Claude Code 2.1.288 brings back the prompt you Ctrl+C’d

Anthropic shipped Claude Code 2.1.288 on October 2. You can now recover a prompt you cleared with Ctrl+C, and the release fixes a long run of permission, resume and plugin bugs, according to the release notes on Github. If your team runs Claude Code in bypassPermissions mode or under a shell allow rule, the permission fixes are the ones to update for. Until this release, a dangerous rm on / or the home directory went through with no prompt in those setups as long as it sat inside a bash -c or sh -c script.
That fix has company. PreToolUse and PermissionRequest hooks used to be skipped when matching them failed or when the tool’s input could not be serialized to JSON. Now the tool call is blocked instead. The Bash permission check now prompts before a BASHPID assignment whose value the shell would evaluate as arithmetic, where it used to allow it silently. Permission asks that ended unanswered, in -p or on an interrupted turn, now emit a tool_decision event. If you audit approvals through OpenTelemetry, the claude_code.tool.blocked_on_user spans in -p and SDK sessions now report their source and decision, which used to show up as unknown.
Auto mode gets its own set of changes. When a conversation grows too long for the client-side safety classifier to review, the conversation is now compacted. Before, every tool call either prompted or failed. On Bedrock and Mantle, a single request to an older model, such as a WebFetch summary or a sonnet subagent, used to switch auto mode to the local classifier for the rest of the session. That is fixed. And the client-side classifier now ignores an ANTHROPIC_DEFAULT_SONNET_MODEL pin that names Claude Sonnet 5.5 or Opus 5.5, and uses Claude Sonnet 5 instead.
What to try first
Draft recovery works on an empty prompt. After a Ctrl+C, press Up and the draft comes back, pasted text and images included. The other new controls:
/code-review --max-findings <n>|allreports more or fewer findings than the usual limit. Your choice sticks until you pass--max-findings default.- When an MCP server asks for more OAuth scope in the middle of a tool call, Claude Code prompts you to re-authenticate.
- In the agents view, Ctrl+F finds a session by name and Alt+↑/↓ jumps between groups. Both shortcuts, along with rename, can be rebound in
keybindings.json. Enter in then:filter or in Ctrl+F search opens the session whose name matches best, where it used to open the top row. /autocompactsaves the auto-compact window per model, so each model keeps its own setting when you switch.claude project purgeis nowclaude purge. The old name still works and prints a notice.
If you run on Mantle or behind a gateway that rejects structured outputs, session titles, memory recall and prompt hooks used to fail there. The fix comes with a new variable, CLAUDE_CODE_DISABLE_STRUCTURED_OUTPUTS, which turns structured outputs off.
Headless and unattended runs
A mid-response API timeout used to fail the whole turn. Non-interactive sessions and subagents now pick up from the partial response, and responses that contain only thinking are retried. Unattended sessions running with CLAUDE_CODE_RETRY_WATCHDOG could keep retrying for hours after a very long response stream failed. Claude Code now streams again and gives up after three timeouts.
The background command time limit now applies only in unattended sessions: -p, the Agent SDK, CI and cloud. Terminal, desktop app and VS Code sessions have no limit. Headless sessions also respond to SIGTERM when a supervisor such as timeout or systemd sends SIGCONT along with it, a case where they occasionally ignored the signal. MCP tool calls could run twice when a remote server’s result was over 16 MB or failed to parse. That is fixed. LSP requests that used to hang now time out after 60 seconds, which you can change per server with requestTimeout.
Resume, and a 1.5-second cost
Four resume bugs are gone. --resume sometimes dropped files and other context that a compaction had just restored. A resumed session sometimes failed to save the last response of a turn, so the next resume showed the prompt unanswered. A resume could also load a truncated transcript if the same session rewrote the file mid-load. And conversations started on 2.1.286 or earlier lost the model’s earlier thinking when resumed. Separately, long conversations that failed with “Prompt is too long” when the last reply reported zero token usage now auto-compact.
One fix adds latency. Until now, the first request in a fresh environment or after a model switch used the built-in output limit and auto-compact window in place of the server’s values. To get the server’s values, that request may now wait up to 1.5 seconds.
Plugins and mods
Mod authors get $.ui.selection(), which returns the text you last selected in fullscreen mode. When the selection sits within one transcript row, it returns that row too. It comes one release after mods opened to anyone in v2.1.287. A mod’s button could also run a different button’s action when pressed on a view drawn before Claude Code restarted. That is fixed.
Several install failures are fixed too. On git older than 2.39, such as the 2.34 that ships with Ubuntu 22.04, git-subdir plugin installs failed or cached an incomplete plugin. claude plugin install for GitHub-source plugins failed on macOS and Linux machines with no GitHub SSH key. The clone now falls back to HTTPS and prints a notice. Plugin LSP servers used to receive literal ${user_config.*} and ${CLAUDE_PLUGIN_ROOT} placeholders where substituted values or manifest defaults belonged. And path-scoped .claude/rules and nested CLAUDE.md files now load when Write or Edit creates or changes a file in their scope. Previously only Read loaded them.
Two fixes cover reports of success that weren’t true. /login showed “Login successful” when the credentials could not be saved to secure storage. It now shows the failure and offers a retry. The npm auto-updater reported success when the platform-native binary failed to download and only the placeholder claude stub got installed.
Up next
Anthropic trains AI engineers like medical residents
“Follows the medical model for building expertise.” That’s how Anthropic describes the residency at its $100 million Claude Frontier Academy, which aims to train 10,000 engineers by the end of 2027.
Related stories
- Claude Code 2.1.300 splits cache writes by lifetime in OTel
- Claude Code 2.1.290 gives mod hooks the subagent’s ID
- Claude Code ships mods, the same tool behind its /diff
- Claude Code 2.1.286 stops resume from losing your turns
- Claude Code on desktop folds an iPhone Duo to test layouts
- Claude Code 2.1.318: login expiry warning, runner log change
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.

