claude-headless-mode（Claude headless mode 命令行）

Claude Headless Mode

Use claude -p (or --print) for non-interactive execution in scripts, CI, and automation.

Basic Usage

# Simple prompt
claude -p "Summarize this repo"

# With file input
claude -p "Review this code" < file.py

# Pipe output
claude -p "List all functions" | grep "def "

Output Formats

Text Output (Default)

claude -p "What is 2+2?"
# Output: 4

JSON Output

claude -p "What is 2+2?" --output-format json

Returns wrapper with metadata:

{
  "type": "result",
  "subtype": "success",
  "result": "4",
  "duration_ms": 1234,
  "total_cost_usd": 0.01,
  "session_id": "...",
  "structured_output": null
}

Extract just the result:

claude -p "What is 2+2?" --output-format json | jq -r '.result'

Stream JSON Output

claude -p "Long analysis" --output-format stream-json --verbose

• Streams newline-delimited JSON as execution progresses
• Shows tool calls, messages, progress in real-time
• Final line contains the result
• Requires --verbose flag

Extract final result:

claude -p "Analyze code" --output-format stream-json --verbose \
  | tee /dev/stderr \
  | tail -1 \
  | jq -r '.result'

Structured Output with JSON Schema

Force Claude to return data matching a specific schema:

claude -p "Extract function names from auth.py" \
  --output-format json \
  --json-schema '{
    "type": "object",
    "properties": {
      "functions": { "type": "array", "items": { "type": "string" } }
    },
    "required": ["functions"]
  }'

Result appears in structured_output field:

{
  "result": "...",
  "structured_output": {
    "functions": ["login", "logout", "verify_token"]
  }
}

Schema Examples

Simple object:

{
  "type": "object",
  "properties": {
    "answer": { "type": "integer" }
  },
  "required": ["answer"]
}

Task result (success/failure):

{
  "type": "object",
  "properties": {
    "status": { "type": "string", "enum": ["success", "failure"] },
    "summary": { "type": "string" },
    "files_changed": { "type": "array", "items": { "type": "string" } },
    "error_category": { "type": "string" },
    "suggestion": { "type": "string" }
  },
  "required": ["status", "summary"]
}

Extract from schema result:

OUTPUT=$(claude -p "$PROMPT" --output-format json --json-schema "$SCHEMA")
STATUS=$(echo "$OUTPUT" | jq -r '.structured_output.status')
SUMMARY=$(echo "$OUTPUT" | jq -r '.structured_output.summary')

Common Options

Permission Modes

# Skip all permission prompts (trusted environments only)
claude -p "Fix all linting errors" --permission-mode bypassPermissions

# Plan mode (read-only exploration)
claude -p "Analyze architecture" --permission-mode plan

Examples

CI/CD Integration

# Run tests and get structured result
RESULT=$(claude -p "Run tests and report results" \
  --output-format json \
  --json-schema '{"type":"object","properties":{"passed":{"type":"boolean"},"failures":{"type":"array","items":{"type":"string"}}},"required":["passed"]}')

if [ "$(echo $RESULT | jq '.structured_output.passed')" = "true" ]; then
  echo "Tests passed"
else
  echo "Tests failed"
  exit 1
fi

Batch Processing

for file in src/*.py; do
  claude -p "Review $file for security issues" \
    --output-format json \
    --json-schema '{"type":"object","properties":{"issues":{"type":"array"}},"required":["issues"]}' \
    | jq ".structured_output.issues"
done

Fresh Context Task Execution

# Each invocation is independent (no conversation history)
claude -p "Task 1: Create migration" --output-format json
claude -p "Task 2: Add model" --output-format json
claude -p "Task 3: Write tests" --output-format json

Task Coordination

Share TaskList across headless invocations using CLAUDE_CODE_TASK_LIST_ID.

Environment Variable

# All invocations share the same TaskList
export CLAUDE_CODE_TASK_LIST_ID="my-project"
claude -p "Use TaskCreate to add: Setup database"
claude -p "Use TaskCreate to add: Write migrations"
claude -p "Use TaskList to show all tasks"  # Shows both tasks

Pattern: Orchestrator + Workers

TASK_LIST_ID="epic-$(date +%Y%m%d)"
export CLAUDE_CODE_TASK_LIST_ID="$TASK_LIST_ID"

# Orchestrator creates tasks
claude -p "Use TaskCreate for each: Task 1, Task 2, Task 3"

# Workers execute (each sees shared TaskList)
for task_id in 1 2 3; do
  claude -p "Use TaskUpdate to mark task #$task_id as in_progress, implement it, then mark completed"
done

# Check final state
claude -p "Use TaskList to show status"

Task Tools

Dependencies

# Task 2 depends on Task 1
claude -p "Use TaskUpdate on task #2 to set blockedBy: [1]"

Tasks persist to ~/.claude/tasks/ and survive session restarts.

Notes

• Each -p invocation starts fresh (no context from previous runs)
• Use --output-format json for programmatic parsing
• structured_output field contains schema-validated data
• result field contains raw text response
• Stream JSON requires --verbose flag
• Cost info available in JSON output: total_cost_usd