AI Coding News Roundup — August 9, 2026: Claude Code, Copilot, Codex & OpenCode Updates

August 9, 2026 — This week's AI coding tool cycle leaned heavily on infrastructure and workflow polish rather than headline model launches. Claude Code shipped self-hosted runners and cross-session messaging, GitHub Copilot brought the open-weight Kimi K3 model into general availability, OpenAI's Codex CLI added portable agent plugins that sync with Cursor and Claude, and OpenCode pushed four releases in seven days. Here's everything that changed.

Claude Code: Self-Hosted Runners and Cross-Session Messaging

The biggest feature of the week landed in v2.1.224 (August 7): claude self-hosted-runner, which lets Team and Enterprise plans run Claude Code web, mobile, and desktop sessions on their own infrastructure. The same release added an archive plugin source for installing plugins from HTTPS zip files with optional SHA-256 pinning, cross-machine SendMessage so Claude Code sessions can message each other across machines, and new sandbox credential-masking options — including JWT-aware masking and AWS SigV4 re-signing — that keep secrets out of logs during agentic runs.

The following day's v2.1.225 (August 8) update focused on reliability: gateway spend-limit support now shows the cap, reset time, and operator message directly in usage warnings, and a new workspace trust prompt appears before claude agents touches an untrusted directory. Anthropic also patched a transient 401 error that broke headless sessions when a long-lived CLAUDE_CODE_OAUTH_TOKEN got silently replaced by a short-lived stored-login token, and fixed MCP OAuth servers on macOS failing intermittently after Keychain read timeouts. v2.1.226 followed hours later with further bug fixes and reliability improvements.

Other fixes worth knowing about:

• Auto mode no longer counts safety-filter refusals toward the consecutive-block limit.
• Cross-session messages in headless sessions no longer sit parked indefinitely without notice or expiry.
• Conversation history no longer breaks when a Remote Control session resumes after a large conversation was compacted.
• Photos attached from the Claude mobile app now show directly to Claude during Remote Control sessions instead of requiring a separate tool call.

GitHub Copilot: Kimi K3 Goes GA, CLI Gets a Sessions Sidebar

Moonshot AI's open-weight Kimi K3 model became generally available inside GitHub Copilot this week, priced at $3 per million input tokens, $15 per million output tokens, and $0.30 per million cached input tokens — a notably cheap option for agentic coding workloads across every Copilot surface.

The August 3 weekly release touched the Copilot desktop app, CLI, and VS Code 1.132. Auto mode now shows which model handled each completed request along with AI credit and cache details. The CLI picked up multi-session management from a sidebar, an experimental /worktree command for isolated workspaces, and a git-free /rewind to restore prior conversations and files. VS Code 1.132 added browser element-level feedback with inline annotations, on-device multilingual dictation, a /btw side-chat command that runs a parallel question without losing the main thread's context, and a hybrid markdown diff viewer with gutter change indicators.

• The standalone Copilot Billing Preview app was retired — spend now lives directly in GitHub billing settings.
• GitHub confirmed a round of model deprecations across all Copilot experiences landing September 1, 2026, though Claude Sonnet 4.6 stays available to individual subscribers on annual plans.

OpenAI Codex CLI: Portable Plugins and Cross-Tool Skill Sync

Codex CLI 0.147.0 (August 7) introduced Portable Agent Plugins that can be installed and searched across multiple plugin catalogs, plus persistent conversation sections for organizing long-running threads with incremental browsing. A new --approve-for-me flag automates review approvals for trusted workflows, and Codex can now import Cursor-managed skills and sync Claude and Cursor conversations without creating duplicates. The MCP protocol implementation moved to spec version 2026-07-28, adding paginated tool discovery and non-blocking server startup, and Amazon Bedrock now supports cached web search and remote conversation compaction.

On the security side, 0.147.0 redacts secrets from displayed commands and conversation history and fixes terminal input loss during focus changes. The prior patch, 0.146.1 (August 5), applied safer automatic-review defaults for cyber-capable models and clarified permission-change explanations in the terminal UI. Note: the deprecated codex exec --full-auto flag has been removed in favor of --sandbox workspace-write, and GPT-5.4 / GPT-5.4 mini stop being available inside Codex for ChatGPT-authenticated users on August 31 — API-key sessions are unaffected.

OpenCode: Four Releases in a Week

OpenCode shipped four point releases between August 4 and August 7. v1.18.15 improved message ordering, revert/fork actions, and file-truncation cleanup, added JSON session-transcript exports and broader desktop locale coverage, and fixed blob-based attachments failing to load in the web UI. v1.18.14 streamlined xAI login with a device-code flow for headless environments and improved provider/network error handling. v1.18.13 added PR numbers and URLs to GitHub pull request reviews and fixed right-to-left layout issues across desktop components, and v1.18.12 fixed Azure GPT-5.5+ completion requests with reasoning enabled while reducing composer lag on large pasted images.

Also Worth Watching

Cursor rolled out Cursor Router, pairing Auto Intelligence with a new Auto Balance mode that adapts model routing from live production traffic — Opus 5 has joined the rotation, and Cursor says Auto Intelligence now delivers above-Fable-level satisfaction at 68% lower cost, an 18% improvement since launch.

Google continued migrating users off Gemini CLI toward Antigravity CLI, which hit v1.1.10 (August 3) with Gemini Enterprise Business sign-in, Workforce Identity Federation and Application Default Credentials auth support, read-only .git sandbox access, and fixes for subagent tree termination and --model/--effort flag resolution. Gemini CLI itself is now down to nightly maintenance builds as the sunset continues.

TL;DR — What Actually Matters

• Claude Code v2.1.224–2.1.226 — self-hosted runners for Team/Enterprise, cross-machine SendMessage, gateway spend-limit warnings, and a workspace trust prompt for claude agents.
• GitHub Copilot — Kimi K3 now GA at $3/$15 per million tokens; CLI sessions sidebar, /worktree, and /rewind shipped August 3.
• OpenAI Codex CLI 0.147.0 — portable plugins across catalogs, --approve-for-me, Cursor skill import, MCP spec 2026-07-28.
• OpenCode — four releases (v1.18.12–v1.18.15) this week focused on MCP/provider fixes and desktop polish.
• Also: Cursor Router's new Auto Balance mode, and Antigravity CLI v1.1.10 as the Gemini CLI replacement matures.

Resources & References

• Claude Code Official Changelog
• Claude Code Release Tracker — Havoptic
• GitHub Copilot Weekly Releases — August 3
• Kimi K3 Now Available in GitHub Copilot
• ChatGPT & Codex CLI Changelog
• OpenCode Changelog
• Cursor Changelog
• Google Antigravity CLI Changelog
• Transitioning Gemini CLI to Antigravity CLI — Google Developers Blog