Constructing end-to-end service solutions is a key challenge in service computing. Traditional methods (service recommendation, composition, discovery) each address only a partial workflow, and none provide a complete pipeline from user requirement to executable solution. Recent breakthroughs in large language models (LLMs) offer new opportunities. This paper presents DHGL, an LLM-guided dual-channel heuristic greedy framework for end-to-end service solution construction, comprising three phases: Decompose, Recall, and Construct. In Decompose, an LLM breaks an ambiguous user query into primary subtasks. Recall uses a hierarchical dynamic switching mechanism to iteratively refine the subtask sequence via reward-guided regeneration and counterexample learning. Construct employs dual-channel agents (completeness and accuracy agents) with a greedy strategy to build executable directed acyclic graphs (DAGs) of services that covers both core and special requirements. DHGL integrates LLM’s task planning and natural language understanding into solution construction. Experiments on two real-world datasets show it outperforms state-of-the-art methods in accuracy, executability, and conciseness (The code is available at git@github.com:Felicity155/DHGL.git.).