fix(provider): stream-inactivity timeout + lower read_timeout (stall resilience)

crates/forge_config/src/http.rs

@@ -64,7 +64,7 @@ mod tests {
     fn test_http_config_fields() {
         let config = HttpConfig {
             connect_timeout_secs: 30,
-            read_timeout_secs: 900,
+            read_timeout_secs: 180,
             pool_idle_timeout_secs: 90,
             pool_max_idle_per_host: 5,
             max_redirects: 10,

crates/forge_domain/src/result_stream_ext.rs

@@ -7,6 +7,13 @@ use crate::{
     ToolCallFull, ToolCallPart, Usage,
 };
 
+/// Maximum time to wait for a single event from the provider stream before
+/// treating it as a stall. Generous enough to tolerate slow reasoning gaps,
+/// short enough that a hung backend fails fast and `retry_with_config`
+/// re-issues instead of the session freezing -- HTTP/2 keepalive can otherwise
+/// keep reqwest's `read_timeout` from ever firing on a content-stalled stream.
+const STREAM_IDLE_TIMEOUT: std::time::Duration = std::time::Duration::from_secs(120);
+
 /// Extension trait for ResultStream to provide additional functionality
 #[async_trait::async_trait]
 pub trait ResultStreamExt<E> {
@@ -64,7 +71,23 @@ impl ResultStreamExt<anyhow::Error> for crate::BoxStream<ChatCompletionMessage,
         let mut xml_tool_calls = None;
         let mut tool_interrupted = false;
 
-        while let Some(message) = self.next().await {
+        loop {
+            let next = match tokio::time::timeout(STREAM_IDLE_TIMEOUT, self.next()).await {
+                Ok(item) => item,
+                Err(_) => {
+                    // The provider stream went silent for longer than
+                    // STREAM_IDLE_TIMEOUT (e.g. Codex stalled mid-stream while
+                    // HTTP/2 keepalive PINGs keep the socket alive, so reqwest's
+                    // read_timeout never fires). Surface a Retryable error so
+                    // retry_with_config re-issues instead of hanging forever.
+                    return Err(crate::Error::Retryable(anyhow::anyhow!(
+                        "provider stream inactivity: no event for {}s",
+                        STREAM_IDLE_TIMEOUT.as_secs()
+                    ))
+                    .into());
+                }
+            };
+            let Some(message) = next else { break };
             let message =
                 anyhow::Ok(message?).with_context(|| "Failed to process message stream")?;
             // Process usage information

crates/forge_infra/src/http.rs

@@ -39,7 +39,7 @@ impl<F: forge_app::FileWriterInfra + 'static> ForgeHttpInfra<F> {
     pub fn new(config: ForgeConfig, file_writer: Arc<F>) -> Self {
         let http = config.http.unwrap_or(forge_config::HttpConfig {
             connect_timeout_secs: 30,
-            read_timeout_secs: 900,
+            read_timeout_secs: 180,
             pool_idle_timeout_secs: 90,
             pool_max_idle_per_host: 5,
             max_redirects: 10,