Claude Code Week 28: the desktop browser and /doctor checklist I run before enabling it org-wide

Introduction

Week 28 is not a model launch. It is a capability launch that changes the blast radius of a coding agent. A sandboxed browser that can click through docs and vendor portals is useful. It is also a new way for untrusted pages to enter the same session that can edit your repo.

I have already seen one team enable Desktop browsing for everyone on Monday and spend Wednesday explaining why an agent followed a "quick fix" link from a random Stack Overflow mirror. The features are good. Defaults without policy are not.

What shipped in Week 28

Desktop browser: govern it like Computer Use

The browser does not inherit your Chrome logged-in profile. That is good for isolation and bad for people who assume "it can just open Jira." Plan for explicit auth paths, MCP connectors, or human paste for private systems.

1. 01
   Decide persistence before rollout

   Persistent browsing sessions are convenient for multi-step research and dangerous if cookies accumulate across tasks. Default to non-persistent for regulated teams until you have a written exception.
2. 02
   Allowlist domains by team

   Docs sites and your staging URLs are fine. Random marketing blogs and pastebins are not. Match the discipline you use for Cowork connector allowlists.
3. 03
   Separate browse-research agents from write-fix agents

   A session that browses the public web should not also hold production deploy credentials. Split roles the same way you split triage vs fix after Agentjacking.
4. 04
   Log browser actions into observability

   If you cannot see which domains an agent visited during a PR session, you cannot audit an incident. Wire events into the agent observability stack.
5. 05
   Disable entirely where policy says so

   Some enterprises will turn browser tools off for contractors or for repos that touch customer PII. That is a supported outcome, not a failure to adopt Week 28.

/doctor: run it, but do not auto-apply blindly

The upgraded /doctor is the part I want every team using weekly. Unused MCP servers burning context, local CLAUDE.md files shadowing the checked-in one, and slow hooks that add latency to every turn are real cost and reliability bugs.

• Run /doctor at the start of a migration week (Sonnet 5, MCP SDK bumps, new plugins).
• Read the full finding list before accepting any fix.
• Never pipe /doctor into a headless auto-apply in CI without a human-named allowlist of safe repairs.
• Treat CLAUDE.md trim proposals as suggestions. Claude can derive import structure; it cannot derive your compliance constraints.

Auto mode hardening you should not undo

Transcript tampering protection closes a class of attacks where a malicious tool result rewrites history so a later turn "remembers" a fake approval. That pairs with governing agent autonomy: classifiers and Auto-review are speed bumps, not a security boundary, but removing a free speed bump is malpractice.

The rm -rf confirmation on unresolved variables is the same story. Keep it. Teach engineers why the prompt appeared instead of hunting for a "trust me" toggle.

Rollout checklist for eng leadership

1. 01
   Pin minimum Claude Code version

   Require v2.1.205+ so /doctor repair and auto-mode protections land together. Older Desktop builds without the browser are fine; mixed versions create support noise.
2. 02
   Publish a one-page browser policy

   Persistence default, domain allowlist owner, and who can request exceptions. Link it from onboarding.
3. 03
   Schedule a /doctor week

   Every squad runs /doctor once, files tickets for slow hooks and unused MCP servers, and merges CLAUDE.md dedupe PRs.
4. 04
   Reconfirm Auto-review block_instructions

   Especially for commands that originate from browsed page text or issue trackers. Week 28 makes browsing easier; it does not make untrusted HTML safer.
5. 05
   Update spend and usage dashboards

   Browser-heavy research sessions look different from local refactors. Attribute them in the unified agent spend dashboard so finance does not blame "Claude" as one blob.

Conclusion

Week 28 makes Claude Code better at reading the outside world and cleaning its own setup. That only helps if you treat browser access and repair tools as governed production features. Ship the version bump. Write the allowlist. Keep auto-mode protections on. Then let /doctor earn its keep every Monday.

Sources: Claude Code Week 28 changelog at Claude blog (releases v2.1.202–v2.1.206).