## What's changed
- Added `--hold-idle-file` and `--max-hold-min` to `self-hosted-runner`, to keep an idle session while your host reports processes it left running
- Added a tip for the Explanatory output style, shown when you ask Claude for explanations
- Fixed long freezes at startup and after reading or editing files when a CLAUDE.md file holds a very long list item
- Fixed interrupting Claude not taking effect until a timeout while an MCP server had not answered a resource read or a skill was still downloading
- Fixed standard-context sessions stopping with "Usage credits required for 1M context" when a request ran over 200K tokens; Claude Code now compacts the conversation and continues
- Fixed Write and Edit stalling for up to 5 seconds after they ran, and reporting 0 lines changed, when changing thousands of lines in a file
- Fixed subagents being unable to use MCP tools from servers that finish connecting after the subagent starts
- Fixed Claude seeing stale MEMORY.md and CLAUDE.md content right after automatic compaction when those files were edited earlier in the session
- Fixed `claude -p` and SDK sessions saving transcript messages out of order or keeping messages that were taken back
- Fixed `SubagentStop` hooks running about twice a minute for each running subagent instead of once when it finishes
- Fixed the @ file picker ignoring `.ignore`/`.rgignore` negation rules that re-include git-ignored files
- Fixed a subagent's isolated worktree being removed while a background subagent it started was still working in it
- Fixed background agents in auto mode that wait on their own shells starting a new main conversation turn on every wake
- Fixed the Bash tool description claiming the working directory always persists between calls
- Fixed the environment block naming a shell commands don't run in, such as fish when the Bash tool uses zsh
- Fixed `claude mcp add`, `claude mcp remove` and `/mcp` naming the current folder instead of the project entry in `~/.claude.json` that holds a local-scope server when run below a repository root
- Fixed `claude mcp list` and `claude mcp get` showing the error code twice on failed-connection lines (for example `ENOENT: ENOENT: …`)
- Fixed Remote Control sessions staying on "Connecting" for about two minutes when a network proxy blocks a startup request
- Fixed messages that Claude sends to a cloud or Remote Control session appearing as your own message in the Claude mobile apps
- Fixed Claude not being told the cause when a file attached to a message sent over Remote Control or to a cloud session could not be saved because the disk was full, or that a message's attachments past the first 20 were left out
- Fixed WebFetch and WebSearch hanging on a stalled response in Cowork cloud sessions; an unfinished request now fails after 300 seconds
- Fixed a background session that crashed between turns not being restarted while it was attached, pinned, or had a scheduled task or wake-up set
- Fixed the "background service did not respond" error suggesting the disabled `claude daemon restart`; it now suggests `claude daemon stop --any` and says it ends background sessions
- Fixed `claude respawn <id>` starting the background service before reporting that no background session matches the id
- Fixed `/effort` not warning about a prompt cache miss in plan mode when the plan model differs from the session model
- Fixed `/ide` saying the VS Code extension was installed before the install finished or when it failed
- Fixed the Write tool's file preview skipping rows and miscounting hidden lines when a line is wider than the terminal
- Fixed ↑ in an agent's chat tab recalling a message that was still waiting to be delivered, so resending it sent a second copy
- Fixed Cowork cloud sessions losing their connectors and most tools after a restart when a plugin or your settings named an `agent`; Cowork cloud sessions now ignore that setting
- Fixed agent teams sessions showing an empty task list when coming back with `--resume`, `--continue`, `/resume` or `--session-id`, including after `/clear` or a crash, and `/branch` writing to the original session's list
- Fixed pasting text that mentions an image file, such as `! open /tmp/shot.png` or `/review look at /tmp/shot.png`, attaching the image and dropping its path from the prompt
- Fixed the feedback draft card showing its 1/2/0 and Esc shortcuts while the prompt has text, when those keys act on the prompt instead of the card
- Fixed a lone backslash in quotes or brackets, like `"\"` or `(\)`, disappearing from Claude's responses
- Fixed agent teammates compacting their history when `DISABLE_COMPACT` is set
- Fixed the rewind menu offering "Summarize from here" and "Summarize up to here" when `DISABLE_COMPACT` is set
- Fixed a finished `/code-review`, or another skill that runs in a sub-agent, running a second time when a cloud or background session restarts
- Fixed long hidden thinking being cut off as idle on Google Vertex AI, Microsoft Foundry and Claude Platform on AWS, where the endpoint sends thinking progress updates
- Fixed `/ide` saying it had installed the JetBrains plugin when nothing was installed
- Fixed Claude Pro and Max subscribers being told to ask their admin when their subscription can't be used; the message now points to billing settings
- Fixed the diff pane opening by itself after an edit in wide fullscreen terminals when it had nothing to show
- Fixed `/mcp` showing a remote server as failed after "Clear authentication" when the server works without sign-in
- Fixed `/mcp` showing "authenticated" for a remote server that Claude Code sends no credentials to
- Fixed a `--plugin-dir` folder of plugins loading as one plugin when an empty `.mcp.json` was at its top
- Fixed `/usage` never showing large prompt-cache misses under "What's contributing to your limits usage?"
- Fixed status line links staying plain text in background sessions after you open them from a terminal that supports links
- Fixed the /rewind confirmation saying the conversation "will be forked" when the later messages are removed from the conversation
- Fixed the /rewind warning that wrongly said manual and Bash edits are never affected by a code restore
- Fixed plan approval in screen reader mode sending feedback you had dismissed with Escape
- Fixed Claude Code retrying requests the API marked not to retry, which could hold a turn for minutes or, in unattended retry mode (`CLAUDE_CODE_RETRY_WATCHDOG`), indefinitely
- Linux: Fixed sandboxed Bash commands failing with `bwrap: Can't create file` when a `--plugin-dir` folder was also an added directory and `permissions.blockReadsOutsideWorkingDirectories` was on
- Self-hosted runner: Fixed a session failing to start when a rate limit (HTTP 429, or a git wrapper that exits 75) blocks an added read-only repository; it now starts without that repository
- Improved startup time for Pro and Max accounts by skipping two network requests on launches within an hour of the last check
- Improved `/teleport` and `claude --teleport` to warn when a compacted cloud session loads from its last compaction, so the messages it summarized aren't in the local copy
- Improved the messages shown when a background session's process exits: they no longer promise a restart that isn't coming, and they say how to resume the session
- Improved artifacts with several views: Claude now routes each view on the link's `#name` and no longer says the address bar stays the same as you switch views
- Improved shell commands that reach their timeout: the result now shows the output so far and says the command is still running in the background
- Improved compaction: when the model's reply has no summary in it, Claude Code now usually asks for the summary once more before it replaces the conversation
- Improved the Artifact tool's choice between a page and a document: explainers and visual or interactive work become artifact pages; documents are for writing people keep editing or pass around
- Improved startup time for launches whose only arguments are options and their values, such as `claude --model opus`
- Improved the "Connection dropped" API error: it now says where to look first (the proxy or gateway in use, else VPN or firewall; Anthropic API only, Bedrock, Vertex and Foundry are unchanged)
- Improved screen reader mode: deleting a rule in /permissions says "Deleted allow rule …" right away
- Improved the message shown when a pasted sign-in code is refused: it now says the code may have expired or been used, and to sign in again
- Improved Slack connector sends: Claude now checks a channel's topic for posting rules before messaging a channel you didn't name
- Improved screen reader mode: Claude's questions are listed as "Questions: Color (answered), Pages", without arrows or symbols
- Changed `ultrathink` in the interactive terminal to run its turn at xhigh effort on Opus 5.5, Sonnet 5.5, Haiku 5.5 and Fable 5.1 through Anthropic's API; typed while Claude is working, it raises the rest of that turn
- Changed the `ultrathink` rainbow to appear only where the keyword can raise effort, and not when the word is quoted or part of a file path
- Changed Ctrl+D in agent view to match the prompt: on an empty input, pressing it twice exits
- Changed `--output-format json` and `stream-json` output to write DEL, C1 control and Unicode format characters inside strings as `\uXXXX` escapes; parsed values are unchanged
- Changed model names and IDs to drop the `[1m]` suffix and its label on models where the suffix has no effect; the suffix is still accepted
- Changed startup warnings about permission rules in settings files to show once per user until the rule changes
- Changed `/code-review` to tell Claude to put the effort level first and use max only when asked
- Changed `DISABLE_COMPACT` in settings files and in a mod's `$.env.set` so they can only turn compaction off, not back on
- Changed Claude Code on Bedrock, Mantle, Vertex, and Foundry to load model settings from a signed file on downloads.claude.ai; set CLAUDE_CODE_MODEL_CATALOG=off to turn this off
- Changed sessions that set `CLAUDE_CODE_SYNC_SKILLS` with feature-flag fetching off to reuse a claude.ai skills list up to 10 minutes old that another session on the machine fetched
- Changed clicks on scrollbar arrows and bars in fullscreen lists to act when the mouse button is released, so moving off the scrollbar before letting go cancels the click
- Changed `claude purge` to say it deletes data only on this computer, not anything stored by Anthropic
- [VSCode] Added syntax coloring to code blocks in Claude's replies, in light, dark and high contrast dark themes
- [VSCode] Added a "Running UserPromptSubmit hooks" spinner status and screen reader announcement for slow UserPromptSubmit hooks
- [VSCode] Fixed custom skills and custom slash commands sent from the VS Code or JetBrains panel leaving out the open file or selected lines that ordinary messages include (anthropics/claude-code#102799)
- [VSCode] Fixed the extension host freezing while the login shell ran its start-up files to read PATH on macOS and Linux
- [VSCode] Fixed clicking an image in a tool result, or closing its preview, also opening the result's text in an editor tab
- [VSCode] Fixed voice dictation continuing to record after a message was sent with the Send button; sending now stops dictation first, as pressing Enter already did
- [VSCode] Changed the extension to start Claude Code sessions with `DISABLE_TELEMETRY` set while VS Code's telemetry setting is off, so they send no usage metrics or error reports and fetch no feature flags
- [VSCode] Changed the session list side bar to load Web sessions when the Web tab is first opened, not at startup
- [Cloud sessions] Fixed routines in Anthropic-hosted environments staying locked after their owner left the organization and rejoined; the owner can now turn them back on
- [Cloud sessions] Fixed a failed routine API token delete or regenerate showing nothing if the edit dialog had already closed; it now shows an error toast
- [Cloud sessions] Fixed cloud environment dialogs, including the admin dialog, accepting duplicate names; they now refuse a name another listed environment already has
- [Claude Tag] Added a Skills view to Claude Tag analytics showing how often Claude runs each attached skill in Slack channels and which go unused, with Usage links in admin settings
- [Claude Tag] Improved Claude Tag routines that run in a new session: a run with nothing to report can now end without posting a throwaway message to the channel
- [Claude Tag] Changed Claude Tag admin settings to show a lock instead of # for private Slack channels that your own Slack account can see
- [Code Review] Improved Code Review's check run to say no further automatic review will start when the model still can't analyze a change on the automatic retry
- [Code Review] Improved Code Review reliability: a review whose session fails to start because of a transient server fault is now retried once automatically, and its check run says so