Claude Code · 7 min read

ultrathink now runs at xhigh effort, and subagent effort has its own setting

ultrathink in the terminal and VS Code now runs its turn at xhigh effort, and CLAUDE_CODE_SUBAGENT_EFFORT sets what subagents run at. The week also brings a write guard for paths outside your working directories, idle compaction in VS Code and the Desktop Code tab, and the new Ultra plan.

Key takeaways 6
  • ultrathink runs its turn at xhigh effort on Opus 5.5, Sonnet 5.5, Haiku 5.5 and Fable 5.1 through Anthropic's API. Typed mid-turn, it raises the rest of that turn.
  • CLAUDE_CODE_SUBAGENT_EFFORT sets the effort subagents run at when nothing else names one, and CLAUDE_CODE_DISABLE_ULTRATHINK turns the keyword off.
  • permissions.blockWritesOutsideWorkingDirectories makes Write, Edit and NotebookEdit refuse paths outside your working directories in every permission mode.
  • Idle compaction now covers VS Code and the Desktop Code tab, and cloud sessions gain a form card for MCP server requests and start even when a setup script fails.
  • Ultra is the new name for Max 20x, with the same Claude Code features as Max.
  • Fast mode for Opus 4.8 is deprecated as of October 30, 2026 and will be removed on November 30, 2026.

Effort settings usually apply to a whole session, while the work inside it varies a lot from turn to turn. This week gives you a keyword to raise effort for one turn and a variable to set it for subagents. Alongside come a write guard that holds in every permission mode and idle compaction for long conversations in VS Code. The Ultra plan, cloud session fixes and a Fast mode deadline need your attention too.

Week 44 (October 26-31, 2026) covers Claude Code versions 2.1.311 to 2.1.316: 4 releases and 305 changelog entries. Below are the changes that affect daily work, then the smaller items grouped by area.

Raise effort for one turn, and set it for subagents

What changed. In the interactive terminal, ultrathink now runs its turn at xhigh effort. This applies to Opus 5.5, Sonnet 5.5, Haiku 5.5 and Fable 5.1 through Anthropic’s API (2.1.311). The VS Code chat box behaves the same way and shows the keyword in rainbow colors (2.1.316).

Using it mid-turn. If you type ultrathink while Claude is already working, it raises the rest of that turn. You don’t have to wait for the turn to end or restart it. The change applies to that turn only, so a long session doesn’t stay at xhigh.

Subagents get their own knob. CLAUDE_CODE_SUBAGENT_EFFORT sets the effort level subagents run at when nothing else names one. Before, they inherited the session’s level. If you run a cheap, fast main session and delegate hard work, or the reverse, you can now set the subagent level independently of the session (2.1.316).

Switching the keyword off. CLAUDE_CODE_DISABLE_ULTRATHINK turns the ultrathink keyword off. That helps if the word shows up in your prompts as plain text and you don’t want it to change the effort level.

Conditions.

  • The xhigh behavior is listed for Opus 5.5, Sonnet 5.5, Haiku 5.5 and Fable 5.1 through Anthropic’s API.
  • It applies in the interactive terminal and in the VS Code chat box.
  • The subagent variable only applies when nothing else names an effort level for the subagent.

The model configuration page covers effort in more detail: model-config.


Changes for everyday work

A write guard that ignores permission mode. The new permissions.blockWritesOutsideWorkingDirectories setting makes Write, Edit and NotebookEdit refuse paths outside the working directories, whatever the permission mode (2.1.314). Details are in the permissions docs. Mode switching is exactly where a stray path would otherwise slip through, so this is worth a look if you use permissive modes.

Idle compaction reaches VS Code and the Desktop Code tab. A conversation of 200K+ tokens that you leave idle is now summarized. If the prompt cache has expired by the time you come back, the summary is used for your next message (2.1.314). It also runs for subscribers with telemetry turned off, and on Bedrock, Vertex, Foundry and LLM gateways once the one-hour prompt cache is on. See the context window docs.

Cloud sessions get three fixes. A form card now appears when an MCP server asks for a form, so you can fill it in or decline instead of the request being canceled unseen (2.1.316). Cloud sessions from the web, desktop and mobile apps start even when a setup script exits with an error, instead of failing to start (2.1.314). A ‘Catching up…’ line shows while a cloud or Remote Control session you switch back to is still updating its messages (2.1.316). One caution: a session that starts despite a failed setup script may be missing whatever that script was meant to install, so check your script output. More in the cloud sessions docs.

Ultra is the new name for Max 20x. Max is now the name for Max 5x. Ultra subscribers get the same Claude Code features as Max. That includes bypass permissions mode in Desktop (Settings > Code or Settings > Claude Code) and extra usage bought with /usage-credits. The docs also list computer use, Dispatch and cloud sessions as available on Ultra. Computer use is a research preview on macOS and Windows. Dispatch is not available on Team or Enterprise plans.

Gateway sign-in from the VS Code chat panel. You can sign in to a Claude apps gateway from the chat panel without opening a terminal (2.1.316). Code blocks in replies now have syntax coloring in light, dark and high contrast dark themes (2.1.311). When organization settings can’t be loaded or are refused at startup, the tab now says why and offers Retry or Sign in again, instead of the session ending on a bare exit code (2.1.312). Details are in the VS Code docs.

Agent SDK: per-tool policies and mid-session plugins. Plugin mods can now approve or decide some tool calls, and the permissions docs place their decisions among rules, hooks and the callback. In the TypeScript SDK, in-process MCP tools take a permission policy of always_allow, always_ask or always_deny. The docs set it through a per-tool permission_policy field on the config object that createSdkMcpServer() returns. reloadPlugins() accepts an addPlugins option, which loads new plugin folders mid-session and reloads existing plugins in one pass. The paths must be absolute. And startup() now delivers the newest hook_progress message from a SessionStart or Setup hook, plus hook_started, once you call .query() on the WarmQuery.


Technical notes

Models. Fast mode for Opus 4.8 is deprecated as of October 30, 2026 and will be removed on November 30, 2026. After that, requests return an error. Fast mode is available for Opus 5.5, so migrate to it to keep the speedup (2.1.316).

Terminal.

  • The new /restart command restarts Claude Code and resumes the session to apply an installed update. /update is an alias.
  • At the prompt you can start one message with up to six skills, such as /skill-a /skill-b do XYZ. All are loaded, and the trailing text is passed to each as arguments.
  • In fullscreen, clicking the ‘to send now’ hint under a queued message sends it right away (2.1.312).
  • /usage adds a ‘Prompt cache (main)’ line to the Session block. Status line scripts can read the same figures from the prompt_cache object.

Cloud routines and runners. The model picker for cloud routines now includes an effort level when you create or edit a routine (2.1.312). self-hosted-runner gets --hold-idle-file and --max-hold-min, which keep an idle session while your host reports processes it left running (2.1.311).

Claude apps gateway (admin).

  • oidc.user_id_claim identifies people by an id_token claim other than sub, such as Entra ID’s oid (2.1.312).
  • Pooled group budgets give a group’s members one shared monthly budget, set with pooled_budget on the spend limits admin API (2.1.316).
  • managed.policy_mode: layered makes every gateway policy that matches a developer apply, merged over the ones listed below it (2.1.316).
  • Upstreams accept an optional groups list, so only developers in the listed IdP groups are sent to that upstream. A Bedrock upstream also gets guardrail_trace, which logs which Bedrock guardrail filter fired (2.1.314).

Analytics. The Code Review analytics page gets a findings-by-severity card. It shows total findings, a daily stacked chart and each severity’s count (2.1.316).


What to take away

If you work in the terminal.

  • Try ultrathink on a hard turn, including mid-turn while Claude is already working, and see whether the rest of that turn improves.
  • Set CLAUDE_CODE_SUBAGENT_EFFORT if your subagents should run at a different level than your session.
  • If the keyword gets in your way, CLAUDE_CODE_DISABLE_ULTRATHINK turns it off.
  • Try /restart after an update, and start a message with several skills.
  • Check the new ‘Prompt cache (main)’ line in /usage.
  • If you use Fast mode on Opus 4.8, plan the move to Opus 5.5 before November 30, 2026.

If you run parallel and cloud sessions.

  • Watch for the form card when an MCP server asks for input. You can now answer or decline it.
  • Check your setup scripts, since sessions now start even when one exits with an error.
  • Use the ‘Catching up…’ line as your cue that a session you switched back to is still updating.
  • Add an effort level to your cloud routines in the model picker.
  • In VS Code or the Desktop Code tab, expect long idle conversations of 200K+ tokens to come back summarized when the prompt cache has expired.

If you administer integrations and company deployments.

  • Consider permissions.blockWritesOutsideWorkingDirectories as a baseline, since it holds in every permission mode.
  • If you use Bedrock, Vertex, Foundry or an LLM gateway, idle compaction applies once the one-hour prompt cache is on.
  • On the Claude apps gateway, review oidc.user_id_claim (especially with Entra ID), pooled_budget, managed.policy_mode: layered, and the groups list on upstreams.
  • Use guardrail_trace on Bedrock upstreams to see which guardrail filter fired.
  • Note that Ultra now appears next to Pro, Max, Team and Enterprise in the plan availability table.
  • If you build on the Agent SDK, look at per-tool always_ask and always_deny policies and at addPlugins.

Sources: the Claude Code changelog, versions 2.1.311 to 2.1.316 (October 26-31, 2026), and the Claude Code documentation, including the pages for model configuration, permissions, context window, cloud sessions, VS Code and the Agent SDK. Computer use is documented as a research preview, so its availability may change.

Patrick Zandl trains teams in Claude Code and agentic software development in Prague and writes about what actually works in real-world AI-assisted projects. Follow on LinkedIn

Read next

Get the next one by email

Every Tuesday: what changed in Claude Code, Codex and other coding agents, plus new articles. No spam.