RUSH Deep Task

RUSH is LingXizhi's deep task engine: describe your requirement in one sentence, and the engine automatically runs the full pipeline of task planning → parallel coding → quality verification → convergent delivery — ideal for cross-file, multi-step feature development and refactoring.

Quick Start

How It Works

RUSH splits a task into four stages, each verified independently; failures are automatically diagnosed and re-planned:

• Requirement understanding & planning — the Planner analyzes the blast radius with the Semantic Code Graph and outputs an ordered subtask DAG (including step dependencies, affected-file lists, and potential risks).
• Parallel coding — multiple Coders implement in parallel in isolated git worktrees (up to 4 by default, with branches like agent/<taskID>), without interfering with each other.
• Quality verification — the Verifier runs static checks and real tests, with five verification levels from V1 static checks to V5 benchmarking; quality gates automatically reject pseudocode and TODO placeholders. When standards are not met, the Diagnoser locates the failure cause and triggers full or incremental re-planning.
• Convergent delivery — iteration continues based on the convergence score (0–1, threshold 0.95); the best result is merged, a change report is generated, and task experience is written to long-term memory for future reuse.

Unless you specify a model manually, RUSH automatically schedules the optimal model across models by task type. Execution limits: 120 seconds per LLM call, 30 minutes globally by default, 4 hours maximum.

Configuration

Tune RUSH behavior in lxz.config.json at the project root:

{
  "agent": {
    "max_iterations": 50,
    "parallel_max_workers": 4
  },
  "cache": {
    "enabled": true,
    "l2_similarity_threshold": 0.92
  },
  "safety": {
    "snapshot_before_edit": true
  },
  "memory": {
    "auto_learn": true
  }
}

Best Practices

• Include the three elements — goal + constraints + acceptance criteria — in your requirement description; planning quality improves significantly.
• For large refactors, review the Planner's subtask DAG and affected-file list first; run only after confirming they are correct.
• Pair with the Semantic Code Graph: for cross-repo or large-project tasks, let index building finish before launching.