What happened
Anthropic released Claude Code v2.1.314 on October 30. The release extends idle compaction to two graphical Claude Code surfaces: the VS Code extension and the Code tab in Claude Desktop. The release notes describe the trigger precisely: when a conversation has exceeded 200K tokens, is left idle, and its prompt cache has expired, Claude Code summarizes the conversation. That summary is then used on the user’s next message.
This is a continuation mechanism for long-running work rather than a larger stated context window. A developer can accumulate a large conversation while investigating a repository, making changes, reviewing generated output, or working through a sequence of failures. When the session becomes idle under the stated conditions, the product creates a compressed representation so that the next interaction can continue from a summary. In this version, that behavior is available in the VS Code and Claude Desktop interfaces as well as the product’s existing terminal-oriented workflows.
The release also adds permissions.blockWritesOutsideWorkingDirectories. When enabled, Write, Edit, and NotebookEdit refuse paths outside the working directories in every permission mode. The wording matters: this is not framed as a prompt-level preference or a mode-specific default. It is a setting that applies to those three write-capable tools across permission modes, defining an explicit filesystem boundary for their operations.
Several additions target organizations routing Claude Code through the Claude apps gateway. Gateway upstreams can now have an optional list of identity-provider groups; only developers in those groups are sent to the upstream, including during failover. For Amazon Bedrock upstreams, guardrail_trace logs which Bedrock guardrail filter fired on a request without logging the matched text.
The remainder is a large set of fixes across session reliability, managed settings, image-heavy Bedrock sessions, Microsoft Foundry fallback behavior, MCP authentication, Remote Control, hooks, sandbox protections, CI trust handling, and interface behavior. Among the fixes, Claude Code now lowers an effort level and retries when an organization’s administrator caps a model’s effort, then reports the level it used. It also fixes compaction failures or repetition caused by very large PreCompact-hook output by limiting each hook’s output to 10,000 characters.
What the archive adds
This is a new fact in an already covered release sequence. Claude Code 2.1.311 speeds Pro/Max startup and keeps compacting past 200K tokens reported that compaction could continue beyond 200K tokens. Version 2.1.314 extends that long-context behavior into VS Code and Claude Desktop’s Code tab, rather than changing the stated threshold.
Trend signal
The new filesystem setting supports Agent security moves to runtime boundaries. permissions.blockWritesOutsideWorkingDirectories constrains named write tools by path and applies in every permission mode. That is a runtime control over what the agent can alter, not an instruction asking it to avoid sensitive locations.
Why now
The timing follows closely after the October 27 note on v2.1.311’s compaction beyond 200K tokens. The source does not state Anthropic’s motivation, but it shows a concrete product step: a continuation behavior previously documented in the Claude Code release stream is now present in its VS Code and desktop graphical surfaces.
Connections to other events
The archive also records Claude Code 2.1.312’s OIDC user-id claim and CLAUDE.md safeguard. Together, these adjacent releases point to two operational concerns in the same product: keeping extended sessions usable and making enterprise or local controls more reliable. The current release adds group routing and Bedrock guardrail tracing alongside the local write-boundary control.
Why it matters
Teams often experience a long coding-agent session as both useful and fragile. The useful part is accumulated task context; the fragile part is that a long session can become costly or difficult to resume once its cached prompt state is gone. Idle compaction in the interfaces where developers edit and inspect code gives teams a more consistent continuation path across terminal, editor, and desktop use.
The more consequential operational addition may be the write boundary. A coding agent that can edit files needs a clear definition of where its authority ends. Restricting Write, Edit, and NotebookEdit to working directories can reduce accidental changes outside a repository or designated workspace. It does not establish that every tool action is safe, but it makes one important capability boundary enforceable.
Limits and uncertainty
The release notes do not quantify the quality, size, latency, or fidelity of the generated summary, so they do not establish how well a compacted session preserves every technical detail. They also do not claim that the new write setting constrains shell commands, MCP tools, remote services, or every possible route by which an agent might affect a system. The notes identify the covered tools as Write, Edit, and NotebookEdit. Finally, the source lists release changes but provides no comparative benchmarks, adoption data, or explanation of how broadly the graphical compaction behavior has been tested in production workflows.
Skill to build now
Create a disposable repository with a clearly defined working directory and enable permissions.blockWritesOutsideWorkingDirectories. Run a task that asks Claude Code to edit an in-repository file and then attempts to target a path outside that directory. Record the observed refusal and verify that your managed configuration applies in the editor and desktop workflows your team actually uses. Separately, resume a deliberately long GUI session after it has been idle to see how the generated summary preserves task state.
Reader takeaway
Claude Code 2.1.314 makes long-session compaction available in VS Code and Claude Desktop after a 200K-token conversation is left idle with an expired prompt cache. It also gives teams a direct setting to stop its core write tools from editing outside designated working directories.