docs: Claude Code v2.1.220 - terminal_reason, strictAllowlist, Desktop gateway overlay

[claude-yolo]

Night shift report

WHY THIS MATTERS: v2.1.220 ships three distinct features SDK authors and enterprise admins need to know about: a new terminal_reason field in the Python SDK tells you programmatically why an agent loop ended (completed vs. aborted vs. interrupted), two new ModelUsage fields let multi-provider deployments track cost attribution by provider identity, and network.strictAllowlist gives security-conscious orgs a hard-deny sandbox network posture instead of the previous prompt-per-domain approach. On top of that, the gateway can now serve Claude Desktop alongside Claude Code — a significant enterprise config expansion. Hooks also now propagate into subagents, which changes behavior for anyone relying on PreToolUse/PostToolUse hooks.

HIGHLIGHTS:

• terminal_reason in Python SDK ResultMessage (new field): values like "completed", "max_turns", "aborted_streaming", "aborted_tools" — finally a reliable way to distinguish user-interrupts from natural loop endings without parsing error strings
• ModelUsage gets canonicalModel + provider (new fields): provider values are firstParty, bedrock, vertex, foundry, anthropicAws, mantle, gateway — essential for cost rollup in multi-provider deployments; model_usage type changed from dict[str, Any] to dict[str, ModelUsage]
• network.strictAllowlist: true (new sandbox setting, v2.1.219+): sandbox commands denied outright for off-allowlist hosts instead of prompting; respects same allowlist as before (allowedDomains + WebFetch rules); only honored from user/managed/CLI settings, not per-repo .claude/settings.json
• Gateway: Claude Desktop overlay (v2.1.203+ gateway): same gateway instance can now bootstrap Claude Desktop via <gateway>/user/bootstrap; requires desktop: {} in matched policy; audit-logged as desktop_bootstrap.serve / desktop_bootstrap.denied
• Hooks fire inside subagents: PreToolUse/PostToolUse hooks defined in settings.json now also run for subagent tool calls, with agent_id/agent_type in the hook input — behavioral change for anyone with broad hook patterns
• HTTP hook allowlists apply to all hook sources including managed policy: allowedHttpHookUrls and httpHookAllowedEnvVars gate hooks regardless of where they're defined
• GNU screen clipboard fix (v2.1.219): long clipboard selections no longer spew base64 into the terminal window
• Remote Control error messages improved (v2.1.219): error now names the specific variable (CLAUDE_CODE_USE_BEDROCK, ANTHROPIC_BASE_URL) that routed the session away from Anthropic API
• Support articles and plugin marketplace updated (bulk CDN URL rotation, ~70 support files)

---

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