Summary
The Week 29 digest (July 13–17, 2026, versions v2.1.207–212) introduced eight major features and eight bug‑fix classes. Key additions include live‑per‑viewer MCP connector calls from published artifacts, a screen‑reader‑optimized mode, improved session forking and subagent management, configurable session budgets, and new safety tools. The fix‑log address glob‑rule anchoring, permission oversight on bash redirects, long‑command analysis, plan‑mode bash execution, symlink‑following file creation, and remote prompt confirmation.
Key Points
- Published artifacts can now invoke MCP connectors live based on each viewer’s own configured connectors.
- A screen‑reader mode (
--ax-screen-reader) replaces the TUI with linear, labeled text for VoiceOver / NVDA. /forkcopies a conversation into a background session; the old in‑session subagent behavior is renamed/subtask.- Session‑wide WebSearch cap defaults to 200 (
CLAUDE_CODE_MAX_WEB_SEARCHES_PER_SESSION); subagent spawn cap also 200./clearresets budgets. - MCP tool calls exceeding 2 minutes automatically move to background (
CLAUDE_CODE_MCP_AUTO_BACKGROUND_MS). /resumeshows a picker of past sessions, including deleted ones.- New
auto‑mode resetcommand andEndConversationtool for terminating abusive sessions. - Eight bug classes were fixed, covering glob anchoring, bash redirect parsing, long‑command analysis, zsh variable subscripts, help/man smuggling, plan‑mode bash execution, symlink‑following file creation, and remote prompt confirmation.
Concepts
- MCP Connector: A mechanism that allows artifacts to call external tools or services; now per‑viewer in published artifacts.
- Screen Reader Mode: A TUI alternative that outputs text linearly with labels, enabling assistive technology compatibility.
- /fork: Splits a conversation into a new background session, leaving the original intact.
- Session Budgets: Configurable caps (e.g., WebSearch count, subagent spawns) to prevent runaway usage; reset via .