research-task-executor

Academic coding tasks with strict research workflow are executed by inputting the exact phrase 科研启动-执行 followed by your task description. The system then decomposes, implements, and verifies your code while performing training and inference configuration checks.

To ensure training and inference configuration alignment, the system performs automated consistency checks that verify model naming and internal fields. This prevents mismatches during research execution by validating environment configurations before completing code changes.

Tracking experiment progress and updating scientific documentation happens automatically after task completion. The system reads existing PROGRESS.md, PIPELINE.md, EXAMPLES.csv, and ISSUES.md files when available, then manages git workflow and documentation updates.

The research execution workflow does support multi-agent planning, using multiple expert agents to decompose, design, implement, and verify tasks. This structured approach ensures no steps are skipped during academic coding execution.

Project-aware execution can read existing pipeline and progress files before running experiments. It automatically detects and reads PROGRESS.md, PIPELINE.md, EXAMPLES.csv, and ISSUES.md files when present in your project directory.

Accidental execution prevention matters for scientific code changes because premature runs can corrupt experiment tracking. It is controlled by requiring the exact activation phrase 科研启动-执行, ensuring tasks only start when explicitly intended.