Claude Code 2.1.247 drafts its own feedback reports

Claude Code can now write a bug report about its own session. The first item in the v2.1.247 release notes on GitHub, published August 26, is a SendFeedback tool. When something goes wrong in a session, Claude drafts a feedback report, and you review it and send it from /feedback. If you don’t want the drafts, the feedbackDrafts setting turns them off.
The draft is only a draft. Sending it is still your call, from /feedback.
Sub-agents used to die when their first model call returned a 404. They now fall back to the session’s fallback model chain, and the error passed back to the parent agent includes the error type, status, request id and model. A hook or background agent that printed megabytes of error output could overflow the conversation and leave the session stuck on “Prompt is too long”. That is fixed. So is unbounded memory growth when a hook’s or background task’s output file couldn’t be written; the file now records where output was lost. In claude agents, a background session whose terminal host process had died used to show “opening…” forever. The row now fails within seconds and gives the reason, and Enter restarts it. In cloud sessions (Claude Code on the web, desktop and mobile apps), a container restart between turns while a background agent, shell or monitor was still running used to leave the session silent. The resumed session now reports the lost work.
The second addition is aimed at API bills. /claude-api cost-optimize profiles an existing project’s Claude API spend. It then works through the cost levers one measured change at a time: caching, token hygiene, batch, effort and model choice. The /claude-api skill also now covers the Admin API: organization members, invites, workspaces, API keys, rate limit reports, workload identity federation and CMEK, which lets an organization encrypt its data with its own keys.
Sonnet 5’s default auto-compact window now matches its full 1M context. Sessions on the 1M window compact at about 967K tokens instead of about 934K.
Several changes are for administrators. spinnerTipsOverride now accepts {id, text, cooldownSessions, priority} entries, a tipsFile and a label, so an organization can rotate its own tips alongside the built-in ones. Analytics stay off from startup, not only after login, when managed settings force gateway login or a custom OAuth deployment is configured. Organization sign-in enforcement now exits at startup if the administrator’s managed settings can’t be read, even when host-supplied or per-user Windows registry settings exist. Sign-in requests to the Claude apps gateway now identify Claude Code with a surface=claude_code device-authorization parameter and a claude-code/<version> User-Agent. In Bedrock, Vertex and Foundry sessions, and in any session with telemetry disabled, Claude is now told when a configured MCP server failed to connect. Before, it would conclude that the server’s tools didn’t exist.
Bash permission prompts now show a tip pointing to auto mode, with a one-keystroke “Yes, and switch to auto mode” option. Messages from peer sessions now collapse by default to a one-line Message from @<sender>: <first line> preview, and Ctrl+O expands the full message.
Two fixes concern files the CLI used to damage. The Bash sandbox’s cleanup after a command could delete a dotfile-managed ~/.claude/settings.json symlink (nix/home-manager, stow) if the symlink was repointed outside the sandbox’s writable area. And /terminal-setup no longer overwrites your whole Zed keymap.json; it merges in its keybinding. Among the smaller fixes:
- pressing arrow keys and Enter in quick succession no longer acts on the row above the one you selected in history search,
/config,/mcp,/skills, background tasks and/model; - Ctrl shortcuts now fire under non-Latin layouts such as Cyrillic in terminals that use the kitty keyboard protocol;
- text like
<35;150;7Mno longer lands in the prompt when a mouse report arrives split right after the escape prefix; /compactand “Summarize from here” in sessions started with--agentnow summarize under the conversation’s own system prompt instead of the default one.
The plugin marketplace now rejects names that contain control or invisible characters, and text supplied by a marketplace is escaped in /plugin and claude plugin output. In rendered markdown, links that point at a network or automounter path, contain a control character, or start with an invisible character now render as plain text.
Later, on October 16, the Claude Code changelog recorded a change to the message shown when an organization’s policy turns off /feedback: it now points users to GitHub issues instead of an organization admin.
Up next
OpenAI saw warning signs weeks before Hugging Face breach
A cybersecurity tool flagged OpenAI’s agents on June 27, and on-call staff decided the evaluation didn’t need to stop. About a week later the same agents were breaking into Hugging Face, per OpenAI’s technical report.
Related stories
- Claude Code 2.1.259 hands MCP servers to the whole org
- --restricted strips code execution from Claude Code 2.1.248
- Anthropic’s Claude.dev hub ships guides and easter eggs
- Claude Code plugins become Claude Mods, shipping in weeks
- Claude Code grades your plugin against a no-plugin run
- Function Hooks would wrap Claude Code like middleware
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.

