Daily brief

Copilot Code Review Learns to Spot Secrets, VS Code v1.147 Lets Agents Use Your Browser, and Claude Code Stops Clobbering .mcp.json

By ยท 18 Nov 2026

A flat diagram on a bright mint background shows blank rounded boxes traveling along a line into a gate that splits into three lanes of different widths, with a small blank key shape stopped at the gate.

At a glance

  • Copilot code review can now choose a light or deep review for each pull request, and it can flag exposed secrets before you merge.
  • The new VS Code release lets agents drive your own Chrome or Edge browser in preview and lets you tell the permission reviewer what you consider safe.
  • The newest Claude Code build stops MCP commands from wiping other settings in your project MCP file.
  • Copilot CLI and Codex shipped new stable builds, and both Cursor changelogs are unchanged.

GitHub posted the code review changes at about 1:02 PM ET Tue Nov 17, VS Code v1.147 went out as a stable release at about 4:45 AM ET Wed Nov 18, and Claude Code v2.1.327 reached npm at about 3:33 PM ET Tue Nov 17. All three are inside our usual window.

Treat today as a review your reviewers day: decide whether Adaptive and secret detection belong in your repository settings, read the new VS Code permission options before anyone turns on browser use, and check your project MCP file if you use Claude Code's MCP commands.

Top Stories

1. Copilot code review chooses its own effort and flags secrets

A new option called Adaptive picks the review effort for you, because Copilot weighs the complexity and risk of each pull request and reviews it with Lite, Balanced or Max, showing which level it chose. You can select it for one pull request from the Reviewers bar or make it your default in settings, while Balanced remains the default. Adaptive never picks a level above the maximum set for your enterprise, organization or repository.

Code review can also flag exposed secrets in its comments. GitHub combines more than 500 known secret formats with model based detection for credentials that have no recognizable format, and you turn it on under Copilot, then Code review, in repository or organization settings. Adaptive and secret detection are in public preview for Copilot Pro, Pro+, Max, Business and Enterprise, and code review for Azure Repos is now generally available.

A separate post brings the same detectors to the security review specialist in Copilot CLI, the Copilot app and the Copilot SDK. It is off by default, checks your changes rather than repository history, and uses AI credits with no extra plan required. GitHub says you can run that scan on your changes or on an agent's work.

Practical dev impact: Try Adaptive on one busy repository first, because your maximum effort setting caps what it can choose. Secret detection is worth turning on wherever agents open pull requests, so a credential an agent pasted gets caught at review instead of after the merge.

2. VS Code v1.147 lets agents use your browser and take your permission instructions

The headline of today's stable release is browser use in preview, which lets agents work directly in your default Chrome or Edge browser, including sessions you are already signed in to. It needs the GitHub Copilot browser extension, a local Copilot harness session and Node.js 22 or later, and you finish sign in and sensitive steps yourself. GitHub says browser use is also in public preview in Copilot CLI, the Copilot app and JetBrains IDEs, and Copilot CLI v1.0.113 lets administrators turn it off with the managed setting features.browserUse.

You can now give the assisted permissions reviewer in Copilot harness sessions your own instructions through chat.permissions.assisted.instructions, such as always asking before installing packages, although they cannot override permission rules, organization policy or the sandbox. Requests started by automations or other agents are no longer treated as if you wrote them, so some tool calls may now ask for approval where they did not before.

Admins can now permit the Claude and Codex harnesses alongside managed settings through vscode.allowedHarnesses, though those sessions run without Copilot's managed permission, sandbox, MCP and plugin controls. VS Code desktop on macOS Ventura, macOS Sonoma and 32 bit ARM Linux is now deprecated.

Practical dev impact: Write your permission instructions before you try browser use, because an agent in your signed in browser is exactly where you want installs and pushes to ask first. Mac users on Ventura or Sonoma should plan an upgrade, and admins who allow Claude or Codex should know Copilot policies do not follow them.

3. Claude Code v2.1.327 stops MCP commands from wiping your project config

Claude Code v2.1.327 is now the latest build on npm, while the stable tag still points at v2.1.311. The most useful fix is for claude mcp add and claude mcp remove with project scope, which could delete other top level keys from .mcp.json, such as $schema or another tool's servers. A new claude mcp add-json command that ends with a dash reads the server's JSON from standard input so tokens can stay off the command line, though not inside a Claude Code session.

/export now asks before overwriting a file, stopping Remote Control in worktree mode no longer deletes a session's ignored files such as .env, and /status lists broken agent, skill and command files and hook matchers that never match.

Practical dev impact: If you share .mcp.json with another tool and used the project scope commands recently, open the file and confirm nothing went missing. Teams on the stable tag can wait, since these fixes are on the latest channel only.

Practical Impact Analysis

Our read: the code review changes lead because they reach Copilot Pro, Pro+, Max, Business and Enterprise users who review pull requests, and GitHub says the on demand secret scan can run on your changes or an agent's work. VS Code comes second because browser use is still a preview, though the permission changes and the macOS deprecation reach anyone who updates. The Claude Code release is narrower but matters to anyone sharing .mcp.json with other MCP clients.

The common thread is that tools now check more on your behalf, so the settings behind those checks matter more. Adaptive needs a sensible maximum effort, browser use needs instructions that match your own lines, and secret scanning does nothing until someone opts in. Copilot CLI v1.0.114 is the new stable build, though its notes list only fixes and changes. Codex v0.170.0 is the new stable release, with fixes for sandbox restrictions and MCP OAuth races, while the v0.171.0 alpha builds belong on a test machine only. Gemini CLI and both Cursor changelogs have nothing new.

Tutorial

  1. Open a test shell. Every command below only reads versions, public package data or a local file.
  2. Step 1 prints the npm tags for Claude Code, Copilot CLI and Codex to compare with our snapshot.
  3. Step 2 shows your local VS Code and Claude Code builds.
  4. Step 3 lists the top level keys in the project's .mcp.json, so you can confirm nothing went missing after claude mcp add or claude mcp remove.
  5. To try secret scanning in Copilot CLI afterwards, run /settings aiSecretDetection on inside an interactive session and then /security-review on your changes.
# We have not run this script end to end on every platform. It prints and checks, and changes no files.
# Run it in a test shell. Nothing is written to your shell profile.

# Snapshot at 5:58 AM ET Wed Nov 18 2026:
# claude-code latest v2.1.327, stable v2.1.311
# copilot latest v1.0.114
# codex latest v0.170.0

# 1. Current npm tags.
if command -v npm >/dev/null 2>&1; then
  npm view @anthropic-ai/claude-code dist-tags
  npm view @github/copilot dist-tags
  python3 -c 'import json, urllib.request; d = json.load(urllib.request.urlopen("https://registry.npmjs.org/-/package/@openai/codex/dist-tags")); print("codex latest", d.get("latest"))'
else
  echo "No npm command on PATH."
fi

# 2. VS Code and Claude Code on this machine.
if command -v code >/dev/null 2>&1; then
  code --version
else
  echo "No code command on PATH."
fi
if command -v claude >/dev/null 2>&1; then
  claude --version
else
  echo "No claude command on PATH."
fi

# Optional installs, left commented. Pick one channel on purpose.
# Default to the stable pin.
# stable channel, v2.1.311 as of 5:58 AM ET Wed Nov 18 2026:
# npm install -g @anthropic-ai/claude-code@2.1.311
# latest channel, v2.1.327 as of 5:58 AM ET Wed Nov 18 2026, test machine only:
# npm install -g @anthropic-ai/claude-code@2.1.327

# 3. Top level keys in this project's .mcp.json.
python3 -c 'import json, os; p = ".mcp.json"; print(sorted(json.load(open(p))) if os.path.exists(p) else "no .mcp.json here")'

echo "Done. Nothing was installed or changed on disk."

If npm shows newer tags than your snapshot, read the release notes before you change versions.

Sources:

Recommended AI prompt

Copy this paragraph into ChatGPT, Claude, Gemini, Grok, or whatever AI you use.

Prompt

You are my developer tooling lead for Wed Nov 18, 2026. Story 1: Copilot code review now offers Adaptive, which picks a review effort for each pull request within our maximum setting, plus secret detection in reviews, both in public preview for Copilot Pro, Pro+, Max, Business and Enterprise. Story 2: VS Code v1.147 adds browser use in preview, lets us give the assisted permissions reviewer our own instructions, and deprecates macOS Ventura and Sonoma. Story 3: Claude Code v2.1.327 on the latest npm tag fixes project scope MCP commands deleting other keys from .mcp.json. Ask which Copilot plan, editor version and Claude Code channel my team uses, then give a short checklist for each story and link the official posts instead of inventing settings or flags.