[BUG] language setting ignored during auto-accept execution phase after plan mode clear context (regression in 2.1.71)

Preflight

• [x] I have searched for existing issues
• [x] This is a single bug report
• [x] I am using the latest version of Claude Code

What's Wrong?

After exiting plan mode with option 1 ("Yes, clear context and auto-accept edits"), Claude responds in English during the auto-accept execution phase, ignoring the language setting configured in settings.json. Once execution completes and Claude returns to interactive communication, it reverts to the configured language.

The language setting works correctly in all other contexts:

• Plan phase (plan agent, explore agent): Language is correct ✅
• Execution phase (auto-accept after clear context): Language switches to English ❌
• After execution completes: Language reverts to configured language ✅

What Should Happen?

Claude should respond in the configured language consistently throughout the entire execution phase, matching the behavior observed during the plan phase and post-execution communication.

Steps to Reproduce

1. Configure language in settings.json to any non-English language
2. Enter plan mode (Shift+Tab or /plan)
3. Discuss and finalize a plan with Claude — observe that language is correct at this stage, including Explore and Plan subagent outputs
4. When Claude calls ExitPlanMode, select option 1: "Yes, clear context and auto-accept edits"
5. Observe Claude's text output during the auto-accept execution phase — it switches to English
6. After execution completes and Claude communicates interactively again, it reverts to the configured language

Note: The plan file itself is written in the configured language (not English), so the issue is not caused by plan content language.

Is this a regression?

Yes. Version 2.1.69 and earlier did not exhibit this behavior.

Claude Code Version

2.1.71 (Claude Code)

Platform

Anthropic API

Operating System

Windows 11

Terminal/Shell

bash (Git Bash)

Claude Model

Opus

Last Working Version

2.1.69

Additional Information

The issue appears specific to the combination of:

1. Non-English language setting
2. Plan mode exit with clear context (option 1)
3. Auto-accept execution mode

During auto-accept execution, the model makes continuous tool calls with minimal user-facing text. The built-in system instructions (tool descriptions, agent specs, coding guidelines) are all in English. Without interactive user messages to reinforce the language preference, the language directive may lose priority against the larger volume of English system context.

It would be worth verifying whether option 3 ("Yes, auto-accept edits" without clearing context) also exhibits this issue, or if it is specific to the clear context path.