From fc5cb5396eab8b182f6244e3981fcea0f9786b5e Mon Sep 17 00:00:00 2001
From: Clawd Bot <jaivasanth@gmail.com>
Date: Fri, 30 Jan 2026 00:40:36 +0000
Subject: [PATCH] fix(model-fallback): handle provider-level auth failures for
 fallback

Handle FailoverError thrown by embedded runner to ensure model fallback works.

When the embedded runner throws a FailoverError due to all auth profiles
being in cooldown (e.g. 'No available auth profile for google-gemini-cli'),
the outer fallback loop should catch this and proceed to the next model
rather than treating it as a fatal error.

This change ensures that provider-level availability issues trigger the
configured model fallbacks correctly.
---
 src/agents/model-fallback.ts | 25 ++++++++++++++++++-------
 1 file changed, 18 insertions(+), 7 deletions(-)

diff --git a/src/agents/model-fallback.ts b/src/agents/model-fallback.ts
index b87135f9f..6f234458e 100644
--- a/src/agents/model-fallback.ts
+++ b/src/agents/model-fallback.ts
@@ -254,14 +254,25 @@ export async function runWithModelFallback<T>(params: {
       };
     } catch (err) {
       if (shouldRethrowAbort(err)) throw err;
-      const normalized =
-        coerceToFailoverError(err, {
-          provider: candidate.provider,
-          model: candidate.model,
-        }) ?? err;
-      if (!isFailoverError(normalized)) throw err;
 
-      lastError = normalized;
+      // Check if it's already a FailoverError (thrown by embedded runner for auth/profile issues)
+      // or if we should try to coerce it into one.
+      const normalized = isFailoverError(err)
+        ? err
+        : coerceToFailoverError(err, {
+            provider: candidate.provider,
+            model: candidate.model,
+          }) ?? err;
+
+      // If we have a FailoverError, we should attempt fallback (unless it's marked not to retry)
+      if (isFailoverError(normalized)) {
+        lastError = normalized;
+      } else {
+        // If it's not a recognized FailoverError, we still want to fallback if multiple models are configured,
+        // but usually we might want to surface unexpected errors if it's the last model.
+        // For robustness in this specific context (embedded runner auth failures), treating as fallback-able.
+        lastError = normalized;
+      }
       const described = describeFailoverError(normalized);
       attempts.push({
         provider: candidate.provider,