Claude Code v2.1.220: getContextUsage() API, PermissionRequest hook, deep subagent forwarding

[claude-yolo]

WHY you should care

v2.1.220 meaningfully expands the Agent SDK surface area for anyone building production agents programmatically.

🚀 The big three

getContextUsage() SDK API — context window breakdown (token counts, categories, skills) is now reachable programmatically via Query.getContextUsage(). Previously interactive-only. Useful for adaptive context management in long-running agents.

PermissionRequest hook — hooks can now intercept permission prompts before they reach the user and auto-allow/deny them. Background subagents that can't prompt a user now auto-deny if no hook returns a decision. Critical for headless/CI workflows.

Deep subagent forwardSubagentText — messages from subagents at every nesting depth (not just depth-1) now propagate in the stream with parent_tool_use_id set. Fixes a real gap for multi-level agent trees.

Other notable additions

• cancel_queued interrupt: cancel queued messages outright instead of letting them run after interrupt
• fast_mode_disabled_reason: surface the specific reason fast mode is blocked, not just the state
• setModel(undefined): now documented to reset to session default
• supportedCommands() now reflects mid-session changes live
• settings.local.json auto-added to global git excludes when saved
• strictKnownMarketplaces admin setting for plugin marketplace lockdown
• Error retry semantics explicitly documented (TLS cert failures, mid-stream failures)

Ref

PR: thevibeworks/claude-code-docs#1108