AI Coding Roundup — September 15, 2026: Claude Code Sandboxes Bash by Domain, Copilot Adds Cost Tiers

Yesterday's roundup reported Claude Code holding at 2.1.270 with nothing new. That turned out to be a timing gap, not a quiet day: Claude Code 2.1.271 shipped later on September 14 with one of the year's denser changelogs — domain-scoped Bash sandboxing, self-hosted runner fixes, and dozens of reliability improvements — followed by a small 2.1.272 patch on September 15. Alongside it, GitHub Copilot added cost/quality tiers to auto model selection, and OpenCode shipped v1.18.31 fixing ACP session state. Cursor, Gemini CLI, Cline, and OpenAI Codex CLI show nothing new past what was already reported. Here's what actually changed as of September 15, 2026.

Executive Summary

• Claude Code 2.1.271 (Sep 14) adds per-command allowed_domains sandboxing for Bash, PowerShell, and Monitor in auto mode, plus fast mode in Remote sessions, a self-hosted runner drain marker, a pricing multiplier for internal chargeback, and fixes to MCP OAuth client registration and duplicate Code Review findings.
• Claude Code 2.1.272 (Sep 15) is a small follow-up: bug fixes and reliability improvements, no new features listed.
• GitHub Copilot (Sep 14) added efficiency, balance, and intelligence tiers to auto model selection in VS Code, Copilot CLI, and the Copilot app, so cost/quality tradeoffs are now configurable instead of fixed.
• OpenCode v1.18.31 (Sep 14) restores ACP session model, effort, mode, and reasoning boundaries on resume/fork, surfaces remote-auth failures in the TUI at startup, and requests summarized adaptive thinking from GitHub Copilot models.

Claude Code 2.1.271: Domain-Scoped Bash Sandboxing and Runner Fixes

The headline change in 2.1.271 is per-command allowed_domains for Bash, PowerShell, and Monitor when running in auto mode with sandboxing: the hosts a specific command needs are reviewed and opened for that command alone, and every other host is refused. That's a meaningfully tighter default than an all-or-nothing network allowlist — it limits what a single approved command can reach even if the rest of the session's tools have broader access. The same release also fixes several Bash permission-check gaps: files read by fmt, column-style commands after an unrecognized option, wildcard-expanded files inside a pattern or option value, and shell variable-declaration flags that could misrepresent the command actually being run. Full details are on Anthropic's own changelog.

Beyond sandboxing, the release covers a lot of ground: fast mode now works in Remote sessions (cloud and self-hosted runners); a new claude self-hosted-runner --drain-marker-file flag reports a graceful SIGTERM drain to the server as telemetry; omitClaudeMd lets custom and plugin subagents run without user, project, and local CLAUDE.md files while managed policy files still load; and a multiplier above 1 (up to 10) in the modelPricing managed setting and the Claude apps gateway pricing block enables marked-up internal chargeback rates for organizations reselling usage internally.

On the fix side, three stand out for teams running MCP or Code Review at scale: MCP OAuth no longer mishandles client registrations (denying consent forced a spurious new one, a registration for another redirect URI was wrongly reused, and a concurrent write could delete a valid registration or keep a mismatched one); Code Review no longer occasionally posts the same finding two or three times when GitHub reports an error for a review it actually created; and a follow-up review no longer re-posts a security finding a person already resolved after a later push moves the anchoring lines.

GitHub Copilot: Cost and Quality Tiers for Auto Model Selection

GitHub shipped three configurable tiers for Copilot's auto model selection on September 14: efficiency (prioritizes low cost for fast, straightforward tasks), balance (weighs cost, quality, and latency for routine work), and intelligence (prioritizes quality for complex tasks). All three draw from the same model pool, and auto still evaluates each prompt individually — a trivial request in intelligence mode can still land on a small, fast model. The feature is rolling out in VS Code, Copilot CLI, and the GitHub Copilot app; billing follows whichever model auto actually picks, and the standard 10% discount for paid subscribers on auto-billed usage still applies regardless of tier.

This is a small but useful control: teams that found auto mode too conservative on complex refactors, or too expensive on boilerplate, finally get a lever instead of switching to a pinned model.

OpenCode v1.18.31: ACP Session State Restored

OpenCode reached v1.18.31 on September 14, restoring ACP session model, effort, mode, and reasoning chunk boundaries when a session is loaded, resumed, or forked — a fix that matters if you routinely branch or resume long-running agent sessions and had them silently lose their configured reasoning settings. The release also surfaces remote-configuration authentication errors in the TUI at startup instead of failing silently, and has GitHub Copilot model support request "summarized adaptive thinking." Full entries are on OpenCode's own changelog.

No Change: Cursor, Gemini CLI, Cline, and OpenAI Codex

Re-checking each project's own source confirms no movement since Cursor's September 10 Projects launch, Gemini CLI's September 8 v0.59.0 SSRF fix, Cline's September 14 v4.1.17 default-model swap, and OpenAI's Codex CLI changelog, which lists nothing past its early-September entries. Nothing here warrants its own section today, but it's worth checking back tomorrow — Codex in particular is overdue for a release given its recent cadence.

What to Prioritize

• Update Claude Code for the Bash sandboxing alone. Per-command allowed_domains in auto mode is a real reduction in blast radius for any workflow that runs unattended commands — worth adopting even if nothing else in this release applies to you.
• If you run MCP servers with OAuth, the client-registration fix matters. The bug could delete a valid registration or keep a mismatched one under concurrent writes — check your MCP server connections after upgrading.
• Try the efficiency tier in Copilot on boilerplate-heavy repos. If auto mode has felt expensive for routine completions, the new tiers are the first lever GitHub has given you without pinning a model manually.
• OpenCode users forking or resuming sessions should upgrade to v1.18.31. Losing reasoning-effort settings silently on resume is the kind of bug that's easy to misattribute to model behavior rather than session state.
• Verify claims against each vendor's own changelog. Every version, date, and fix above was cross-checked against Anthropic's, GitHub's, and OpenCode's own changelog pages linked throughout.