Version 2.1.329 is the newest entry in the official Claude Code changelog, and I read its full bullet list from the raw file. I did not run it, and the changelog gives no release date. Everything below is a reading of release notes, not a field test.
The release has about a hundred bullets. Most are fixes. The pattern across them is more interesting than any single line: Anthropic is putting limits and honest failure messages around features that let the agent run unattended. That is where agentic tooling has to mature, and it is the part IDE wrappers barely have to think about, because a human is clicking “accept” on every step.
Code Mode: from feature to something you can leave running#
Code Mode, where Claude writes a script that makes many tool calls, takes the largest share of this release. The fixes read like a list of how it can go wrong:
- A runaway loop could fill the machine’s memory. It is now stopped at a limit.
- Code Mode kept every large text file the code read in memory, and memory grew each time a background subagent woke up. Both are fixed.
- A script reading a line range over 256 MiB, which could use gigabytes, is now refused.
- Scripts that hit the ten-minute limit while an MCP form or prompt was waiting for an answer no longer die mid-question.
- MCP calls over two minutes used to hand the script a “moved to the background” message as if it were data. The script now waits up to ten minutes, then gets a proper error.
- A script’s
let x;no longer inherits the value an earlier script gavex. - Scripts lost access to tools in sessions with more than 4,096 tools. Fixed.
Two behavior changes matter for anyone building policy around hooks. First, a PreToolUse hook that errors no longer refuses a script’s tool call, which matches behavior outside Code Mode. If a hook must refuse, set "onFailure": "block" on a command, HTTP or MCP tool hook. That is the same fail-closed switch added for mcp_tool hooks in 2.1.328, now generalized. Second, Code Mode now starts even when telemetry, nonessential traffic or feature flags are turned off, and an administrator can keep it off in managed settings. It can also now run alongside OpenTelemetry events and metrics, with each tool call the code makes sending its events. Tracing still keeps it off.
If you run Code Mode in CI or on a shared machine, the practical advice is simple: put the “keep it off” decision in managed settings, and put onFailure: block on every hook that enforces something.
/tasks becomes the one place to look#
The /tasks view now includes the session’s scheduled tasks, /loop wakeups, the goal and the conversation’s forks. Pressing x stops one after asking. The header and the footer’s background count are now one number covering shells, monitors, MCP calls, agents, workflows and scheduled tasks.
That sounds cosmetic until you have three forks, a /loop and a monitor running. One count and one list is what lets you trust an agent you are not watching.
Related changes:
- “Exit and stop tasks” now ends the session’s scheduled tasks so they don’t run again on resume, unless you exit while Claude is responding.
- Saved scheduled tasks from earlier versions mostly need scheduling again in new signed-in sessions. Resuming their own session still runs them. If you rely on scheduled tasks, check yours after upgrading.
- Stop and SubagentStop hooks now list a Monitor’s script as
monitor, notshell, inbackground_tasks. If your hooks parse that field, update them.
/code-review: fewer silent mistakes#
I covered the narrowing of /code-review in the previous release. This one fixes how it behaves:
/code-review <pr number>sometimes reviewed the checked-out branch instead of the pull request, without saying so. It now reads the PR withgh pr diff, or says when it can’t.- With
--fix, or given a PR URL, it now stops without reviewing ifgh pr diffis refused or fails. - With no target it reviews the whole branch against the default branch, and no longer loses the diff on multi-commit branches.
- It no longer halts at a permission prompt to look for a
~/.claude/CLAUDE.mdthat doesn’t exist. - The findings card no longer often shows “high” for a review that ran at medium.
The first item is the important one. A reviewer that quietly reviews the wrong code is worse than none, because it produces confidence. Failing loudly is the right design.
Hooks, trust and MCP#
- A trust dialog now warns where a folder’s hooks are declared. Given the poisoned project config attacks this ecosystem has already seen, showing the source of a hook before it runs is overdue and welcome.
- Hooks in managed settings that fail to load now produce a warning in Claude Desktop and other non-interactive hosts, instead of silently not applying. This is the same fail-loudly principle.
/mcpgets a “Server message” row showing the JSON-RPC error from a remote HTTP MCP server that refuses the connection./statusgets a “Marketplace auto-update” row naming managed marketplaces that get no automatic updates, and why.- Claude is now told when a connected MCP server’s tools could not be loaded in sessions with tool search, instead of assuming the server has no tools.
- Files that MCP servers or websites send as
application/octet-stream, such as PDFs, images and Office documents, were saved as.bin, so Claude couldn’t read them. Fixed. - Claude Code now reconnects most HTTP MCP servers by itself when one is left on an older protocol version.
Auto mode and permissions#
Auto mode runs git status less often, so requests wait less in large repositories. The release notes also say a git command that discards changes may be blocked until the working tree is checked. That is a deliberate trade, and it leans toward safety. Auto mode also no longer asks for manual approval to rm a symlink that points outside the working directory or at itself, and it handles a short reply such as “yes” better when telemetry is off, behind a gateway, or on Microsoft Foundry. /auto-mode-setup now considers recent auto mode denials when suggesting rules, and you review each suggestion before it is saved.
On Linux, /sandbox was listing allowUnixSockets entries as allowed. They don’t open those sockets there, and /sandbox now says so. A security display that overstated what was allowed is exactly the kind of bug that matters.
Self-hosted runner and platform fixes#
The self-hosted runner gets several fixes: Git LFS repositories cloned through Anthropic-managed git now start on pointer files, and git lfs pull fetches the content. SSH host aliases, GitHub SSH certificates and GHE.com SSH users now work. --push-outcome-on-release could delete a session’s branch on the git host, and that is fixed. Some changes need attention: a second host name’s clone of the same owner/name goes under _repos/clones/<host>/, and repositories with unusual characters in their address path clone anew, so rules naming the old folder stop applying.
Elsewhere:
~/.claude.jsoncould be reset to defaults when a network stall or file lock made it unreadable during a save.- On macOS, repeated sign-outs happened when Claude Code could read the Keychain but not save to it.
- 2.1.326’s automatic retry at the limit named in a
max_tokenserror was removed. Behind a gateway whose model takes fewer output tokens, setCLAUDE_CODE_MAX_OUTPUT_TOKENSto that limit. - Computer use on macOS moves to newer safety rules, so some actions are refused and some apps have less access.
- The VS Code extension gets a long list of input-method, accessibility and diff fixes.
What to take from it#
Nothing here is a headline feature, and that is the point. The agent is increasingly something you leave running: scheduled tasks, forks, monitors, scripts making thousands of tool calls. The work that matters is bounding memory, making failures visible, and making the control surface honest. A tool built for a human clicking “accept” does not need most of this. A tool built for autonomy cannot ship without it.
What I did not check#
- I read the changelog only. I did not run 2.1.329 or test any flag.
- The changelog gives no release date, so I can’t place it relative to other releases.
- Some behavior changes, such as which saved scheduled tasks need rescheduling, are described tersely. Test them before depending on them.



