Claude Code 2.1.139: Agent View, Goal Commands, MCP Fixes

Claude Code 2.1.139 ships agent view for session management, /goal command for multi-turn objectives, and 30+ fixes. Major improvements to MCP handling, credential auth, and UI polish across scrolling, plugins, and transcript navigation.

Claude Code 2.1.139: Agent View, Goal Commands, MCP Fixes

TL;DR

  • Agent view lists all Claude Code sessions in one place — run claude agents
  • /goal command sets completion conditions; Claude keeps working until met
  • Major MCP and credential handling fixes, plus 30+ bug fixes across UI, scrolling, and plugin management

New

  • Agent view (Research Preview) — single list of every session (running, blocked, done); run claude agents to access
  • /goal command — set a completion condition and Claude continues across turns until met; shows live elapsed time, turn count, and token usage overlay
  • /scroll-speed command — tune mouse wheel scroll speed with live preview
  • claude plugin details — shows plugin component inventory and projected per-session token cost
  • Transcript view navigation? for shortcuts, {/} to jump between prompts, v to toggle panel
  • Hook args field (exec form) — spawns commands directly without shell, eliminating quoting issues for path placeholders
  • Hook continueOnBlock option — set to true to feed rejection reasons back to Claude and continue the turn
  • MCP stdio servers now receive CLAUDE_PROJECT_DIR — matches hooks; plugin configs can reference ${CLAUDE_PROJECT_DIR}
  • Compaction prompt improvement — now asks model to preserve sensitive user instructions
  • /mcp Reconnect enhancements — picks up .mcp.json edits without restart; shows HTTP status and URL on reconnect failures
  • /context all token estimates — now account for model's tokenizer and show rounded values
  • Plugin install auto-refreshclaude plugin install <name>@<marketplace> now auto-refreshes marketplace and retries before reporting not found
  • Remote MCP server retry — transient failure retry now enabled for all users
  • Subagent API headers — requests now carry x-claude-code-agent-id and x-claude-code-parent-agent-id; OTEL spans include agent attributes
  • VSCode shortcut — Cmd/Ctrl+Shift+T reopens most recently closed session tab (configurable)

Fixed

  • Credential deadlock blocking claude auth login/logout/status when expired credentials and forceRemoteSettingsRefresh policy were set
  • autoAllowBashIfSandboxed not auto-approving commands with shell expansions like $VAR and $(cmd)
  • Hook writing to terminal corrupting on-screen interactive prompts; hooks now run without terminal access
  • Unbounded memory growth from HTTP/SSE MCP servers streaming non-protocol data — response bodies now capped at 16 MB per frame
  • Skill(name *) permission rules now work as prefix match, matching Bash(ls *) behavior
  • Settings hot-reload not detecting edits to symlinked ~/.claude/settings.json
  • Plugin details failing when marketplace key differs from manifest name
  • /model pickerSource: Claude Code