Switchback 2.4.0

Added

• Asking for subagents gets you subagents. "Use 4 subagents to…" starts four, in parallel, each with its own brief and files; with no number the model uses as few as the work needs, and when a number looks wrong for the work it says so and you agree on one. Two subagents never edit the same file: each gets its own files or its own worktree
• Big fan-outs don't flood the parent's context. The subagents started in one step share 60,000 characters of report, and each is told its share, so 150 of them cost the parent about what two do. Subagents get a short note on reporting instead of the whole Switchback guide
• A built-in review agent. "Have a reviewer check it" sends a read-only reviewer over the change, which lists bugs and missed cases with file and line
• Subagents waiting for a slot show as waiting. Past subagents.maxConcurrent, a subagent's row says waiting for a free slot until it starts (subagent.queued in the protocol), and a running row shows the model it's on
• The mock provider fans out use 3 subagents to <goal>, so --mock demos orchestration without a model
• The prompt-cache hit rate is in the status bar. The TUI's status bar and VS Code's composer show how much of the session's input came from the provider's cache (cache 91%), with the latest turn's when it differs. call.stats events carry each call's usage. --mock simulates a prefix cache, so the rate shows there too
• /rename names a session. Up to 60 characters, in the TUI and VS Code (session.rename in the protocol). The title shows in /sessions, --resume, the VS Code view, the terminal's title (restored on exit), and ACP editors. /rename alone shows the current title
• Sessions are retitled at compaction. The summary that compaction already makes suggests a title, so it costs no extra model call; a title you set is never replaced
• /goal <objective> keeps working until it's done. The agent starts turn after turn until the model reports the goal done or blocked with the new goal tool, a goal.maxTurns or goal.maxMinutes limit is reached, three turns in a row fail, or you stop it with Esc or /goal stop. Each new turn is an appended reminder, so routing, budgets, and permission prompts apply as usual. Both clients show the goal above the prompt (ADR 0024)

Changed

• 16 subagents run at once by default, up from 4 (subagents.maxConcurrent), so a large fan-out finishes sooner. Local servers queue what they can't run at once
• Tool results and @ mentions fit the model's context window. A step's tool results share a quarter of the calling model's window (2,000 to 30,000 characters each), and a prompt's attachments half the session's, so a small local model isn't swamped by one read and pushed to escalate. Cut results keep their last lines, so a command's exit code survives. The cap is applied in one place for every tool, MCP tools included
• More of each request hits the cache. Claude requests add a cache breakpoint on the system prompt, so tools and system stay cached after compaction or past the 20-block lookback. OpenAI and Azure OpenAI requests send the session ID as prompt_cache_key, so a session's calls reach the same cache

Fixed

• Copying text works on Windows. The TUI no longer turns on mouse reporting there, so dragging selects text and Ctrl+C or Enter copies it in Windows Terminal, conhost, and VS Code's terminal. /mouse on brings back wheel scrolling and tab clicks, and /mouse off gives plain selection on macOS and Linux; the choice is remembered
• PowerShell commands don't flash a console window on Windows. Commands run outside the sandbox, hooks, git, ripgrep, the clipboard helpers, and the engine VS Code starts now run with their window hidden
• Status bar marks have their space back. The permission mode's mark (» accept edits, ║ plan, ! bypass) and the remote model's (▲, was ☁) are one cell wide in every terminal; the emoji-capable ones they replace (⚠, ⏸, ☁) were drawn two cells wide on Windows and covered the space after them
• Setup with one model doesn't ask which model does what. It used to offer an escalation ladder, reviewers, and a subagent model that a single model can't have

Install

On macOS and Linux: curl -fsSL https://switchback.sh/install.sh | sh (add -s -- --vscode after sh for the VS Code extension too). On Windows, in PowerShell: irm https://switchback.sh/install.ps1 | iex. Or download the binary for your platform, verify it against SHA256SUMS, make it executable, and put it on your PATH as switchback. Then run switchback init.
Install the VS Code extension from the Marketplace or Open VSX (code --install-extension isaiah-harville.switchback), or from the switchback-vscode-2.4.0-<platform>.vsix here. Each platform's extension includes the engine, so no separate install is needed; the universal .vsix uses switchback from your PATH.

Binaries are not yet signed or notarized (#33). On macOS, clear the quarantine flag with xattr -d com.apple.quarantine switchback.