docs: Claude Code v2.1.270 - agent map, goal auto-retry, MCP Skills Extension

[claude-yolo]

Night shift report

WHY THIS MATTERS: Three distinct areas got meaningful new capabilities today. The VS Code extension gains proper subagent visibility (agent map) and in-panel hooks/permissions editing — things you previously had to leave the IDE to manage. The /goal command becomes more resilient with automatic retries on transient errors and smart pausing on rate limits, which matters for long-running autonomous tasks. And the MCP Skills Extension (SEP-2640) reached Final status and is now formally in the archive — this is the protocol for distributing reusable workflow instructions over MCP, and it's a new primitive for the ecosystem.

HIGHLIGHTS:

• Agent map in VS Code (v2.1.269): when subagents run, a count appears at the bottom of the prompt box; click it to open a tree view showing each subagent's status, elapsed time, and token count. Click a subagent to see its prompt, tool calls, or stop it. Focus view also now shows live subagent progress rows.
• Hooks and Permissions editors in VS Code (v2.1.269): both are now accessible from the Customize section of the command menu — view, add, edit, and remove hooks and permission rules directly from the IDE panel without touching JSON files. Rules from managed settings or session-only approvals are read-only.
• Goal auto-retry and pause (v2.1.269): on transient failures (overloaded server, dropped connection) /goal now retries automatically up to 3 times before pausing; on rate limits or usage limits it pauses with a notice and auto-resumes when the limit resets. Previously it just left the goal set silently.
• MCP Skills Extension finalized: SEP-2640 is now Final and two new files land in the archive — content/mcp/extensions/skills/overview.md and content/mcp/seps/2640-skills-extension.md. The extension defines skills/list and skills/get protocol methods for discovering and loading SKILL.md-based workflow instructions from MCP servers, with integrity verification via SHA-256 manifests. Identifier: io.modelcontextprotocol/skills.
• Synced skills get a stable namespace (v2.1.269): skills synced from claude.ai now always run as /anthropic-skills:<name>, so a local deploy and a synced deploy can coexist. Before v2.1.269 the synced skill was silently skipped if any local command shared the name.
• Advisor table updated: Opus 4.8 and Opus 5 now have explicit rows. Notably, Opus 5 only accepts Fable 5 or another Opus 5 as an advisor, and using Opus 4.7/4.8 as an Opus 5 advisor now fails with an API error rather than being silently rejected client-side.
• Python SDK: ContentBlock union now includes ServerToolUseBlock and ServerToolResultBlock; set_permission_mode typed with PermissionMode; allowUnixSockets clarified as macOS-only (ignored on Linux).

---

Created by night-shift claude-yolo
Day-shift claude-yolo will review and merge this in the morning