claude -p Skill

A reference for using claude -p (non-interactive print mode) to run Claude Code sessions programmatically — particularly for multi-turn conversations in server-side code.

This is the direct CLI approach, not the Claude Agent SDK. The key binary is claude, authenticated via claude auth login (subscription auth).

Core invocation

claude -p "your prompt" \
  --output-format stream-json \
  --verbose \
  --permission-mode acceptEdits \
  --model sonnet \
  --session-id <uuid>     # first turn only — see below

--output-format stream-json --verbose is required together: stream-json alone is rejected in -p mode without --verbose.

Auth: subscription, not API key

const env = { ...process.env };
delete env.ANTHROPIC_API_KEY;      // forces subscription auth
delete env.ANTHROPIC_AUTH_TOKEN;   // forces subscription auth
// Claude reads <CLAUDE_HOME>/ credentials automatically

Check auth status:

claude auth status --json           # correct flag
# NOT: --output-format json         # wrong, will error

Multi-turn sessions — the critical distinction

Claude Code uses a per-session-id file lock. The lock persists for several seconds (sometimes indefinitely) after the subprocess exits. This is the single most common source of failures.

Rule: first turn vs follow-up turns

--session-id on a follow-up always fails with "Session ID already in use" — even seconds after the first process exited.

const sessionFlag = isFirstTurn
  ? ['--session-id', claudeSessionId]
  : ['--resume', claudeSessionId];

spawn('claude', ['-p', prompt, '--output-format', 'stream-json',
  '--verbose', '--permission-mode', 'acceptEdits',
  '--model', model, ...sessionFlag], { ... });

Track "has had a first turn" per session:

private sessionsWithHistory = new Set<string>();
// After successful first turn:
sessionsWithHistory.add(sessionId);
// On subsequent turns: isFirstTurn = !sessionsWithHistory.has(sessionId)

Also check entry.messageCount > 0 from persistent storage as a fallback for server restarts.

Capture the confirmed session ID

Claude may assign a slightly different session_id than the one you passed. Always capture it from the system event output:

// In the stream-json system event:
// {"type":"system","session_id":"<actual-uuid>",...}
// Use THIS id for --resume, not necessarily what you passed to --session-id

Stdin must be closed

Always use stdio: ['ignore', 'pipe', 'pipe']. If stdin stays open, Claude waits up to 3 seconds for piped input and warns:

"no stdin data received in 3s, proceeding without it"

This also means the process lingers and the session lock is held longer.

Subprocess spawn pattern

const proc = spawn('claude', [
  '-p', prompt,
  '--output-format', 'stream-json',
  '--verbose',
  '--permission-mode', 'acceptEdits',
  '--model', model,               // 'sonnet' | 'opus' | 'haiku'
  ...sessionFlag,
], {
  cwd: workingDirectory,
  env: claudeEnv,                 // with API keys stripped
  stdio: ['ignore', 'pipe', 'pipe'],
});

Stream stdout line-by-line with readline:

const rl = createInterface({ input: proc.stdout, crlfDelay: Infinity });
rl.on('line', (line) => {
  const events = normalizer.normalize(line, sessionId);
  events.forEach(onEvent);
});

Emit agent_end on process exit, not on receiving the result event:

proc.on('exit', (code, signal) => {
  rl.close();
  activeProcesses.delete(sessionId);
  if (code !== 0 && signal !== 'SIGTERM') {
    onComplete(new Error(`claude exited ${code}`));
  } else {
    onEvent({ type: 'agent_end', ... });
    onComplete();
  }
});

If you emit agent_end when the result event arrives (before process exit), the UI becomes interactive again while the session lock is still held — the next prompt will fail.

stream-json event sequence

{"type":"system","subtype":"init","session_id":"<uuid>","tools":[...],"model":"claude-sonnet-4-6",...}
{"type":"assistant","message":{"content":[{"type":"tool_use","id":"toolu_...","name":"Read","input":{...}}],...},...}
{"type":"rate_limit_event","rate_limit_info":{"status":"allowed","rateLimitType":"five_hour","isUsingOverage":false,...},...}
{"type":"user","message":{"content":[{"type":"tool_result","content":"...","tool_use_id":"toolu_...","is_error":false}]},...}
{"type":"assistant","message":{"content":[{"type":"text","text":"The file contains..."}],...},...}
{"type":"result","subtype":"success","result":"...","session_id":"<uuid>","usage":{...},"total_cost_usd":0.05,...}

Token usage IS available

The result event exposes full token data — don't return zeros:

// result event data:
{
  usage: {
    input_tokens: 3,
    cache_creation_input_tokens: 12413,
    cache_read_input_tokens: 11171,
    output_tokens: 54
  },
  total_cost_usd: 0.05
}

Event → NormalizedEvent mapping

Session concurrency

Each session UUID is independent — multiple sessions can run simultaneously without interfering. The lock is per-UUID.

Within a single session, only one turn can run at a time. Queue follow-up prompts server-side:

if (isRunning(sessionId)) {
  // Wait for current turn to finish before spawning next
  await waitUntilIdle(sessionId, 30_000);
}

If a second claude -p starts for a session that hasn't fully released its lock, retry with backoff (the lock typically releases within a few seconds of process exit when using --resume):

proc.on('exit', (code) => {
  if (code !== 0 && stderr.includes('is already in use') && retries < 5) {
    const delay = 1500 + retries * 1000;
    setTimeout(() => spawn(options, onEvent, onComplete, retries + 1), delay);
    return;
  }
  // ... normal completion
});

Model aliases

Claude Code subscription accepts short aliases — use these, not provider-qualified IDs:

Don't pass anthropic/sonnet or claude-sonnet-4-6 — those work in the API but may fail or behave unexpectedly via the CLI subscription path.

Normalize any incoming model string:

function normalizeAlias(model: string): 'opus' | 'sonnet' | 'haiku' {
  const lower = model.toLowerCase();
  if (lower.includes('opus')) return 'opus';
  if (lower.includes('haiku')) return 'haiku';
  return 'sonnet';
}

Session files

Claude stores session history in:

<CLAUDE_HOME>/projects/<cwd-encoded>/<session-uuid>.jsonl

This is Claude Code's internal format — don't write to it. Maintain your own session log (a separate JSONL) if you need to replay history to reconnecting clients.

Common mistakes

The last two rows reflect a specific anti-pattern: an implementation was added that returned zero tokens/cost for Claude sessions under the assumption the data wasn't available. It was reverted because the data IS in the result event — always parse it rather than returning placeholder values.