[BUG] `claude -p` (print mode) returns empty output in v2.1.83 — regression from v2.1.81

Environment

• Claude Code version: 2.1.83 (broken), 2.1.81 (works), 2.1.79 (works)
• OS: macOS Darwin 25.3.0 (Apple Silicon, M-series)
• Shell: zsh (oh-my-zsh)
• Node: installed via Homebrew (/opt/homebrew/bin)
• Auth: OAuth (claude.ai), Team subscription
• Install method: npm install -g @anthropic-ai/claude-code

Description

After upgrading from v2.1.81 to v2.1.83, claude -p (print/headless mode) returns empty stdout for all prompts — including trivially short ones. Interactive mode continues to work normally. Downgrading to v2.1.81 immediately resolves the issue.

Reproduction Steps

# Upgrade to 2.1.83
npm install -g @anthropic-ai/claude-code@2.1.83

# All of these return empty output:
claude -p "say hi"
claude -p "say hi" < /dev/null
echo "say hi" | claude -p
claude -p --model sonnet "say hi"

# Debug flag reveals error on stderr, but exit code is still 0:
claude -p --debug "say hi" 2>debug.log
# => Exit code: 1
# => stderr: "Error: Input must be provided either through stdin or as a prompt argument when using --print"

# Even piping (which should satisfy stdin requirement) produces empty output:
echo "say hi" | claude -p
# => (empty, exit code 0)

# --bare mode shows auth issue:
claude --bare -p "your prompt here"
# => "Not logged in - Please run /login"

# Downgrade fixes it immediately:
npm install -g @anthropic-ai/claude-code@2.1.81
claude -p "who are you"
# => Normal response returned

Observed Behavior

| Method | v2.1.81 | v2.1.83 |
|--------|---------|---------|
| claude -p "prompt" | Normal output | Empty, exit 0 |
| echo "prompt" \| claude -p | Normal output | Empty, exit 0 |
| claude -p --debug "prompt" | Normal output | "Error: Input must be provided either through stdin or as a prompt argument", exit 1 |
| claude --bare -p "prompt" | N/A | "Not logged in" |
| Interactive claude | Works | Works |

Impact

This is a breaking change for all claude -p automation pipelines. In our case, it silently broke a scheduled report generation system (launchd cron jobs calling claude -p --model sonnet --mcp-config ...):

• Reports generated as 0-1 line empty files
• Exit code 0 masked the failure — no error handling triggered
• Empty PDFs were generated and uploaded to Slack
• Multiple retry attempts all produced the same empty result
• Only discovered when end users reported blank reports

Key Observations

1. Not a prompt size issue — even "say hi" (6 chars) fails
2. Not an auth issue — claude auth status returns loggedIn: true, and interactive mode works fine with the same session
3. Not a stdin issue — both positional argument and stdin pipe fail
4. Exit code is misleading — returns 0 (success) despite producing no output, except when --debug is enabled (then returns 1 with error message)
5. Regression is narrow — v2.1.81 works, v2.1.83 broken (v2.1.82 was not published)

Expected Behavior

claude -p "prompt" should return the model's response on stdout, as it does in v2.1.81 and earlier.

Workaround

Downgrade to v2.1.81:

npm install -g @anthropic-ai/claude-code@2.1.81