fix(cli, api): round trip the overall extraction instruction in ixp packages

CHANGELOG.md

@@ -1,4 +1,10 @@
 # Unreleased
+- Fix `re package download`/`re package upload` silently dropping an IXP project's overall
+  extraction instruction, so an uploaded project ran a different prompt from the one it was
+  downloaded from and scored differently in Measure. `LabelGroup` gains an `instructions` field and
+  `UpdateDataset` a `default_label_group_instructions` field
+- Breaking (`reinfer-client` API): `LabelGroup` and `UpdateDataset` each gain a field, so struct
+  literals of them need an extra initializer
 - Update `--help` text, the READMEs and the crate descriptions to refer to UiPath IXP rather than
   Re:infer, and print `re` rather than `reinfer-cli` in `re --help` and `re --version`. The config
   file location (`~/.config/reinfer`), the `REINFER_CLI_NUM_THREADS` environment variable, the

api/src/resources/dataset.rs

@@ -570,6 +570,16 @@ pub struct UpdateDataset<'request> {
     #[serde(rename = "_model_config", skip_serializing_if = "Option::is_none")]
     pub model_config: Option<ModelConfig>,
 
+    /// The instructions for the dataset's `default` label group, surfaced in
+    /// IXP projects as the "overall extraction instruction". Applied to the
+    /// group that already exists on the dataset, so this does not need to be
+    /// sent after the group's label defs.
+    #[serde(
+        rename = "_default_label_group_instructions",
+        skip_serializing_if = "Option::is_none"
+    )]
+    pub default_label_group_instructions: Option<String>,
+
     #[serde(skip_serializing_if = "Vec::is_empty")]
     pub entity_defs: Vec<NewEntityDef>,
 }
@@ -898,4 +908,53 @@ mod tests {
             r#"{"kind":"gpt_ixp","model_version":"a_model_from_the_future","flags":["a_flag_from_the_future"],"attribution_method":"table_formatted_word_ids"}"#
         );
     }
+
+    /// The "overall extraction instruction" is fed to the model as part of its
+    /// prompt, so `re package download` has to capture it for
+    /// `re package upload` to be able to reproduce a project's scores.
+    #[test]
+    fn test_deserialize_default_label_group_instructions() {
+        let dataset: Dataset = serde_json::from_str(
+            r#"{"id":"aaaaaaaaaaaaaaaa","name":"ixp-one","owner":"proj","title":"IXP",
+                "description":"","created":"2026-01-01T00:00:00Z",
+                "last_modified":"2026-01-01T00:00:00Z","model_family":"english",
+                "source_ids":[],"has_sentiment":false,"entity_defs":[],"general_fields":[],
+                "label_defs":[],
+                "label_groups":[{"name":"default","instructions":"Extract from invoices only.",
+                                 "label_defs":[]}],
+                "_dataset_flags":["ixp"],"_model_config":{"kind":"gpt_ixp","flags":[]}}"#,
+        )
+        .expect("an ixp dataset must parse");
+
+        assert_eq!(
+            dataset.label_groups[0].instructions,
+            "Extract from invoices only."
+        );
+    }
+
+    /// Groups other than `default` have their instructions round-tripped by
+    /// the label defs themselves, so only the default group's are sent here,
+    /// and only when set — an absent field must leave the new dataset's own
+    /// default in place rather than blanking it.
+    #[test]
+    fn test_serialize_update_dataset_default_label_group_instructions() {
+        let with_instructions = UpdateDataset {
+            source_ids: None,
+            title: None,
+            description: None,
+            model_config: None,
+            default_label_group_instructions: Some("Extract from invoices only.".to_owned()),
+            entity_defs: Vec::new(),
+        };
+        assert_eq!(
+            serde_json::to_string(&with_instructions).unwrap(),
+            r#"{"_default_label_group_instructions":"Extract from invoices only."}"#
+        );
+
+        let without_instructions = UpdateDataset {
+            default_label_group_instructions: None,
+            ..with_instructions
+        };
+        assert_eq!(serde_json::to_string(&without_instructions).unwrap(), "{}");
+    }
 }

api/src/resources/label_group.rs

@@ -13,6 +13,14 @@ pub static DEFAULT_LABEL_GROUP_NAME: Lazy<Name> = Lazy::new(|| Name("default".to
 pub struct LabelGroup {
     pub name: Name,
 
+    /// For the `default` group of an IXP project this is the "overall
+    /// extraction instruction", which is fed to the model as part of its
+    /// prompt. Set it with [`UpdateDataset::default_label_group_instructions`].
+    ///
+    /// [`UpdateDataset::default_label_group_instructions`]: crate::UpdateDataset::default_label_group_instructions
+    #[serde(default)]
+    pub instructions: String,
+
     #[serde(default)]
     pub label_defs: Vec<LabelDef>,
 }

cli/src/commands/package/upload.rs

@@ -128,13 +128,28 @@ fn wait_until(
     Err(anyhow!("Timeout waiting for {what}"))
 }
 
+/// The packaged dataset's "overall extraction instruction", or `None` when it
+/// is unset so that we leave the new dataset's own default in place.
+///
+/// The instruction lives on the `default` label group rather than on the
+/// dataset, so it is not covered by the label defs the group is created with.
+fn default_label_group_instructions(dataset: &Dataset) -> Option<String> {
+    dataset
+        .label_groups
+        .iter()
+        .find(|label_group| label_group.name == *DEFAULT_LABEL_GROUP_NAME)
+        .map(|label_group| label_group.instructions.clone())
+        .filter(|instructions| !instructions.is_empty())
+}
+
 fn create_ixp_dataset(
     name: DatasetName,
     label_defs: Vec<LabelDef>,
     client: &Client,
     timeout_s: u64,
     model_config: ModelConfig,
     entity_defs: Vec<NewEntityDef>,
+    default_label_group_instructions: Option<String>,
 ) -> Result<Dataset> {
     let mut new_label_defs = Vec::new();
 
@@ -154,6 +169,7 @@ fn create_ixp_dataset(
             source_ids: None,
             title: None,
             description: None,
+            default_label_group_instructions,
             entity_defs,
         },
     )?;
@@ -841,6 +857,8 @@ fn unpack_ixp(
     let packaged_sources = get_ixp_source(&dataset, SourceProvider::Packaged(package))
         .context("Could not get ixp source from package")?;
 
+    let default_label_group_instructions = default_label_group_instructions(&dataset);
+
     // We use title here as the name will already have a hex appended, the api will normalize
     // the title into an api name
     let new_dataset = create_ixp_dataset(
@@ -876,6 +894,7 @@ fn unpack_ixp(
                 instructions: def.instructions,
             })
             .collect(),
+        default_label_group_instructions,
     )
     .context("Could not create dataset")?;
 

cli/src/commands/update/dataset.rs

@@ -59,6 +59,7 @@ pub fn update(client: &Client, args: &UpdateDatasetArgs, printer: &Printer) -> R
                 title: title.as_deref(),
                 description: description.as_deref(),
                 model_config: None,
+                default_label_group_instructions: None,
                 entity_defs: Vec::new(),
             },
         )

cli/tests/test_datasets.rs

@@ -289,6 +289,7 @@ fn test_create_update_dataset_custom() {
         ],
         label_groups: vec![LabelGroup {
             name: LabelGroupName("default".to_owned()),
+            instructions: String::new(),
             label_defs: vec![
                 LabelDef {
                     name: LabelName("bar".to_owned()),