diff --git a/src/api_hooks.py b/src/api_hooks.py
index a895e02e..ccda56d5 100644
--- a/src/api_hooks.py
+++ b/src/api_hooks.py
@@ -211,8 +211,12 @@ class HookHandler(BaseHTTPRequestHandler):
     self.send_response(200)
     self.send_header("Content-Type", "application/json")
     self.end_headers()
+    # flat_config returns a ProjectContext dataclass; json.dumps cannot serialize
+    # dataclasses directly, so call .to_dict() first. The endpoint is the wire
+    # boundary; consumers expect the legacy dict[str, Any] shape.
     flat = project_manager.flat_config(_get_app_attr(app, "project"))
-    self.wfile.write(json.dumps({"project": flat}).encode("utf-8"))
+    flat_dict = flat.to_dict() if hasattr(flat, "to_dict") else flat
+    self.wfile.write(json.dumps({"project": flat_dict}).encode("utf-8"))
    elif self.path == "/api/project_switch_status":
     # Determinstic signal for tests waiting on a project switch to complete.
     # Polling /api/project returns derived state that may be stale from prior
diff --git a/src/app_controller.py b/src/app_controller.py
index 2505b27c..0e3100b5 100644
--- a/src/app_controller.py
+++ b/src/app_controller.py
@@ -4025,6 +4025,9 @@ class AppController:
   self.save_config()
   track_id = self.active_track.id if self.active_track else None
   flat = project_manager.flat_config(self.project, self.active_discussion, track_id=track_id)
+  # flat_config returns a frozen ProjectContext dataclass; convert to a mutable dict
+  # before overriding files.paths with the runtime context_files list.
+  flat = flat.to_dict() if hasattr(flat, "to_dict") else dict(flat)
   
   import copy
   flat["files"] = copy.copy(flat.get("files", {}))
diff --git a/tests/test_app_controller_result.py b/tests/test_app_controller_result.py
index 6cb1cd96..b683bc91 100644
--- a/tests/test_app_controller_result.py
+++ b/tests/test_app_controller_result.py
@@ -460,6 +460,9 @@ def test_fetch_models_aggregates_per_provider_errors():
    # do_fetch is the inner function; we need to access it. Easiest: call _fetch_models
    # and inspect the resulting side effect on all_available_models.
    ctrl._fetch_models("anthropic")
+   # _fetch_models submits do_fetch to the IO pool; wait for it to complete
+   # before asserting on the side effect.
+   assert ctrl.wait_io_pool_idle(timeout=5.0), "do_fetch did not complete"
  # Per-provider errors should be accumulated in self._model_fetch_errors
  assert "gemini" in ctrl._model_fetch_errors
  assert isinstance(ctrl._model_fetch_errors["gemini"], ErrorInfo)
diff --git a/tests/test_warmup_canaries.py b/tests/test_warmup_canaries.py
index b9045434..56ccb4d6 100644
--- a/tests/test_warmup_canaries.py
+++ b/tests/test_warmup_canaries.py
@@ -255,9 +255,12 @@ def test_app_controller_init_start_ts_is_set() -> None:
 def test_app_controller_warmup_done_ts_none_until_completed() -> None:
  """warmup_done_ts is None before wait, float after."""
  from src.app_controller import AppController
- ctrl = AppController(log_to_stderr=False)
+ # defer_warmup=True avoids the race where warmup completes before we can
+ # observe the initial state (warmup runs in a background thread).
+ ctrl = AppController(defer_warmup=True, log_to_stderr=False)
  initial = ctrl.warmup_done_ts
  assert initial is None
+ ctrl.start_warmup()
  assert ctrl.wait_for_warmup(timeout=60.0) is True
  assert isinstance(ctrl.warmup_done_ts, float)
  assert ctrl.warmup_done_ts > 0