AI Coding Roundup — June 28, 2026: Claude Code /rewind, OpenCode Snapshots & More

Five days of dense shipping: between June 23 and June 26, Anthropic released Claude Code v2.1.187, v2.1.191, v2.1.193, and v2.1.195 — a rapid cadence that brought a new /rewind command, a 37 per cent cut in streaming CPU usage, shell command routing through the auto-mode classifier, and tighter security controls. OpenCode matched the pace with v1.17.10 and v1.17.11, adding MCP server instructions to the session context, a --mini terminal mode, and — most notably — session snapshots with full file-level revert support.

Claude Code v2.1.187–2.1.195: Four Releases in Five Days

v2.1.187 — Credential Sandboxing & Org Model Restrictions (June 23)

Released late on June 23, v2.1.187 introduced a sandbox.credentials setting that blocks the model from reading credential files on disk — a meaningful hardening step for teams running Claude Code with broad file access. Organisations can now configure model restrictions from their admin dashboard, and a clear messaging string is shown to users when a model is unavailable under their plan. A smaller UX improvement: mouse click support in fullscreen select menus means you can pick list items with a click rather than arrow keys, and a fix to structured output prevents it from looping indefinitely on schema mismatches.

v2.1.191 — /rewind & 37% Faster Streaming (June 24)

The /rewind slash command is the headline: calling it after /clear walks the session back to the state before the clear, letting you recover a conversation branch without losing your reset intent. Complementing that UX win is a streaming CPU reduction of roughly 37 per cent, measured internally — noticeable on longer agentic runs generating heavy text output. The release also fixed a reliability edge case where background agents would resurface after being explicitly stopped. Remote MCP tool calls that previously hung silently for five minutes now abort with a clear error; the timeout is configurable via CLAUDE_CODE_MCP_TOOL_IDLE_TIMEOUT. Sandbox network permission dialogs also gained memory for previously approved hosts, reducing repeated confirmation prompts during long agentic runs.

v2.1.193 — Auto-Mode Shell Routing & OpenTelemetry Logging (June 25)

The new autoMode.classifyAllShell setting routes all bash and PowerShell commands through the auto-mode classifier, giving organisations granular visibility into which shell operations are auto-approved versus held for confirmation. When auto mode denies a command, the reason now appears in the transcript and in /permissions — previously a black box. On the observability side, a new OpenTelemetry log event claude_code.assistant_response is emitted for every assistant turn, making it straightforward to pipe Claude Code sessions into Grafana, Loki, or any OTLP-compatible backend. Developers in bash mode also gain live file path autocomplete mid-command — a small but frequently requested ergonomic improvement.

v2.1.195 — Voice Dictation Fixes & Hook Matcher Accuracy (June 26)

The latest Claude Code release as of June 26 patches two separate voice issues: macOS dictation no longer drops words mid-sentence, and languages without word-boundary spaces — Japanese, Chinese, Korean, Thai — now auto-submit correctly. A subtle but previously breaking bug in hook matchers is fixed: identifiers containing hyphens (code-reviewer, mcp__brave-search) were matched as substrings, causing hooks to fire on unintended events — they are now exact-matched. External plugins no longer bypass the install consent screen on some loader paths. The new CLAUDE_CODE_DISABLE_MOUSE_CLICKS environment variable lets users disable fullscreen mouse interaction while preserving mouse-wheel scrolling — useful in environments where accidental clicks cause unwanted selections.

OpenCode v1.17.10 & v1.17.11: MCP Instructions, Session Snapshots & –mini Mode

v1.17.10 — MCP Server Instructions & –mini CLI Mode (June 24)

OpenCode v1.17.10 injects MCP server instructions directly into the session context window, letting tools self-describe their usage requirements without the user needing to consult external documentation. The --mini CLI flag launches OpenCode in a compact terminal mode suited for smaller screens or split-pane workflows — think of it as a lightweight shell companion rather than a full IDE replacement. Desktop users get a new session progress indicator showing step-by-step task completion at a glance, plus collapsible server sections to keep long-running multi-tool sessions readable.

v1.17.11 — Session Snapshots & Draggable Tabs (June 25)

The headline feature in v1.17.11 is session snapshots with full revert: OpenCode now creates a checkpoint at each message, capturing both the conversation state and any file changes made up to that point. Rolling back undoes both the conversation and the file edits — a meaningful safety net for long agentic sessions where a sequence of file changes needs to be cleanly unwound without manual git resets. On the desktop side, tabs are now draggable and support Chrome-style keyboard shortcuts for tab cycling, making multi-project and multi-session workflows significantly more ergonomic.

Why This Week Matters

The convergence between Claude Code and OpenCode updates this week is telling. Both tools zeroed in on MCP reliability, auto-mode transparency, and session recovery in the same five-day window — suggesting a maturing phase where teams running AI agents in production are surfacing identical pain points: unpredictable shell command approval, opaque auto-mode decisions, and the absence of a reliable rollback path in long-running sessions. Claude Code's /rewind and OpenCode's session snapshots address the same core developer anxiety: committing to an AI-driven workflow without a meaningful undo. If you are evaluating either tool for team adoption, the week of June 24 significantly strengthened both safety stories.

Resources & References

Claude Code Official Changelog · Claude Code GitHub Releases · OpenCode Changelog · OpenCode GitHub Releases · OpenTelemetry Documentation · Model Context Protocol (MCP)