Summary
Claude Code v2.1.198 to v2.1.201 (early July 2026) introduced twelve new features spanning shell execution, session recovery, browser integration, background agent workflows, persistent memory, sub‑agent recursion, permission management, MCP authentication, parameter‑based permission rules, inline configuration, directory changes without cache rebuild, and shareable session artifacts. The release also achieved parity with two previously attempted‑and‑reverted features from the competing tool yoyo (issues #563, #564), and matched yoyo spawn --pr with automated commit‑push‑PR on background agent completion.
Key Points
- Shell Mode (
! npm test): run commands directly from the prompt and let the model respond to output without a second prompt. /rewind: resume a conversation from before/clearwas run.- Chrome integration (GA): read console errors + live DOM, drive local dev servers, record GIFs.
- Background agent hook notifications:
agent_needs_input,agent_completed— plus automatic commit‑push‑PR on agent completion. - Auto Memory: learned context persists across workflows automatically.
- Sub‑agent nesting: sub‑agents can spawn further sub‑agents, capped at 5 levels.
- Permission mode renamed: default permission mode is now “Manual”.
claude mcp login/logout: shell‑based MCP authentication.Tool(param:value)permission rules: e.g.Agent(model:opus)./config key=value: settable from prompt and-pmode./cd: change working directory mid‑session without rebuilding prompt cache.- Artifacts beta: session output becomes a live shareable page.
- Competitive significance: Items 1 (shell mode) and 2 () are direct parity targets with (issues 563, 564), now validated by the market. Item 4 mirrors .