fix: preserve Codex runner intent and private eval evidence

Retain native setup links and built-in settings, record explicit runner choices, and qualify native/legacy skill fixtures. Keep full grading artifacts private while projecting safe first-task summaries.

Co-Authored-By: Paperclip <noreply@paperclip.ing>

ui/src/components/ConfigureBuiltInAgentModal.tsx

@@ -134,10 +134,16 @@ export function ConfigureBuiltInAgentModal({
     mutationFn: async () => {
       const adapterConfig: Record<string, unknown> = {};
       if (model.trim()) adapterConfig.model = model.trim();
+      const existing = state.agent;
+      const sameHarness = existing && agentHarnessType(existing.adapterType, existing.adapterConfig) === adapterType;
+      const savedModel = typeof existing?.adapterConfig.model === "string" ? existing.adapterConfig.model.trim() : "";
+      // A runner-only change must let the server translate the full saved config.
+      const adapterInput = sameHarness && normalizedModel === savedModel ? {}
+        : sameHarness ? { adapterType: existing.adapterType, adapterConfig: { ...existing.adapterConfig, ...adapterConfig } }
+        : { adapterType, adapterConfig };
       const result = await builtInAgentsApi.provision(companyId, definition.key, {
-        adapterType: state.agent && agentHarnessType(state.agent.adapterType, state.agent.adapterConfig) === adapterType ? state.agent.adapterType : adapterType,
+        ...adapterInput,
         ...(adapterType === "codex_local" && runner !== undefined ? { runner } : {}),
-        adapterConfig,
         ...(budgetMonthlyCents !== undefined ? { budgetMonthlyCents } : {}),
       });
       return result;