Version 2.1.327 is the newest entry at the top of the official Claude Code changelog. I read the raw file, which is over a million characters, and only its first 100,000. This post is a summary of release notes, not a field test. I have not run this version, and the entry lists 172 bullets, so I cover the ones that matter for autonomous workflows and skip the terminal-rendering fixes.
The headline: things that used to disappear#
Most releases are dominated by features. This one reads like a cleanup of ways the tool could quietly destroy your files. Per the changelog:
- Project-scope
claude mcp addandclaude mcp removecould delete other top-level keys in.mcp.json. That is fixed, andclaude mcp add --scope projectnow stops instead of overwriting when the file lists servers at the top level without anmcpServerskey. /exportcould overwrite an existing file without asking. It now asks, and/export <name>keeps the file.- In Remote Control, stopping a worktree-mode server could delete a running session’s git-ignored files, such as
.env. - Self-hosted runners started with
--use-anthropic-git-proxycould delete a non-emptyGIT_CONFIG_GLOBALdirectory at startup. It is now kept, with a warning. - Compaction after a reply cut off at the output token limit could lose that reply’s tool call and result.
ExitWorktreenow asks before removing a worktree that may hold unsaved work.
If you run agents unattended, this is the category that hurts. A permission prompt is a visible failure. A deleted .env or a half-written .mcp.json is the kind you find on Thursday. These are the changelog’s claims; I haven’t reproduced any of the original bugs.
Path checks get stricter#
The changelog lists several changes to how Claude Code decides what is inside your project:
- Path checks now refuse some symbolic links.
- What counts as “inside a working directory” changed, so some paths now ask for permission where they didn’t before.
permissions.blockReadsOutsideWorkingDirectoriesnow asks about more Bash commands in bypass and auto modes, denies some in auto mode, and refuses more Bash commands naming paths under Read or Edit deny rules.- Bash commands whose variables expand to very long text are denied.
- Claude in Chrome now refuses to upload some credential files.
The changelog doesn’t say which commands or paths are affected, so I can’t tell you what will newly prompt. The practical takeaway: if you have unattended jobs that read through symlinks or outside the repo, run them once interactively after upgrading and watch for new prompts.
Auto mode: fewer allow rules honored, fallback going away#
Two entries concern auto mode:
- Auto mode now ignores more package-running allow rules, with
Bash(pnpm dlx *)as the example, along with path or version forms of the ignored rules. The reasoning is plausible: an allow rule for a package runner is effectively an allow rule for arbitrary code from a registry. But note the consequence. A rule you wrote may silently stop doing anything in auto mode. - The notice for the billed-classifier fallback changed, and the changelog says an upcoming release removes the fallback. It doesn’t explain what replaces it. I’d read this as a warning to check how your auto mode setup behaves when the classifier isn’t available, rather than assuming the fallback will keep covering you.
Also fixed: auto mode ignoring some wildcard PowerShell allow rules naming a .\ script, and Code Mode scripts getting prompts and denials in auto mode when a second script started. Code Mode itself now starts even when an installed mod hooks tool calls, though administrators can still disable it in managed settings.
Subagents and background work#
For people running agent teams, several fixes target model and permission leakage:
- A finished subagent could resume on a sibling subagent’s model and effort. Subagents launched without a model now stay on their starting model when resumed, including after
/model. - In-process teammates on a different model used to compact at the lead’s limit. They now use their own limit and model.
- A subagent resumed by a sibling’s message now follows session permission settings.
- Background agents: fixes for the main conversation not learning that one had finished, and for Claude reporting one as running while it sat on a permission prompt.
This is the kind of plumbing that decides whether multi-agent runs are trustworthy. Nothing here is flashy, and all of it is claimed rather than tested by me.
Operational details worth knowing#
- MCP via stdin:
claude mcp add-json <name> -reads the server JSON from stdin, so tokens stay off the command line and out of shell history. It is not available inside a Claude Code session. - Remote MCP retries: most remote servers that stayed disconnected after long outages are now retried in the background for about a day.
/statusas a linter: it now flags broken agent, skill, command and output style files, keybinding errors, an unavailable sandbox, ignored permission settings and hook matchers that never match. Hook matchers that never match are a real silent failure for anyone enforcing policy through hooks.- Gateways: a managed-settings checksum header (
x-claude-code-managed-settings-checksum) is added to gateway hint headers whenCLAUDE_CODE_GATEWAY_HINT_HEADERS=1, andattributionProjectnames a session’s billing group on Claude apps gateways. - Shell head start: the first Bash command is faster, and shell start-up files may now run before approval.
CLAUDE_CODE_DISABLE_SHELL_HEAD_START=1turns this off. If your shell profile has side effects, that is worth a look. - Secrets in commits: on newer models and in PowerShell, Claude is told not to commit likely secret files. That is an instruction to the model, not an enforcement mechanism, so a pre-commit hook is still the right backstop.
- WebFetch: URLs over 2,000 characters are refused before any prompt, and only the first 100
@mentions in a message are looked up. - Telemetry: in Anthropic-hosted cloud sessions,
CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFICnow acts at most asDISABLE_TELEMETRY. /tui classic: the new name for the non-fullscreen renderer./tui defaultstill works.
What I’d do with this release#
- Upgrade, then re-run one unattended job interactively. The symlink and working-directory changes can introduce new prompts that would stall a headless run.
- Audit your auto mode allow rules. Anything shaped like
Bash(<package-runner> *)may no longer be honored. - Commit your
.mcp.jsonand.envbackups. The old data-loss bugs are reportedly fixed, but a version-controlled.mcp.jsoncosts nothing. - Run
/status. If it finds a hook matcher that never matches, you’ve learned that a policy you thought was enforced wasn’t.
The pattern#
This is the third changelog post in a row on the 2.1.32x line, and the theme is consistent: the product’s trust boundary is the permission system, and most of the work is closing gaps in it. That is the right work. A terminal-native agent with real shell access is only as good as the boundary around it, and a changelog full of boundary fixes is evidence the team treats it as the product. It is also a reminder that the boundary is still being found. Treat each release as something to verify against your own workflows rather than trust.


