and API keys, rate limit reports, workload identity federation, and customer-managed encryption keys (CMEK). Teams using the Admin API to manage Claude Enterprise deployments can now work through it directly from Claude Code rather than switching to the console.

Bash permission prompt gets an auto mode shortcut

The Bash permission prompt now shows a tip pointing to auto mode, with a one-keystroke option to approve the current command and switch to auto mode at the same time. A common pattern when starting a session is dismissing multiple permission prompts before realizing auto mode would have handled all of them. This collapses that sequence into a single confirmation.

Custom spinner tips for organizations

Organizations can now configure their own tips to rotate alongside the built-in spinner messages during long operations. The new spinnerTipsOverride setting accepts entries with id, text, cooldownSessions, and priority fields, plus tipsFile and label options. For teams deploying Claude Code internally, this is a way to surface team-specific documentation or reminders during waits.

Bug fixes

A number of specific issues are resolved in this release:

Arrow-key navigation in menus: Fast arrow-key and Enter sequences were acting on the wrong row in history search, /config, /mcp, /skills, background tasks, and /model. This has been a source of accidental selections in all of those contexts.

Sub-agent 404 fallback: Sub-agents that hit a 404 on their first model call were dying instead of falling through to the session’s fallback model chain. They now retry with full error context.

Background agent error overflow: Background agents printing large volumes of error output could overflow and wedge the parent session. Error output is now bounded.

Ctrl shortcuts on non-Latin keyboard layouts: Keyboard shortcuts using Ctrl were broken in kitty-protocol terminals when the active layout was non-Latin (such as Japanese, Korean, Chinese, Arabic). Fixed.

Bash sandbox symlink cleanup: The sandbox cleanup step was deleting symlinks managed by dotfile tools like nix, home-manager, and stow. Those symlinks are now preserved.

Zed keymap overwrite: Running /terminal-setup was overwriting the entire Zed keymap.json rather than merging changes in. If you ran this recently and lost Zed keybindings, re-run /terminal-setup to restore correct behavior.

Mouse report split across reads: Mouse event reports arriving split across multiple reads could corrupt the prompt text display. Fixed.

The full changelog is at the Claude Code GitHub releases page.