Claude Code 2.1.329: What Every Professional Needs to Know About This Sweeping Update

Reading Time: 6 minutes

Claude Code 2.1.329 delivers sweeping changes to Code Mode safety and performance, MCP server reliability, scheduled task lifecycle, and macOS computer use — along with dozens of bug fixes. Non-technical professionals and enterprise teams should review the breaking change to saved scheduled tasks and the new restrictions on computer use before updating.

What Just Shipped in Claude Code 2.1.329

According to the Claude Code changelog at code.claude.com, version 2.1.329 is one of the most wide-ranging updates the tool has received in a single release. It touches Code Mode, scheduled tasks, MCP (Model Context Protocol) server handling, computer use on macOS, memory management, the self-hosted runner, and dozens of bug fixes that affect everyday reliability. If you use Claude Code — whether through Claude Desktop, a VS Code extension, or a managed enterprise setup — this update will change how the tool behaves in ways that are worth understanding before you encounter them.

This post breaks down the most consequential changes in plain language, explains what they mean in practice, and flags the tradeoffs and limitations you should be aware of.


Code Mode Gets Smarter and Safer

Code Mode — the setting that lets Claude run scripts, make tool calls, and automate multi-step coding tasks — received several important changes in 2.1.329.

First, the changelog notes that Code Mode can now start even when telemetry, nonessential traffic, or feature flags are turned off. Previously, turning off certain monitoring options would also disable Code Mode, which was frustrating for teams with strict data policies. Administrators can still keep Code Mode off in managed settings, but the default is now less restrictive.

Second, Code Mode now handles PreToolUse hook errors differently. The changelog specifies that a PreToolUse hook that errors no longer refuses a script’s tool call — this brings behaviour in line with how hooks work outside Code Mode. If you need a hook to actually block a tool call, you now have to explicitly set "onFailure": "block" on the relevant hook configuration. This is a subtle but important change for teams that rely on hooks to enforce security rules.

Third, there is a new memory protection: Code Mode will now refuse a script’s attempt to read a line range over 256 MiB, which could otherwise consume gigabytes of RAM. The changelog also notes a fix for a runaway loop in Code Mode that could fill a machine’s memory — this is now capped at a limit. For anyone running Claude Code on a laptop or a mid-range cloud VM, this matters.

Finally, the changelog records that Code Mode’s terminal now freezes for far less time after a script makes thousands of tool calls — a performance improvement that will be noticeable in heavy automation workflows.


Scheduled Tasks: A Breaking Change to Be Aware Of

The changelog introduces an important behavioural change for scheduled tasks that qualifies as a breaking change for some users. In new signed-in sessions, most saved scheduled tasks from earlier versions will need to be scheduled again. The exception is --resume of their own session, which will still run them.

Additionally, ‘Exit and stop tasks’ now ends a session’s scheduled tasks so they do not run again on resume — unless you exit while Claude is responding. This changes the mental model around background automation: tasks are now more tightly coupled to the session lifecycle.

The changelog also adds scheduled tasks and loop wakeups to the /tasks command, so you can see everything running in a session at once. The /tasks header and the footer’s background count now show a single number covering shells, monitors, MCP calls, agents, workflows, and scheduled tasks together.


MCP Server Reliability: Fewer Silent Failures

For teams using Model Context Protocol servers to connect Claude Code to external tools and databases, 2.1.329 adds meaningful visibility and reliability improvements.

The changelog notes a new “Server message” row on a server’s page in /mcp, which shows the JSON-RPC error message of a remote HTTP MCP server that refuses the connection. Previously, a refused connection could appear as a silent failure. Now you get the actual error.

There is also a fix for background MCP tasks that stalled after a slow reconnect: Claude Code now reconnects most HTTP MCP servers automatically when one is left on an older protocol version. And a fix for PDFs, images, and .docx, .xlsx, or .pptx files that MCP servers send as application/octet-stream being saved as .bin — which previously prevented Claude from reading those files at all.

A Scenario: A Finance Team in Pune

Consider a finance analyst team in Pune that uses Claude Code connected via MCP to an internal document server holding quarterly reports, compliance PDFs, and Excel models. Before 2.1.329, if the MCP server sent a compliance PDF with the wrong MIME type, Claude Code would save it as a .bin file and fail to read it — with no clear error. The analyst would see Claude claiming it could not access the document, with no explanation.

After 2.1.329, two things improve. The file-type fix means those PDFs are now correctly identified and readable. And if the MCP server itself refuses a connection — say, because a session token expired — the new “Server message” row in /mcp shows the actual JSON-RPC error, so the analyst (or their IT support) knows exactly what to fix rather than guessing.

This is the kind of reliability improvement that does not make headlines but saves real working time for teams whose workflows depend on external data sources.


Computer Use on macOS: Tighter Safety Rules

The changelog records that computer use on macOS has been updated to newer safety rules, meaning some actions are now refused and some apps have less access than before. Additionally, an app launch that has not finished after 30 seconds will now be abandoned rather than hanging indefinitely.

This is a tradeoff worth naming honestly: if you have workflows that rely on Claude controlling specific macOS applications, some of those may now fail or behave differently. The changelog does not enumerate which apps are affected, so you will need to test your specific setup after updating.


VS Code-Specific Changes

For VS Code users, the changelog adds a line above the message box after you decline the trust prompt, with a “Review trust” option that brings the prompt back without requiring a full window reload. This is a usability improvement for anyone who accidentally dismisses the trust dialog.

The changelog also notes a change in how command output handles characters: hidden characters and control sequences other than plain colours are now shown as symbols. The practical consequence, stated explicitly in the changelog, is that Persian, Khmer, Lao, Sinhala, and Malayalam text can look broken or out of order in command output. This is a known limitation introduced by the change, not an accidental side effect — and it is relevant for Indian professionals working in Kannada-adjacent scripts or teams handling Malayalam-language content in their repositories.


Gateway and Self-Hosted Runner: Infrastructure Notes

For teams running Claude Code behind a gateway or on a self-hosted runner, the changelog makes several changes worth flagging:

  • The start-up warning shown while access_control.allow_cidrs is empty has been changed to a notice that appears at every start, boxed in text logs. The changelog is explicit: never expose the gateway to the public internet.
  • The --repository-folder option now takes most folders in the base folder at --capacity 1; above capacity 1, one there stops start-up.
  • Repositories with unusual characters in their address path will now clone anew in a new folder. Rules and volumes naming the old folder stop applying, and the old clone stays in place. This is a migration concern for self-hosted runner operators.
  • The changelog also removes 2.1.326’s automatic retry at the token limit named in a max_tokens error. Behind a gateway whose model takes fewer output tokens, you now need to set CLAUDE_CODE_MAX_OUTPUT_TOKENS to that limit manually.

Limitations and Honest Tradeoffs

Several things are worth naming clearly:

  • Scheduled task migration is real work. If your team has built automation on saved scheduled tasks, you will need to reschedule most of them in new sessions after updating. Plan for this.
  • Malayalam and other Indic script rendering in VS Code command output is broken as a known consequence of the control-character change. This is not fixed in this release.
  • Computer use on macOS is more restricted, and the changelog does not specify exactly which apps or actions are now refused. You will need to discover limitations by testing.
  • The automatic retry on max_tokens errors is removed. If you rely on a gateway with a lower token limit, you need to set the environment variable manually or requests may fail silently.
  • Tracing keeps Code Mode off. The changelog notes that OpenTelemetry tracing still keeps Code Mode off, even though enabling telemetry events and metrics no longer does. This distinction matters for teams with observability requirements.

What to Watch For Next

The breadth of this release — spanning Code Mode, MCP, computer use, scheduled tasks, memory management, and infrastructure — signals that Claude Code is maturing rapidly as an enterprise tool. The pattern of explicit changelog entries for memory limits, hook failure modes, and gateway exposure warnings suggests Anthropic is paying attention to production reliability, not just feature addition.

If you are evaluating Claude Code for a team in India — whether a software consultancy in Hyderabad, a fintech team in Mumbai, or an IT services firm in Chennai — version 2.1.329 is worth reviewing against your specific use of MCP servers, scheduled automation, and any macOS computer-use workflows. The full changelog is available at code.claude.com and reading it against your own setup is the most reliable way to anticipate what changes.

Related stories