General Discussion

[tech-cove01]

Self Checks

• I have searched for existing issues search for existing issues, including closed ones.
• I confirm that I am using English to submit this report (我已阅读并同意 Language Policy).
• [FOR CHINESE USERS] 请务必使用英文提交 Issue，否则会被关闭。谢谢！:)
• Please do not modify this template :) and fill in all the required fields.

Content

Title: Chatflow Bug: Variable Not Found during branch convergence & Aggregator limitations

Hey guys, wanted to share a major headache I ran into with Dify’s Chatflow engine regarding branch convergence and the Variable Aggregator.

The Flaw:
If you have multiple conditional branches that eventually merge into a single downstream LLM node, you'll run into a breaking issue. If any upstream branch is skipped, its output variables become Undefined. When the downstream LLM tries to read the template, Dify throws a Variable Not Found error and completely freezes the workflow.

Why the Variable Aggregator doesn't help:
The Aggregator node is too limited. Without grouping, you can only add a single variable. Even with grouping turned on, it still only outputs one variable from whichever path was executed. It acts as a selector, but it cannot merge or fuse multiple variables together into a single object for downstream nodes.

Our Workaround:
To bypass this variable-swallowing bug, we had to dump the single-merge design and use a "Multi-Endpoint Architecture." Basically, each functional path (QA, Refactoring, Coding) gets its own dedicated LLM and Answer node to close the loop independently. It 100% fixes the crash, but it makes the canvas massive and messy.

Has anyone found a cleaner way to fuse variables from parallel branches without multiplying downstream nodes?

[dosubot[bot]]

Hi @tech-cove01! I'm Dosu and I'm helping the Dify team.

This is a well-known pain point with branch convergence in Dify workflows. Here's what I found that may help:

Why it happens: In v1.15.0, Dify's pre-rendering validation checks variable availability before Jinja2 templates evaluate, so even {% if var is defined %} guards can be rejected before they get a chance to run [1]. When a branch is skipped, its output variables don't exist at all for downstream nodes.

Variable Aggregator — it actually can help, with the right setup:

The Variable Aggregator is designed exactly for mutually-exclusive branches (if/else, question classifier). It automatically outputs the value from whichever branch actually executed and silently ignores skipped branches [2]. The key detail is that it always produces a single output variable that downstream nodes can safely reference [3].

If you need to merge multiple different variables from your branches, the pattern is:

• Use one Variable Aggregator per variable you want to merge (e.g., one for the "answer" variable, one for "sources," etc.)
• Each aggregator collects the same-purpose variable from each branch
• Your downstream LLM then references the aggregator outputs instead of the branch outputs directly

There's also an "Aggregate All" mode (added in a 2025 update) that collects ALL resolved inputs into a list rather than just picking the first valid one — useful if multiple branches can execute in parallel [4].

Active improvements coming:

• A PR for advisory warnings on unsafe cross-branch variable references is in progress — it will surface warnings instead of hard-blocking the workflow, and explicitly exempts Variable Aggregator readers since they handle missing variables safely.
• Another PR to relax the duplicate End node variable restriction for conditional branches is also open, since only one End node executes per run anyway.

Recommended pattern for your use case:

start → if-else → [Branch A (LLM), Branch B (LLM), Branch C (LLM)] 
    → Variable Aggregator (collects outputs from A/B/C) 
    → single downstream LLM/Answer node

This should eliminate the "Variable Not Found" crash without needing separate Answer nodes per branch. If each branch produces multiple variables you need downstream, add one aggregator per variable.

Would you like me to dig deeper into the Variable Aggregator configuration or the grouping feature specifically?

To reply, just mention @dosu.

---

Share context across your team and agents. Try Dosu.