diff --git a/desktop/src-tauri/src/managed_agents/relay_mesh.rs b/desktop/src-tauri/src/managed_agents/relay_mesh.rs
index 03b43fe01..5ea453d82 100644
--- a/desktop/src-tauri/src/managed_agents/relay_mesh.rs
+++ b/desktop/src-tauri/src/managed_agents/relay_mesh.rs
@@ -42,15 +42,12 @@ pub fn apply_relay_mesh_env(
         RELAY_MESH_PREFER_MESH_FOR_AUTO_ENV.to_string(),
         "1".to_string(),
     );
-    // Keep the entire turn inside smaller local-model context windows, and
-    // spend a bounded output budget on an answer/tool call instead of hidden
-    // reasoning. The ordinary buzz-agent defaults (200K context / 32K output)
-    // are appropriate for frontier APIs but can wedge a local Mesh runtime.
+    // Keep the requested response inside smaller local-model context windows,
+    // and spend that budget on an answer/tool call instead of hidden reasoning.
     // These are defaults, not policy: the effective agent/persona/global env
     // may deliberately choose a smaller cap or enable thinking. This function
     // runs after those layers during readiness, so never clobber their values.
-    insert_default_if_unset(env, "BUZZ_AGENT_MAX_CONTEXT_TOKENS", "32768");
-    insert_default_if_unset(env, "BUZZ_AGENT_MAX_OUTPUT_TOKENS", "1024");
+    insert_default_if_unset(env, "BUZZ_AGENT_MAX_OUTPUT_TOKENS", "4096");
     insert_default_if_unset(env, "BUZZ_AGENT_THINKING_EFFORT", "none");
 }
 
@@ -76,11 +73,7 @@ pub fn relay_mesh_process_env(
     model: &str,
 ) -> std::collections::BTreeMap<String, String> {
     let mut env = std::collections::BTreeMap::new();
-    for key in [
-        "BUZZ_AGENT_MAX_CONTEXT_TOKENS",
-        "BUZZ_AGENT_MAX_OUTPUT_TOKENS",
-        "BUZZ_AGENT_THINKING_EFFORT",
-    ] {
+    for key in ["BUZZ_AGENT_MAX_OUTPUT_TOKENS", "BUZZ_AGENT_THINKING_EFFORT"] {
         if let Some(value) = effective_env.get(key) {
             env.insert(key.to_string(), value.clone());
         }
@@ -104,13 +97,9 @@ mod tests {
             Some(RELAY_MESH_AUTO_MODEL_ID),
         );
 
-        assert_eq!(
-            env.get("BUZZ_AGENT_MAX_CONTEXT_TOKENS").map(String::as_str),
-            Some("32768")
-        );
         assert_eq!(
             env.get("BUZZ_AGENT_MAX_OUTPUT_TOKENS").map(String::as_str),
-            Some("1024")
+            Some("4096")
         );
         assert_eq!(
             env.get("BUZZ_AGENT_THINKING_EFFORT").map(String::as_str),
@@ -126,10 +115,6 @@ mod tests {
     #[test]
     fn native_provider_preserves_explicit_generation_controls() {
         let mut env = BTreeMap::from([
-            (
-                "BUZZ_AGENT_MAX_CONTEXT_TOKENS".to_string(),
-                "65536".to_string(),
-            ),
             (
                 "BUZZ_AGENT_MAX_OUTPUT_TOKENS".to_string(),
                 "2048".to_string(),
@@ -142,10 +127,6 @@ mod tests {
             Some(RELAY_MESH_AUTO_MODEL_ID),
         );
 
-        assert_eq!(
-            env.get("BUZZ_AGENT_MAX_CONTEXT_TOKENS").map(String::as_str),
-            Some("65536")
-        );
         assert_eq!(
             env.get("BUZZ_AGENT_MAX_OUTPUT_TOKENS").map(String::as_str),
             Some("2048")
@@ -159,10 +140,6 @@ mod tests {
     #[test]
     fn process_env_seeds_controls_without_restoring_unrelated_credentials() {
         let effective_env = BTreeMap::from([
-            (
-                "BUZZ_AGENT_MAX_CONTEXT_TOKENS".to_string(),
-                "16384".to_string(),
-            ),
             (
                 "BUZZ_AGENT_MAX_OUTPUT_TOKENS".to_string(),
                 "1024".to_string(),
@@ -172,10 +149,6 @@ mod tests {
 
         let env = relay_mesh_process_env(&effective_env, "Gemma-4");
 
-        assert_eq!(
-            env.get("BUZZ_AGENT_MAX_CONTEXT_TOKENS").map(String::as_str),
-            Some("16384")
-        );
         assert_eq!(
             env.get("BUZZ_AGENT_MAX_OUTPUT_TOKENS").map(String::as_str),
             Some("1024")