November 15 agent guide: Claude Code 2.1.325 and Code Mode ops
November 15, 2026 closes the window with Claude Code fleet hygiene. Version 2.1.325 tightens sandbox defaults, plugin safety, and Code Mode edge cases. This Bacancy Technology brief turns Anthropic’s experimental Code Mode docs into an operational checklist.

In this post (8 sections)
November 15, 2026 closes the Bacancy Technology window that opened on November 13. After Poppy, ChatGPT Wallet, and Cursor’s security findings, the remaining high-value builder work is Claude Code fleet hygiene. Version 2.1.325 landed November 13 with sandbox default changes, plugin path hardening, and Code Mode fixes. This brief is the operational companion to the short Code Mode mention in the November 12 Hydrogen post.
Claude Code 2.1.325 at a glance
| Area | Change | Ops impact |
|---|---|---|
| Sandbox | Sandboxed Bash follows normal permission prompts by default | Set sandbox.autoAllowBashIfSandboxed: true only where silent execution is still approved |
| Plugins | Safer unpack paths, skipped unsafe links, stronger install warnings | Re-validate private marketplaces and prepend/append plugin lists |
| Code Mode | Fixes for worktree paths, permission Yes text, and resume-during-script | Safe to re-pilot Code Mode on 2.1.325+ after 2.1.323 introduction |
| MCP | SSH/WSL redirect paste, task Esc = answer later, reconnect fixes | Retest MCP OAuth from VS Code remote windows |
| Secrets in env | Sandbox helpers and ripgrep/git run without sign-in tokens or mTLS passphrases | Good for least privilege; update any brittle env assumptions |
Sandbox and permission defaults
The highest-impact behavioral change is that sandboxed Bash commands no longer auto-run without prompts. Managed settings that enable the sandbox must also set sandbox.autoAllowBashIfSandboxed: true if the prior silent behavior is still desired. Read-only git in untrusted or non-project directories now asks first unless the path is added with --add-dir. Scheduled tasks and /loop ask before running skills that declare allowed-tools or hooks unless Skill(name) is in permissions.allow. Sandboxed writes to .claude settings, skills, agents, commands, and hooks are refused in added directories.
Code Mode: what it is
Code Mode changes the tool loop. Instead of Claude requesting Bash, Read, or Edit one at a time and reading each result, Claude writes JavaScript that can run several tool calls, including parallel work, and returns only the summary Claude needs. Users still see diffs and tool activity in the conversation; the generated code is visible in the transcript viewer (Ctrl+O). Anthropic marks the feature experimental and off by default.
Requirements and hard limits
- Claude Code v2.1.323 or later; native install; interactive terminal session (not -p, not Agent SDK, not VS Code panel, not Desktop, not cloud sessions).
- Models: Opus 5.5, Sonnet 5.5, Haiku 5.5, or Fable 5.2.
- Anthropic API account path; not Bedrock, Vertex-style gateways, Foundry, HIPAA, or zero-data-retention orgs.
- Incompatible with agent teams, Remote Control, Claude in Chrome, computer use, API key helpers, and OpenTelemetry export settings listed in the docs.
- Resume only with Code Mode still on; sessions cannot move to Desktop or background copies via /fork.
When Code Mode helps
Use it for multi-file edits, search-then-filter pipelines, and batches of independent commands. Leave it off for short Q&A, for hosts that need Remote Control, and for any workflow that depends on per-tool auto-mode classification of computed paths. In auto mode the classifier reviews the JavaScript block, not each runtime path, so PreToolUse hooks and deny rules remain essential.
2.1.325 Code Mode and MCP fixes worth knowing
- Code Mode scripts respect EnterWorktree; ~ paths no longer write into the wrong project root.
- Words typed with Yes at permission or plan prompts reach Claude during Code Mode.
- Resuming a session that ended mid Code Mode script no longer auto-continues work; Claude waits for the next prompt.
- MCP: Esc on a background task question means answer later; reconnect and elicitation edge cases fixed; VS Code remote MCP sign-in can paste a redirect URL.
Builder checklist
- 01Pin 2.1.325+ on coding fleetsUpdate stable or latest channels deliberately. Confirm changelog behavior for sandbox and plugins before Friday deploys.
- 02Document sandbox.autoAllowBashIfSandboxedIf silent sandboxed Bash is required, set it in the highest-ranking managed source and record who approved it.
- 03Keep Code Mode off in managed settingsSet CLAUDE_CODE_EXPERIMENTAL_CODE_MODE to 0 org-wide. Allow named pilots to set 1 in user settings only.
- 04Pilot with hooks firstWrite PreToolUse coverage for REPL (the Code Mode runner) and keep deny rules for secrets and production hosts.
- 05Retest MCP OAuth after upgradeEspecially SSH, WSL, and container VS Code windows that need the new redirect paste field.
Conclusion
November 15 is a Claude Code operations day. Version 2.1.325 hardens sandboxes and plugins after a week of major model and commerce news. Code Mode remains a powerful experimental accelerator for multi-tool tasks, not a default for enterprise fleets. Bacancy Technology’s rule: upgrade, re-approve silent sandbox settings, and keep Code Mode behind an explicit pilot gate.
Related reading
- November 14 Cursor security and Coding Agent Index: /blog/november-14-cursor-security-coding-agent-index-2026
- November 13 Poppy and ChatGPT Wallet: /blog/november-13-poppy-chatgpt-wallet-2026
- November 12 Hydrogen, Inkling, and first Code Mode note: /blog/november-12-gemini-4-hydrogen-inkling-sglang-2026
- November 9 Fable 5.2: /blog/november-09-claude-fable-5-2-muse-ember-2026
- November 6 Claude Code 2.1.319 and memory: /blog/november-06-claude-code-2-1-319-agent-memory-2026
- MCP topic map: /topics/mcp
Need Claude Code fleet hardening?
Bacancy Technology designs Claude Code managed settings, sandbox policy, MCP allowlists, and Code Mode pilots. Consulting: /consulting/ . Contact: /contact/ .
Sources: Claude blog ; Claude blog .
Agentic AI patterns, delivered Thursdays
What I am shipping, watching, and pruning out of client stacks each week. One email. No fluff.