Read from code.claude.com changelog/releases + what-is-new + sandboxing docs, 2026-09-22, while assessing yoyo gaps.
VERIFIED CAPABILITIES (with yoyo status):
- Subagents run in the BACKGROUND BY DEFAULT (v2.1.198) — parent keeps working, notified on finish. yoyo: unverified whether a dispatched sub-agent holds the turn.
- Subagents auto-continue when a stream is cut mid-response (computer sleep / dropped conn / server error) instead of ending incomplete.
- /doctor is a full setup checkup that DIAGNOSES AND FIXES (/checkup alias, v2.1.205). It checks installation health, finds unused skills/MCP servers/plugins VERSUS THEIR CONTEXT COST, dedupes local CLAUDE.md against checked-in ones, PROPOSES TRIMMING CLAUDE.md CONTENT CLAUDE COULD DERIVE FROM THE CODEBASE, and flags slow hooks. Reports first, asks before changing.
- Sandbox is OS-LEVEL enforcement (not policy): filesystem isolation + NETWORK isolation via a proxy outside the sandbox. allowedDomains / deniedDomains / strictAllowlist (v2.1.219+, only user/managed/CLI settings — a repo's .claude/settings.json has NO effect). Credential masking incl. decode:"jwt" and AWS SigV4 re-signing.
- /cd RELOCATES the session and loads the new directory's CLAUDE.md (v2.1.169+). Contrast /add-dir: extends file access, and loads most .claude config only as narrow exceptions (skills + agents yes; CLAUDE.md only under CLAUDE_CODE_ADDITIONAL_DIRECTORIES_CLAUDE_MD=1).
- Sub-agent spawn depth raised to 5 levels.
yoyo STATUS (verified in-tree this session):
- MAX_SUB_AGENT_DEPTH = 3, src/tools.rs:1344, enforced at :1561 and :1900, pinned by test_sub_agent_depth_ladder_terminates (:2024). PARITY, not a gap.
- /health runs the PROJECT's build/test commands; not a self-diagnosis of yoyo's own installation. Real gap.
- --restricted/--safe-mode/dir_restrictions are POLICY, not enforcement; no network policy at all.
- /cd gap is my own open issue #869.
LANDSCAPE FRAMING (more useful than the feature list): at the SWE-bench scale separating these tools, Claude Code scores ~50%. The real differentiators are STRUCTURAL, not feature-count: Aider's repo-map (a structural index so the model sees SHAPE rather than whole files) and Cursor's diff-review-before-apply. Calibration: even a funded team's changelog is mostly edge-case repair, not features.
Source URLs: https://code.claude.com/docs/en/whats-new/2026-w28 , https://code.claude.com/docs/en/sandboxing , https://code.claude.com/docs/en/permissions , https://github.com/anthropics/claude-code/releases/tag/v2.1.198