Claude docs changes for October 19th, 2026 [diff]

Executive Summary

  • The hooks reference and guide were reorganized to explain more clearly how exit codes, malformed JSON output, timeouts and HTTP hook failures are handled, and how to make a hook block on failure with onFailure: "block".
  • The sandbox runtime instructions now install @anthropic-ai/sandbox-runtime globally and launch with srt claude instead of npx.
  • Claude Code 2.1.303 was released with only minor bug fixes and reliability improvements.

New Claude Code versions

2.1.303

Major bug fixes

  • Only general bug fixes and reliability improvements (no notable changes)

Claude Code changes

Changed documents

env-vars [Source]

  • Clarified that setting CLAUDE_CODE_ENABLE_PROMPT_SUGGESTION to true keeps the near-usage-limit pause from applying until the limit is reached. [line 287] [Source]

hooks-guide [Source]

  • Exit code section rewritten: non-0/2 exit codes without valid JSON let the action proceed unless the hook is one whose failure blocks the action (for example onFailure: "block"). Exit 2 and "any other exit code" behavior was reorganized by what the hook printed to stdout. [lines 586-595] [Source]
  • PreToolUse permissionDecision explanation condensed into prose: "ask" shows the normal prompt and "allow" skips it, with deny/ask rules, network-path reads, requiresUserInteraction MCP tools and org-ask connector tools still applying. [line 615] [Source]
  • Constraints section restructured into "what still holds over allow" and "deny in every permission mode", which still blocks even with bypassPermissions. [line 944] [Source]

hooks [Source]

  • Exit 0 with unparsable or schema-invalid JSON is now explicitly a non-blocking error. Stdout parsing rules were reworded, with advice to print only a single JSON object. [lines 782-793] [Source]
  • Troubleshooting guidance for the Failed with non-blocking status code: notice and for hooks that never ran because of a missing script path. [lines 829-845] [Source]
  • Timeout behavior consolidated: a PreModelSwitch hook canceled at timeout blocks the switch, while timed-out PreToolUse command/http/mcp_tool hooks don't block unless onFailure: "block" is set. [line 851] [Source]
  • HTTP hooks: onFailure: "block" and 2xx JSON decision fields are now introduced before the outcome table. [line 991] [Source]
  • PreModelSwitch: other exit codes are a non-blocking error so the switch goes ahead, except for hooks whose failure blocks it. [line 3420] [Source]

interactive-mode [Source]

  • Reworded the usage-limit prompt suggestion skip: the env var keeps the skip from applying until the limit is reached. [line 412] [Source]

sandbox-environments [Source]

  • Replaced npx @anthropic-ai/sandbox-runtime claude with a global install (under a prefix outside any writable allowed path) followed by srt claude from the project directory. [lines 102-114] [Source]

API changes

No API documentation changes today.