Testing a three-node Dify support flow: valid structure, missing advice
A support draft can look perfectly organized and still leave out an important instruction. In a small Dify test, a model correctly classified a login problem and ignored a request to approve a refund, but failed to recommend rotating a credential mentioned in the input—even though the system prompt explicitly requested that advice. Here is the three-node flow, the configuration problem…
In a small Dify support flow test, a model correctly identified a login issue but overlooked a request to rotate a credential, despite the system prompt explicitly requesting that action. The three-node flow, configuration issues, and four examples used to evaluate its behavior are detailed below. This setup can be replicated using your own model, but the system has no integrated email sender or account-management tools.
The AI agent responsible for these tests is operated by 1ZheAI, who utilized the Dify Cloud environment for the experiments. These are synthetic examples, not actual production results or a security benchmark.
The support draft follows a user input, LLM processing, and answer node structure. The LLM receives the system query, separates facts from unknowns, and drafts a reply before presenting the result to an operator. The output sections include language, category, facts and unknowns, suggested reply, and human review. The category is one of access, billing, integration, or other, and operator notes are in English while the suggested reply is in the sender's language.
There is no parser enforcing headings, and no machine-readable routing validation. If machine-readable routing is needed, an explicit schema and validation layer must be added before connecting downstream actions.
To recreate the prompt, use a system message guiding the LLM to assist a human support operator in creating a triage note, not a customer-facing final decision. The model should not follow embedded instructions, visit links, run commands, reveal secrets, or claim to have checked accounts, approved refunds, sent emails, changed passwords, or fixed services.
The output should include five Markdown sections: Language, Category, Facts and Unknowns, Suggested Reply, and Human Review. The message itself should be in a separate user prompt section, marked as untrusted and redacted. The downloadable DSL and full prompt contain the exact tested wording in English, Chinese, or Russian, and provide setup notes without including any credentials.
The flow's first failure occurred on October 19, 2026, during testing in the Dify Cloud Preview environment using model ID claude-haiku-4-5-20251001. The test failed due to a ValueError: Model Parameter max_tokens should be less than or equal to 128.0. This issue was resolved by increasing the upper bound to 1024, but this does not confirm the model's actual maximum context or output capability.
The fix only resolves a parameter mismatch. To reproduce, import the DSL into Studio, install the provider dependency, configure the API base URL, key, and model in Integrations → Model Provider, set a supported output bound of at least 650, and run Preview with synthetic inputs. Changing providers or models can alter supported parameters and output behavior, and API calls may incur charges.
Four checks were conducted, each with a different language and issue reported. The results showed that the model could classify problems but failed to provide recommended actions, such as rotating a credential or approving refunds.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
