symforge 2.0.8

SymForge — in-memory code intelligence MCP server for Claude Code, Codex, and Gemini CLI
Documentation
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
diff --git a/src/main.rs b/src/main.rs
index 2459926..0bf7e86 100644
--- a/src/main.rs
+++ b/src/main.rs
@@ -58,10 +58,9 @@ fn run_daemon() -> anyhow::Result<()> {
     tokio::runtime::Builder::new_multi_thread()
         .enable_all()
         // Tool handlers acquire std::sync::RwLock on the index, which blocks
-        // the OS thread. With the default worker_threads (= num_cpus), 10+
-        // concurrent subagent calls can starve the runtime. Scale to 4x CPU
-        // cores (min 16, max 64) so blocked threads don't exhaust the pool.
-        .worker_threads(std::thread::available_parallelism().map_or(16, |n| (n.get() * 4).clamp(16, 64)))
+        // the OS thread. Reserve 2 threads for the OS/system, use the rest
+        // for tokio workers so blocked threads don't exhaust the pool.
+        .worker_threads(std::thread::available_parallelism().map_or(4, |n| n.get().saturating_sub(2).max(4)))
         .build()?
         .block_on(async {
             observability::init_tracing()?;