Claude Code this week: October 30 to November 4, 2026
The changes worth knowing from Claude Code 2.1.314 to 2.1.318, from a new write guard to subagent effort and sign-in expiry.
This is the first weekly entry: a short, plain read of what changed in Claude Code over the last week, so you do not have to scan the full release notes. It covers four releases, 2.1.314 (October 30), 2.1.316 (October 31), 2.1.317 (November 3) and 2.1.318 (November 4), all from the official changelog. Each section says what changed, who it matters to, and what to do about it, if anything.
How to get these changes
Run claude --version to see what you have (changelog). Anything at 2.1.318 or later includes everything below. If you rely on auto-update, 2.1.318 also makes it retry right away when a brief network failure hits while a release is being verified.
A guard against writes outside your project
New in 2.1.314: the permissions.blockWritesOutsideWorkingDirectories setting. With it on, the Write, Edit and NotebookEdit tools refuse any path outside your working directories, in every permission mode (changelog).
Who it is for: anyone who runs Claude in auto mode or with few prompts and wants a hard line around the project folder. It is a simple setting worth considering for team projects. Note that the changelog names only those three file tools, so do not treat it as a limit on shell commands.
Idle compaction reaches VS Code and Desktop
Also in 2.1.314: VS Code and the Code tab in Claude Desktop now do idle compaction. When a conversation of 200K or more tokens sits idle, Claude Code summarises it, and uses that summary on your next message if the prompt cache has expired (changelog). The same release extends idle compaction to subscribers with telemetry off, and to Bedrock, Vertex, Foundry and LLM gateways once the one-hour prompt cache is on.
Why it matters: coming back to a big, cold session no longer means paying to reprocess the whole history. If you would rather keep the full history, the idleCompaction setting set to false turns it off, on v2.1.290 or later (prompt caching).
Cloud scheduled tasks load your claude.ai plugins
In 2.1.314, scheduled task runs in Anthropic's cloud started loading the plugins enabled on your claude.ai account (changelog). If a scheduled task needs a plugin's skills, turning the plugin on for your account is now enough.
A related change in 2.1.317: /schedule now warns you when a new cloud scheduled task picks up connectors you did not ask for, and tells you how to remove them.
Subagents get their own effort level
Two changes this week make subagent effort easier to control.
2.1.316 added CLAUDE_CODE_SUBAGENT_EFFORT. It sets the effort level subagents run at when nothing else names one, instead of them inheriting the session's level (changelog).
2.1.317 changed subagents and teammates that run on a different model from the main conversation. They now use that model's saved effort level or its default, not the session's (changelog).
What to do: if a subagent should always run at a set level, the most explicit option is still the effort field in its definition (sub-agents docs). Use the new variable for a blanket default.
Background subagents keep the cache warm
From 2.1.318, a background subagent moves a shell command to the background after 3 minutes on all providers. The changelog gives the reason: so the prompt cache does not expire while the command runs (changelog).
Why it matters: long test or build commands inside a subagent were a quiet source of cache misses. You do not need to change anything.
Ultrathink: an off switch, and VS Code catches up
2.1.316 added CLAUDE_CODE_DISABLE_ULTRATHINK, which turns the ultrathink keyword off (changelog). It is useful if you or your team trigger it by accident and want effort to stay where you set it.
The same release made ultrathink in VS Code run its turn at xhigh effort, as it already did in the terminal. 2.1.317 then fixed ultrathink in VS Code not raising effort when telemetry is off.
Fast mode on Opus 4.8 is going away
2.1.316 deprecated fast mode for Opus 4.8 as of October 30, 2026, with removal on November 30, 2026. After that date, fast mode requests on Opus 4.8 return an error rather than falling back to standard Opus 4.8. The changelog's advice: "Migrate to Opus 5.5 to keep the speedup" (changelog).
What to do: if any script, setting or alias pins Opus 4.8 with fast mode, move it to Opus 5.5 before the end of November.
Rereading a file returns its content again
2.1.317 removed the Read tool's "file unchanged since your last Read" stub. Reading an unchanged file a second time now returns its content (changelog). Claude no longer has to work from a note saying the file has not changed.
See when your sign-in expires
2.1.318 shows the date a stored claude.ai sign-in ends. It appears as loginExpiresAt in claude auth status and as credential_expires_at in claude -p JSON results. claude -p text output also prints a one-line warning on stderr starting 10 days before the sign-in ends (changelog).
Who it is for: anyone running Claude Code in scripts or CI on a stored login. Add a check for that field and renew before a nightly job fails.
Also worth a glance
Ctrl+Rprompt history search now matches text inside collapsed[Pasted text]pills (2.1.314).- Sessions that start with a skill or a prompt-based command, such as
/security-review, now get an automatic title from it (2.1.314). - Long headless and SDK turns that run many subagents use less memory, because the parent no longer keeps each subagent's tool results until the turn ends (2.1.314).
All three come from the changelog.
Next
- Subagents: when to split the work: set up the subagents whose effort and cache behaviour changed this week.
- The multi-agent orchestration playbook: a full config pack for running several agents at once.
Sources
Last checked: 2026-11-06