docs: Claude Code v2.1.251 - agent_needs_input for agent teams, policyHelper failure docs, server-side auto mode toggle

[claude-yolo]

Night shift report

WHY THIS MATTERS: This update documents three things enterprise admins and hook authors need to know. The agent_needs_input hook expanded its trigger scope in v2.1.248 — anyone using it for notifications now gets alerts for agent team setup questions too. The policyHelper reference finally has a formal "Helper failures" section spelling out exactly when the helper fails and why Claude Code refuses to start — critical for enterprise rollouts where a misconfigured policy helper silently blocks everyone. And auto mode now has a documented server-side kill switch: Anthropic can disable it per-account, and a session that gets that answer stays in Manual until restarted — which explains user-facing "auto mode unavailable" reports that no local configuration change was causing.

HIGHLIGHTS:

• Hooks (v2.1.248): agent_needs_input now fires when the current session asks an agent-team teammate's terminal setup question and you haven't typed for ~6 seconds — not just for background sessions in agent view. Hook authors may want to update matchers/handlers.
• policyHelper "Helper failures" section (settings-reference.md): New section enumerates all failure conditions — path rules, file not found, non-zero exit, timeout, >1 MiB output, schema violations — and documents that startup failures print stderr and block Claude Code from starting. Background refresh failures keep the last good policy instead of blocking.
• Auto mode server-side toggle: permission-modes.md now says Anthropic may turn auto mode off server-side and a session that received that answer keeps auto mode off until the session ends. Previously the docs implied "unavailable" was always a local config issue; now users know to start a new session and try later.
• Bedrock proxy fallback: Clarified that a proxy that drops Content-Type but passes the binary body through still works; one that also re-emits as SSE triggers silent fallback to non-streaming (no error, just slow). New advice: if your proxy converts to Anthropic Messages API format, use ANTHROPIC_BASE_URL as an LLM gateway instead of CLAUDE_CODE_USE_BEDROCK.
• errors.md: New: Claude Code doesn't treat every /-prefixed prompt as a command — /-- (Lean doc comment) and /var/log/syslog (paths) go to Claude as plain messages. Also: policy-disabled commands now return their own message rather than the generic "unknown command."
• Sandbox write paths: sandbox.filesystem and allowWrite docs updated to mention permissions.additionalDirectories as a third source of default writable directories alongside --add-dir and /add-dir.
• Quickstart: Duplicate theme={null} attributes on code blocks — likely an upstream rendering artifact, not user-visible.

---

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