Merge branch 'main' into jsdw-redo-modes

crates/core/src/driver/mod.rs

@@ -4,6 +4,7 @@ use std::collections::HashMap;
 use std::marker::PhantomData;
 use std::path::PathBuf;
 
+use alloy::consensus::EMPTY_ROOT_HASH;
 use alloy::hex;
 use alloy::json_abi::JsonAbi;
 use alloy::network::{Ethereum, TransactionBuilder};
@@ -26,9 +27,12 @@ use revive_dt_format::traits::{ResolutionContext, ResolverApi};
 use semver::Version;
 
 use revive_dt_format::case::{Case, CaseIdx};
-use revive_dt_format::input::{Calldata, EtherValue, Expected, ExpectedOutput, Method};
+use revive_dt_format::input::{
+    BalanceAssertion, Calldata, EtherValue, Expected, ExpectedOutput, Input, Method,
+    StorageEmptyAssertion,
+};
 use revive_dt_format::metadata::{ContractInstance, ContractPathAndIdent};
-use revive_dt_format::{input::Input, metadata::Metadata};
+use revive_dt_format::{input::Step, metadata::Metadata};
 use revive_dt_node::Node;
 use revive_dt_node_interaction::EthereumNode;
 use tracing::Instrument;
@@ -70,6 +74,32 @@ where
         }
     }
 
+    pub async fn handle_step(
+        &mut self,
+        metadata: &Metadata,
+        case_idx: CaseIdx,
+        step: &Step,
+        node: &T::Blockchain,
+    ) -> anyhow::Result<StepOutput> {
+        match step {
+            Step::FunctionCall(input) => {
+                let (receipt, geth_trace, diff_mode) =
+                    self.handle_input(metadata, case_idx, input, node).await?;
+                Ok(StepOutput::FunctionCall(receipt, geth_trace, diff_mode))
+            }
+            Step::BalanceAssertion(balance_assertion) => {
+                self.handle_balance_assertion(metadata, case_idx, balance_assertion, node)
+                    .await?;
+                Ok(StepOutput::BalanceAssertion)
+            }
+            Step::StorageEmptyAssertion(storage_empty) => {
+                self.handle_storage_empty(metadata, case_idx, storage_empty, node)
+                    .await?;
+                Ok(StepOutput::StorageEmptyAssertion)
+            }
+        }
+    }
+
     pub async fn handle_input(
         &mut self,
         metadata: &Metadata,
@@ -78,7 +108,7 @@ where
         node: &T::Blockchain,
     ) -> anyhow::Result<(TransactionReceipt, GethTrace, DiffMode)> {
         let deployment_receipts = self
-            .handle_contract_deployment(metadata, case_idx, input, node)
+            .handle_input_contract_deployment(metadata, case_idx, input, node)
             .await?;
         let execution_receipt = self
             .handle_input_execution(input, deployment_receipts, node)
@@ -93,8 +123,36 @@ where
             .await
     }
 
+    pub async fn handle_balance_assertion(
+        &mut self,
+        metadata: &Metadata,
+        _: CaseIdx,
+        balance_assertion: &BalanceAssertion,
+        node: &T::Blockchain,
+    ) -> anyhow::Result<()> {
+        self.handle_balance_assertion_contract_deployment(metadata, balance_assertion, node)
+            .await?;
+        self.handle_balance_assertion_execution(balance_assertion, node)
+            .await?;
+        Ok(())
+    }
+
+    pub async fn handle_storage_empty(
+        &mut self,
+        metadata: &Metadata,
+        _: CaseIdx,
+        storage_empty: &StorageEmptyAssertion,
+        node: &T::Blockchain,
+    ) -> anyhow::Result<()> {
+        self.handle_storage_empty_assertion_contract_deployment(metadata, storage_empty, node)
+            .await?;
+        self.handle_storage_empty_assertion_execution(storage_empty, node)
+            .await?;
+        Ok(())
+    }
+
     /// Handles the contract deployment for a given input performing it if it needs to be performed.
-    async fn handle_contract_deployment(
+    async fn handle_input_contract_deployment(
         &mut self,
         metadata: &Metadata,
         case_idx: CaseIdx,
@@ -462,6 +520,128 @@ where
         Ok((execution_receipt, trace, diff))
     }
 
+    pub async fn handle_balance_assertion_contract_deployment(
+        &mut self,
+        metadata: &Metadata,
+        balance_assertion: &BalanceAssertion,
+        node: &T::Blockchain,
+    ) -> anyhow::Result<()> {
+        let Some(instance) = balance_assertion
+            .address
+            .strip_prefix(".address")
+            .map(ContractInstance::new)
+        else {
+            return Ok(());
+        };
+        self.get_or_deploy_contract_instance(
+            &instance,
+            metadata,
+            Input::default_caller(),
+            None,
+            None,
+            node,
+        )
+        .await?;
+        Ok(())
+    }
+
+    pub async fn handle_balance_assertion_execution(
+        &mut self,
+        BalanceAssertion {
+            address: address_string,
+            expected_balance: amount,
+            ..
+        }: &BalanceAssertion,
+        node: &T::Blockchain,
+    ) -> anyhow::Result<()> {
+        let address = Address::from_slice(
+            Calldata::new_compound([address_string])
+                .calldata(node, self.default_resolution_context())
+                .await?
+                .get(12..32)
+                .expect("Can't fail"),
+        );
+
+        let balance = node.balance_of(address).await?;
+
+        let expected = *amount;
+        let actual = balance;
+        if expected != actual {
+            tracing::error!(%expected, %actual, %address, "Balance assertion failed");
+            anyhow::bail!(
+                "Balance assertion failed - Expected {} but got {} for {} resolved to {}",
+                expected,
+                actual,
+                address_string,
+                address,
+            )
+        }
+
+        Ok(())
+    }
+
+    pub async fn handle_storage_empty_assertion_contract_deployment(
+        &mut self,
+        metadata: &Metadata,
+        storage_empty_assertion: &StorageEmptyAssertion,
+        node: &T::Blockchain,
+    ) -> anyhow::Result<()> {
+        let Some(instance) = storage_empty_assertion
+            .address
+            .strip_prefix(".address")
+            .map(ContractInstance::new)
+        else {
+            return Ok(());
+        };
+        self.get_or_deploy_contract_instance(
+            &instance,
+            metadata,
+            Input::default_caller(),
+            None,
+            None,
+            node,
+        )
+        .await?;
+        Ok(())
+    }
+
+    pub async fn handle_storage_empty_assertion_execution(
+        &mut self,
+        StorageEmptyAssertion {
+            address: address_string,
+            is_storage_empty,
+            ..
+        }: &StorageEmptyAssertion,
+        node: &T::Blockchain,
+    ) -> anyhow::Result<()> {
+        let address = Address::from_slice(
+            Calldata::new_compound([address_string])
+                .calldata(node, self.default_resolution_context())
+                .await?
+                .get(12..32)
+                .expect("Can't fail"),
+        );
+
+        let storage = node.latest_state_proof(address, Default::default()).await?;
+        let is_empty = storage.storage_hash == EMPTY_ROOT_HASH;
+
+        let expected = is_storage_empty;
+        let actual = is_empty;
+
+        if *expected != actual {
+            tracing::error!(%expected, %actual, %address, "Storage Empty Assertion failed");
+            anyhow::bail!(
+                "Storage Empty Assertion failed - Expected {} but got {} for {} resolved to {}",
+                expected,
+                actual,
+                address_string,
+                address,
+            )
+        };
+
+        Ok(())
+    }
+
     /// Gets the information of a deployed contract or library from the state. If it's found to not
     /// be deployed then it will be deployed.
     ///
@@ -651,38 +831,55 @@ where
             return Ok(0);
         }
 
-        let mut inputs_executed = 0;
-        for (input_idx, input) in self.case.inputs_iterator().enumerate() {
-            let tracing_span = tracing::info_span!("Handling input", input_idx);
+        let mut steps_executed = 0;
+        for (step_idx, step) in self.case.steps_iterator().enumerate() {
+            let tracing_span = tracing::info_span!("Handling input", step_idx);
 
-            let (leader_receipt, _, leader_diff) = self
+            let leader_step_output = self
                 .leader_state
-                .handle_input(self.metadata, self.case_idx, &input, self.leader_node)
+                .handle_step(self.metadata, self.case_idx, &step, self.leader_node)
                 .instrument(tracing_span.clone())
                 .await?;
-            let (follower_receipt, _, follower_diff) = self
+            let follower_step_output = self
                 .follower_state
-                .handle_input(self.metadata, self.case_idx, &input, self.follower_node)
+                .handle_step(self.metadata, self.case_idx, &step, self.follower_node)
                 .instrument(tracing_span)
                 .await?;
+            match (leader_step_output, follower_step_output) {
+                (
+                    StepOutput::FunctionCall(leader_receipt, _, leader_diff),
+                    StepOutput::FunctionCall(follower_receipt, _, follower_diff),
+                ) => {
+                    if leader_diff == follower_diff {
+                        tracing::debug!("State diffs match between leader and follower.");
+                    } else {
+                        tracing::debug!("State diffs mismatch between leader and follower.");
+                        Self::trace_diff_mode("Leader", &leader_diff);
+                        Self::trace_diff_mode("Follower", &follower_diff);
+                    }
 
-            if leader_diff == follower_diff {
-                tracing::debug!("State diffs match between leader and follower.");
-            } else {
-                tracing::debug!("State diffs mismatch between leader and follower.");
-                Self::trace_diff_mode("Leader", &leader_diff);
-                Self::trace_diff_mode("Follower", &follower_diff);
+                    if leader_receipt.logs() != follower_receipt.logs() {
+                        tracing::debug!("Log/event mismatch between leader and follower.");
+                        tracing::trace!("Leader logs: {:?}", leader_receipt.logs());
+                        tracing::trace!("Follower logs: {:?}", follower_receipt.logs());
+                    }
+                }
+                (StepOutput::BalanceAssertion, StepOutput::BalanceAssertion) => {}
+                (StepOutput::StorageEmptyAssertion, StepOutput::StorageEmptyAssertion) => {}
+                _ => unreachable!("The two step outputs can not be of a different kind"),
             }
 
-            if leader_receipt.logs() != follower_receipt.logs() {
-                tracing::debug!("Log/event mismatch between leader and follower.");
-                tracing::trace!("Leader logs: {:?}", leader_receipt.logs());
-                tracing::trace!("Follower logs: {:?}", follower_receipt.logs());
-            }
-
-            inputs_executed += 1;
+            steps_executed += 1;
         }
 
-        Ok(inputs_executed)
+        Ok(steps_executed)
     }
 }
+
+#[derive(Clone, Debug)]
+#[allow(clippy::large_enum_variant)]
+pub enum StepOutput {
+    FunctionCall(TransactionReceipt, GethTrace, DiffMode),
+    BalanceAssertion,
+    StorageEmptyAssertion,
+}

crates/core/src/main.rs

@@ -1,6 +1,6 @@
 use std::{
     collections::HashMap,
-    path::Path,
+    path::{Path, PathBuf},
     sync::{Arc, LazyLock},
     time::Instant,
 };
@@ -18,7 +18,7 @@ use revive_dt_common::iterators::FilesWithExtensionIterator;
 use revive_dt_node_interaction::EthereumNode;
 use semver::Version;
 use temp_dir::TempDir;
-use tokio::sync::{Mutex, RwLock};
+use tokio::sync::{Mutex, RwLock, mpsc};
 use tracing::{Instrument, Level};
 use tracing_subscriber::{EnvFilter, FmtSubscriber};
 
@@ -32,7 +32,7 @@ use revive_dt_core::{
 use revive_dt_format::{
     case::{Case, CaseIdx},
     corpus::Corpus,
-    input::Input,
+    input::{Input, Step},
     metadata::{ContractInstance, ContractPathAndIdent, Metadata, MetadataFile},
     mode::Mode,
 };
@@ -41,15 +41,28 @@ use revive_dt_report::reporter::{Report, Span};
 
 static TEMP_DIR: LazyLock<TempDir> = LazyLock::new(|| TempDir::new().unwrap());
 
-type CompilationCache<'a> = Arc<
+type CompilationCache = Arc<
     RwLock<
         HashMap<
-            (&'a Path, Mode, TestingPlatform),
+            (PathBuf, Mode, TestingPlatform),
             Arc<Mutex<Option<Arc<(Version, CompilerOutput)>>>>,
         >,
     >,
 >;
 
+/// this represents a single "test"; a mode, path and collection of cases.
+#[derive(Clone)]
+struct Test {
+    metadata: Metadata,
+    path: PathBuf,
+    mode: Mode,
+    case_idx: usize,
+    case: Case,
+}
+
+/// This represents the results that we gather from running test cases.
+type CaseResult = Result<Option<usize>, anyhow::Error>;
+
 fn main() -> anyhow::Result<()> {
     let args = init_cli()?;
 
@@ -120,7 +133,7 @@ fn collect_corpora(args: &Arguments) -> anyhow::Result<HashMap<Corpus, Vec<Metad
 
 async fn run_driver<L, F>(
     args: &Arguments,
-    tests: &[MetadataFile],
+    metadata_files: &[MetadataFile],
     span: Span,
 ) -> anyhow::Result<()>
 where
@@ -129,10 +142,25 @@ where
     L::Blockchain: revive_dt_node::Node + Send + Sync + 'static,
     F::Blockchain: revive_dt_node::Node + Send + Sync + 'static,
 {
-    let leader_nodes = NodePool::<L::Blockchain>::new(args)?;
-    let follower_nodes = NodePool::<F::Blockchain>::new(args)?;
+    let (report_tx, report_rx) = mpsc::unbounded_channel::<(Test, CaseResult)>();
 
-    let test_cases = tests
+    let tests = prepare_tests::<L, F>(metadata_files);
+    let driver_task = start_driver_task::<L, F>(args, tests, span, report_tx)?;
+    let status_reporter_task = start_reporter_task(report_rx);
+
+    tokio::join!(status_reporter_task, driver_task);
+
+    Ok(())
+}
+
+fn prepare_tests<L, F>(metadata_files: &[MetadataFile]) -> impl Iterator<Item = Test>
+where
+    L: Platform,
+    F: Platform,
+    L::Blockchain: revive_dt_node::Node + Send + Sync + 'static,
+    F::Blockchain: revive_dt_node::Node + Send + Sync + 'static,
+{
+    metadata_files
         .iter()
         .flat_map(
             |MetadataFile {
@@ -177,191 +205,188 @@ where
                 Some(false) | None => true,
             },
         )
-        .collect::<Vec<_>>();
+        .filter(|(metadata_file_path, metadata, ..)| match metadata.required_evm_version {
+            Some(evm_version_requirement) => {
+                let is_allowed = evm_version_requirement
+                    .matches(&<L::Blockchain as revive_dt_node::Node>::evm_version())
+                    && evm_version_requirement
+                        .matches(&<F::Blockchain as revive_dt_node::Node>::evm_version());
 
-    let metadata_case_status = Arc::new(RwLock::new(test_cases.iter().fold(
-        HashMap::<_, HashMap<_, _>>::new(),
-        |mut map, (path, _, case_idx, case, solc_mode)| {
-            map.entry((path.to_path_buf(), solc_mode.clone()))
-                .or_default()
-                .insert((CaseIdx::new(*case_idx), case.name.clone()), None::<bool>);
-            map
-        },
-    )));
-    let status_reporter_task = {
-        let metadata_case_status = metadata_case_status.clone();
-        let start = Instant::now();
-        async move {
-            const GREEN: &str = "\x1B[32m";
-            const RED: &str = "\x1B[31m";
-            const RESET: &str = "\x1B[0m";
-
-            let mut entries_to_delete = Vec::new();
-            let mut number_of_successes = 0;
-            let mut number_of_failures = 0;
-            loop {
-                let metadata_case_status_read = metadata_case_status.read().await;
-                if metadata_case_status_read.is_empty() {
-                    break;
+                if !is_allowed {
+                    tracing::warn!(
+                        metadata_file_path = %metadata_file_path.display(),
+                        leader_evm_version = %<L::Blockchain as revive_dt_node::Node>::evm_version(),
+                        follower_evm_version = %<F::Blockchain as revive_dt_node::Node>::evm_version(),
+                        version_requirement = %evm_version_requirement,
+                        "Skipped test since the EVM version requirement was not fulfilled."
+                    );
                 }
 
-                for ((metadata_file_path, solc_mode), case_status) in
-                    metadata_case_status_read.iter()
-                {
-                    if case_status.values().any(|value| value.is_none()) {
-                        continue;
-                    }
-
-                    let contains_failures = case_status
-                        .values()
-                        .any(|value| value.is_some_and(|value| !value));
-
-                    if !contains_failures {
-                        eprintln!(
-                            "{}Succeeded:{} {} - {:?}",
-                            GREEN,
-                            RESET,
-                            metadata_file_path.display(),
-                            solc_mode
-                        )
-                    } else {
-                        eprintln!(
-                            "{}Failed:{} {} - {:?}",
-                            RED,
-                            RESET,
-                            metadata_file_path.display(),
-                            solc_mode
-                        )
-                    };
-
-                    number_of_successes += case_status
-                        .values()
-                        .filter(|value| value.is_some_and(|value| value))
-                        .count();
-                    number_of_failures += case_status
-                        .values()
-                        .filter(|value| value.is_some_and(|value| !value))
-                        .count();
-
-                    let mut case_status = case_status
-                        .iter()
-                        .map(|((case_idx, case_name), case_status)| {
-                            (case_idx.into_inner(), case_name, case_status.unwrap())
-                        })
-                        .collect::<Vec<_>>();
-                    case_status.sort_by(|a, b| a.0.cmp(&b.0));
-                    for (case_idx, case_name, case_status) in case_status.into_iter() {
-                        if case_status {
-                            eprintln!(
-                                "  {GREEN}Case Succeeded:{RESET} {} - Case Idx: {case_idx}",
-                                case_name
-                                    .as_ref()
-                                    .map(|string| string.as_str())
-                                    .unwrap_or("Unnamed case")
-                            )
-                        } else {
-                            eprintln!(
-                                "  {RED}Case Failed:{RESET} {} - Case Idx: {case_idx}",
-                                case_name
-                                    .as_ref()
-                                    .map(|string| string.as_str())
-                                    .unwrap_or("Unnamed case")
-                            )
-                        };
-                    }
-                    eprintln!();
-
-                    entries_to_delete.push((metadata_file_path.clone(), solc_mode.clone()));
-                }
-
-                drop(metadata_case_status_read);
-                let mut metadata_case_status_write = metadata_case_status.write().await;
-                for entry in entries_to_delete.drain(..) {
-                    metadata_case_status_write.remove(&entry);
-                }
-
-                tokio::time::sleep(std::time::Duration::from_secs(3)).await;
+                is_allowed
             }
+            None => true,
+        })
+        .map(|(metadata_file_path, metadata, case_idx, case, solc_mode)| {
+            Test {
+                metadata: metadata.clone(),
+                path: metadata_file_path.to_path_buf(),
+                mode: solc_mode,
+                case_idx,
+                case: case.clone(),
+            }
+        })
+}
 
-            let elapsed = start.elapsed();
-            eprintln!(
-                "{GREEN}{}{RESET} cases succeeded, {RED}{}{RESET} cases failed in {} seconds",
-                number_of_successes,
-                number_of_failures,
-                elapsed.as_secs()
-            );
-        }
-    };
-
+fn start_driver_task<L, F>(
+    args: &Arguments,
+    tests: impl Iterator<Item = Test>,
+    span: Span,
+    report_tx: mpsc::UnboundedSender<(Test, CaseResult)>,
+) -> anyhow::Result<impl Future<Output = ()>>
+where
+    L: Platform,
+    F: Platform,
+    L::Blockchain: revive_dt_node::Node + Send + Sync + 'static,
+    F::Blockchain: revive_dt_node::Node + Send + Sync + 'static,
+{
+    let leader_nodes = Arc::new(NodePool::<L::Blockchain>::new(args)?);
+    let follower_nodes = Arc::new(NodePool::<F::Blockchain>::new(args)?);
     let compilation_cache = Arc::new(RwLock::new(HashMap::new()));
-    let driver_task = futures::stream::iter(test_cases).for_each_concurrent(
-        None,
-        |(metadata_file_path, metadata, case_idx, case, solc_mode)| {
+    let number_concurrent_tasks = args.number_of_concurrent_tasks();
+
+    Ok(futures::stream::iter(tests).for_each_concurrent(
+        // We want to limit the concurrent tasks here because:
+        //
+        // 1. We don't want to overwhelm the nodes with too many requests, leading to responses timing out.
+        // 2. We don't want to open too many files at once, leading to the OS running out of file descriptors.
+        //
+        // By default, we allow maximum of 10 ongoing requests per node in order to limit (1), and assume that
+        // this number will automatically be low enough to address (2). The user can override this.
+        Some(number_concurrent_tasks),
+        move |test| {
+            let leader_nodes = leader_nodes.clone();
+            let follower_nodes = follower_nodes.clone();
             let compilation_cache = compilation_cache.clone();
-            let leader_node = leader_nodes.round_robbin();
-            let follower_node = follower_nodes.round_robbin();
-            let tracing_span = tracing::span!(
-                Level::INFO,
-                "Running driver",
-                metadata_file_path = %metadata_file_path.display(),
-                case_idx = case_idx,
-                solc_mode = ?solc_mode,
-            );
-            let metadata_case_status = metadata_case_status.clone();
+            let report_tx = report_tx.clone();
+
             async move {
+                let leader_node = leader_nodes.round_robbin();
+                let follower_node = follower_nodes.round_robbin();
+
+                let tracing_span = tracing::span!(
+                    Level::INFO,
+                    "Running driver",
+                    metadata_file_path = %test.path.display(),
+                    case_idx = ?test.case_idx,
+                    solc_mode = ?test.mode,
+                );
+
                 let result = handle_case_driver::<L, F>(
-                    metadata_file_path.as_path(),
-                    metadata,
-                    case_idx.into(),
-                    case,
-                    solc_mode.clone(),
+                    &test.path,
+                    &test.metadata,
+                    test.case_idx.into(),
+                    &test.case,
+                    test.mode.clone(),
                     args,
                     compilation_cache.clone(),
                     leader_node,
                     follower_node,
                     span,
                 )
+                .instrument(tracing_span)
                 .await;
-                let mut metadata_case_status = metadata_case_status.write().await;
-                match result {
-                    Ok(None) => {
-                        tracing::info!("Execution skipped");
-                    }
-                    Ok(Some(inputs_executed)) => {
-                        tracing::info!(inputs_executed, "Execution succeeded");
-                        metadata_case_status
-                            .entry((metadata_file_path.clone(), solc_mode))
-                            .or_default()
-                            .insert((CaseIdx::new(case_idx), case.name.clone()), Some(true));
-                    }
-                    Err(error) => {
-                        metadata_case_status
-                            .entry((metadata_file_path.clone(), solc_mode))
-                            .or_default()
-                            .insert((CaseIdx::new(case_idx), case.name.clone()), Some(false));
-                        tracing::error!(%error, "Execution failed")
-                    }
-                }
-                tracing::info!("Execution completed");
+
+                report_tx
+                    .send((test, result))
+                    .expect("Failed to send report");
             }
-            .instrument(tracing_span)
         },
+    ))
+}
+
+async fn start_reporter_task(mut report_rx: mpsc::UnboundedReceiver<(Test, CaseResult)>) {
+    let start = Instant::now();
+
+    const GREEN: &str = "\x1B[32m";
+    const RED: &str = "\x1B[31m";
+    const GRAY: &str = "\x1B[90m";
+    const COLOUR_RESET: &str = "\x1B[0m";
+    const BOLD: &str = "\x1B[1m";
+    const BOLD_RESET: &str = "\x1B[22m";
+
+    let mut number_of_successes = 0;
+    let mut number_of_failures = 0;
+    let mut number_of_ignored = 0;
+    let mut failures = vec![];
+
+    // Wait for reports to come from our test runner. When the channel closes, this ends.
+    while let Some((test, case_result)) = report_rx.recv().await {
+        let case_name = test.case.name.as_deref().unwrap_or("unnamed_case");
+        let case_idx = test.case_idx;
+        let test_path = test.path.display();
+        let test_mode = test.mode.clone();
+
+        match case_result {
+            Ok(Some(_inputs)) => {
+                number_of_successes += 1;
+                eprintln!(
+                    "{GREEN}Case Succeeded:{COLOUR_RESET} {test_path} -> {case_name}:{case_idx} (mode: {test_mode:?})"
+                );
+            }
+            Ok(None) => {
+                number_of_ignored += 1;
+                eprintln!(
+                    "{GRAY}Case Ignored:{COLOUR_RESET} {test_path} -> {case_name}:{case_idx} (mode: {test_mode:?})"
+                );
+            }
+            Err(err) => {
+                number_of_failures += 1;
+                eprintln!(
+                    "{RED}Case Failed:{COLOUR_RESET} {test_path} -> {case_name}:{case_idx} (mode: {test_mode:?})"
+                );
+                failures.push((test, err));
+            }
+        }
+    }
+
+    eprintln!();
+    let elapsed = start.elapsed();
+
+    // Now, log the failures with more complete errors at the bottom, like `cargo test` does, so
+    // that we don't have to scroll through the entire output to find them.
+    if !failures.is_empty() {
+        eprintln!("{BOLD}Failures:{BOLD_RESET}\n");
+
+        for failure in failures {
+            let (test, err) = failure;
+            let case_name = test.case.name.as_deref().unwrap_or("unnamed_case");
+            let case_idx = test.case_idx;
+            let test_path = test.path.display();
+            let test_mode = test.mode.clone();
+
+            eprintln!(
+                "---- {RED}Case Failed:{COLOUR_RESET} {test_path} -> {case_name}:{case_idx} (mode: {test_mode:?}) ----\n\n{err}\n"
+            );
+        }
+    }
+
+    // Summary at the end.
+    eprintln!(
+        "{} cases: {GREEN}{number_of_successes}{COLOUR_RESET} cases succeeded, {RED}{number_of_failures}{COLOUR_RESET} cases failed, {GRAY}{number_of_ignored}{COLOUR_RESET} cases ignored in {} seconds",
+        number_of_successes + number_of_failures + number_of_ignored,
+        elapsed.as_secs()
     );
-
-    tokio::join!(status_reporter_task, driver_task);
-
-    Ok(())
 }
 
 #[allow(clippy::too_many_arguments)]
-async fn handle_case_driver<'a, L, F>(
-    metadata_file_path: &'a Path,
-    metadata: &'a Metadata,
+async fn handle_case_driver<L, F>(
+    metadata_file_path: &Path,
+    metadata: &Metadata,
     case_idx: CaseIdx,
     case: &Case,
     mode: Mode,
     config: &Arguments,
-    compilation_cache: CompilationCache<'a>,
+    compilation_cache: CompilationCache,
     leader_node: &L::Blockchain,
     follower_node: &F::Blockchain,
     _: Span,
@@ -455,9 +480,13 @@ where
         // doing the deployments from different accounts and therefore we're not slowed down by
         // the nonce.
         let deployer_address = case
-            .inputs
+            .steps
             .iter()
-            .map(|input| input.caller)
+            .filter_map(|step| match step {
+                Step::FunctionCall(input) => Some(input.caller),
+                Step::BalanceAssertion(..) => None,
+                Step::StorageEmptyAssertion(..) => None,
+            })
             .next()
             .unwrap_or(Input::default_caller());
         let leader_tx = TransactionBuilder::<Ethereum>::with_deploy_code(
@@ -504,11 +533,9 @@ where
         );
 
         let Some(leader_library_address) = leader_receipt.contract_address else {
-            tracing::error!("Contract deployment transaction didn't return an address");
             anyhow::bail!("Contract deployment didn't return an address");
         };
         let Some(follower_library_address) = follower_receipt.contract_address else {
-            tracing::error!("Contract deployment transaction didn't return an address");
             anyhow::bail!("Contract deployment didn't return an address");
         };
 
@@ -538,8 +565,16 @@ where
         .any(|(code, _)| !code.chars().all(|char| char.is_ascii_hexdigit()));
     let (leader_compiled_contracts, follower_compiled_contracts) =
         if metadata_file_contains_libraries && compiled_contracts_require_linking {
-            let leader_key = (metadata_file_path, mode.clone(), L::config_id());
-            let follower_key = (metadata_file_path, mode.clone(), L::config_id());
+            let leader_key = (
+                metadata_file_path.to_path_buf(),
+                mode.clone(),
+                L::config_id(),
+            );
+            let follower_key = (
+                metadata_file_path.to_path_buf(),
+                mode.clone(),
+                F::config_id(),
+            );
             {
                 let mut cache = compilation_cache.write().await;
                 cache.remove(&leader_key);
@@ -599,15 +634,19 @@ where
     driver.execute().await.map(Some)
 }
 
-async fn get_or_build_contracts<'a, P: Platform>(
-    metadata: &'a Metadata,
-    metadata_file_path: &'a Path,
+async fn get_or_build_contracts<P: Platform>(
+    metadata: &Metadata,
+    metadata_file_path: &Path,
     mode: Mode,
     config: &Arguments,
-    compilation_cache: CompilationCache<'a>,
+    compilation_cache: CompilationCache,
     deployed_libraries: &HashMap<ContractInstance, (Address, JsonAbi)>,
 ) -> anyhow::Result<Option<Arc<(Version, CompilerOutput)>>> {
-    let key = (metadata_file_path, mode.clone(), P::config_id());
+    let key = (
+        metadata_file_path.to_path_buf(),
+        mode.clone(),
+        P::config_id(),
+    );
     if let Some(compilation_artifact) = compilation_cache.read().await.get(&key).cloned() {
         let mut compilation_artifact = compilation_artifact.lock().await;
         match *compilation_artifact {