Claude Code v2.1.323 adds experimental Code Mode
The release adds an opt-in mode where Claude calls its tools by writing JavaScript, a plugin version range in plugin.json, stricter plugin validation and a long list of permission and sandbox fixes.
Claude Code v2.1.323 ships an experimental Code Mode
Anthropic published Claude Code v2.1.323 on 12 November. The release notes for v2.1.323 read mostly as a changelog: new settings, one experiment behind an opt-in, and a long list of fixes, many of them to permission prompts and the Bash sandbox.
The headline addition is Code Mode, which the notes describe as an experimental opt-in where Claude writes JavaScript to call its tools. The notes point to a documentation page for turning it on, or for an organization to turn it off. That is as far as the source goes: it does not show what the generated JavaScript looks like, which tools are reachable through it, or how a Code Mode call differs from an ordinary tool call.
Plugins declare a Claude Code version range
Plugins can now set engines["claude-code"] in plugin.json. A plugin outside its stated range is not installed, updated or loaded, and install and update pick the newest version that fits. The claude plugin validate command gained type checking for a plugin's TypeScript hooks module, with each finding reported as a warning and --strict failing on it. It also checks the $ nouns a mod's engine.create hook adds, withholds or replaces. Separately, claude plugin install --json returns unsetOptionCount, the number of userConfig options not yet set.
Heartbeats and settings for host programs
Several changes target programs that host Claude Code. A progress heartbeat runs every 30 seconds while long-running hooks execute in claude -p and SDK sessions, so a host such as Claude Desktop will not treat the turn as hung, though the notes say this applies once the host includes this version. CLAUDE_CODE_AUTO_MODE_MODEL is now supported in managed settings for Claude Desktop sessions on third-party providers, with the same caveat. CLAUDE_CODE_RESUME_HINT_COMMAND lets a program that starts Claude Code name its own resume command in the exit hint. CLAUDE_CODE_DISABLE_AGENT_PROGRESS_SUMMARIES, noted as available since 2.1.318, turns off subagent progress summaries and the requests that write them when set to 1.
CLAUDE_CODE_PERFORCE_MODE gets a one-time startup notice when Perforce settings are found and the mode is off. Artifact tool asset uploads accept .docx. The diff panel gained a hint for closing with /diff, a startup warning flags Bash allow rules with a * inside a variable's value, and a plugin language server that cannot start now produces a warning naming the server and what to check.
The fixes concentrate on permissions and the sandbox
A large share of the release closes gaps where a command or edit could act without a prompt. The notes list fixes for permission checks that missed part of a command, for deny and ask rules missing commands run through sudo or env, for allow rules with attached wildcards skipping the prompt, and for sandbox.excludedCommands letting an unnamed command run outside the sandbox. Write and Edit could write to a path covered by a Read deny rule, and sandboxed commands could read Claude Code's file checkpoint copies; both are listed as fixed. When auto mode is unavailable, sessions now start in Manual mode rather than a mode that could allow more.
Other fixes cover /compact summarizing away a message that had not been answered yet, a memory spike after a very large file change, a setting saved in a running session undoing a hand edit to settings.json, and several cases where a background session or a profile change could leave the previous account's credentials in place.
Limits and open questions
The notes are cumulative and terse, with many changes described in a single line. Some items cannot be judged from them at all: what Code Mode offers beyond writing JavaScript, whether it is switched on per user or per organization, and which hosts already include the version needed for the heartbeat and managed-settings changes. The opt-in label and the organization-level off switch suggest the feature is still early, but the source gives no timeline for it leaving that stage. There are no benchmarks or performance figures in the source.
What this means for building apps with AI
If you run Claude Code alongside a web app builder, the plugin version range and the stricter validation likely matter more than individual fixes: they make it harder for an out-of-date plugin to load unnoticed. If you would rather describe an app in plain language and have the code written, built and hosted, Draftship does that, with a built-in database and user sign-in. Draftship is not affiliated with Anthropic, and nothing in this release says it supports Draftship or the reverse.
Primary sources
- v2.1.323 · Claude Code releases, Nov 12, 2026
This article was drafted with AI assistance from the primary sources listed below and checked automatically against them. An editor hasn't reviewed it yet.