Claude Code 2.1.139 adds `claude agents` view and `/goal` runs with live turn counters

• ClaudeCodeLog's release thread says Claude Code 2.1.139 shipped 50 CLI changes, led by a new claude agents control plane and a /goal command that keeps working across turns until a completion condition is met.
• According to claudeai's agent view announcement and claudeai's availability note, agent view is a research preview on all paid plans that puts running, blocked, and finished sessions into one list.
• ClaudeCodeLog's changelog post says /goal works in interactive mode, -p, and Remote Control, and surfaces live elapsed time, turn count, and token count while it runs.
• ClaudeCodeLog's system prompt note and ClaudeCodeLog's release thread both flag a quieter compaction path: Claude is no longer explicitly told long chats trigger auto-compaction, and the trimming alert is gone.
• The release also changes who gets certain remote features, because ClaudeCodeLog's changelog post says Remote Control, /schedule, MCP connectors, and notification preferences are disabled whenever an API key or auth token is set.

You can read Anthropic's agent view post, skim the agent view docs, inspect the official 2.1.139 changelog entry, and compare prompt and CLI diffs in marckrenn's tracked release. The interesting bits split cleanly: a built-in session control plane, a long-run goal loop with visible counters, a silent compaction tweak, and a few under-the-hood changes that make plugins, hooks, and subagents easier to inspect.

Anthropic is turning multi-session Claude Code into a first-party workflow. The official pitch from claudeai's agent view announcement is simple: one list for every session, while claudeai's follow-up post adds inline replies, quick jumping between sessions, and state labels for what is running, waiting, or done.

The changelog and docs line up on the mechanics:

• Run claude agents to open the view, according to ClaudeCodeLog's changelog post and the agent view docs.
• Sessions keep running without occupying a terminal tab, per claudeai's dispatch note.
• The view is a research preview on paid plans, according to claudeai's availability note.
• Anthropic describes it as a list of sessions across states, while Boris Cherny's post frames the jump as a way to move from one agent to many without tab-cycling.

Practitioner reaction immediately snapped to terminal metaphors. trq212's tmux comparison called it "tmux built for CC," and Cat Wu's setup tip suggested running it from a root code directory so one control plane can track agents across repos.

/goal is the other big ship, and it is a very specific one: tell Claude Code what done looks like, then let it keep taking turns until it gets there. ClaudeCodeLog's changelog post says it works in interactive mode, print mode, and Remote Control, with a live overlay for elapsed time, turns, and tokens.

The strongest detail here is not just autonomy, it is instrumentation. Kevin Kern's long-run post shows a /goal session still running after two hours, while Matt Lam's screenshot shows a Codex /goal run pursuing the same kind of long objective for more than a day.

That makes Claude Code's addition feel like Anthropic catching up to a workflow that Codex users were already stress-testing over the weekend. Adam G's post says /goal had already landed in the Codex app, not just the CLI, and OpenAI's own follow-goals guide defines the pattern as work with a clear target and validation loop.

The spiciest change is a prompt change, not a UI feature. ClaudeCodeLog's system prompt note says Claude is no longer explicitly told that long chats trigger automatic context compaction, while ClaudeCodeLog's release thread adds that the old trimming alert is now gone.

ClaudeCodeLog's diff thread describes two parts:

• the harness no longer reminds Claude that long chats will be compressed, per ClaudeCodeLog's system prompt note
• the compaction prompt now asks the model to preserve sensitive user instructions, per ClaudeCodeLog's changelog post

That is a subtle trade. The model gets less explicit warning that compression happened, but the compaction pass gets a stronger instruction to preserve high-value user intent. In the release metadata tracked by ClaudeCodeLog's additional updates, system prompt tokens fell by 781 overall, with the mix shifting from system tokens toward system-reminder tokens.

One change in the changelog is easy to miss because it sits under safety, but it materially changes who can use the new remote surfaces. ClaudeCodeLog's changelog post says Remote Control, /schedule, MCP connectors, and notification preferences are disabled whenever ANTHROPIC_API_KEY, apiKeyHelper, or ANTHROPIC_AUTH_TOKEN is set, even if the user also has a Claude login.

That means the new long-run and multi-session story is cleaner for logged-in Claude users than for setups anchored on raw API credentials. The same changelog entry also says an auth deadlock was fixed, specifically a case where expired credentials plus forceRemoteSettingsRefresh could block claude auth login, logout, and status with no recovery path.

The quieter engineering work in 2.1.139 is about inspectability and attribution. ClaudeCodeLog's changelog post adds claude plugin details <name> with component inventory and projected per-session token cost, and ClaudeCodeLog's additional updates says /context now shows the providing plugin's name for plugin-sourced skills.

A few more plumbing changes stand out:

• MCP stdio servers now receive CLAUDE_PROJECT_DIR, matching hooks, per ClaudeCodeLog's changelog post.
• Hook commands can now use an args: string[] exec form, which avoids shell quoting issues, according to ClaudeCodeLog's changelog post.
• Subagent API requests now carry x-claude-code-agent-id and x-claude-code-parent-agent-id headers, and OTEL spans include matching agent_id attributes, per ClaudeCodeLog's changelog post.
• The CLI surface added CLAUDE_CODE_MAX_TURNS and CLAUDE_CODE_SUPERVISED, according to ClaudeCodeLog's additional updates.

Those are the kind of changes that make the new control-plane story legible when teams start stacking plugins, MCP servers, and nested agent runs in the same session graph.