From ae1f13402b233a2be564917a7a1da839dff3f4c8 Mon Sep 17 00:00:00 2001
From: Bryan Helmkamp <bryan@brynary.com>
Date: Sun, 8 Mar 2026 10:03:50 -0400
Subject: [PATCH] Add 5-minute stream read timeout and increase retry backoff
 to 5s

A stalled Anthropic SSE stream caused a run to hang for 10 minutes
until the stall watchdog killed it. The HTTP connection succeeded but
no SSE events arrived, and with stream_read_timeout defaulting to None
the read blocked indefinitely.

- Set AdapterTimeout default stream_read to 300s (5 min), matching
  the idle timeout used by OpenAI Codex
- Increase BackoffConfig default initial_delay_ms from 200ms to 5s
  so transient LLM failures get meaningful recovery time before retry

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---
 crates/arc-llm/src/types.rs        |  4 ++--
 crates/arc-workflows/src/engine.rs | 10 +++++-----
 2 files changed, 7 insertions(+), 7 deletions(-)

diff --git a/crates/arc-llm/src/types.rs b/crates/arc-llm/src/types.rs
index e7438e878..4a6a81ca7 100644
--- a/crates/arc-llm/src/types.rs
+++ b/crates/arc-llm/src/types.rs
@@ -703,7 +703,7 @@ impl Default for AdapterTimeout {
         Self {
             connect: 30.0,
             request: None,
-            stream_read: None,
+            stream_read: Some(300.0),
         }
     }
 }
@@ -1194,7 +1194,7 @@ mod tests {
         let timeout = AdapterTimeout::default();
         assert!((timeout.connect - 30.0).abs() < f64::EPSILON);
         assert!(timeout.request.is_none());
-        assert!(timeout.stream_read.is_none());
+        assert!((timeout.stream_read.unwrap() - 300.0).abs() < f64::EPSILON);
     }
 
     #[test]
diff --git a/crates/arc-workflows/src/engine.rs b/crates/arc-workflows/src/engine.rs
index a8fcc0b36..a994e7e7d 100644
--- a/crates/arc-workflows/src/engine.rs
+++ b/crates/arc-workflows/src/engine.rs
@@ -71,7 +71,7 @@ pub struct BackoffConfig {
 impl Default for BackoffConfig {
     fn default() -> Self {
         Self {
-            initial_delay_ms: 200,
+            initial_delay_ms: 5_000,
             backoff_factor: 2.0,
             max_delay_ms: 60_000,
             jitter: true,
@@ -125,13 +125,13 @@ impl RetryPolicy {
         }
     }
 
-    /// Standard retry policy: 5 attempts, 200ms initial, 2x factor.
+    /// Standard retry policy: 5 attempts, 5s initial, 2x factor.
     #[must_use]
     pub const fn standard() -> Self {
         Self {
             max_attempts: 5,
             backoff: BackoffConfig {
-                initial_delay_ms: 200,
+                initial_delay_ms: 5_000,
                 backoff_factor: 2.0,
                 max_delay_ms: 60_000,
                 jitter: true,
@@ -2322,7 +2322,7 @@ mod tests {
     fn retry_policy_standard() {
         let policy = RetryPolicy::standard();
         assert_eq!(policy.max_attempts, 5);
-        assert_eq!(policy.backoff.initial_delay_ms, 200);
+        assert_eq!(policy.backoff.initial_delay_ms, 5_000);
     }
 
     #[test]
@@ -2399,7 +2399,7 @@ mod tests {
         let policy = build_retry_policy(&node, &graph);
         assert_eq!(policy.max_attempts, 4); // 3 retries + 1 initial
                                             // Should use default backoff, not a preset's backoff
-        assert_eq!(policy.backoff.initial_delay_ms, 200);
+        assert_eq!(policy.backoff.initial_delay_ms, 5_000);
     }
 
     #[test]