Claude Code's Stable Build Carries a Gateway Bug, GitHub Puts a Firewall in Front of npm, and VS Code Lets Models Race

At a glance

• Claude Code's stable build can fail every request behind some gateways, and only the faster channels have the fix.
• An MCP allowlist change in Claude Code may load a server you thought was blocked.
• GitHub can now screen npm installs by coding agents and CI before they land.
• VS Code can run one task across several models and let a judge pick the winner.

Tuesday's stable jump looked like good news, and today it comes with a catch, because the build it landed on has a regression the newer channels already fixed. Treat today as a check what you trusted day: see whether your Claude Code machines sit behind a gateway or cloud provider, reread your MCP allowlist before the next upgrade, and decide whether your agents' npm installs should pass through a firewall.

Top Stories

1. Claude Code v2.1.305 fixes a gateway regression that stable still ships

Claude Code v2.1.305 was published to npm at 3:29 PM ET Tuesday 20 October and v2.1.306 at 10:28 PM ET, and at our 5:49 AM ET check Wednesday latest and next served v2.1.306 while stable stayed on v2.1.304. That matters because the v2.1.305 changelog lists a v2.1.304 regression in which every request failed on Vertex and other providers when the endpoint or a gateway rejected tool search, and behind gateways with ToolSearch disallowed or left out of --tools. Claude Code now retries without it.

The same release changes an allowlist. allowedMcpServers no longer filters managed-mcp.json stdio servers that use variable expansion only in their env block, so one your allowlist left out now loads on upgrade, and Anthropic says to use deniedMcpServers to keep it off. A new sandbox.allowNestedUserNamespaces setting lets sandboxed Linux commands create user namespaces for tools such as Chromium and rootless podman, and the changelog warns that this lets them get around the sandbox's file protections. Other fixes cover bare repository worktrees starting from local HEAD instead of the default branch, and Workflow resume rerunning completed agents.

v2.1.306 then fixes claude -p reporting "Not logged in" inside cloud sessions started without the session's environment variables, and a false "prompt is too long" error in sessions with many MCP tools.

Practical dev impact: If you run Claude Code through Vertex or a corporate gateway on machines that follow stable, they are on the broken build, so in our view pin them to v2.1.306 until stable moves. Before anyone upgrades, admins with a managed MCP allowlist should check for stdio servers that rely on variable expansion in env and add them to deniedMcpServers if they should stay off.

2. GitHub puts a package firewall between coding agents and npm

In a GitHub Changelog post at 12:11 PM ET Tuesday, GitHub put a package firewall into public preview for all GitHub Enterprise Cloud customers, including those with data residency. GitHub says a growing number of malware attacks harvest credentials at install time without the package's code even running. The firewall sits between endpoints such as Actions runners or coding agents and the upstream registry, and it inspects both metadata requests and package downloads.

Admins set policy first, with a configurable cooldown period, checks against the GitHub Advisory Database for malware or vulnerability alerts, and supply chain checks such as attestation downgrades. Each endpoint then has to be configured to send its requests through the firewall instead of straight to the registry, and if a package is blocked by mistake, a user can request an exception. Only npm is supported, the preview is free, and GitHub says general availability will bring metered billing similar to Actions.

Practical dev impact: In our reading, a cooldown period is the most useful control for agent heavy teams, because an agent that installs whatever was published an hour ago is exactly what install time malware targets. Nothing changes until you point an endpoint at it, so pilot one runner or agent environment while the preview is free.

3. VS Code v1.143 lets models race on one task

VS Code v1.143 was on the stable update feed by our check, with a build timestamp of 5:38 PM ET Tuesday 20 October, while its release notes page reads last updated October 19. The headline is Compare model results, a preview in the Agents window behind sessions.chat.compareAgents.enabled. You pick two or more models, mixing Copilot, Claude and Codex harnesses and reasoning levels, send one prompt, and each attempt runs in its own session. An optional Judge model scores the attempts with time and token usage, and Synthesize builds on the recommended one only when you ask.

The release also adds experimental on-device models: MAI Code 1.1 Flash runs locally in Copilot harness sessions through a llama.cpp llama-server, using no AI credits, but it needs an NVIDIA GPU with 96 GB or more of GPU memory. For admins, new Copilot harness chats now follow organization managed model defaults even when a user sets chat.defaultModel.

Practical dev impact: Comparing models runs several full attempts, so in our view try it on one bounded task and read the Judge's token column before going further. If a developer says their default model setting stopped working, check your managed defaults.

Practical Impact Analysis

In our view the Claude Code regression is the most urgent, since stable exists for teams that want safety and it is currently the channel with the gateway bug, while latest and next have the fix. The MCP allowlist change is easy to miss, because it widens what loads on upgrade. The firewall and model comparison both wait for you to opt in, so in our reading each deserves one measured trial.

Elsewhere at our 5:49 AM ET check, Copilot CLI v1.0.101, published 2:49 PM ET Tuesday, had moved to latest, bringing Tuesday's prerelease items to the default channel. On prerelease, v1.0.102-0 at 10:47 PM ET Tuesday makes sandboxed network traffic honor NO_PROXY and ALL_PROXY, and v1.0.102-1 at 12:46 AM ET Wednesday makes -p runs honor managed Auto model selection. Gemini CLI v0.65.0 reached latest at 5:26 PM ET Tuesday, enforcing read only workspace settings in untrusted folders and supporting custom OTLP telemetry headers, while preview moved to v0.66.0-preview.0, which adds a banner saying Gemini CLI is not optimized for Gemini 4. Codex latest held at v0.163.0, with alpha at v0.164.0-alpha.8 from 6:37 PM ET Tuesday and no notes beyond the version name. Neither Cursor changelog had anything new. GitHub also said the business field in some secret scanning audit log events switches from the enterprise display name to the enterprise slug on November 12.

Tutorial

1. Run step 1 on each machine and compare versions with this brief.
2. Run step 2 to see whether the machine points Claude Code at Vertex or a custom gateway.
3. Run step 3 from the folder where you keep managed Claude Code settings.
4. Uncomment step 4 only on an npm install that is stuck on v2.1.304 behind a gateway.

# We have not run this script end to end on every platform. It only prints and searches unless you uncomment an install.
# 1. Claude Code. At 5:49 AM ET Wed 21 Oct npm showed stable v2.1.304, latest and next v2.1.306.
npm view @anthropic-ai/claude-code dist-tags --json
if command -v claude; then claude --version; else echo "claude not on PATH"; fi

# 2. Provider hints. Prints whether each variable is set, never its value.
for name in CLAUDE_CODE_USE_VERTEX ANTHROPIC_BASE_URL; do
  if [ -n "${!name:-}" ]; then echo "$name is set"; else echo "$name is not set"; fi
done
#    The Vertex setup wizard can save these in the env block of your user settings instead.
if [ -f "$HOME/.claude/settings.json" ]; then
  echo "matching names in user settings:"
  grep -c -e CLAUDE_CODE_USE_VERTEX -e ANTHROPIC_BASE_URL "$HOME/.claude/settings.json"
fi

# 3. MCP allowlist check, current folder only.
if grep -rln -e allowedMcpServers -e deniedMcpServers --include="*.json" .
then echo "review the lists in the files above"; else echo "no MCP allow or deny lists here"; fi
if grep -rln --include="managed-mcp.json" '\${' .
then echo "review these variable references"; else echo "no variable references in managed-mcp.json here"; fi

# 4. Optional, npm installs only. Pinned at 5:49 AM ET Wed 21 Oct.
#    Claude Code latest and next channel, while stable was still v2.1.304:
# npm install -g @anthropic-ai/claude-code@2.1.306

On Vertex, /status inside Claude Code also shows an API provider line. Step 3 searches only the current folder. If a server with variable references shows up and your allowlist left it out, add it to deniedMcpServers before upgrading. Do not mix npm and native installs of Claude Code on one machine.

Sources:

• Claude Code changelog
• Claude Code on Google Vertex AI docs
• GitHub package firewall in public preview
• VS Code v1.143 release notes
• Copilot CLI v1.0.101 release notes
• Copilot CLI v1.0.102-0 release notes
• Copilot CLI v1.0.102-1 release notes
• Gemini CLI v0.65.0
• Gemini CLI v0.66.0-preview.0
• Updates to GitHub secret scanning
• Codex releases
• Cursor changelog
• Cursor CLI changelog

Recommended AI prompt

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