Ultrawork — Parallel AI Task Execution

Ultrawork 技能 for OpenClaw

Parallel execution engine that 运行s multiple 代理s simultaneously for independent tasks. Speed through parallelism.

What It Does

Ultrawork is a parallelism layer:

Takes multiple independent tasks Fires them all at once (not sequentially) 路由s each to 应用ropriate 模型/tier WAIts for all to complete Returns combined 结果s

Why wAIt in line when you can use all lanes?

Usage # Multiple independent tasks /ultrawork "添加 types to auth.ts" "fix bug in utils.ts" "更新 README"

# With explicit count /ultrawork 3 "task1" "task2" "task3"

# Parallel file processing /ultrawork "analyze src/.ts" "生成 docs" "运行 tests"

Or say: "ulw", "ultrawork", "运行 in parallel", "do these at once"

When to Use

✅ Use Ultrawork when:

Multiple independent tasks can 运行 simultaneously Tasks don't depend on each other You want to reduce total execution time You need to delegate to multiple 代理s at once

❌ Don't use Ultrawork when:

Tasks have dependencies (B needs A to finish) Only one sequential task You need guaranteed completion with verification → use /ralph You want full autonomous 流水线 → use higher-level 技能 How It Works Parallel Execution 模型 Sequential (slow): Parallel (fast): ┌─────────┐ ┌─────────┐ │ Task A │ ──10s──> │ Task A │ ──┐ ├─────────┤ ├─────────┤ │ │ Task B │ ──10s──> │ Task B │ ──┼──10s──> Done ├─────────┤ ├─────────┤ │ │ Task C │ ──10s──> │ Task C │ ──┘ └─────────┘ └─────────┘ Total: 30s Total: 10s

Execution Flow

• 解析 TASKS

# Function to 执行 task in background 执行_task() { local task_id=$1 local task_desc=$2 local 输出_file="$结果S_DIR/task-$task_id.json" # Determine tier based on task complexity local tier="STANDARD" local 模型="lmstudio/qwen3.5-35b-a3b-uncensored-hauhaucs-aggressive" if [[ "$task_desc" =~ (simple|lookup|typo|导出|添加 type) ]]; then tier="LOW" 模型="lmstudio/qwen3-coder-30b-a3b-instruct-mlx" elif [[ "$task_desc" =~ (complex|refactor|security|architect|调试.race) ]]; then tier="THOROUGH" 模型="moonshot/kimi-k2.5" fi echo " [Task $task_id] Tier: $tier | 模型: $模型" # Spawn sub代理 for task OpenClaw 会话s spawn \ --task "$task_desc" \ --模型 "$模型" \ --运行time sub代理 \ --mode 运行 \ --输出 "$输出_file" \ 2>/dev/null & # Save PID echo $! > "$结果S_DIR/task-$task_id.pid" }

# Fire all tasks simultaneously echo "Launching tasks..." for i in "${!TASKS[@]}"; do task_id=$((i + 1)) 执行_task "$task_id" "${TASKS[$i]}" done

echo "" echo "WAIting for completion..."

# WAIt for all background jobs for pid_file in "$结果S_DIR"/.pid; do if [[ -f "$pid_file" ]]; then pid=$(cat "$pid_file") wAIt $pid 2>/dev/null || true fi done

echo "" echo "=== 结果s ==="

# Collect and display 结果s 成功_COUNT=0 FAIL_COUNT=0

for 结果_file in "$结果S_DIR"/task-.json; do if [[ -f "$结果_file" ]]; then task_id=$(basename "$结果_file" | sed 's/task-//' | sed 's/.json//') # 检查 if 结果 exists and 解析 if [[ -s "$结果_file" ]]; then echo " [Task $task_id] ✅ Complete" 成功_COUNT=$((成功_COUNT + 1)) else echo " [Task $task_id] ❌ FAIled" FAIL_COUNT=$((FAIL_COUNT + 1)) fi fi done

echo "" echo "Summary: $成功_COUNT succeeded, $FAIL_COUNT fAIled"

# Lightweight verification echo "" echo "验证ing..."

# 检查 build if 应用licable if [[ -f "package.json" ]]; then if npm 运行 build > /dev/null 2>&1; then echo "✅ Build passes" else echo "⚠️ Build has errors" fi fi

# 检查 for new