Claude Code 2.1.139: Agent View Arrives in Research Preview, Plus a /goal Command

Claude Code 2.1.139 landed on May 11. The headline feature is agent view, a research preview that puts every Claude Code session, running, waiting on you, or finished, into a single terminal interface. The second is a native /goal command that drives Claude across as many turns as it takes until a condition you specify is met.

Agent View: One Place for All Your Sessions

Until now, running multiple Claude Code sessions meant juggling terminal tabs with no single view of what each was doing. Agent view changes that.

Run claude agents from the terminal, or press the left arrow on an empty prompt inside any session, to open it. The full terminal becomes a list of every session under your config directory — across projects and worktrees. Each row shows the session name, what Claude is currently doing or last said, and how long ago it last changed. Sessions that need your input rise to the top.

From agent view you can kick off a new background session, jump into one that’s waiting, or just watch. To push an existing interactive session into the background, run /bg inside it or press ← on an empty prompt, which backgrounds the session and opens agent view at the same time. From the shell, claude --bg "<prompt>" starts a new background session directly.

The usual rate limits apply. Background sessions consume quota at the same rate as interactive ones, so running ten agents at once uses roughly ten times as much.

Agent view is a research preview, requires v2.1.139 or later, and is available on Pro, Max, Team, Enterprise, and API plans.

/goal: Run Until Done

The /goal command takes a completion condition in plain text. After each turn, Claude’s configured small fast model (Haiku by default) reads the condition and the conversation so far and returns a yes/no decision plus a short reason. If the condition isn’t met, Claude starts another turn automatically instead of handing control back to you.

A live overlay panel shows elapsed time, current turn count, and tokens spent. Run /goal with no arguments inside an active goal session to see the current totals at any point.

A few things worth knowing: the evaluator judges against what Claude surfaces in the conversation text. It doesn’t run commands or read files independently, so the condition needs to be something Claude’s own output can demonstrate. /goal works in interactive mode, with the -p flag, and through Remote Control.

Transcript Navigation and Other Additions

The release also adds keyboard shortcuts for navigating transcript view — useful for long sessions where scrolling by hand gets tedious — and a new /scroll-speed command that lets you tune mouse wheel speed with a live preview.

claude plugin details <n> shows a plugin’s component inventory and its projected per-session token cost. For teams with large plugin sets, that cost estimate is a practical way to understand where token usage is going before a session starts.

There’s also a new args: string[] field for hooks in exec form. It spawns the command directly without a shell, which means path placeholders in hook arguments no longer need quoting to survive shell expansion.

Fixes

The fix list covers Windows keyboard input issues, a silent exit 1 on startup when 10 or more MCP servers are configured and the cache directory isn’t writable, memory leaks from MCP servers, a bug where /unauthorized MCP connectors showed as “failed” instead of “needs auth,” and several display issues including the typing cursor blinking in the wrong state, model picker rendering, and hyperlink visibility on dark themes.

Bash-mode up-arrow history, multiple image pasting, and letter shortcuts in transcript view after a mouse click were also corrected.

Why It Matters

Agent view is the practical complement to the parallel-agents work Anthropic has been shipping across recent releases. The infrastructure for running background sessions, remote control, and worktree isolation has been building for months. Agent view is the surface where you actually manage it. Running it as a research preview signals Anthropic is still iterating on the interaction model, but the core function, a unified session list with attention routing, is already useful.

The /goal command closes a loop Codex CLI opened with its own /goal support. Both tools now let you specify a target state and step away, rather than supervising turn by turn.

---

Sources: Claude Code v2.1.139 release, Agent view docs, Anthropic blog: agent view in Claude Code, Claude Code changelog