Claude Code 2.1.200: Manual Mode Is Now the Default — What Breaks and How to Fix It

Three Claude Code releases landed between July 2 and July 4, 2026. Two are incremental. One — 2.1.200 — contains a breaking change that will silently stall any unattended pipeline running on the old defaults. Here is what happened and how to adapt.

The Breaking Change: Manual Mode Is Now the Default

Before 2.1.200, Claude Code shipped in what was effectively an “allow” default: tool calls and shell commands ran without user confirmation unless you explicitly set a stricter mode.

Starting with 2.1.200, the default is Manual — every tool call now pauses and waits for your explicit “Allow” before executing.

This affects:

• claude in the CLI
• The VS Code extension
• The JetBrains extension
• Anything that passes --permission-mode default (which is now aliased to manual)

Who is affected most: background agents running in automated pipelines or via cron where no human is watching the terminal. Those sessions will now stall silently at the first tool call instead of proceeding.

To restore the previous behavior

In your project’s .claude.json or global config, set:

{
  "defaultMode": "acceptEdits"
}

Or pass it on the command line:

claude --permission-mode acceptEdits

The accepted values are manual (new default), acceptEdits (auto-approve file edits only), and bypassPermissions (allow everything, requires elevated trust flag).

If you run Claude Code via a script or CI job, add --permission-mode acceptEdits or --permission-mode bypassPermissions (only if the environment is already hardened) to your invocation to avoid the stall.

The Second Breaking Change: AskUserQuestion No Longer Auto-Continues

Also in 2.1.200: AskUserQuestion dialogs no longer auto-continue after an idle timeout. Previously, if a background agent asked the user a question and nobody answered, it would eventually self-continue and pick a default path.

Now it stalls indefinitely until you respond.

Why this matters for builders: any prompt in your agent’s skill chain that calls AskUserQuestion — including many third-party skills — will block the pipeline until a human responds. If you had relied on the old idle-timeout behavior (e.g., “if nobody answers in 30 seconds, proceed with X”), that behavior is gone.

To opt back in: set an idle timeout in /config:

/config → idle-timeout → 30

This re-enables auto-continuation after 30 seconds of no input. The value is in seconds.

What Fixed in 2.1.199 (Released July 2–3)

2.1.199 was a 24-change reliability release focused on streaming and background agent stability. Key fixes:

Streaming / API errors:

• Partial streaming responses are now preserved when the API emits a mid-stream overload or server error, rather than being discarded. You now receive the partial output with an “incomplete response” notice instead of nothing.
• Subagents cut off by a rate limit or server error now return partial work to the parent instead of silently failing.
• Subagents that hit a usage limit now report the error to the parent agent rather than reporting success.

Linux daemon crash:

• A Linux-specific bug where the background-agent daemon killed itself (and all running agents) roughly every 50 seconds after an unclean shutdown has been fixed. The root cause was a corrupted worker record from a prior crash.

SSL / proxy errors:

• SSL certificate errors — common behind TLS-inspecting proxies or when NODE_EXTRA_CA_CERTS is not set — previously burned all retries before surfacing a useful error message. Now they fail immediately with a clear fix hint.

IDE and terminal:

• JetBrains terminals (IntelliJ, PyCharm, WebStorm) no longer flicker.
• Shift+non-ASCII characters (e.g., Shift+ä → Ä) are no longer dropped in terminals using the Kitty keyboard protocol (WezTerm, Ghostty, kitty).

Skills:

• Stacked slash-skill invocations — \skill-a \skill-b do XYZ — now load all leading skills (up to 5). Previously only the first skill loaded.

What Fixed in 2.1.200 (Beyond the Breaking Changes)

Alongside the permission defaults, 2.1.200 also shipped a batch of background-session stability fixes:

• Sleep/wake crash: Background sessions that silently stopped mid-turn after a macOS or Linux sleep/wake cycle are now stable through suspend-resume.
• Daemon lock file: Daemon crashes that left a stale daemon.lock with a reused PID are fixed. Previously this prevented the daemon from restarting until the lock file was manually removed.
• MCP server config crash: Starting Claude Code with a non-array value for disabledMcpServers or enabledMcpServers in .claude.json no longer crashes at startup.
• Git worktree plugins: Project-scoped plugins now load correctly when working inside a git worktree.
• Control bytes: Control bytes from background-agent output no longer reach the terminal (they caused garbled display in some terminal emulators).
• Voice dictation: “Voice connection failed” false positives are fixed.
• tmux flicker: Rendering flicker under tmux 3.4+ is resolved.
• Accessibility: Decorative glyphs are now hidden from screen readers; transcript symbols are read as short labels; nested tables are read as “Header: value” lines.

What Fixed in 2.1.201 (July 4)

A single targeted fix: Claude Sonnet 5 sessions no longer receive mid-conversation system roles for harness reminders. This corrects an inconsistency in how the harness injected internal instructions during long Sonnet 5 sessions, which could subtly affect behavior partway through a conversation.

Summary: What You Need to Do Right Now

The 2.1.199 stability improvements (streaming, Linux daemon, SSL) apply automatically — no config changes required to benefit from them.

Claude Code 2.1.199 shipped July 2–3, 2026. Claude Code 2.1.200 shipped July 3, 2026. Claude Code 2.1.201 shipped July 4, 2026. Version numbers are visible via claude --version.