Resume keeps every turn
Claude Code v2.1.286 keeps every turn when you resume a crashed session
- claude --resume / --continue
- keep the turns after parallel tool calls
- When
- the earlier session crashed or was killed
Claude Code 2.1.286 fixes resumed sessions. When a session crashed or was killed, resume and continue could lose every turn after a batch of parallel tool calls. Now they keep them.
01 Models and retries
>Models
and retries
Models and retries.
A refused model falls back
>A refused model now falls back instead of failing every turn
- Fallback
- one retry on the previous model of the same tier
When the Anthropic API refused the model your default or an alias resolves to, every turn failed. Now Claude Code retries once, on the previous model of the same tier.
One retry limit per call
>A failing model call now sends at most 14 requests
- Retry limit
- one limit covers the whole model call
- Default settings
- at most 14 requests
One limit now covers a whole model call, so with the default retry settings, a failing call sends at most 14 requests.
Fallbacks at standard speed
>Fallback retries now run at standard speed instead of failing
- Notice
- once, in interactive sessions
- Context
- the notice says when 1M dropped to 200K tokens
A fallback model that can't run fast now runs at standard speed. Interactive sessions get a one-time notice that says when the fallback dropped your context window from 1 million to 200 thousand tokens.
02 Prompts and lists
>Prompts
and lists
Permission prompts and lists.
Permission prompt count
>Stacked permission requests now show a count
The prompt reads 2 of 5 when several requests stack up.
When several permission requests stack up, the prompt now shows a count, such as 2 of 5, so you can see how many are waiting.
One look for permission prompts
>Permission prompts now share the look of file edit prompts
- Matched
- fetch, skill, file read, sandbox network, Claude in Chrome, workflow script, notebook edit
- Bash, PowerShell, Monitor
- the command sits between dashed lines
Prompts for fetches, skills, file reads, notebook edits and more now match the file edit prompt. Bash, PowerShell and Monitor prompts show the command between dashed lines.
Clickable list overflow
>Fullscreen lists now respond to clicks on their overflow rows
- ↑ N more · ↓ N more
- click to jump to that end of the list
- Scrollbar arrows
- click or hold to scroll
- Steady bar
- in most lists, it holds still as those rows come and go
In fullscreen mode, click an up or down more row to jump to that end of a list. The scrollbar gains arrows you can click or hold, and in most lists it holds still.
03 Agents and background work
>Agents and
background work
Agents and background work.
verify runs before commits
>Claude now runs a skill named verify right before committing
- Where
- your project or user skills
- Skipped for
- docs-only and tests-only commits
If your project or user skills include one named verify, Claude is now told to run it right before each commit. Commits that only touch docs or only touch tests skip it.
Commands in an agent's view ask first
>/compact, /clear and /rewind ask first inside a background agent's view
A dialog names the conversation they will act on.
Typed while you viewed a background agent's or a teammate's transcript, /compact, /clear and /rewind silently acted on your main conversation. Now a dialog names the target and asks first.
Send now in a subagent's view
>Ctrl+Enter in a subagent's view now reaches the subagent at once
- Subagent's command
- moves to the background
- Skill's shell command
- moves to the background instead of ending
Press control enter in a subagent's view, and its running command moves to the background so your message is read right away. A skill's shell command moves there too, instead of ending.
--bare starts less
>--bare now starts only what you name on the command line
- MCP servers
- only those named on the command line
- System reminders
- off
- Background tasks
- off, and a shell command that times out stops
The bare flag now connects only the MCP servers you name on the command line, and skips system reminders and background tasks. Under it, a shell command that hits its timeout stops.
04 Secrets and plugins
>Secrets
and plugins
Secrets and plugins.
More secrets masked
>Error messages, logs and transcripts now mask more secrets
- MCP errors
- a credential after “Bearer” or “Basic”
- Encoded tokens
- percent-encoded Bearer tokens, masked in full
- URL passwords
- punctuation such as ) ] & or a second @
MCP errors could show a credential written after Bearer or Basic, percent-encoded tokens were only partly masked, and logs could show part of a URL password with punctuation in it.
Plugin installs refuse git and folder sources
>Plugin installs now refuse npm sources that are git repositories or folders
Plugin dependencies install only from registry packages.
Plugin installs now refuse an npm source that is a git repository or a folder. A plugin's dependencies now install only from registry packages.
05 The VS Code extension
>The VS Code
extension
The VS Code extension.
Bookmarks and question cards
>VS Code adds bookmarks and keeps your answers to Claude's questions
- Bookmarks
- save responses to a side panel
- Question cards
- preview the highlighted option
- Questions row
- each question with the answers you picked
In VS Code, you can bookmark Claude's responses into a side panel. Question cards preview the highlighted option, and a Questions row keeps each question with your picks.
Stop ends only the turn
>Stop and Escape in VS Code now end only the current turn
Background agents keep running. Stop each one from the agent map.
Stop and Escape in VS Code now end only the current turn. Background agents keep running, and you can stop them one by one from the agent map.
The full changelog
Every change in v2.1.286 is in the Claude Code changelog
CHANGELOG.md in anthropics/claude-code
The release also fixes cloud sessions and Claude Tag in Slack. Every change in 2.1.286 is in the Claude Code changelog on GitHub.
