docs: Claude Code v2.1.238 - classifierContext hooks, keybindingFlavor, cross-session safety

[claude-yolo]

Night shift report

WHY THIS MATTERS: Three new developer-facing features shipped with fresh documentation, plus behavioral clarifications for auto mode and cross-session messaging that change how agents handle edge cases. Developers embedding Claude Code via the Agent SDK, writing hooks, or customizing keyboard behavior need to know about these now.

HIGHLIGHTS:

• classifierContext in PostToolUse hooks (v2.1.236+): hooks can now annotate a tool call's result for the auto mode permission classifier without touching what Claude sees. In-process TypeScript SDK callbacks get elevated trust — a user statement relayed in the note can satisfy a classifier consent check. Shell hooks get "application-provided context" trust. Hard limits: 2000 chars shared across all hooks on the same call, synchronous responses only, ignored for read-only tool calls. New classifierContext field documented in hooks.md, hooks reference table, and agent-sdk/hooks.md and typescript.md.
• keybindingFlavor setting (v2.1.238+): new ~/.claude/settings.json option switches Ctrl+W from "delete previous word stopping at punctuation" (classic) to "delete back to whitespace" (readline/bash-style). Useful for path editing. Docs in interactive-mode.md.
• Cross-session messaging safety checks (v2.1.236+): new error section documents three new refusal reasons — symlinked socket paths, unreadable endpoints, and process-identity mismatches. Also documents rate-limiting of rapid bursts to a session inbox. Previously these silently dropped; now sending session gets an explicit refusal in the tool result.
• Auto mode behavioral change (v2.1.236): Monitor allow rules are now dropped in auto mode (restored on exit), fixing a gap where a broad Monitor rule could approve shell commands without classifier review. Pre-v2.1.236 behavior was to leave them in effect.
• Auto mode classifier scope clarified: docs now explicitly state the classifier omits read-only tool results (file reads, searches) and that git status is injected automatically before destructive git commands, even when status.showUntrackedFiles=no.
• Agent SDK additionalDirectories clarified: now documented that --add-dir also loads skills, commands, and subagents from each added directory when the project setting source is active. Settings source table updated to reflect commands and subagents alongside skills.
• Interactive-mode keyboard shortcuts expanded: new table covers Ctrl+A/E/K/U/W/Y, Alt+Y, Alt+B/F, and Ctrl+_ with platform-specific notes (macOS iTerm2 Cmd+Backspace, Windows Ctrl+Backspace).
• dialogExpiry behavior for consent prompts: new docs explain what happens when Fable 5 usage-credit consent goes unanswered in Remote Control, background, or agent team sessions — turn ends without sending, model unchanged, prompts again on next message.
• Support articles: routine freshness updates (no signal).

---

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