Pipe Mode

Overview

Pipe Mode activates automatically when Autohand detects piped input (non-TTY stdin). It reads the entire stdin stream, processes it with the LLM, and outputs structured results to stdout.

This makes Autohand composable with other Unix tools:

# Pipe a file for analysis
cat src/auth.ts | autohand -p "find bugs in this code"

# Chain with other tools
git diff HEAD~1 | autohand -p "summarize these changes" | jq '.response'

I/O behavior

Because stdout contains only structured data, you can safely redirect or pipe it without capturing progress noise.

NDJSON output

Output uses Newline-Delimited JSON (NDJSON) format. Each line is a valid JSON object:

{"type":"message","content":"The authentication module has..."}
{"type":"tool_use","tool":"read_file","args":{"path":"src/auth.ts"}}
{"type":"tool_result","tool":"read_file","output":"..."}
{"type":"message","content":"I found 3 potential issues..."}

Common message types:

• message — Agent text responses
• tool_use — Tool invocations with arguments
• tool_result — Tool execution results
• error — Error messages
• status — Session status updates

Usage examples

Code review pipeline

# Review staged changes
git diff --staged | autohand -p "review this diff for bugs"

# Review and extract just the response text
git diff --staged | autohand -p "review this diff" | \
  jq -r 'select(.type=="message") | .content'

Batch file processing

# Add type annotations to all TypeScript files
for f in src/**/*.ts; do
  cat "$f" | autohand -p "add missing type annotations" \
    --output "$f"
done

CI integration

# Generate PR description from commits
git log main..HEAD --oneline | \
  autohand -p "write a PR description for these commits" | \
  jq -r 'select(.type=="message") | .content' > pr-body.md

Combining with flags

# Use a specific model and generate a patch
cat buggy.ts | autohand -p "fix the bug" \
  --model anthropic/claude-4-sonnet \
  --patch --output fix.patch

Tips

• Use jq to filter NDJSON output by message type
• Combine with --model to use faster models for batch tasks
• Use --dry-run in pipe mode to preview what the agent would do
• Pipe mode respects --restricted and --unrestricted flags
• Set --timeout to prevent runaway executions in CI