Code Node Output Validation#

Overview#

Code Node output validation in Dify enforces that values returned by sandboxed user code match the declared output schema. Validation runs after code execution, against the outputs map declared in CodeNodeData. It is strict: no implicit type coercion is performed. If a Python function returns an int for an output declared as string, the node fails immediately.

The implementation lives in the external graphon package (≥ v0.5.2), which Dify instantiates via DifyNodeFactory. The validation entry point is CodeNode._transform_result(), called by CodeNode._run().

Execution & Response Parsing#

Before validation, the code runs in dify-sandbox (a seccomp-isolated Go process). Inputs are JSON-serialized and base64-encoded by TemplateTransformer.serialize_inputs(), and the sandbox wraps its return value between <<RESULT>> tags. TemplateTransformer.transform_response() extracts and JSON-parses that response, then:

• Requires the result to be a dict with all-string keys 1
• Post-processes scientific-notation strings back into float (e.g. -8.0E-5 → -8e-05) 2

Declared Output Types#

Outputs are declared as CodeNodeData.Output objects keyed by variable name. Supported SegmentType values 3:

For OBJECT outputs, children can be nested recursively, up to CODE_MAX_DEPTH levels (default: 5) 4. Exceeding the depth limit raises DepthLimitError 5.

Strict Type Enforcement#

No implicit coercion is applied. The per-type check methods (_check_string, _check_number, _check_boolean) accept only exact Python types 6:

• _check_string(value, name) — passes only str (or None); strips null bytes (\x00) but does not call str() on non-string values 7
• _check_number(value, name) — passes int or float; does not coerce string representations
• _check_boolean(value, name) — passes bool only

The practical consequence: if your code returns {"result": 42} but the declared output type is string, the node fails with "Output result must be a string, got int instead." 8.

The integration test test_execute_code_output_validator specifically covers this case. The deeper test_execute_code_output_validator_depth 9 also tests:

• Type mismatches in nested objects and arrays (number_validator: "1" against number type)
• Strings exceeding CODE_MAX_STRING_LENGTH (default: 400,000 chars) 10
• Arrays exceeding length limits

Size & Depth Limits#

All limits are configured via CodeExecutionSandboxConfig 11 and passed into CodeNodeLimits by DifyNodeFactory 12:

These are overridable via environment variables.

Exception Hierarchy#

All output validation errors are ValueError subclasses 13:

ValueError
└── CodeNodeError (graphon.nodes.code.exc)
    ├── OutputValidationError — type mismatch / limit exceeded
    └── DepthLimitError — nested structure exceeds max_depth

The CodeNode._run() method catches these and maps them to WorkflowNodeExecutionStatus.FAILED with the error message surfaced on NodeRunResult.error 14.

Agent Node Contrast#

The Agent v2 node has a separate output checker (PerOutputTypeChecker) that handles required vs. optional presence semantics (outputs missing from the backend payload are flagged as NOT_PRODUCED vs. TYPE_CHECK_FAILED). The Code node's graphon validation does not have a separate "optional" concept — an output key present in outputs must appear in the returned dict unless the type check allows None.

Key Source Files#