AI Coding Roundup — September 16, 2026: Claude Code 2.1.273 Ships a Mega-Update, Gemini CLI Hardens Extensions

Yesterday's roundup covered Claude Code 2.1.271 and a minimal 2.1.272 patch. Both were quickly overtaken: Claude Code 2.1.273 shipped later on September 15 with one of the month's largest changelogs — gateway telemetry headers, Remote Control session forking, and dozens of Code Review, MCP, and artifact-publishing fixes spanning VS Code, Claude Code on the web, and Claude Tag. The same day, Google shipped Gemini CLI v0.60.0, a security-focused release adding consent prompts for extension-driven environment changes and hardened sandboxing on macOS. GitHub Copilot added AI-suggested custom property definitions for enterprise governance, and Cline shipped a coordinated SDK/CLI/Desktop release built around hub-managed agent plugins. OpenAI also posted a deprecation notice touching Codex: GPT-5.5 retires from ChatGPT, ChatGPT Work, and Codex on October 14, 2026. OpenCode and Cursor show nothing new past what was already reported. Here's what actually changed as of September 16, 2026.

Executive Summary

• Claude Code 2.1.273 (Sep 15) adds gateway hint headers for LLM gateways, Remote Control session forking from the Claude app, an MCP disconnect notification, and fixes across Code Review duplicate-posting, artifact publish retries, Bash permission checks, and authentication error messages.
• Gemini CLI v0.60.0 (Sep 15) adds explicit consent prompts for extension-driven environment changes, sanitizes runtime-altering environment variables, hardens macOS sandboxing and directory isolation, and enforces RFC 9207 issuer checks in MCP OAuth.
• GitHub Copilot (Sep 15) now suggests allowed values when admins define custom properties for repository governance, in public preview for Business and Enterprise plans.
• Cline shipped SDK v0.0.83, CLI v3.0.62, and Desktop v0.0.28 (Sep 15) together, introducing hub-managed agent plugins, retry logic for transient provider failures, checkpoint performance gains, and a refreshed model catalog spanning 44+ providers.
• OpenAI (Sep 14) will retire GPT-5.5 from ChatGPT, ChatGPT Work, and Codex on October 14, 2026; scripts, agents, and scheduled tasks pinned to gpt-5.5 need to move to gpt-5.6-sol.

Claude Code 2.1.273: Gateway Telemetry, Session Forking, and a Long Fix List

The headline additions in 2.1.273 are aimed at teams running Claude Code through LLM gateways and Remote Control. New x-claude-code-request-class, x-claude-code-agent-type, x-claude-code-prev-tool-durations, and x-claude-code-context-compacted request headers (opt-in via CLAUDE_CODE_GATEWAY_HINT_HEADERS=1) give gateway operators visibility into what a request actually is, and gateways now warn at startup if access_control.allow_cidrs is empty. Sessions started with claude --remote-control or /remote-control from the Claude app can now be forked into a background session running on your own computer, and Claude now tells you when an MCP server disconnects mid-session and automatic reconnection has given up, pointing you at /mcp.

The fix list is long and mostly aimed at reliability rather than new capability. Code Review no longer occasionally double-posts the same finding, and merging the base branch into a PR no longer triggers a full re-review just because an earlier review listed “Additional findings.” Artifact publishing is now safe to retry: a publish whose connection drops after reaching claude.ai is re-sent instead of failing or creating a duplicate version. Authentication error messages on Bedrock, Vertex, Foundry, and the Claude apps gateway now name the specific credential to refresh instead of a generic 401/403. Full details, including the platform-specific VS Code, Claude Code on the web, and Claude Tag fixes, are on Anthropic's own changelog.

Gemini CLI v0.60.0: Consent Prompts and Sandbox Hardening

Google's v0.60.0 release, also dated September 15, is entirely about tightening the CLI's security posture. Extensions that want to alter the runtime environment now require explicit user consent instead of applying changes silently, and Gemini CLI sanitizes runtime-altering environment variables before executing tools. On macOS specifically, the release hardens sandboxing and directory isolation; more broadly it strengthens workspace boundary validation, symlink resolution, and mitigation for NTFS 8.3 short-name path tricks that could otherwise escape a restricted directory. On the network side, MCP OAuth connections now enforce strict RFC 9207 issuer identification, and web-fetch utilities get improved destination validation and connection routing — closing off a class of SSRF-adjacent risk similar to the one patched in v0.59.0 earlier this month. Release notes are on the Gemini CLI changelog.

GitHub Copilot: AI-Suggested Custom Properties for Governance

GitHub shipped a narrower but still notable change on September 15: Copilot now suggests allowed values when an admin defines a custom property for repository governance — compliance-style options for a multi-select property like FedRAMP, or a simple yes/no for a binary property like internet-facing. It's in public preview for Copilot Business and Copilot Enterprise, controlled by the “Repository custom property suggestions” Copilot policy setting, and aimed at the common problem of inconsistent governance metadata across large repository fleets making rulesets harder to apply consistently. Details are on GitHub's changelog and in the custom properties docs.

Cline's Coordinated Release: Hub-Managed Agent Plugins

Cline shipped SDK v0.0.83, CLI v3.0.62, and Desktop v0.0.28 together on September 15, all built on the same underlying SDK bundle. The headline feature is hub-managed agent plugins — plugin discovery and management moved into Cline's Hub system rather than being bundled per client. The release also adds retry logic for transient provider failures, strips invisible characters from stored credentials, improves checkpoint performance, and refreshes the model catalog across 44+ providers, alongside default-model changes for Cline Pass subscribers and desktop-specific fixes for IME text input and model-picker refresh. Full notes are on Cline's GitHub releases page.

OpenAI Retires GPT-5.5 Across ChatGPT and Codex

OpenAI posted a deprecation notice on September 14: GPT-5.5 retires from ChatGPT, ChatGPT Work, and Codex on all plans effective October 14, 2026 — consumer, Business, Enterprise, and Edu subscriptions are affected, though the API is not. Anyone with workspace defaults, saved model settings, managed configurations, custom agents, scheduled tasks, or scripts still pointing at gpt-5.5 needs to migrate to gpt-5.6-sol before the deadline. No Codex CLI-specific changelog entries shipped this week, but this notice directly affects any Codex workflow pinned to the retiring model. Full details are on OpenAI's changelog.

No Change: OpenCode and Cursor

Re-checking each project's own source shows no movement since OpenCode's September 14 v1.18.31 ACP session-state fix and Cursor's September 10 Projects launch. Neither has anything new as of September 16, 2026.

What to Prioritize

• Update Claude Code for the Code Review and artifact-publishing fixes. Teams running automated review or publishing pipelines were hit hardest by the duplicate-posting and failed-retry bugs 2.1.273 resolves.
• Gemini CLI users on macOS should upgrade to v0.60.0 now. The sandboxing and directory-isolation hardening closes real escape paths, not just theoretical ones.
• Cline users juggling many model providers should look at hub-managed plugins. Centralizing plugin management in the Hub is a workflow change worth understanding before your next update.
• Audit for gpt-5.5 references before October 14. Scheduled tasks and custom agents pinned to a specific model string are the easiest thing to miss in a retirement like this.
• Verify claims against each vendor's own changelog. Every version, date, and fix above was cross-checked against Anthropic's, Google's, GitHub's, Cline's, and OpenAI's own changelog and release pages linked throughout.