Claude Code v2.1.273 Major Updates - Fix for early auto-compact triggering and security fixes around permissions

This is Ishikawa from the Cloud Business Division. Claude Code v2.1.273 (released 2026-09-15) has been released. This felt like a release with a lot of changes for a single update, and it took some time to go through. Today, I tried forking a Remote Control session. (But it didn't work.)

The previous update article is here.

https://dev.classmethod.jp/articles/20260915-cc-updates-v2-1-272/

Update Summary

v2.1.273 includes 64 changes. In my classification, fixes are the most numerous at 31, followed by improvements at 19, security at 8, new features at 5, and performance at 1.

Of the 64 changes, 24 are changes for peripheral platforms tagged with [VSCode] / [Claude Code on the web] / [Claude Tag] / [Code Review], and the remaining 40 are directly relevant to the terminal version.

Highlighted Updates

Forking Remote Control Sessions (New Feature)

Sessions started with claude --remote-control or /remote-control can now be forked from the Claude app. The fork runs as a background session on your local machine.

Since you can now create branches from a location away from the terminal, I feel this is a meaningful change for those who work back and forth between the Claude app and the terminal.

Fix for auto-compact Triggering at Half the Actual Window Size (Bug Fix)

The context meter and auto-compact were counting advisor tool turns at approximately twice the actual context size. As a result, auto-compact was triggering at roughly half the actual window size.

For those who have been running long sessions and felt "there should still be room left but compression is running," I expect the experience will change with this version.

Fix for Sub-agent Results Not Being Delivered (Bug Fix)

When streaming responses didn't include token usage at the end or lacked a model ID, sub-agents and background agents were treated as failures and results were not being delivered.

For those building workflows that heavily use sub-agents and background agents, issues that were previously handled as unexplained failures may now be resolved.

Fix for Permission Check and Management Setting Application Gaps (Security)

Three fixes were made for cases where configured restrictions were not actually taking effect.

• Bash commands that the permission checker couldn't fully parse were skipping confirmation prompts under permissions.blockReadsOutsideWorkingDirectories. Additionally, a problem where subshells in bypass mode could hide dangerous rm commands has also been fixed.
• Due to another gap in applying permissions.blockReadsOutsideWorkingDirectories, memory directories specified in repository settings were being used for loading into prompts, recall, indexing, and memory extraction.
• allowManagedMcpServersOnly, deniedMcpServers, and disableClaudeAiConnectors configured via MDM or managed-settings.json were being ignored when server management settings also existed.

For organizations managing settings via MDM or managed-settings.json, I think it's worth verifying after updating that the intended settings are actually taking effect.

Reversion of Applying Deny Rules to Unparseable Bash Lines (Security)

The v2.1.268 change that applied Read/Edit deny rules to Bash lines that the permission checker couldn't parse (such as eval, env -C) has been reverted. Commands like time -p make build will no longer be rejected and will again show a confirmation prompt.

For those restricting Read/Edit with deny rules, behavior will differ from between v2.1.268 and v2.1.272, so I recommend checking how your local rules will apply.

Update Details

New Features

• New request headers for LLM gateways have been added: x-claude-code-request-class, x-claude-code-agent-type, x-claude-code-prev-tool-durations, x-claude-code-compaction, and x-claude-code-context-compacted. Opt in with CLAUDE_CODE_GATEWAY_HINT_HEADERS=1.
• A notification has been added to suggest /mcp when an MCP server disconnects mid-session and gives up on automatic reconnection.
• Forking of sessions started with Remote Control has been added (as mentioned above).

Security

• Fixed the issue where Bash commands that the permission checker couldn't fully parse were skipping confirmation prompts, and the issue where subshells in bypass mode could hide dangerous rm commands (as mentioned above).
• Fixed the gap in applying permissions.blockReadsOutsideWorkingDirectories that allowed memory directories specified in repository settings to be used (as mentioned above).
• Fixed the issue where MDM / managed-settings.json settings were being ignored when coexisting with server management settings (as mentioned above).
• Reverted the application of Read/Edit deny rules to unparseable Bash lines (added in v2.1.268) (as mentioned above).
• Fixed an issue where synced skills from claude.ai remained usable after an organization disabled Skills. They are now moved to a recoverable trash.
• Signing in with a Claude account now also requires access to claude.ai plugins.
• /bug and /feedback reports now include only model behavior-related parameters (model, system prompt, tools) from the most recent API request, excluding request metadata and CLAUDE_CODE_EXTRA_BODY fields.
• Improved network path permission checks for UNC paths when adding an assigned network drive to --add-dir on Windows.

Improvements

• Responsiveness in long sessions has been improved. Hook progress display and sub-agent activity no longer reprocess the entire conversation on each update.
• Auto mode for Bedrock, Vertex, and Foundry now uses the local classifier by default for the time being. Setting CLAUDE_CODE_AUTO_MODE_SERVER=1 will use the server-side classifier for each platform.
• SSL certificate and proxy connection errors during a session now indicate the error code and what needs to be fixed (such as NODE_EXTRA_CA_CERTS for untrusted internal CAs).
• OTEL_LOG_TOOL_DETAILS=1 now also includes the actual agent name, skill name, plugin name, and MCP server name in cost and token metrics.
• Artifact database writes can now delete a single field without rewriting the entire document.
• Artifact publishing now safely retries requests that were cut off after reaching claude.ai, preventing failures and duplicate version creation.
• Additionally, improvements have been made to provide more specific causes and remediation for /web-setup, /autofix-pr, MCP server sign-in expiration, cloud session GitHub errors, and Artifact tool error displays.

Fixes

• Fixed double-counting in context meter and auto-compact: Advisor tool turns were being counted at approximately twice the actual size (as mentioned above).
• Fixed sub-agent and background agent results not being delivered: Missing token usage or model ID in streaming responses caused them to be treated as failures (as mentioned above).
• Fixed Read rejecting files on macOS: For files that the system also reports under different paths, such as screenshots passed by drag-and-drop, Read was rejecting them as "symlink resolution changed after permission was checked."
• Fixed authentication error display for Bedrock, Vertex, and Foundry: 401/403 errors and Claude app gateway 403s were uniformly prompting users to run /login. They now specify which credentials need to be updated or guide users to the gateway administrator.
• Fixed full prompt cache rewrite: /login, /upgrade, and /extra-usage were discarding all thinking up to that point in the conversation, forcing a full cache rewrite on the next request.
• Fixed scheduled task execution session mix-up: After copying .claude/scheduled_tasks.json to a different folder such as a new worktree, saved tasks were being executed under a different session.
• Fixed missing sub-agent messages in SDK output: When a sub-agent was moved to the background during execution (such as with CLAUDE_AUTO_BACKGROUND_TASKS), SDK and --output-format stream-json output was dropping the remaining messages and final report.
• A quietly appreciated fix: The issue where a ! typed at the beginning of a prompt when already in shell mode would disappear has been fixed, making it possible to enter negation commands like ! grep …. I personally find this a welcome fix, as I had unknowingly retyped commands multiple times due to this behavior.
• Additionally, numerous minor bugs have been fixed, including /tui restart, .git/info/exclude recreation, /install-github-app error display, context usage retrieval from Remote Control, and spinner display.

Peripheral Platform Changes

The following are changes outside the terminal version. You can skip this section if you use Claude Code in the terminal.

VS Code Extension

• Fixed an issue where "Report a problem" kept appearing in organizations with product feedback disabled, and /bug and /feedback were opening the report form.
• Fixed an issue where a red "Claude Code process exited with code 4294967295" banner appeared after turn completion on Windows.

Claude Code on the web

• The new routine page and routine editing dialog now prompt with "Discard unsaved changes?" before discarding entered names, prompts, and edits.
• Fixed an issue where, after an administrator deleted and re-added an organization's connector, routines lost access to that connector and kept calling the old one.
• Fixed an issue where creating a self-hosted environment from organization settings could fail with a server error and leave a partially created environment.
• The "Share cloud sessions" setting for administrators has moved from the Claude Code page to under Data and privacy. Data and privacy administrators can also manage it.
• The routine detail page layout has been reorganized, and the full-screen desktop app download display that appeared for new users without a cloud environment on Mac and Windows has been removed.

Claude Tag (Slack Integration)

• Fixed an issue where AWS connections were rejecting region-less endpoints like Budgets, Savings Plans, WAF Classic, and Import/Export. Requests to Global Accelerator are now signed correctly. Additionally, when signing fails, the reason and remediation are now communicated to Claude.
• Fixed an issue where Claude replying in an old Slack thread during task execution would cause the task to restart from the beginning, potentially losing work before a push.
• Fixed an issue where scheduled tasks configured in a private Slack channel shared within an organization would complete without posting anything. They now continue operating in the created thread.
• Claude now proactively checks related public channels, such as the incident channel that is the basis of a conversation, rather than only when asked.
• Additionally, fixes have been made for unresponsiveness during Enterprise Grid disconnection, missing messages immediately after token refresh, OAuth connection failures with providers that return token types in lowercase, rejection when adding channel managers, and missing display on the admin Memory page.

Code Review

• Fixed an issue where @ mentions, multi-line code spans, and HTML tags wrapped in backticks caused the entire REVIEW.md to be ignored. Now only lines containing links to changed files are excluded.
• Fixed an issue where merging the base branch into a PR that had "Additional findings" from a previous review would trigger a full re-review. It now performs a lightweight follow-up review.
• Fixed an issue where retrying after a GitHub error with /ultrareview --post would result in finding comments not being posted or being posted twice. Comments now explicitly state the commit being reviewed.
• Fixed an issue where empty or identical-content pushes were being re-reviewed for GitHub repositories with uppercase letters in the owner or repository name.
• Suggested fixes now indicate what must not be broken when other code depends on the behavior being changed, and comments pointing to a second affected location are now written in complete sentences.

Trying to Fork a Remote Control Session (But It Didn't Work)

I attempted to verify whether a forked session would start as a background session on my local machine, but the fork failed.

First, on the Claude app (iPhone), the message was /fork isn't available over Remote Control. I had somewhat expected this since there was no update on the App Store.

Next, I tried the Claude desktop app (Claude 2.110.0 (dfb2ba) 2026-09-15T05:50:55.000Z) and the web app, with the same result.

The original terminal also displayed /fork isn't available over Remote Control., so this seems to be fairly confirmed.

I hope to verify this again at a later date.

Closing Thoughts

v2.1.273 feels like a release that leans more toward fixing bugs that are easy to encounter in day-to-day work rather than new features. In particular, the fix for auto-compact triggering at half the actual window size should produce a noticeable difference in experience for those running long sessions. It was unfortunate that forking Remote Control sessions didn't work, but from the perspective of someone who always writes "tried it out" posts shortly after release, this kind of thing happens.

For those running long sessions or organizations managing permissions and administrative settings, I encourage you to update and check how the behavior has changed.

References

https://github.com/anthropics/claude-code/blob/main/CHANGELOG.md

https://code.claude.com/docs/en/changelog

https://dev.classmethod.jp/articles/20260915-cc-updates-v2-1-272/