Claude Code 2.1.323: experimental Code Mode (Claude writes JS to call tools), plugin version pinning, permission fixes
retained text
2.1.323 came out on Nov 12. The main items: \- Code Mode (experimental, opt-in): Claude writes JavaScript to call its tools instead of calling them one at a time. Org admins can turn it on or off, see the docs. \- engines\["claude-code"\] in plugin.json: plugins outside their stated Claude Code version range won't install, update or load. Plugin authors can finally stop breaking older setups. \- claude plugin validate now type-checks TypeScript hooks modules. Warnings by default, --strict fails on them. \- Unrecognized model IDs now get a 1M context window instead of 200K. Set CLAUDE\_CODE\_DISABLE\_1M\_CONTEXT=1 to keep 200K. Matters if you use gateways or custom model names. \- /security-review now asks before Claude runs it. \- Fix: --permission-mode auto (or defaultMode: auto) could start in a looser mode when auto was unavailable. It now starts in Manual and asks first. Worth updating for this one alone. \- 30-second progress heartbeat for long hooks in claude -p and SDK sessions. Code Mode is the one I want to try. Has anyone turned on Code Mode yet? Does it cut down the back-and-forth tool calls on bigger tasks, or does it add new failure modes? Source: https://github.com/anthropics/claude-code/releases/tag/v2.1.323
document fetched 2026-11-15 00:35 · canonical https://reddit.com/r/ClaudeCode/comments/1xvw576/claude_code_21323_experimental_code_mode_claude