AI Coding Tools Weekly, October 1-4, 2026: Claude Code Mods and Copilot Computer Use
Part of AI Coding Tools News, a recurring series.

Executive Summary
- Claude Code moved from 2.1.286 to 2.1.289. Version 2.1.287 introduced Mods, and 2.1.289 closed several cases where deny and ask rules were not enforced.
- GitHub Copilot gained computer use and dynamic workflows in public preview, and HydraFusion, a research preview, coordinates several models inside one turn.
- GitHub deprecated four Copilot models on October 2 and gave Copilot code review a REST and GraphQL API.
- Codex CLI 0.160.0 is the latest stable build, with an agent command center history and opt-in Guardian review. The 0.162.0 builds are alphas with no notes.
- OpenCode stayed on v1.18.34, a maintenance release.
Claude Code: Mods and Permission Fixes
Mods (2.1.287 and 2.1.288)
According to the Claude Code changelog, the headline of 2.1.287 is Claude Mods: plugins can now modify deeper behavior. The first built-in example is "You should know", a side agent that flags things you or Claude might miss. Enable it with /plugin enable cc-plugin-you-should-know@builtin. It is available for first-party sessions with telemetry on.
Version 2.1.288 (see the GitHub releases) adds $.ui.selection() so mods can capture transcript selections in fullscreen mode, built-in gh api support in cloud sessions without the GitHub CLI installed, and a fix for control characters being transmitted incorrectly. Version 2.1.289 adds agent.spawn for teammates plus idle and waiting states in the mods agent list, and many of its other fixes target plugin panes and mods.
Permission rule fixes (2.1.289)
Version 2.1.289 was published on GitHub on October 3 as an immutable release. According to the changelog, the most important fixes concern deny and ask rules:
- Bash deny and ask rules could miss a command placed behind an environment variable prefix with an expanded value, such as
TZ="$HOME" rm -rf build, when the sandbox auto-allows commands. - A Bash deny or ask rule could also be skipped under sandbox auto-allow when a bare variable assignment came before the command.
- Read deny rules were not applied to files that were @-mentioned, changed or selected in the IDE through a symlink.
- A deny or ask rule on a nested part of a compound shell command did not hold over a user-installed mod's approval on managed machines.
- A user-installed plugin could rewrite the descriptions of an organization-managed MCP server's sign-in tools.
Earlier in the week, 2.1.287 restored safeguards for dangerous rm commands that used output redirection to ~ or wildcards.
Other changes
- 2.1.286 (September 30): a permission prompt counter such as "2 of 5" when several requests are stacked, mouse support for list navigation in fullscreen mode, and fixes for
gcpAuthRefreshandawsAuthRefreshcredential expiration. It also fixesclaude --resumeand--continuelosing turns after parallel tool calls, API errors when tools return non-text objects, cloud sessions with large histories failing to wake, and Remote Control sessions not disconnecting when org policy disables the feature. It follows 2.1.285 (plugin configuration and a WebFetch kill switch), covered in the September 30 roundup. - 2.1.287 (October 1): an
n:<text>filter in the agents view matches session names and tasks, and MCP servers can send URL prompts (for example for sign-in) on the 2025-11-25 protocol. Remote Control no longer goes minutes without receiving messages, and its reconnect timeout is now 30 seconds with retry. There are also many screen reader mode fixes and VS Code extension additions such as a "Run in background" option. - 2.1.289: fixes terminal freezes on short code blocks with many unclosed script tags, and reverts a 2.1.288 change to
claude auth statusin VS Code that may have made sign-outs more frequent.
GitHub Copilot: Computer Use, Workflows and Multi-Model Turns
Computer use
Per the GitHub changelog, computer use is in public preview for the Copilot CLI and desktop app on macOS and Windows. Copilot can read app content, click controls, enter text, press keys, scroll and drag, including in legacy software that has no API or command line. Copilot asks for approval before controlling apps, macOS needs Accessibility and Screen Recording permissions, and organizations can disable the feature.
- CLI:
/computer onto enable,/computer showfor status,/computer offto disable. - Copilot app: Settings > Computer Use, then toggle it on.
Dynamic workflows
The dynamic workflows announcement describes a way to define how a task is carried out by combining automated steps with one or more agents. Workflows can run commands and call external services, run independent tasks in parallel, pass structured results between stages, pause for user input and let agents verify findings. Good fits include release validation, parallel review of changed files and codebase pattern detection. The feature is in public preview on all Copilot plans, and in the CLI you must enable experimental mode with the --experimental flag or /experimental on.
HydraFusion
According to the GitHub changelog, HydraFusion is a research preview that treats workflow selection as an optimization problem. For each request it picks one of three approaches:
- Single: one model solves the task directly.
- Cascade: an efficient model drafts a solution, and it escalates to a stronger model only if the draft misses the quality bar.
- Critique: one model drafts, an independent model reviews, and the original model revises based on that feedback.
This differs from the Auto model, which picks a single model per request. The September 30 update adds more transparency into what happens at each step, real-time progress updates and clearer status indicators on long tasks. It requires VS Code 1.140 or later or the GitHub Copilot desktop app, and you enable it through settings or the model picker. It is available on Copilot Pro, Pro+, Business and Enterprise, and Business and Enterprise need an administrator to approve preview features.
Model deprecations and code review
On October 2 GitHub deprecated four models:
- Gemini 3.5 Flash and Gemini 3.6 Flash: replaced by Gemini 3.8 Flash.
- Kimi K2.7 Code: replaced by Kimi K3.
- Claude Opus 4.7: replaced by Claude Opus 5.5.
Enterprise admins should enable the replacement models in Copilot settings. Separately, Copilot code review now has REST and GraphQL API support with a selectable effort level. Balanced became the default effort level from September 28, and people who chose Lite keep that setting.
Copilot CLI 1.0.91 and the 1.0.92 pre-releases
Per the Copilot CLI release page, version 1.0.91 (October 1) added copilot sandbox ca commands for managing proxy certificate authority trust and improved shell pipeline analysis. The 1.0.92 pre-releases (October 2) add a Ctrl+E environment picker for switching between local and cloud runs, network bypass prompts for sandboxed shell commands, MCP reconnect fixes, Windows fixes for sandboxed temp files, and fixes for Git authentication with masked credentials.
OpenAI Codex: CLI 0.160.0
The Codex changelog and the GitHub releases page list 0.160.0 as the latest stable build with notes, dated October 1 (Releasebot lists it as October 2). It follows 0.159.3 (September 30), which lets eligible local sessions signed in with ChatGPT receive optional reminders to finish account security setup.
What 0.160.0 adds:
- Browsing older tasks in the agent command center with a keyboard-accessible "Show more" action.
- Workspace defaults for sessions where policy permits.
- Opt-in Guardian review that can retrieve earlier user instructions.
- Better Windows sandbox PowerShell fallbacks, a fix for SQLite connection setup stalls, and X11 primary selection and middle-click paste for Linux terminals.
Builds of 0.162.0 are alphas with no detailed notes, so stay on the latest stable build for now. Last week's models, including GPT-6.1 Sol, were covered in the September 30 roundup.
OpenCode: v1.18.34
The OpenCode changelog still shows v1.18.34 from September 30 as the latest release. Per the release notes, it sends namespaced session and parent-session identity headers with model requests and improves macOS binary signing. A maintenance release rather than a feature drop.
What to Do
- Update Claude Code to 2.1.289, especially if you rely on deny or ask rules together with sandbox auto-allow, or on IDE selections through symlinks. Then test your rules against commands with variable prefixes.
- Plugin authors should read up on Mods, since they reach deeper than earlier plugin hooks.
- If you pin Copilot models in workflows or integrations, move off the deprecated ones now.
- Treat Copilot computer use as a preview: keep approvals on and test on non-critical apps first.
- Teams scripting code review can now call it by API. Check the effort level your tooling expects.
- Copilot admins should decide whether to approve preview features for Business and Enterprise seats.
- Stay on stable Codex until notes appear for 0.162.0, and test Guardian review before relying on it.
Sources
- Claude Code changelog
- Claude Code CHANGELOG.md
- Claude Code releases on GitHub
- GitHub Changelog: HydraFusion in VS Code and the GitHub Copilot app
- GitHub Changelog: Copilot can now interact with desktop apps
- GitHub Changelog: Dynamic workflows in Copilot CLI and the Copilot app
- GitHub Changelog: Selected models deprecated
- GitHub Changelog: Copilot code review API support
- GitHub Copilot changelog
- GitHub Copilot CLI releases
- OpenAI Codex changelog
- OpenAI Codex releases on GitHub
- Releasebot: OpenAI Codex updates
- OpenCode changelog
- OpenCode releases
Comments
Share your thoughts and join the conversation
Leave a Comment
Keep reading.

Claude Fable 5.2 Guide: Pricing, Benchmarks and When It Beats Opus 5.5

Node.js HIGH Severity Release Due Nov 11, Elysia 2.0 RC4: Dev Stack Release Audit, November 9, 2026
