Claude Code Clears Another Gateway Snag, Cursor CLI Finally Reads the AGENTS.md Next to Your Edit, and Copilot Shows Which Tools Spend Its Tokens

At a glance

• Claude Code's newest build retries when a gateway rejects part of its request, but stable has not picked it up yet.
• The same build stops background cleanup from deleting recent files when you keep history for a very long time.
• Cursor CLI now applies the AGENTS.md rules in a subfolder as soon as the agent edits a file there.
• Copilot usage reports can now show which outside tools, such as OpenCode, Claude Code and Codex, send requests through Copilot.

It is a Saturday, so sources are thinner than a weekday, and Friday afternoon brought the kind of news you fix before Monday rather than the kind that reshapes a roadmap. Because all three stories touch how your agents reach their models and which rules they follow, today's work is short and mostly checking.

Treat today as a weekend tune up day: find out whether your Claude Code machines sit behind a gateway, update Cursor CLI if your repos keep rules in nested folders, and pull one Copilot usage report to see which tools your developers actually use.

Top Stories

1. Claude Code v2.1.310 retries past gateways that reject its requests

Claude Code v2.1.310 was published to npm at 5:20 PM ET Friday 23 October, and at our 5:53 AM ET check it served latest and next, while stable stayed on v2.1.308. The headline fix is for requests that failed behind some gateways and compatible servers that reject the effort-2025-11-24 beta value or system messages sent in the middle of a conversation, and Claude Code now retries without them. This is a different bug from the tool search regression we covered Wednesday.

The rest is housekeeping. Background cleanup could delete some recent files, such as pasted images, when cleanupPeriodDays was set very large. Compaction after a "Prompt is too long" error could keep too much dense text, such as base64 or SVG, so the next request failed again. File read permission prompts now show where a symbolic link leads, and on Linux, /sandbox and the startup warning now name Ubuntu's AppArmor restriction on user namespaces, and how to lift it, when it keeps bubblewrap from starting the sandbox.

Practical dev impact: If you route Claude Code through a company gateway or a compatible server and have seen unexplained request failures this week, try v2.1.310 before you file a ticket, because the fix is only on latest and next today. Anyone who set cleanupPeriodDays very high should update too, since that setting was the trigger for losing recent files.

2. Cursor CLI 2026.10.22 applies nested AGENTS.md rules when the agent edits

Cursor's CLI changelog lists a 2026.10.22 release, and its most useful change is for monorepos. When the agent edits or creates a file in a subdirectory, it now sees that directory's AGENTS.md and nested always apply rules, even if it never read a file there first. Each rule is shown once per conversation, and glob scoped and root rules behave as before.

Custom subagents also stop failing when their model cannot run. If a subagent is pinned to a model that your team's admin settings block, that requires Max Mode, or that is not available, it now runs on the parent agent's model, and its result notes which model it ran on.

Practical dev impact: If your repo keeps an AGENTS.md inside packages or services, the CLI agent now follows those rules on its first edit there, so expect slightly different and, in our reading, usually better behavior in nested folders. Teams that pin subagents to specific models should know that a blocked model now falls back quietly, so check subagent results for the model note.

Correction: Friday's brief said Cursor had nothing new at our check, and the Cursor CLI changelog now also lists versions 2026.10.15 and 2026.10.17.

3. Copilot usage metrics now show which tools spend its tokens

GitHub said at 7:42 PM ET Friday that the Copilot usage metrics API now reports model usage from more of the tools developers use with Copilot, including tools like OpenCode that were not captured before. A new totals_by_client array appears in the enterprise, organization, enterprise user and organization user reports, in both the 1 day and 28 day versions. Each entry names a client, such as opencode, claude_code, codex or zed, with request_count, session_count, token_usage and a per model breakdown that says whether the model was user selected, chosen by Auto, or chosen by the product.

GitHub spells out the limits. request_count counts model calls, not prompts, so one prompt can show up as several requests, and clients are identified by the integration ID they send, so the grouping is best effort, not a verified identity. Access needs an enterprise or organization owner, an enterprise billing manager, or a custom role with the View Copilot Metrics permission, plus the usage metrics policy enabled.

Practical dev impact: If your developers use Copilot as the model provider behind other tools, this is the first report that shows that traffic by tool, which helps you decide which clients to support. Do not compare request_count with prompt or interaction counts from other dashboards, because the units differ.

Practical Impact Analysis

All three stories are about seeing how your agents actually run. In our view the Claude Code fix is the most time sensitive, because a failing gateway request looks like an outage, and since stable lacks the fix, machines that follow stable behind a strict gateway may keep failing until it moves. The Cursor CLI change is quieter, but stale rules in a subfolder now matter more than they did last week. The Copilot metrics cost nothing to try, as long as whoever reads them knows that requests are not prompts.

Elsewhere at our check, Copilot CLI latest stayed on v1.0.103, while prerelease v1.0.104-0 makes Assisted Permissions available without experimental mode and v1.0.104-2 keeps Autopilot sessions waiting for attached background tasks. GitHub launched a consolidated GitHub app for Microsoft Teams where you mention @GitHub to turn an issue into a pull request, and the legacy app retires on November 16. Codex latest stayed on v0.164.0, Gemini CLI had only a new nightly, VS Code was still on v1.143, and Cursor's main changelog had nothing newer than October 16.

Tutorial

Read this script before you run it. It only prints and searches.

1. Run step 1 on each Claude Code machine, in the shell where you start it.
2. Run step 2 from the root of a repo you use with Cursor CLI.
3. If you administer Copilot, copy a report path from GitHub's Copilot usage metrics API docs into report_endpoint first.

# We have not run this script end to end on every platform. It only prints and searches unless you uncomment a line.
# 1. Claude Code. At 5:53 AM ET Sat 24 Oct npm showed latest and next on v2.1.310, stable on v2.1.308.
npm view @anthropic-ai/claude-code dist-tags --json
if command -v claude >/dev/null; then claude --version; else echo "claude not on PATH"; fi
# Says only whether a custom base URL is set in this shell, never its value.
if [ -n "${ANTHROPIC_BASE_URL:-}" ]; then echo "custom base URL set here"; else echo "no custom base URL in this shell"; fi
#    Optional, npm installs only. Pinned at 5:53 AM ET Sat 24 Oct, latest and next channel:
# npm install -g @anthropic-ai/claude-code@2.1.310

# 2. Cursor CLI. The changelog lists 2026.10.22 as the newest CLI release.
if command -v agent >/dev/null; then agent --version; else echo "agent not on PATH"; fi
#    Nested AGENTS.md files the agent now applies on its first edit there. Paths only.
find . -mindepth 2 -name AGENTS.md -not -path "*/node_modules/*" -not -path "*/.git/*"

# 3. Copilot usage metrics. Paste a 1 day or 28 day report path from GitHub's
#    Copilot usage metrics API docs. We do not guess the path, so it starts empty.
report_endpoint=""
if [ -z "$report_endpoint" ]; then
  echo "report_endpoint is empty, so step 3 is skipped"
elif command -v gh >/dev/null; then
  out="$(mktemp)"
  gh api "$report_endpoint" > "$out" && grep -o '"client": *"[^"]*"' "$out"
  rm -f "$out"
else
  echo "gh not on PATH"
fi

If step 1 says a custom base URL is set, our reading is that the machine likely talks to a gateway, so it is a good candidate for v2.1.310. If you installed Claude Code natively, update it the native way and do not mix npm and native installs on one machine. If step 2 lists files, check that their rules are still current, and if agent --version is older than 2026.10.22, the changelog says agent update upgrades in place.

Sources:

• Claude Code changelog
• Claude Code on npm
• Cursor CLI changelog
• Usage metrics API adds usage from tools like OpenCode
• Turn notifications into code with GitHub Copilot in Teams
• Copilot CLI releases
• Codex releases
• Gemini CLI releases
• Visual Studio Code release notes
• Cursor changelog

Recommended AI prompt

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