At a glance
- Claude Code v2.1.273 notifies when MCP reconnect gives up, forks remote-control sessions to local background work, and adds opt-in gateway hint headers.
- Gemini CLI v0.60.0 hardens extension path and env consent, isolates sandbox settings and temp dirs, and tags untrusted tool-output provenance.
- Copilot can suggest allowed values when you define organization or enterprise repository custom properties (Business and Enterprise public preview).
- Copilot usage metrics now include dedicated VS Code Agents window fields for 1-day and 28-day enterprise and org reports.
Wednesday is an agent-ops and governance morning, so the useful work is how these four moves fit together. Anthropic ships a Claude Code reliability release aimed at MCP dropouts and remote-control handoffs, Google tightens Gemini CLI extension and sandbox boundaries, GitHub helps admins stand up ruleset metadata faster, and Copilot metrics finally separate the Agents window from editor Agent Mode.
Treat today as a reconnect, harden, and measure day. Pin Claude Code to v2.1.273 and smoke-test an MCP disconnect path plus one remote-control fork, bump Gemini CLI to v0.60.0 before installing new extensions, try Copilot’s property-value suggestions on one new custom property aimed at a ruleset, and pull the new VS Code Agents metrics fields so adoption dashboards stop mixing Agents-window traffic into generic Copilot totals.
Top Stories
Claude Code v2.1.273: MCP reconnect notify, remote-control forks, gateway hints
Practical dev impact: Pin `@anthropic-ai/claude-code` to `2.1.273` so mid-session MCP failures surface a notification when automatic reconnection gives up (with a pointer to `/mcp`), and so a session started with `claude –remote-control` or `/remote-control` from the Claude app can be forked into a background session that keeps running on your computer. The same release adds opt-in request headers for LLM gateways (`x-claude-code-request-class`, `x-claude-code-agent-type`, `x-claude-code-prev-tool-durations`, `x-claude-code-compaction`, `x-claude-code-context-compacted`) when you set `CLAUDE_CODE_GATEWAY_HINT_HEADERS=1`, which matters if your org routes Claude Code through a gateway that needs compaction and agent-class telemetry. Pair that with the managed-settings fixes for MCP allowlists and clearer Bedrock, Vertex, or Foundry credential errors so operators stop chasing generic `/login` dead ends. Start with one `/mcp` reconnect drill and one remote-control fork smoke test before rolling the pin fleet-wide.
Gemini CLI v0.60.0: extension consent, sandbox isolation, tool-output provenance
Practical dev impact: Upgrade `@google/gemini-cli` to `0.60.0` before you install or refresh extensions, because the September 15 stable channel now prompts for consent on environment changes, sanitizes runtime-altering environment variables, hardens extension path boundary checks, isolates settings and temporary directories inside sandbox containers (including macOS Seatbelt), and enforces envelope metadata provenance for untrusted tool outputs. Official announcements also call out destination validation in web-fetch utilities and RFC 9207 issuer identification in the MCP OAuth flow, so treat this as a security pin rather than a feature chase. Run `gemini –version`, re-approve any extension that mutates the environment, and confirm sandboxed sessions still see only the isolated settings path your policy expects.
Copilot suggests allowed values for repository custom properties
Practical dev impact: When you create a new custom property at the enterprise or organization level, Copilot can now propose allowed values (for example compliance-flavored multi-select options, or yes/no for an `internet-facing` single-select) so ruleset targeting does not start from a blank taxonomy. The feature is in public preview for Copilot Business and Copilot Enterprise, and owners can gate it with the Repository custom property suggestions Copilot policy. Accept a suggestion in one click, then wire the property into a ruleset that scopes which repos get which agent or branch protections. Use it on one real governance property this week rather than inventing a parallel spreadsheet of allowed values.
Copilot usage metrics add the VS Code Agents window
Practical dev impact: Enterprise and organization Copilot usage metrics for both 1-day and 28-day windows now expose optional Agents-window fields, so you can measure that surface without folding it into editor Agent Mode or generic Copilot totals. Aggregate reports add `daily_active_vscode_agent_users` plus `totals_by_vscode_agent` (`session_count`, `total_user_messages`). User reports add `used_vscode_agent` and the same per-user totals. Fields stay absent or null when data is unavailable, so existing consumers remain backward compatible. Access still needs the Copilot usage metrics policy plus owner, billing-manager, or View Copilot Metrics permissions. Point your adoption dashboard at the new fields this week and stop treating Agents-window sessions as anonymous Copilot chat.
Practical Impact Analysis
Today’s through-line is operational control of long-running agents, not another model launch. Claude Code v2.1.273 and Gemini CLI v0.60.0 both tighten what happens when tools and sandboxes misbehave: one makes MCP death and remote-control handoffs visible and forkable, while the other refuses silent env mutation and loosens sandbox path leakage. If your fleet still floats `latest` on either CLI, Wednesday is the day to pin and re-run the failure drills your runbooks already claim to cover.
GitHub’s custom-property suggestions close a quieter gap on the governance side. Rulesets only work when metadata exists, and Copilot proposing allowed values is how large orgs stop shipping empty taxonomies that never match a ruleset filter. Treat the public preview as admin tooling, then connect one property to a real ruleset before celebrating autocomplete.
The September 11 metrics drop is the measurement twin of those pins. Platform teams that already report Copilot adoption can finally chart the dedicated Agents window on its own axis, which matters once you start arguing about whether agent workflows are actually landing. Wire the optional fields into your existing metrics pull and leave editor Agent Mode in its old buckets.
If you only do three things this morning, pin Claude Code to v2.1.273 and confirm `/mcp` after a forced disconnect, upgrade Gemini CLI to v0.60.0 before any extension install, and add the VS Code Agents metrics fields to the dashboard that already tracks Copilot seats.
Tutorial
Pin Claude Code to v2.1.273, enable gateway hint headers only if a gateway will consume them, and smoke-test `/mcp`. Export whatever auth your org already uses. Keep secrets in the environment.
1. Pin `@anthropic-ai/claude-code@2.1.273` and confirm `claude –version`.
2. Set `CLAUDE_CODE_GATEWAY_HINT_HEADERS=1` only if your LLM gateway reads Claude Code hint headers.
3. After a forced MCP disconnect (or a stale server), reconnect via `/mcp`.
4. Next, start a remote-control session from the Claude app and fork it to a local background session once, then document the fork in your on-call notes.
Recommended AI prompt
Copy this paragraph into ChatGPT, Claude, Gemini, Grok, or whatever you use.
You are my staff engineer for Claude Code ops, Gemini CLI security pins, Copilot governance metadata, and Agents-window adoption metrics on 2026-09-16. Claude Code v2.1.273 adds MCP disconnect notifications when auto-reconnect gives up, forking of remote-control sessions into local background work, and opt-in gateway hint headers via `CLAUDE_CODE_GATEWAY_HINT_HEADERS=1`. Gemini CLI v0.60.0 hardens extension env consent, sandbox settings and temp isolation, and untrusted tool-output provenance. Copilot can suggest allowed values when creating org or enterprise repository custom properties (Business and Enterprise public preview) to speed ruleset taxonomies. Copilot usage metrics now expose optional VS Code Agents window fields for 1-day and 28-day enterprise and org reports, separate from editor Agent Mode. Ask which CLIs, gateways, Copilot plan, and metrics pipelines we run. Then produce a v2.1.273 pin plus MCP and remote-control fork checklist, a Gemini CLI v0.60.0 upgrade and extension-consent checklist, a one-property ruleset taxonomy trial using Copilot suggestions, and a metrics-schema patch list for the new Agents-window fields. Keep it concrete and copy-paste ready.