AI Coding News: August 7, 2026 — Claude Code Ships Self-Hosted Runners and Cross-Session Messaging

August 7 was a bigger release day than most: Claude Code jumped a full version to 2.1.224 with self-hosted runners and cross-session agent messaging, OpenAI's Codex CLI hit 0.147.0 with plugin installs and a new auto-approve flag, and GitHub started rolling the open-weight Kimi K3 model into Copilot. Gemini CLI, OpenCode, and GitHub Copilot CLI stayed quiet or shipped only patch-level housekeeping. Here's what actually changed, sourced from each project's own changelog.

Claude Code 2.1.224: Self-Hosted Runners and Cross-Session Messaging

Anthropic's Claude Code changelog jumped from 2.1.223 to 2.1.224 in a single day, its largest release this week. The headline addition is claude self-hosted-runner, which turns your own machines or containers into a place Claude Code web, mobile, and desktop sessions can run — available on Team and Enterprise plans. Alongside it, Claude Code sessions can now message each other across machines: a new cross-session SendMessage works with ListAgents for discovery on macOS and Linux, and crossSessionInbound/dialogExpiry settings decide whether a message to a bypassed-permissions session waits for approval or delivers automatically.

• New: archive plugin source installs plugins from a zip over HTTPS without git or npm, with optional SHA-256 pinning.
• New: the 200-subagent-per-session spawn cap is gone, so long-running sessions no longer refuse new agents (concurrency and depth limits still apply).
• Security: sandbox credential masking adds extract/onExtractNoMatch for structured env values, JWT-aware masking via decode: "jwt" with maskClaims, and awsPairs/sigv4 for AWS SigV4 re-signing.
• Fixed: sandbox filesystem deny entries written with a trailing slash (e.g. denyRead: "~/.aws/") were silently bypassable on Linux and macOS.
• Fixed: sandbox violation details never appeared in Bash tool results, so Claude couldn't see which file or network access was denied and why.

OpenAI Codex CLI 0.147.0: Plugin Installs and an Auto-Approve Flag

OpenAI's Codex CLI releases page shows 0.147.0 landing the same day, one full release past last week's rust-v0.146.1 patch. The release adds portable Agent Plugins you can install and search for across local, personal, workspace, and remote catalogs, plus a new --approve-for-me flag that enables automatically reviewed approvals for agents you're willing to trust with less hand-holding. Codex can now import Cursor-managed skills and keep imported Claude and Cursor conversations in sync, and it adds opt-in support for the MCP 2026-07-28 protocol, including paginated discovery and multi-round requests.

• New: conversations can be organized into persistent, manually ordered sections, and long transcripts browse incrementally instead of loading all at once.
• New: cached web search and remote conversation compaction are now available for Amazon Bedrock.
• Security fix: secrets and bearer tokens are now redacted from displayed commands and replayed conversation history.
• Removed: the deprecated codex exec --full-auto flag — use --sandbox workspace-write instead.

GitHub Copilot: Kimi K3 Goes GA

GitHub's Copilot changelog confirms the open-weight Kimi K3 model is now generally available across Copilot Pro, Pro+, Max, Business, and Enterprise, usable in VS Code, JetBrains, Xcode, and on github.com. GitHub hosts the model on Fireworks AI and prices it at $3 per million input tokens, $15 per million output tokens, and $0.30 per million cached input tokens. Business and Enterprise admins need to turn it on explicitly — it ships off by default for those plans pending further quality and performance monitoring.

Also Watching: Gemini CLI, OpenCode, GitHub Copilot CLI

Google's Gemini CLI releases page shows v0.54.4 published August 7, but the diff is version-bump housekeeping — a cherry-picked patch and a couple of version-number corrections — not a user-facing change; the real fixes (capacity-exhaustion classification, stream error propagation, a macOS seatbelt profile fallback) are still sitting in the v0.55.0 preview channel. OpenCode's GitHub releases hold at v1.18.14 from August 5, and GitHub's Copilot CLI changelog hasn't moved past v1.0.78 from August 3 — both quiet for a second straight day.

Why This Matters

Today's releases point the same direction: agents are being given more autonomy and more reach, not just more raw model capability. Claude Code's cross-session messaging and self-hosted runners let agents coordinate across your own machines instead of staying siloed in one terminal, and Codex's --approve-for-me flag automates a slice of the review loop that used to require a human in the middle. Copilot's Kimi K3 rollout points the same way from the model side — Copilot is becoming a router across providers, not a single-model tool. None of this is dangerous by itself, but each change widens what an agent can do without a human directly watching, which is exactly the kind of thing worth reading the fine print on before you turn it on for a whole team.

Sources & References

• Claude Code changelog docs
• Claude Code CHANGELOG (anthropics/claude-code)
• OpenAI Codex CLI releases (openai/codex)
• GitHub Copilot changelog: Kimi K3 availability
• GitHub Copilot supported models documentation
• Gemini CLI releases (google-gemini/gemini-cli)
• OpenCode releases (sst/opencode)
• GitHub Copilot CLI changelog