AI Coding Tools Weekly, October 12-18, 2026: Git Worktrees and Claude Code Fork Subagents

Executive Summary

• Running several coding agents at once only works if their file changes stay apart. Codex and OpenCode both added managed Git worktree tools, and VS Code 1.141 lets you watch parallel Copilot sessions in a grid and clean up what they leave behind.
• Claude Code reached 2.1.302, with fork subagents on by default in claude -p, the Agent SDK, the desktop app and VS Code. That one can change headless runs without a config edit.
• Codex 0.163.0 is the stable release, and you can now steer it while it responds.
• OpenCode 1.19.1 discontinued session sharing, and 1.19.2 keeps 1.x installs on the 1.x line.
• Gemini 4 Argon is rolling out in Copilot, and GitHub announced another round of model retirements for mid-November.

Git Worktrees: Three Tools, One Answer

OpenAI Codex 0.162

According to the Codex release notes, version 0.162.0 (October 8) adds managed Git worktree tools for trusted projects and lets you pin tasks in the agent Command Center, so one task's edits do not bleed into another's. It also adds:

• /copy and Ctrl+Insert for transcript copy and navigation.
• Clickable URLs across many TUI prompts.
• Live web access and remote compaction settings for custom model providers.
• Opt-in ranked tool search in Code Mode.

Fixes include CRLF preservation in apply_patch, honoring server Retry-After advice, Linux sandbox startup when several files are denied, and Windows drive-letter access. Version 0.162.1 (October 9) followed with a TUI crash fix and a startup fix for mismatched background-server settings.

OpenCode 2.0.27

The OpenCode changelog lists v2.0.27 on October 12. Code Mode can now create, list and remove Git worktrees, and removing one requires edit approval. The first opencode pair --remote now warns that it will expose your server and asks for confirmation, and --yes skips the prompt for scripts.

Other changes:

• Chrome and Edge can install the OpenCode web app, including behind authenticating proxies.
• Desktop and TUI timelines show execution errors that occur between model steps.
• Vertex AI gained routes for Interactions and Mistral models.
• Code Mode's fetch can no longer read file:// URLs.

The prior release, v2.0.26 (October 8), added Google Cloud credentials for Vertex in /connect, lighter and faster session snapshots, and clickable links throughout transcripts.

Copilot in VS Code 1.141

GitHub's Copilot weekly releases for October 5 describe two VS Code 1.141 changes that fit the theme. The Agents window can show multiple sessions side by side in a grid. A new "Open worktree cleanup" option in chat shows how much disk space inactive session worktrees use and lets you choose which to remove. The same notes add support for using one GitHub account for your Copilot license and another for repositories in the Copilot app.

Claude Code: Up to 2.1.302

Fork subagents become the default (2.1.302)

Released on October 16 per the Claude Code GitHub releases and the official changelog:

• Fork subagents are now enabled in claude -p, the Agent SDK, the desktop app and VS Code. Set CLAUDE_CODE_FORK_SUBAGENT=0 to opt out.
• CLAUDE_CODE_AUTO_MODE_MODEL selects the model used by the client-side auto mode classifier.
• Image sources over 128 MB are now refused by the Read tool instead of attempted.
• Plugin marketplaces no longer re-clone on every refresh, and several cloud session issues are fixed.

Reliability and safer defaults (2.1.300 and 2.1.301)

• Retry watchdog cap (2.1.301): unattended retry mode now ends a request when a usage limit resets too far away, controlled with CLAUDE_CODE_RETRY_WATCHDOG_MAX_USAGE_LIMIT_WAIT_MS, instead of waiting.
• Safer worktrees (2.1.301): quitting no longer removes a WorktreeCreate hook's worktree without asking when it may hold unpushed commits.
• Honest errors (2.1.301): Grep and Glob now return an error when a search fails instead of reporting no matches, and the Read tool no longer reads binary files such as .parquet or .heic as text.
• Smoother switching (2.1.301): moving to a model with a smaller context window now summarizes an over-long conversation with the previous model first.
• Observability (2.1.300 and 2.1.301): OpenTelemetry GenAI semantic-convention attributes on trace spans, cache_creation_5m_tokens and cache_creation_1h_tokens on api_request events, and classifier_verdict and permission_mode on tool_decision events.
• Behavior changes (2.1.300): large requests to api.anthropic.com are gzip-compressed by default (disable with CLAUDE_CODE_GZIP_REQUEST_BODIES=0), background commands are force-killed 1.5 seconds after exit is requested, and --sandbox and --no-sandbox were removed from claude remote-control.

Idle compaction and hooks (2.1.295 to 2.1.299)

• Idle compaction (2.1.299): conversations over 200K tokens that sit idle for about 54 minutes are compacted before the one-hour prompt cache expires. Tune with CLAUDE_CODE_IDLE_COMPACT_MIN_TOKENS or disable it with "idleCompaction": false.
• Cost (2.1.296): Sonnet 5.5 cache reads drop to $0.10/M tokens from $0.20, and CLAUDE_CODE_WORKFLOW_SUBAGENT_MODEL runs all workflow agents on one model.
• Hooks (2.1.295): per the CHANGELOG, onFailure: "block" makes a command or HTTP hook that fails to start, times out or exits unexpectedly block the action instead of silently passing. It is opt-in. The same release backs off repeated remote MCP reconnects up to 30 seconds.
• Fixes: WSL2 sandbox failures and scheduled wakeups not firing in Agent SDK and IDE sessions.

OpenAI Codex 0.163.0: Steer While It Thinks

The stable 0.163.0 release (October 15) is on the Codex releases page. The headline is real-time steering: send a follow-up while Codex is responding and it adjusts course, and interrupting a turn also stops active Code Mode cells.

• /config now searches settings across categories and applies supported changes immediately. /keymap adds a configurable leader key (default Ctrl+X), and voice mute moved to F9.
• /daemon can restart a managed background server and warns about work that could be interrupted. Independent read-only tools can now run in parallel.
• The network proxy checks permission before DNS lookups and rejects duplicate Host headers, and resuming a saved thread restores its permissions and network approvals.
• The patched zsh shell backend was removed in favor of the standard shell executor, and packaged binaries are smaller on Linux and macOS.

Pre-release 0.164.0-alpha.1 through alpha.3 were tagged on October 16 and 17 without published notes, so there is nothing to adopt yet.

GitHub Copilot

Gemini 4 Argon

Gemini 4 Argon is rolling out gradually to Pro+, Max, Business and Enterprise, available in the model pickers of VS Code, Visual Studio, Copilot CLI, the coding agent, JetBrains, Xcode and Eclipse. It is billed at provider list pricing under usage-based billing, and admins can control it through the model policy.

Mid-November retirements

The Copilot changelog has an October 15 entry titled "Upcoming deprecation of selected GitHub Copilot models in mid-November". A third-party summary at SessionWatcher says the November 16 retirement covers Claude Opus 4.8, Claude Opus 5 and GPT-5.6 Terra across Copilot experiences. We could not confirm those details on GitHub's own post, so read the primary entry before changing any model policy.

Copilot CLI 1.0.95 to 1.0.98

The GitHub changelog and the Havoptic Copilot CLI tracker show four releases. The tracker is third-party, so check GitHub for the authoritative notes.

• 1.0.98 (October 14): fixes keystrokes lost at startup, adds one-key expansion of AI answers, and adds Claude Haiku 5.5 and Gemini 4 Argon as models.
• 1.0.97 (October 13): more reliable custom-agent discovery in complex repo layouts, agent-scoped declarative hooks, and /usage now shows billed AI Credits.
• 1.0.96 (October 12): the sandbox detects potential secrets and lets you mask them before saving.
• 1.0.95 (October 9): improved macOS authentication and more control over the context your sessions use.

Other changelog items

• Copilot code review gained new organization billing controls.
• The changelog also lists "Agentic autofix for code scanning alerts general availability" (October 15) and "HydraFusion updates and new Max option" (October 14).
• Local sandboxing is generally available, Claude Haiku 5.5 joined Copilot, and Copilot CLI's /model command can discover local models from a running Ollama instance.

OpenCode: The 1.x Line

OpenCode now has two lines, and the 1.x line is still published on its GitHub releases page, so check which one you are on before upgrading. Per the OpenCode releases:

• v1.19.0 (October 11): introduces opencode upgrade --major. Updates are now manual via /update in the TUI, and a security fix validates PTY connection origin and ticket first.
• v1.19.1 (October 15): discontinues session sharing. Existing share links and importing from share URLs no longer work.
• v1.19.2 (October 16): keeps 1.x installs on the 1.x line so update checks stop offering 2.x builds. Moving to OpenCode 2 is an explicit choice via /update or opencode upgrade --major. It also fixes yarn-managed upgrades, skips 1.x updates for Homebrew Core installs, and makes the install script refuse 2.x versions.
• v1.19.3 (October 16): adds model reasoning options that generate reasoning variants. These replace catalog variants, while your own configured variant overrides are kept.

What to Do

• Parallel agent users: give each agent task its own worktree instead of one shared checkout, and prune old ones.
• Scripted Claude Code users: update to 2.1.302 and run your headless jobs once. Set CLAUDE_CODE_FORK_SUBAGENT=0 if behavior differs.
• Running Claude Code unattended? Set a maximum usage-limit wait so jobs fail fast instead of idling, and add onFailure: "block" to security-relevant hooks.
• If you keep long Claude Code sessions open, review idle compaction.
• Codex users: try steering mid-response in 0.163.0, and review the new /keymap leader key if you use Ctrl+X bindings.
• OpenCode users: confirm whether you are on 1.x or 2.x before upgrading, export anything you rely on from shared sessions, and review variant overrides in your config.
• Copilot admins: review the deprecation post and check which retiring models your teams pin.

Sources

• Claude Code releases
• Claude Code changelog
• Claude Code CHANGELOG.md
• OpenAI Codex releases
• OpenCode changelog
• OpenCode releases (sst)
• OpenCode releases (anomalyco)
• GitHub Copilot weekly releases, October 5
• GitHub changelog: Gemini 4 Argon in Copilot
• GitHub changelog: Copilot code review billing controls
• GitHub changelog: Local sandboxing GA
• GitHub changelog: Claude Haiku 5.5 in Copilot
• GitHub changelog: Discover local models in Copilot CLI
• GitHub Copilot changelog
• GitHub Changelog
• Havoptic Copilot CLI tracker (third-party)
• SessionWatcher: Copilot model deprecations (third-party)