diff --git a/swarm-controller/src/forge.rs b/swarm-controller/src/forge.rs
index 17753d21..a21fa422 100644
--- a/swarm-controller/src/forge.rs
+++ b/swarm-controller/src/forge.rs
@@ -414,7 +414,33 @@ impl Client {
     /// check [`Self::seed_agent_config`] uses instead of catching a
     /// create-conflict, since `repo_change_files` has no
     /// already-exists-is-fine status the way `create_org_repo` does.
+    ///
+    /// `create_repo` makes `repo` with `auto_init: false` (see
+    /// [`Self::repo_option`]) — no commits, no `main` ref, nothing a
+    /// contents lookup can resolve. Forgejo doesn't 404 that the way it
+    /// does a missing path on a real ref: it returns `200` with a bare
+    /// JSON `[]`, which fails to deserialize into the single-object
+    /// `ContentsResponse` the crate expects for a file path ("invalid
+    /// length 0, expected struct `ContentsResponse` with 16 elements" —
+    /// serde reading the 0-element array as a positional struct). So a
+    /// freshly created, not-yet-seeded repo trips this as an opaque
+    /// deserialize error, not the `NOT_FOUND` branch below — confirmed
+    /// against a live re-test of the swarm agent-create flow tracked on
+    /// the forge. Short-circuit on `empty` first: an empty repo
+    /// obviously has no `agent.nix`, and skipping
+    /// the contents call for it sidesteps the malformed-response shape
+    /// entirely rather than trying to special-case parsing it.
     async fn file_exists(&self, repo: &str, path: &str) -> Result<bool> {
+        let is_empty = self
+            .api
+            .repo_get(CONFIG_ORG, repo)
+            .await
+            .with_context(|| format!("get repo {CONFIG_ORG}/{repo}"))?
+            .empty
+            .unwrap_or(false);
+        if is_empty {
+            return Ok(false);
+        }
         match self
             .api
             .repo_get_contents(CONFIG_ORG, repo, path, RepoGetContentsQuery::default())