Claude Code releases 2.1.200/2.1.201 with Manual approval fixes

• ClaudeCodeLog's 2.1.200 overview says Claude Code changed its default permission mode to Manual across the CLI, --help, VS Code, and JetBrains, with commands requiring manual approval.
• The 2.1.200 changelog is mostly reliability plumbing: background sessions, stale daemon locks, daemon handover, plugin loading, screen readers, voice dictation, and tmux rendering all got fixes.
• ClaudeCodeLog's 2.1.200 diff reported prompt tokens up 4,817 (+8.3%), prompt files down 3, and a Grep tool description that now explicitly says it is ripgrep-based.
• The 2.1.201 release note followed with one CLI change for Sonnet 5: sessions no longer use the mid-conversation system role for harness reminders.
• ClaudeCodeLog's 2.1.201 metadata put the second update 16 hours after 2.1.200 and measured another 10,607 prompt tokens (+16.8%).

Anthropic's official changelog has both entries, the permission-mode docs spell out Manual as the read-only-by-default mode, and the settings docs put askUserQuestionTimeout at "never" for 2.1.200. In a Hacker News thread, commenters separated AskUserQuestion from action authorization: the timer applied to follow-up planning choices, not command approval.

The headline change was a safer default in the harness. The full 2.1.200 changelog says the old default permission mode is now Manual across CLI, --help, VS Code, and JetBrains.

The mechanics are narrow:

• Surfaces changed: CLI, --help, VS Code, JetBrains.
• Accepted forms: --permission-mode manual and "defaultMode": "manual".
• Behavior: according to Anthropic's permission-mode docs, Manual reviews edits, shell commands, and network requests while reads run without asking.
• Compatibility: Anthropic's permissions docs say default remains the standard mode, labeled Manual in 2.1.200 and later.

2.1.200 also changed unanswered AskUserQuestion dialogs so they no longer auto-continue by default. Anthropic's settings docs define the new askUserQuestionTimeout key with default "never", allowed values of "60s", "5m", "10m", and "never", and a /config label of Question auto-continue timeout.

That setting is user-scoped. The docs say it is not read from project or local settings, which keeps a repository from silently changing how long a personal prompt can wait.

The 2.1.198 note is useful context: background agents had just gained auto-commit, push, and draft PR creation after finishing code. The 2.1.200 changelog reads like a cleanup pass for the failure cases that show up once agents run longer:

• Background sessions silently stopping mid-turn after sleep, wake, or reopening a stalled session.
• Cancelled turns re-running after a stall respawn.
• Stale daemon.lock files blocking future background agents after PID reuse.
• Older builds taking over the daemon after reinstall, now blocked by embedded build timestamp checks.
• Roster corruption disabling orphan cleanup, newer fields being dropped by older binaries, and socket auth tokens being stripped during daemon restarts.
• Subagents cut off by rate limits returning empty results instead of failing cleanly.
• Control bytes from background-agent output leaking into the terminal agent view.

The small release had a measurable prompt-side shift. ClaudeCodeLog's diff counted these 2.1.200 changes since 2.1.199:

• Removed env var: CLAUDE_CODE_CLIENT_KEY_PASSPHRASE.
• Bundle file size: +130.8 kB (+0.4%).
• Prompt files: -3 (-7.9%).
• Prompt tokens: +4,817 (+8.3%).
• System-reminder token share: 48.3% to 53.9%.
• Tool token share: 46.8% to 41.6%.
• Grep description: now identifies Grep as ripgrep-based and explicitly recommends it for search tasks.

Christmas come early for people who diff agent harness prompts for sport.

2.1.201 was a one-line behavioral change for Sonnet 5. The 2.1.201 changelog item says Claude Sonnet 5 sessions no longer use the mid-conversation system role for harness reminders.

The 2.1.201 metadata diff measured the release as small in bundle bytes but large in prompt text:

• Time since 2.1.200: 16h 10m 26s.
• Bundle file size: -1.2 kB.
• Prompt files: +3 (+8.6%).
• Prompt tokens: +10,607 (+16.8%).
• System-reminder token share: 53.9% to 54.6%.

One community read came from nummanali, who called system reminders one of the traits that made Anthropic models feel tied to Claude Code and said the change made Sonnet 5 worth trying in Pi with a modified system prompt.

The same 2.1.200 changelog also shipped user-facing polish outside the headline safety changes:

• /mcp server list now tracks focus for screen readers and magnifiers.
• Decorative glyphs are hidden from screen readers.
• Transcript symbols now read as short labels.
• Nested tables now read as Header: value. lines.
• tmux 3.4+ rendering flicker is fixed through synchronized terminal output.
• Voice dictation no longer shows a misleading Voice connection failed message when a recording captures no audio.
• The install script now explains when installation is killed because the system ran out of memory.