Dev Tool Changelog: Fable 5, Nested Agents, and Bugbot Gets Fast

Eight Claude Code releases. Cursor's review bot got three times faster. Zed shipped agent compaction and a new "Fast mode" for AI queries. Here's what changed across AI dev tools this week.

Claude Code (8 releases, June 5-11)

The big one: Fable 5 landed June 9 (v2.1.170). Anthropic calls it their most capable model ever made generally available. It runs at $10/$50 per million tokens (2x Opus 4.8), with a 1M context window and 128K max output. It is free on Pro, Max, Team, and Enterprise plans through June 22, after which you need usage credits. This is a "try before you buy" play, and the two-week window is tight if you want to benchmark it against Opus on your own codebase.

The second headline: nested sub-agents now go five levels deep (v2.1.172, June 10). You can have an agent spawn an agent that spawns an agent, and so on, up to five layers. This matters for complex multi-repo workflows where one orchestrator delegates to specialists. Whether anyone actually needs five levels is a separate question, but two or three levels unlock real workflows that were awkward before.

Other changes worth knowing about:

• --safe-mode flag (v2.1.169) disables all customizations (CLAUDE.md, plugins, skills, hooks, MCP servers) for troubleshooting. Useful when something breaks and you need to rule out your own config.
• /cd command (v2.1.169) lets you switch working directories mid-session without blowing the prompt cache. Small change, big quality-of-life improvement.
• fallbackModel setting (v2.1.166) lets you configure up to three fallback models tried in order when the primary is overloaded. The CLI also auto-retries on unexpected API errors once before surfacing them.
• Hardened cross-session messaging (v2.1.166). Messages relayed via SendMessage from other Claude sessions no longer carry user authority. Receivers refuse relayed permission requests, and auto mode blocks them entirely. This fixes a real attack surface if you run multiple agents concurrently.
• disableBundledSkills setting (v2.1.169) hides built-in skills and slash commands from the model. Enterprise admins will want this for locked-down environments.
• Post-session lifecycle hook for self-hosted runners (v2.1.169) runs after the session ends, so you can snapshot uncommitted work or export logs before the workspace gets deleted.
• Bedrock region auto-detection (v2.1.172) reads ~/.aws config when AWS_REGION is not set. The /status command now shows where the region came from.

Quick take: 110+ individual changes across eight releases is a lot. The Fable 5 launch and nested agents get the attention, but the security hardening on cross-session messaging is arguably more important for anyone running multiple agents in production. The /cd command should have existed months ago.

Cursor (2 updates, June 5-10)

Bugbot got a major performance upgrade (June 10). Review time dropped from roughly five minutes to about 90 seconds. Cost per run fell 22%. And bugs found per review actually went up 10% (from 0.56 to 0.62 on average). The gains come from training improvements to the Composer 2.5 model.

The new /review slash command runs Bugbot and Security Review before you push code. You can also use /review-bugbot and /review-security directly. If you run /review locally and then open a PR with the same diff, Bugbot recognizes the overlap and skips re-reviewing. There is also a new incremental re-review config option that only reviews what changed since the last review.

Cursor 3.7 (June 5) brought two Design Mode improvements: multi-select elements (click two or more elements in the Cursor browser and the agent sees their code, layout, and visual relationships) and voice input (narrate UI changes while the agent is running).

Quick take: Bugbot going from five minutes to 90 seconds matters because review latency was the main reason people skipped it. The /review command syncing with GitHub/GitLab PRs is the kind of integration detail that makes a tool stick.

Zed (4 releases, June 6-11)

Zed shipped three stable releases and one pre-release this week.

v1.7.1-pre (June 11) added agent auto-compaction and a new /compact command for manually triggering it. The agent panel now shows context window usage and cost metrics for external agents. Agent skills management moved to the settings UI. New OpenCode models were added: MiniMax M3, Qwen 3.7 Plus, DeepSeek V4 Flash, and Nemotron 3 Ultra Free.

v1.6.3 stable (June 10) is the bigger one. It added "Fast mode" for Anthropic and OpenAI models in the agent panel (faster responses at higher token cost), single-file Git diff tabs from the Git panel, shareable agent skill links, Claude Opus 4.8 BYOK support, Grok 4.3 reasoning effort support, agent terminal sandboxing with per-path write access, and a new agent.commit_message_instructions setting.

v1.5.5 (June 9) was a small patch. v1.5.4 (June 6) fixed a crash when unsplitting side-by-side diff view and fixed ACP Registry agent downloads not starting.

Quick take: Zed is moving fast on agent features. Fast mode is a direct answer to the latency problem everyone hits with large models. The /compact command is something Claude Code still does not have as an explicit slash command (though it auto-compacts). Opus 4.8 BYOK matters for teams already paying for Anthropic API access separately.

Windsurf / Devin Desktop (v3.1.7, June 6)

Reminder: Windsurf was acquired by Cognition and rebranded as Devin Desktop. The changelog now lives at docs.devin.ai/desktop/changelog.

v3.1.7 (June 6) was a maintenance release: hardened Windows migration (preserving shortcut icons), improved file context handling in Devin Local agent, increased proxy authentication timeout for corporate networks, and fixed stdio MCP servers on Windows. Enterprise customers got a simplified model picker pricing view.

Quick take: The rebrand is still settling in. This release was almost entirely plumbing and migration fixes. No new agent capabilities this week.

OpenAI Codex (1 alpha release, June 9)

The only release in the June 5-12 window was CLI rust-v0.139.0-alpha.1 on June 9, an alpha/pre-release build. It adds mobile access with branch selection, worktree support, environment setup scripts, and inline review comments for task outputs.

The stable CLI v0.137.0 landed June 4 (just outside the window) with a long list: F13-F24 keybindings, monthly enterprise credit limits, cloud-managed config bundles, remote-control pairing, plugin JSON output, and a batch of multi-agent v2 improvements.

Quick take: Codex stable releases are every one to two weeks, so this is a normal cadence. The alpha build suggests mobile and worktree features are coming to stable soon. If you are waiting on worktree support for parallel Codex tasks, the alpha is available now.

Not on the list

Antigravity had no releases this week. Last release was v2.0.11 on June 3 (dark screen startup fix). The Gemini CLI shutdown is June 18, after which Antigravity CLI is the only option for Google-based agent workflows.

OpenCode remains archived since September 2025. Removed from future changelogs.

The week in perspective

Three things stand out. First, every major tool is converging on nested/multi-agent architectures: Claude Code now supports five levels deep, Cursor added nested subagents to its SDK (June 4), and Zed keeps expanding its agent panel. Second, model resilience is becoming a feature category. Claude Code's fallbackModel and Zed's Fast mode both address the same problem: models go down or get slow, and your workflow should not die when that happens. Third, review automation is getting real infrastructure. Cursor's Bugbot went from a nice-to-have to something you can actually run on every commit without waiting five minutes.

No breaking changes anywhere this week. The Fable 5 free period ends June 22, and the Sonnet 4/Opus 4 model ID retirement hits June 15. If you are still using old model IDs, you have days, not weeks.

Sources: Claude Code changelog, Cursor changelog, Zed releases, Devin Desktop changelog, Codex CLI releases