Claude Sub Agent

Claude Sub-Agent Skill

Launch claude as a headless sub-agent in a tmux pane with live JSONL monitoring. Depends on the tmux skill for pane interaction.

When to use

• Delegating a scoped task to Claude CLI running autonomously
• Parallel execution alongside other sub-agents (gemini-sub-agent, pi-sub-agent)
• Running Ralph or Bart loops where each iteration is a separate claude process
• Tasks where cost tracking (total_cost_usd) per agent run matters

Output format

Claude CLI streams newline-delimited JSON via --output-format stream-json. Key event types:

Tool calls live inside assistant.message.content[] as {"type":"tool_use","name":"Bash","input":{"command":"..."}}.

Completion signal:

{"type":"result","subtype":"success","is_error":false,"total_cost_usd":0.039,"duration_ms":11300,"num_turns":2,"result":"..."}

Non-interactive invocation

Lean / headless flags

claude \
  --print \
  --output-format stream-json \
  --dangerously-skip-permissions \
  --no-session-persistence \
  -p @/tmp/task-prompt.md \
  > /tmp/claude-output.jsonl 2>&1

Additional useful flags:

• --model sonnet / --model opus / --model haiku — model selection
• --system-prompt @file — replace system prompt (e.g. for persona injection)
• --append-system-prompt @file — append persona on top of default prompt
• --max-budget-usd 1.00 — hard spending cap for the run
• --tools "Bash,Edit,Read,Write" — restrict available tools
• --bare — minimal mode: skip hooks, LSP, CLAUDE.md discovery, plugins (fastest cold start)

Bare mode (fastest, for simple tasks)

claude --print --output-format stream-json \
  --dangerously-skip-permissions \
  --no-session-persistence \
  --bare \
  -p @/tmp/task-prompt.md \
  > /tmp/claude-output.jsonl 2>&1

--bare skips CLAUDE.md auto-discovery, extensions, MCP, plugins, and keychain reads. Use when the worktree has no CLAUDE.md persona you need loaded.

Workflow

1. Write the task prompt to a file

cat > /tmp/task-prompt.md << 'EOF'
Your task description here.
EOF

2. Identify a free tmux pane

bash {SKILLS_DIR}/tmux-remote-control/scripts/tmux-list.sh

3. Launch claude + monitor in the pane

TARGET="{session}:{window}.{pane}"

tmux send-keys -t "$TARGET" \
  "cd /path/to/repo && claude --print --output-format stream-json --dangerously-skip-permissions --no-session-persistence -p @/tmp/task-prompt.md > /tmp/claude-output.jsonl 2>&1 &" Enter
sleep 5
tmux send-keys -t "$TARGET" \
  "python3 {SKILLS_DIR}/claude-sub-agent/scripts/monitor.py /tmp/claude-output.jsonl" Enter

4. Poll for completion

python3 {SKILLS_DIR}/claude-sub-agent/scripts/poll.py "$TARGET" --interval 30

5. Verify results

tmux send-keys -t "$TARGET" C-c   # kill monitor
git log --oneline -5
uv run pytest tests/ -x -q        # or project test command

Full copy-paste pattern

cat > /tmp/task-prompt.md << 'EOF'
Your task here.
EOF

TARGET="{session}:{window}.{pane}"

tmux send-keys -t "$TARGET" \
  "cd /path/to/repo && claude --print --output-format stream-json --dangerously-skip-permissions --no-session-persistence -p @/tmp/task-prompt.md > /tmp/claude-output.jsonl 2>&1 &" Enter
sleep 5
tmux send-keys -t "$TARGET" \
  "python3 {SKILLS_DIR}/claude-sub-agent/scripts/monitor.py /tmp/claude-output.jsonl" Enter

python3 {SKILLS_DIR}/claude-sub-agent/scripts/poll.py "$TARGET" --interval 30

tmux send-keys -t "$TARGET" C-c
git log --oneline -5

Agent / persona injection

# Append a persona on top of claude's default system prompt
claude --print --output-format stream-json \
  --dangerously-skip-permissions --no-session-persistence \
  --append-system-prompt @{SKILLS_DIR}/bart-adversarial-reviewer/references/bart.md \
  -p @/tmp/task.md \
  > /tmp/claude-output.jsonl 2>&1 &

In tmux:

tmux send-keys -t "$TARGET" \
  "cd /repo && claude --print --output-format stream-json --dangerously-skip-permissions --no-session-persistence --append-system-prompt @{SKILLS_DIR}/bart-adversarial-reviewer/references/bart.md -p @/tmp/task.md > /tmp/claude-output.jsonl 2>&1 &" Enter

Unlike Gemini (which uses GEMINI.md), Claude supports --system-prompt and --append-system-prompt flags directly — no CWD file needed.

Model selection

claude --model haiku   # fastest, cheapest (claude-haiku-4-5)
claude --model sonnet  # balanced (claude-sonnet-4-6) — default
claude --model opus    # most capable (claude-opus-4-5)

Cost control

# Hard cap — agent stops if budget exceeded
claude --print --output-format stream-json \
  --dangerously-skip-permissions \
  --max-budget-usd 0.50 \
  -p @/tmp/task.md \
  > /tmp/claude-output.jsonl 2>&1

Final cost is in the result line:

grep '"type":"result"' /tmp/claude-output.jsonl | python3 -c "
import json, sys
d = json.loads(sys.stdin.read())
print(f'Cost: \${d[\"total_cost_usd\"]:.4f}  Duration: {d[\"duration_ms\"]/1000:.1f}s  Turns: {d[\"num_turns\"]}')
"

Troubleshooting

Prompt mangled by shell quoting: Always use -p @/tmp/file.md — never inline long prompts.

Permission prompts blocking headless run: Ensure --dangerously-skip-permissions is set. Without it, claude pauses for approval.

--bare missing CLAUDE.md context: If your repo's CLAUDE.md has important guardrails, omit --bare or use --append-system-prompt @CLAUDE.md explicitly.

Agent runs forever: Check if stuck in a retry loop:

tail -3 /tmp/claude-output.jsonl | python3 -c "import json,sys; [print(json.loads(l).get('type','?')) for l in sys.stdin]"

Parse the result stats:

python3 -c "
import json
for line in open('/tmp/claude-output.jsonl'):
    d = json.loads(line)
    if d.get('type') == 'result':
        print('Status:', d['subtype'], '| Error:', d['is_error'])
        print('Cost: \$' + str(d['total_cost_usd']))
        print('Duration:', d['duration_ms'] / 1000, 's')
        print('Turns:', d['num_turns'])
"