Claude Code 2.1.247 drafts its own feedback reports

Claude Code 2.1.247 adds a SendFeedback tool that has Claude draft a feedback report when something goes wrong in a session, for the user to review and send from /feedback. The drafting can be switched off with the feedbackDrafts setting, and it arrives alongside more than 30 other entries in the release notes on GitHub.

At a glance

• The feedbackDrafts setting turns the drafting off, and organizations gain new spinnerTipsOverride fields, {id, text, cooldownSessions, priority} entries, tipsFile and label, to rotate their own tips beside the built-in ones.
• /claude-api cost-optimize profiles an existing project's Claude API spend and works through caching, token hygiene, batch, effort and model choice one measured change at a time, while the skill also gains Admin API coverage.
• Sonnet 5 sessions on the 1M window now auto-compact at about 967K tokens instead of about 934K, after the model's default auto-compact window was moved to its full 1M context.

Much of the release sits in places where a single failure ends a session rather than degrades it: a sub-agent hitting a model 404, a hook printing megabytes of errors, a cloud container restarting mid-turn. Fixing those is less visible than a new command, but it is what makes long-running and unattended sessions usable. The feedback tool appears to be aimed at the same problem from the other side, shortening the path from a broken session to a report someone can act on.

Sub-agents now fall back through the session's model chain after a first-call 404

Sub-agents that hit a model 404 on their first call no longer die: they follow the session's fallback model chain, and the error returned to the parent carries the error type, status, request id and model. Hooks and background agents that printed megabytes of error output could previously overflow the conversation and wedge the session on "Prompt is too long", which no longer happens.

When a hook's or background task's output file cannot be written, memory growth is now bounded and the file records where output was lost. Background sessions in claude agents that showed "opening…" indefinitely after their terminal host process died now fail within seconds with a reason, and Enter restarts them.

Shell commands carried over from the foreground no longer log an internal error or print a misleading [exited with code -1] line when they finish inside background sessions.

Ctrl shortcuts fire again under Cyrillic layouts in kitty-protocol terminals

Fast arrow-key and Enter sequences acted on the row above the one navigated to in history search, /config, /mcp, /skills, background tasks and /model; that is fixed. Ctrl shortcuts now fire under non-Latin keyboard layouts, Cyrillic among them, in kitty-protocol terminals.

Text such as <35;150;7M no longer lands in the prompt when a mouse report arrives split across reads right after the escape prefix. The Bash sandbox's after-command cleanup stopped deleting a dotfile-managed ~/.claude/settings.json symlink, as produced by nix/home-manager or stow, when it is repointed outside the sandbox's writable area.

/terminal-setup now merges its keybinding into an existing Zed keymap.json instead of overwriting the file, and /rename no longer confirms silently when the session registry cannot be updated: it now says other sessions may still show the old name. In sessions started with --agent, /compact and "Summarize from here" summarize under the conversation's own system prompt.

Managed settings now cut analytics from startup and halt sign-in when unreadable

Analytics stay off from startup, not only after login, when managed settings force gateway login or a custom OAuth deployment is configured. Sign-in requests to the Claude apps gateway identify Claude Code through a surface=claude_code device-authorization parameter and a claude-code/<version> User-Agent.

Organization sign-in enforcement exits at start when the administrator's managed settings cannot be read, even where host-supplied or per-user Windows registry settings exist. First-run setup no longer exits with "Unable to connect to Anthropic services" when managed settings configure Claude apps gateway sign-in and Anthropic endpoints are unreachable.

Cloud sessions on the web and in the desktop and mobile apps no longer show the previous permission mode when the mode is switched right after a message is sent. A cloud session whose container restarts between turns while a background agent, shell or monitor runs now reports the lost work on resume.

What the notes leave unstated

The release notes do not state whether SendFeedback drafts are limited to particular plans or deployments, and they give no schedule for further /claude-api work beyond the cost-optimize and Admin API additions shipped here. They also leave open whether the Sonnet 5 auto-compact threshold change reaches sessions already open when the update is installed.