reviewer: limit max concurrency semaphore

docs/configuration.md

@@ -159,6 +159,7 @@ Optional array of additional git remotes to track.
 | Key | Type | Default | Description |
 |-----|------|---------|-------------|
 | `concurrency` | integer | -- | Number of concurrent reviews. |
+| `max_concurrency` | integer | -- | Max number of concurrent review queue. |
 | `worktree_dir` | string | -- | Directory for git worktrees used during review. |
 | `timeout_seconds` | integer | `3600` | Maximum time per review (seconds). |
 | `max_retries` | integer | `3` | Retry count on transient failures. |

src/reviewer.rs

@@ -96,7 +96,8 @@ impl Reviewer {
     /// * `db` - The database connection.
     /// * `settings` - Application settings.
     pub async fn new(db: Arc<Database>, settings: Settings) -> Self {
-        let concurrency = settings.review.concurrency;
+        let mut concurrency = settings.review.concurrency;
+        let max_concurrency = settings.review.max_concurrency;
         let repo_path = PathBuf::from(&settings.git.repository_path);
 
         let baseline_registry =
@@ -129,12 +130,17 @@ impl Reviewer {
         // On average, an active patch review consumes ~3 LLM slots over its execution lifetime.
         // Thus, the global LLM request semaphore is scaled to (concurrency * 3) to fully
         // saturate LLM capacity while gating local processes/worktrees strictly to `concurrency`.
-        let llm_concurrency = if concurrency < 2 {
+        let mut llm_concurrency = if concurrency < 2 {
             1
         } else {
             std::cmp::max(1, concurrency * 3)
         };
 
+        if max_concurrency.is_some() {
+            concurrency = concurrency.min(max_concurrency.unwrap_or(usize::MAX));
+            llm_concurrency = llm_concurrency.min(max_concurrency.unwrap_or(usize::MAX));
+        }
+
         Self {
             db,
             settings,

src/settings.rs

@@ -367,6 +367,7 @@ pub struct GitSettings {
 #[allow(unused)]
 pub struct ReviewSettings {
     pub concurrency: usize,
+    pub max_concurrency: Option<usize>,
     pub worktree_dir: String,
     #[serde(default = "default_review_timeout")]
     pub timeout_seconds: u64,
@@ -469,6 +470,7 @@ fn default_forge() -> ForgeSettings {
 #[derive(Debug, Deserialize, Clone)]
 pub struct LocalReviewReviewSettings {
     pub concurrency: Option<usize>,
+    pub max_concurrency: Option<usize>,
 }
 
 #[derive(Debug, Deserialize, Clone)]