Claude Code 2.1.310, Copilot Autofix GA and Codex 0.164

Executive Summary

Claude Code moved from 2.1.303 to 2.1.310. Version 2.1.309 is a security-focused release that tightens the sandbox and permission model, and 2.1.304 adds guided auto mode setup.
GitHub made agentic autofix for code scanning alerts generally available, and the Copilot usage metrics API now counts tools like OpenCode, Claude Code and Codex.
GitHub retired Gemini 3.7 Flash, GPT-5.5, GPT-5.4, GPT-5.4 mini, GPT-5 mini and Grok 4.5 from Copilot on October 19, with more retirements scheduled for mid-November.
Codex 0.164.0 went stable on October 23 with saved prompt presets and enterprise MCP authentication.
OpenCode 1.19.4 upgrades its bundled Bun runtime with security fixes.

Claude Code: 2.1.303 to 2.1.310

Security and sandbox (2.1.305 and 2.1.309)

Version 2.1.309 (October 22) is mostly about tightening the sandbox and permission model, per the GitHub releases page:

Sandboxed Bash on macOS and Linux no longer runs shell startup files outside the sandbox.
Sandboxed commands and auto-accepted edits can no longer modify the --mcp-config file.
sed -i now goes through the standard Bash permission prompt and is no longer restored by /rewind.
claude ultrareview now requires workspace trust.
In bypass-permissions mode, the Desktop and VS Code apps deny dangerous rm and rmdir commands immediately and suggest a rewrite.
Windows: PowerShell hooks that exit with code 2 now block, and PreToolUse and PermissionRequest hooks no longer fail open when Git for Windows is missing.

If you rely on hooks as a policy layer, the Windows fix is the important one: a missing dependency used to let tool calls through silently.

According to the Claude Code changelog, version 2.1.305 adds a sandbox.allowNestedUserNamespaces setting for sandboxed Linux commands that need user namespaces. The notes warn it can bypass the sandbox's file protections, so enable it only where a tool genuinely requires it.

Auto mode (2.1.302 and 2.1.304)

Version 2.1.304 adds /auto-mode-setup, a guided setup that reads your project and recent sessions to teach auto mode about your environment. It now works on Bedrock, Vertex, Foundry and telemetry-disabled sessions. Pro and Max users also get a warning when resuming an inactive conversation would use a noticeable share of the 5-hour usage limit.

Version 2.1.302 added CLAUDE_CODE_AUTO_MODE_MODEL to pick the model the client-side auto mode classifier uses, an "Idle compaction" toggle in /config, and a limit of 12 prompt and agent hooks of one event at a time. Fork subagents are now on by default in claude -p, the Agent SDK, the desktop app and VS Code; set CLAUDE_CODE_FORK_SUBAGENT=0 to turn them off. The 2.1.302 notes also warn that standalone Claude Design closes on December 14, 2026.

Safer deletes and worktrees (2.1.304)

rm -r on a mounted drive such as /Volumes, /media or WSL's /mnt/c now gets the same safety prompt as deleting /Applications.
claude --worktree no longer removes a worktree and branch without asking when HEAD is detached, which could lose commits.
Subagent responsiveness with many agents running at once also improved.

Telemetry, API version and MCP (2.1.307)

Version 2.1.307 (October 21) has two additions. The /claude-api api-version command moves Claude API code to API version 2026-10-21, and a new claude_code.safety_event OpenTelemetry event is emitted whenever a permission rule, auto mode, safety check, hook or sandbox stops, prompts or sends back a tool call. Don't build on it, though: the 2.1.312 notes from October 27 say the event is no longer sent.

Fixes in the same release:

HTTP MCP tool calls were sent twice when the connection dropped mid-call. Only read-only or idempotent tools are retried now, which matters for tools that create or modify data.
MCP servers no longer report "needs authentication" in claude -p on macOS when the keychain is locked.
Model-not-found errors with CLAUDE_CODE_DISABLE_1M_CONTEXT and [1m] aliases such as opus[1m] are fixed, as are "Autocompact is thrashing" errors in turns that read many PDF pages or images.
Conversations cleared with /clear no longer reappear after a cloud or self-hosted runner restart.

Two behavior changes to check: the axScreenReader setting and CLAUDE_AX_SCREEN_READER now apply only from user, managed or --settings sources, not from a settings env block. Self-hosted runner checkout hooks now stop after 30 minutes by default. Raise the limit with --checkout-hook-timeout-min or set it to 0 to disable the timeout.

Cloud providers and enterprise roles (2.1.310)

Version 2.1.310 (October 23) is aimed at teams running Claude Code behind gateways or cloud providers:

The Monitor tool and plugin monitors now work on Amazon Bedrock, Google Vertex and Foundry.
Requests behind some gateways no longer fail on the effort-2025-11-24 beta value or on mid-conversation system messages. Claude Code retries without them.
Compaction after a "Prompt is too long" error keeps less dense text such as base64 or SVG, so the next request is less likely to fail again.
Cloud sessions and Remote Control permissions can now be set through Enterprise custom roles.

Smaller releases

2.1.305 also adds a variable for unattended retry mode to notice when usage limits lift, makes requests retry without tool search when a provider rejects it, and fixes a batch of issues in worktrees, /usage, screen reader mode and Remote Control.
2.1.306 is a fix release: claude -p in cloud sessions no longer reports "Not logged in" when started without the session's environment variables, an API error is no longer replaced by a misleading "prompt is too long" message, and sessions with many MCP tools avoid an unneeded compaction.
2.1.303 and 2.1.308 list only bug fixes and reliability improvements.

GitHub Copilot

Agentic autofix is GA

Agentic autofix for code scanning alerts is generally available for GitHub Code Security and Advanced Security customers on GitHub Enterprise Cloud. Copilot explores relevant files, proposes a fix, reruns CodeQL to confirm the alert closes, then opens a draft pull request. Alerts on the default branch can be assigned one at a time, in batches of up to 25, or through the REST API. It needs an enterprise policy and a Copilot billing method, and consumes AI Credits and Actions minutes.

Model retirements

GitHub's October 19 deprecation notice, first announced on September 18, lists the retired models and suggested replacements:

Gemini 3.7 Flash: use Gemini 3.8 Flash.
GPT-5.5 and GPT-5.4: use GPT-6.1 Sol.
GPT-5.4 mini and GPT-5 mini: use GPT-6 Luna.
Grok 4.5: use Grok 4.7.

The change covers most Copilot experiences, including chat, inline edits, ask and agent modes, and completions. Enterprise admins may need to enable the replacement models in their model policies before they appear in the model picker. The new model policy REST API covers enterprise, enterprise teams, enterprise apps and organization levels, and can help automate that. More retirements are scheduled, see the mid-November deprecation notice.

Gemini 4 Argon continues its gradual rollout for Pro+, Max, Business and Enterprise, billed at provider list pricing.

Usage metrics now include other tools

The Copilot usage metrics API now counts usage from tools like OpenCode. A new totals_by_client array appears in the enterprise, organization and user reports, with request count, session count, token usage and a per-model breakdown for each client. Named clients include OpenCode, Claude Code, Codex, Zed, Copilot CLI, VS Code and JetBrains. Because clients are inferred from an integration ID, treat the grouping as best-effort. The change is additive, and it needs the Copilot usage metrics policy to be enabled.

Copilot in Teams and GitHub CLI tokens

Turn notifications into code with GitHub Copilot in Teams merges the GitHub notifications app and Copilot into one Teams app. You can mention GitHub on an issue notification and have Copilot investigate, fix and open a pull request. The older app becomes "GitHub (deprecated)" and is retired on November 16, so existing users need to install the consolidated app before then.

Separately, opt-in expiring, auto-refreshing OAuth tokens for GitHub CLI is a security improvement for scripts and agents that authenticate through gh, though it is not a Copilot feature.

OpenAI Codex: 0.164.0 Goes Stable

After days of alpha builds, the Codex 0.164.0 release (October 23) is now marked Latest. It adds:

Saved prompt presets in ten /presets slots, plus inline arguments for /model and /permissions.
Image previews in the fullscreen transcript.
Enterprise MCP authentication and credential masking for remote commands.
The Amazon Bedrock setup wizard on by default.
Fixes for failed compaction, session resume, Windows proxy authentication and sandbox checks.

A 0.165.0-alpha.1 build also appeared the same day, with no release notes yet. The previous stable release, 0.163.0 (October 15), also improved Windows and WSL paste handling, and codex exec no longer repeats unchanged file diffs.

OpenCode: 1.19.3 and 1.19.4

The OpenCode releases page shows v1.19.4 (October 22), which upgrades the bundled Bun runtime to 1.4.4 with security fixes and updated root certificates. Earlier in the 1.19 line, v1.19.1 discontinued session sharing, including sharing commands, v1.19.2 made the move to 2.x an explicit choice through /update or opencode upgrade --major, and v1.19.3 lets you set reasoning options in model config to generate reasoning variants.

What to Do

Update Claude Code to 2.1.310, especially if you use sandboxed Bash, Windows hooks or a Bedrock or Vertex gateway.
Run /auto-mode-setup before relying on auto mode, and leave nested user namespaces off unless a sandboxed command needs them.
Check self-hosted runner jobs that rely on checkout hooks longer than 30 minutes.
Copilot users and admins: check that nothing pins a retired model, that replacements are enabled in policy, and which models the mid-November notice covers.
Pilot agentic autofix on a low-risk repository first.
Copilot admins: check the new totals_by_client data to see usage by client.
Teams users: install the consolidated GitHub app before November 16.
Move from Codex alphas to stable 0.164.0 and try the new presets.
OpenCode users: decide deliberately whether to move to 2.x.

AI Coding Tools Weekly, October 19-25, 2026: Copilot Agentic Autofix GA and Claude Code Sandbox Hardening

Executive Summary

Claude Code: 2.1.303 to 2.1.310