feat(engine): parallelize read-only shell calls

Closes #2983.

Allow exec_shell to opt into the existing parallel tool lane when its concrete input is a conservative read-only shell command. The whitelist covers direct git/list/search/read commands plus simple bash/sh/zsh -c wrappers whose inner command passes the same classifier; anything effectful, interactive, redirected, piped, backgrounded, or stdin-fed remains serial and approval-gated.

Parallel exec_shell fanout is capped at four workers, and explicit multi_tool_use.parallel now preserves request-order results while using the same input-aware checks.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

crates/tui/src/command_safety.rs

@@ -355,6 +355,90 @@ pub fn prefix_allow_matches(pattern: &str, command: &str) -> bool {
     command_norm == pattern_norm || command_norm.starts_with(&format!("{pattern_norm} "))
 }
 
+const PARALLEL_READONLY_PREFIXES: &[&str] = &[
+    "git status",
+    "git log",
+    "git diff",
+    "git show",
+    "git ls-files",
+    "git blame",
+    "git grep",
+    "ls",
+    "pwd",
+    "cat",
+    "head",
+    "tail",
+    "wc",
+    "which",
+    "stat",
+    "file",
+    "du",
+    "df",
+    "grep",
+    "rg",
+    "fd",
+];
+
+/// Return `true` when a shell command is safe to auto-approve and run in a
+/// parallel read-only chunk.
+pub fn is_parallel_readonly_command(command: &str) -> bool {
+    let trimmed = command.trim();
+    if trimmed.is_empty() {
+        return false;
+    }
+    if trimmed.contains("$(")
+        || trimmed
+            .chars()
+            .any(|ch| matches!(ch, '\n' | '\r' | ';' | '&' | '|' | '>' | '<' | '`'))
+    {
+        return false;
+    }
+
+    let tokens = shell_words(trimmed);
+    let Some(start) = primary_token_index(&tokens) else {
+        return false;
+    };
+    let command_tokens = tokens[start..].to_vec();
+
+    if let Some(inner_command) = readonly_shell_wrapper_inner_command(&command_tokens) {
+        return is_parallel_readonly_command(inner_command);
+    }
+
+    let command_refs = command_tokens
+        .iter()
+        .map(String::as_str)
+        .collect::<Vec<_>>();
+    let canonical = classify_command(&command_refs);
+    if canonical == "tail"
+        && command_refs.iter().skip(1).any(|token| {
+            *token == "-f"
+                || *token == "-F"
+                || *token == "--follow"
+                || token.starts_with("--follow=")
+        })
+    {
+        return false;
+    }
+
+    PARALLEL_READONLY_PREFIXES
+        .iter()
+        .any(|prefix| *prefix == canonical)
+}
+
+fn readonly_shell_wrapper_inner_command(tokens: &[String]) -> Option<&str> {
+    let shell = tokens.first()?.as_str();
+    if !matches!(shell, "bash" | "sh" | "zsh") {
+        return None;
+    }
+    if tokens.len() != 3 {
+        return None;
+    }
+    if !matches!(tokens[1].as_str(), "-c" | "-lc") {
+        return None;
+    }
+    Some(tokens[2].as_str())
+}
+
 /// Safety classification of a command
 #[derive(Debug, Clone, Copy, PartialEq, Eq)]
 pub enum SafetyLevel {
@@ -1037,6 +1121,42 @@ mod tests {
         );
     }
 
+    #[test]
+    fn parallel_readonly_command_classifier_is_strict() {
+        for command in [
+            "git status -s",
+            "git log --oneline -5",
+            "rg foo crates/",
+            "ls -la",
+            "cat Cargo.toml",
+            "bash -lc 'git status -s'",
+            "sh -c 'rg foo crates/'",
+        ] {
+            assert!(
+                is_parallel_readonly_command(command),
+                "{command} should be parallel read-only"
+            );
+        }
+
+        for command in [
+            "git status && rm -rf /",
+            "cat a > b",
+            "git push",
+            "cargo build",
+            "tail -f log",
+            "rg foo | head",
+            "find . -delete",
+            "sleep 5 &",
+            "bash -lc 'git status && rm -rf /'",
+            "bash -lc 'rg foo | head'",
+        ] {
+            assert!(
+                !is_parallel_readonly_command(command),
+                "{command} should not be parallel read-only"
+            );
+        }
+    }
+
     #[test]
     fn test_workspace_safe_commands() {
         assert_eq!(

crates/tui/src/core/engine.rs

@@ -2849,6 +2849,8 @@ mod tool_setup;
 mod turn_loop;
 pub(crate) use token_estimate_cache::TokenEstimateCache;
 
+pub(super) const MAX_PARALLEL_SHELL_EXEC: usize = 4;
+
 pub(crate) fn default_active_native_tool_names() -> &'static [&'static str] {
     tool_catalog::DEFAULT_ACTIVE_NATIVE_TOOLS
 }

crates/tui/src/core/engine/tests.rs

@@ -488,6 +488,48 @@ fn tool_execution_batches_use_serial_barriers() {
     }
 }
 
+#[test]
+fn shell_readonly_plans_batch_around_serial_barrier() {
+    let mut shell_a = make_plan_at(0, true, true, false, false);
+    shell_a.name = "exec_shell".to_string();
+    shell_a.input = json!({"command": "git status -s"});
+    let mut shell_b = make_plan_at(1, true, true, false, false);
+    shell_b.name = "exec_shell".to_string();
+    shell_b.input = json!({"command": "git log --oneline -5"});
+    let mut write_shell = make_plan_at(2, false, false, true, false);
+    write_shell.name = "exec_shell".to_string();
+    write_shell.input = json!({"command": "cargo build"});
+    let mut shell_c = make_plan_at(3, true, true, false, false);
+    shell_c.name = "exec_shell".to_string();
+    shell_c.input = json!({"command": "bash -lc 'rg TODO crates/tui/src/core'"});
+
+    let batches = plan_tool_execution_batches(vec![shell_a, shell_b, write_shell, shell_c]);
+    assert_eq!(batches.len(), 3);
+
+    match &batches[0] {
+        ToolExecutionBatch::Parallel(plans) => {
+            assert_eq!(
+                plans.iter().map(|plan| plan.index).collect::<Vec<_>>(),
+                vec![0, 1]
+            );
+        }
+        ToolExecutionBatch::Serial(_) => panic!("first batch should be parallel"),
+    }
+    match &batches[1] {
+        ToolExecutionBatch::Serial(plan) => assert_eq!(plan.index, 2),
+        ToolExecutionBatch::Parallel(_) => panic!("write shell should be a serial barrier"),
+    }
+    match &batches[2] {
+        ToolExecutionBatch::Parallel(plans) => {
+            assert_eq!(
+                plans.iter().map(|plan| plan.index).collect::<Vec<_>>(),
+                vec![3]
+            );
+        }
+        ToolExecutionBatch::Serial(_) => panic!("third batch should be parallel"),
+    }
+}
+
 #[test]
 fn successful_update_plan_ends_plan_mode_turn_immediately() {
     assert!(should_stop_after_plan_tool(

crates/tui/src/core/engine/tool_execution.rs

@@ -185,8 +185,10 @@ impl Engine {
             ));
         };
 
+        let result_count = calls.len();
         let mut tasks = FuturesUnordered::new();
-        for (tool_name, tool_input) in calls {
+        let shell_permits = Arc::new(tokio::sync::Semaphore::new(MAX_PARALLEL_SHELL_EXEC));
+        for (index, (tool_name, tool_input)) in calls.into_iter().enumerate() {
             if tool_name == MULTI_TOOL_PARALLEL_NAME {
                 return Err(ToolError::invalid_input(
                     "multi_tool_use.parallel cannot call itself",
@@ -206,17 +208,17 @@ impl Engine {
                         "tool '{tool_name}' is not registered"
                     )));
                 };
-                if !spec.is_read_only() {
+                if !spec.is_read_only_for(&tool_input) {
                     return Err(ToolError::invalid_input(format!(
                         "Tool '{tool_name}' is not read-only and cannot run in parallel"
                     )));
                 }
-                if spec.approval_requirement() != ApprovalRequirement::Auto {
+                if spec.approval_requirement_for(&tool_input) != ApprovalRequirement::Auto {
                     return Err(ToolError::invalid_input(format!(
                         "Tool '{tool_name}' requires approval and cannot run in parallel"
                     )));
                 }
-                if !spec.supports_parallel() {
+                if !spec.supports_parallel_for(&tool_input) {
                     return Err(ToolError::invalid_input(format!(
                         "Tool '{tool_name}' does not support parallel execution"
                     )));
@@ -227,7 +229,13 @@ impl Engine {
             let lock = tool_exec_lock.clone();
             let tx_event = self.tx_event.clone();
             let mcp_pool = mcp_pool.clone();
+            let shell_permits = shell_permits.clone();
             tasks.push(async move {
+                let _shell_permit = if tool_name == "exec_shell" {
+                    shell_permits.acquire_owned().await.ok()
+                } else {
+                    None
+                };
                 let result = Engine::execute_tool_with_lock(
                     lock,
                     true,
@@ -240,36 +248,39 @@ impl Engine {
                     None,
                 )
                 .await;
-                (tool_name, result)
+                (index, tool_name, result)
             });
         }
 
-        let mut results = Vec::new();
-        while let Some((tool_name, result)) = tasks.next().await {
-            match result {
+        let mut results: Vec<Option<ParallelToolResultEntry>> = Vec::with_capacity(result_count);
+        results.resize_with(result_count, || None);
+        while let Some((index, tool_name, result)) = tasks.next().await {
+            let entry = match result {
                 Ok(output) => {
                     let mut error = None;
                     if !output.success {
                         error = Some(output.content.clone());
                     }
-                    results.push(ParallelToolResultEntry {
+                    ParallelToolResultEntry {
                         tool_name,
                         success: output.success,
                         content: output.content,
                         error,
-                    });
+                    }
                 }
                 Err(err) => {
                     let message = format!("{err}");
-                    results.push(ParallelToolResultEntry {
+                    ParallelToolResultEntry {
                         tool_name,
                         success: false,
                         content: format!("Error: {message}"),
                         error: Some(message),
-                    });
+                    }
                 }
-            }
+            };
+            results[index] = Some(entry);
         }
+        let results = results.into_iter().flatten().collect();
 
         ToolResult::json(&ParallelToolResult { results })
             .map_err(|e| ToolError::execution_failed(e.to_string()))
@@ -381,7 +392,7 @@ impl Engine {
 mod tests {
     use super::*;
     use serde_json::json;
-    use std::{sync::Mutex, time::Duration};
+    use std::{ffi::OsString, path::Path, sync::Mutex, time::Duration};
 
     /// Tests in this module mutate `DEEPSEEK_TOOL_AUDIT_LOG` which is
     /// process-global; serialise through this guard so the parallel
@@ -392,6 +403,43 @@ mod tests {
         AUDIT_TEST_GUARD.lock().unwrap_or_else(|e| e.into_inner())
     }
 
+    struct AuditEnvGuard {
+        previous: Option<OsString>,
+    }
+
+    impl AuditEnvGuard {
+        fn set(path: &Path) -> Self {
+            let previous = std::env::var_os("DEEPSEEK_TOOL_AUDIT_LOG");
+            // SAFETY: serialised by the guard above.
+            unsafe {
+                std::env::set_var("DEEPSEEK_TOOL_AUDIT_LOG", path);
+            }
+            Self { previous }
+        }
+
+        fn unset() -> Self {
+            let previous = std::env::var_os("DEEPSEEK_TOOL_AUDIT_LOG");
+            // SAFETY: serialised by the guard above.
+            unsafe {
+                std::env::remove_var("DEEPSEEK_TOOL_AUDIT_LOG");
+            }
+            Self { previous }
+        }
+    }
+
+    impl Drop for AuditEnvGuard {
+        fn drop(&mut self) {
+            // SAFETY: callers hold AUDIT_TEST_GUARD for this guard's lifetime.
+            unsafe {
+                if let Some(previous) = self.previous.take() {
+                    std::env::set_var("DEEPSEEK_TOOL_AUDIT_LOG", previous);
+                } else {
+                    std::env::remove_var("DEEPSEEK_TOOL_AUDIT_LOG");
+                }
+            }
+        }
+    }
+
     #[tokio::test]
     async fn terminal_guard_queues_resume_when_event_channel_is_full() {
         let (tx, mut rx) = mpsc::channel(1);
@@ -443,29 +491,35 @@ mod tests {
         let _g = audit_test_guard();
         let tmp = tempfile::tempdir().expect("tempdir");
         let path = tmp.path().join("audit.log");
-        // SAFETY: serialised by the guard above.
-        unsafe {
-            std::env::set_var("DEEPSEEK_TOOL_AUDIT_LOG", &path);
-        }
+        let _env = AuditEnvGuard::set(&path);
+        let marker = path.display().to_string();
 
         emit_tool_audit(json!({
             "event": "tool.spillover",
+            "test_marker": marker,
             "tool_id": "call-abc",
             "tool_name": "exec_shell",
             "path": "/tmp/foo.txt",
         }));
         emit_tool_audit(json!({
             "event": "tool.result",
+            "test_marker": marker,
             "tool_id": "call-xyz",
             "success": true,
         }));
 
         let body = std::fs::read_to_string(&path).expect("audit log written");
-        let lines: Vec<&str> = body.lines().collect();
-        assert_eq!(lines.len(), 2, "two emits → two lines");
+        let entries: Vec<serde_json::Value> = body
+            .lines()
+            .map(|line| serde_json::from_str(line).expect("audit line is JSON"))
+            .filter(|entry: &serde_json::Value| {
+                entry.get("test_marker").and_then(|v| v.as_str()) == Some(marker.as_str())
+            })
+            .collect();
+        assert_eq!(entries.len(), 2, "two marked emits -> two lines");
 
         // Each line round-trips as JSON, has the expected event key.
-        let first: serde_json::Value = serde_json::from_str(lines[0]).expect("first line is JSON");
+        let first = &entries[0];
         assert_eq!(
             first.get("event").and_then(|v| v.as_str()),
             Some("tool.spillover")
@@ -475,26 +529,17 @@ mod tests {
             Some("call-abc")
         );
 
-        let second: serde_json::Value =
-            serde_json::from_str(lines[1]).expect("second line is JSON");
+        let second = &entries[1];
         assert_eq!(
             second.get("event").and_then(|v| v.as_str()),
             Some("tool.result")
         );
-
-        // SAFETY: cleanup under the guard.
-        unsafe {
-            std::env::remove_var("DEEPSEEK_TOOL_AUDIT_LOG");
-        }
     }
 
     #[test]
     fn emit_tool_audit_is_noop_when_env_var_unset() {
         let _g = audit_test_guard();
-        // SAFETY: serialised by the guard above.
-        unsafe {
-            std::env::remove_var("DEEPSEEK_TOOL_AUDIT_LOG");
-        }
+        let _env = AuditEnvGuard::unset();
         // Should not panic and should not create any file. We can't
         // assert "no file written" without knowing where one might be
         // written, but the contract is "do nothing", which we verify
@@ -510,16 +555,8 @@ mod tests {
         // Path with a parent that doesn't exist yet — the writer
         // should create it.
         let nested = tmp.path().join("nested").join("dir").join("audit.log");
-        // SAFETY: serialised by the guard above.
-        unsafe {
-            std::env::set_var("DEEPSEEK_TOOL_AUDIT_LOG", &nested);
-        }
+        let _env = AuditEnvGuard::set(&nested);
         emit_tool_audit(json!({"event": "test"}));
         assert!(nested.exists(), "writer should mkdir -p the parent chain");
-
-        // SAFETY: cleanup under the guard.
-        unsafe {
-            std::env::remove_var("DEEPSEEK_TOOL_AUDIT_LOG");
-        }
     }
 }

crates/tui/src/core/engine/turn_loop.rs

@@ -1530,10 +1530,11 @@ impl Engine {
                 } else if let Some(registry) = tool_registry
                     && let Some(spec) = registry.get(&tool_name)
                 {
-                    approval_required = spec.approval_requirement() != ApprovalRequirement::Auto;
+                    approval_required =
+                        spec.approval_requirement_for(&tool_input) != ApprovalRequirement::Auto;
                     approval_description = spec.description().to_string();
-                    supports_parallel = spec.supports_parallel();
-                    read_only = spec.is_read_only();
+                    supports_parallel = spec.supports_parallel_for(&tool_input);
+                    read_only = spec.is_read_only_for(&tool_input);
                 } else if tool_name == CODE_EXECUTION_TOOL_NAME {
                     approval_required = true;
                     approval_description =
@@ -1666,6 +1667,8 @@ impl Engine {
 
                 if parallel_allowed {
                     let mut tool_tasks = FuturesUnordered::new();
+                    let shell_permits =
+                        Arc::new(tokio::sync::Semaphore::new(MAX_PARALLEL_SHELL_EXEC));
                     for plan in plans {
                         if let Some(result) = plan.guard_result.clone() {
                             let result = Ok(result);
@@ -1704,8 +1707,14 @@ impl Engine {
                         let tx_event = self.tx_event.clone();
                         let session_id = self.session.id.clone();
                         let started_at = Instant::now();
+                        let shell_permits = shell_permits.clone();
 
                         tool_tasks.push(async move {
+                            let _shell_permit = if plan.name == "exec_shell" {
+                                shell_permits.acquire_owned().await.ok()
+                            } else {
+                                None
+                            };
                             let mut result = Engine::execute_tool_with_lock(
                                 lock,
                                 plan.supports_parallel,

crates/tui/src/tools/shell.rs

@@ -1768,7 +1768,9 @@ pub fn new_shared_shell_manager(workspace: PathBuf) -> SharedShellManager {
 
 // === ToolSpec Implementations ===
 
-use crate::command_safety::{SafetyLevel, analyze_command, extract_primary_command};
+use crate::command_safety::{
+    SafetyLevel, analyze_command, extract_primary_command, is_parallel_readonly_command,
+};
 use crate::execpolicy::{ExecPolicyDecision, load_default_policy};
 use crate::features::Feature;
 use crate::tools::cargo_failure_summary::summarize_cargo_failure;
@@ -1913,6 +1915,26 @@ fn shell_network_restricted_hint<'a>(
     }
 }
 
+fn exec_shell_input_is_parallel_readonly(input: &serde_json::Value) -> bool {
+    let Some(command) = input.get("command").and_then(serde_json::Value::as_str) else {
+        return false;
+    };
+    if ["background", "interactive", "tty", "combined_output"]
+        .iter()
+        .any(|key| input.get(*key).and_then(serde_json::Value::as_bool) == Some(true))
+    {
+        return false;
+    }
+    if ["stdin", "input", "data"]
+        .iter()
+        .any(|key| input.get(*key).is_some())
+    {
+        return false;
+    }
+
+    is_parallel_readonly_command(command)
+}
+
 async fn execute_foreground_via_background(
     context: &ToolContext,
     command: &str,
@@ -2061,6 +2083,22 @@ impl ToolSpec for ExecShellTool {
         ApprovalRequirement::Required
     }
 
+    fn approval_requirement_for(&self, input: &serde_json::Value) -> ApprovalRequirement {
+        if exec_shell_input_is_parallel_readonly(input) {
+            ApprovalRequirement::Auto
+        } else {
+            self.approval_requirement()
+        }
+    }
+
+    fn is_read_only_for(&self, input: &serde_json::Value) -> bool {
+        exec_shell_input_is_parallel_readonly(input)
+    }
+
+    fn supports_parallel_for(&self, input: &serde_json::Value) -> bool {
+        exec_shell_input_is_parallel_readonly(input)
+    }
+
     async fn execute(
         &self,
         input: serde_json::Value,

crates/tui/src/tools/shell/tests.rs

@@ -148,6 +148,41 @@ fn wait_for_completed_shell(manager: &mut ShellManager, task_id: &str) -> ShellR
     }
 }
 
+#[test]
+fn exec_shell_parallel_flags_are_input_aware() {
+    let tool = ExecShellTool;
+    let readonly = json!({"command": "git status -s"});
+    assert!(tool.supports_parallel_for(&readonly));
+    assert!(tool.is_read_only_for(&readonly));
+    assert_eq!(
+        tool.approval_requirement_for(&readonly),
+        ApprovalRequirement::Auto
+    );
+
+    let bash_readonly = json!({"command": "bash -lc 'rg TODO crates/tui/src/tools'"});
+    assert!(tool.supports_parallel_for(&bash_readonly));
+    assert!(tool.is_read_only_for(&bash_readonly));
+    assert_eq!(
+        tool.approval_requirement_for(&bash_readonly),
+        ApprovalRequirement::Auto
+    );
+
+    for input in [
+        json!({"command": "git status -s", "background": true}),
+        json!({"command": "git status -s", "stdin": ""}),
+        json!({"command": "cargo build"}),
+        json!({"command": "bash -lc 'rg TODO crates | head'"}),
+    ] {
+        assert!(!tool.supports_parallel_for(&input), "{input:?}");
+        assert!(!tool.is_read_only_for(&input), "{input:?}");
+        assert_eq!(
+            tool.approval_requirement_for(&input),
+            ApprovalRequirement::Required,
+            "{input:?}"
+        );
+    }
+}
+
 #[test]
 #[cfg(unix)]
 fn shell_execution_scrubs_parent_env_and_keeps_explicit_env() {

crates/tui/src/tools/spec.rs

@@ -643,6 +643,11 @@ pub trait ToolSpec: Send + Sync {
         }
     }
 
+    /// Returns the approval requirement for this concrete tool input.
+    fn approval_requirement_for(&self, _input: &Value) -> ApprovalRequirement {
+        self.approval_requirement()
+    }
+
     /// Returns whether this tool is sandboxable.
     #[allow(dead_code)]
     fn is_sandboxable(&self) -> bool {
@@ -657,11 +662,21 @@ pub trait ToolSpec: Send + Sync {
             && !caps.contains(&ToolCapability::ExecutesCode)
     }
 
+    /// Returns whether this concrete tool input is read-only.
+    fn is_read_only_for(&self, _input: &Value) -> bool {
+        self.is_read_only()
+    }
+
     /// Returns whether this tool can be executed in parallel with others.
     fn supports_parallel(&self) -> bool {
         false
     }
 
+    /// Returns whether this concrete tool input can run in parallel.
+    fn supports_parallel_for(&self, _input: &Value) -> bool {
+        self.supports_parallel()
+    }
+
     /// Returns whether this tool should be excluded from the model-visible
     /// tool catalog (deferred loading). Tools marked `true` are registered
     /// but not sent to the model until explicitly activated via tool search.