Sent strict tool schemas only to providers that enforce them

app/MindWork AI Studio/Provider/BaseProvider.cs

@@ -1245,6 +1245,7 @@ public abstract class BaseProvider : IProvider, ISecretId
     /// <param name="requestPath">The request path, relative to the provider base URL.</param>
     /// <param name="headersAction">Optional additional headers to add.</param>
     /// <param name="mayAskForSequentialToolCalls">Whether a request which offers tools may ask for one call at a time. False for a provider which rejects the parallel_tool_calls parameter.</param>
+    /// <param name="enforcesStrictToolSchemas">Whether the provider binds the model's tool calls to a strict schema. Only then are tools offered in strict mode; everywhere else a strict schema would only tell the model that every argument is required.</param>
     /// <param name="token">The cancellation token.</param>
     /// <typeparam name="TRequest">The request DTO type.</typeparam>
     /// <typeparam name="TDelta">The delta stream line type.</typeparam>
@@ -1262,6 +1263,7 @@ public abstract class BaseProvider : IProvider, ISecretId
         string requestPath = "chat/completions",
         Action<HttpRequestHeaders>? headersAction = null,
         bool mayAskForSequentialToolCalls = true,
+        bool enforcesStrictToolSchemas = false,
         [EnumeratorCancellation] CancellationToken token = default)
         where TRequest : ChatCompletionAPIRequest
         where TDelta : IResponseStreamLine
@@ -1300,7 +1302,7 @@ public abstract class BaseProvider : IProvider, ISecretId
             if (runnableTools.Count > 0)
             {
                 var adapter = new ChatCompletionToolCallingAdapter<TRequest>(requestFactory, systemPrompt, apiParameters,
-                    runnableTools.Select(x => ProviderToolAdapters.ToChatCompletionTool(x.Definition)).ToList(), mayAskForSequentialToolCalls, runnableTools,
+                    runnableTools.Select(x => ProviderToolAdapters.ToChatCompletionTool(x.Definition, enforcesStrictToolSchemas)).ToList(), mayAskForSequentialToolCalls, runnableTools,
                     (requestDto, requestToken) => this.StreamChatCompletionRequest(requestDto, providerName, requestPath, requestedSecret, headersAction, requestToken),
                     ChatCompletionSourceReader.Read<TDelta, TAnnotation>,
                     this.logger);

app/MindWork AI Studio/Provider/OpenAI/ProviderOpenAI.cs

@@ -168,6 +168,16 @@ public sealed class ProviderOpenAI() : BaseProvider(LLMProviders.OPEN_AI, new Ur
                                },
                                systemPromptRole: systemPromptRole,
                                requestPath: "chat/completions",
+
+                               //
+                               // OpenAI binds a strict function's arguments to its schema -- read on
+                               // 2026-09-24 at https://developers.openai.com/api/docs/guides/function-calling.
+                               // Most other hosts of this API do not: Groq applies strict mode to
+                               // response formats only (https://console.groq.com/docs/structured-outputs,
+                               // same day) and merely validates a tool call afterward, so it rejects
+                               // every call that leaves out an argument the strict schema requires:
+                               //
+                               enforcesStrictToolSchemas: true,
                                token: token))
                 yield return content;
 

app/MindWork AI Studio/Provider/ProviderToolAdapters.cs

@@ -10,33 +10,47 @@ namespace AIStudio.Provider;
 /// <remarks>
 /// The definitions state a tool once, in plain JSON Schema. What differs per API is not only the
 /// field names but how an optional argument is expressed, which is why the OpenAI shapes convert
-/// the schema while Anthropic takes it as written.
+/// the schema while Anthropic takes it as written.<br/><br/>
+/// Strict mode is a promise of the host, not of the definition: only a host which binds the
+/// model's output to the schema keeps it. Everywhere else the model merely reads a schema that
+/// calls every argument required, and then either invents a value for an argument it meant to
+/// leave out, or the host rejects the call for lacking one. That is why the Chat Completions
+/// shape, which reaches every OpenAI-compatible host, only goes strict where the host says so.
 /// </remarks>
 public static class ProviderToolAdapters
 {
     /// <summary>
     /// Builds the nested function tool shape used by Chat Completions compatible APIs.
     /// </summary>
-    public static object ToChatCompletionTool(ToolDefinition definition) => new
+    /// <param name="definition">The tool to describe.</param>
+    /// <param name="hostEnforcesStrict">Whether the host binds the model's tool calls to a strict schema. Only then is the tool sent in strict mode.</param>
+    public static object ToChatCompletionTool(ToolDefinition definition, bool hostEnforcesStrict)
     {
-        type = "function",
-        function = new
+        var isStrict = definition.Function.Strict && hostEnforcesStrict;
+        return new
         {
-            name = definition.Function.Name,
-            description = definition.Function.DescriptionForLLM,
-            parameters = ToOpenAIParameters(definition),
-            strict = definition.Function.Strict,
-        }
-    };
+            type = "function",
+            function = new
+            {
+                name = definition.Function.Name,
+                description = definition.Function.DescriptionForLLM,
+                parameters = ToOpenAIParameters(definition, isStrict),
+                strict = isStrict,
+            }
+        };
+    }
 
     /// <summary>
     /// Builds the flat function tool shape used by the OpenAI Responses API.
     /// </summary>
+    /// <remarks>
+    /// Only OpenAI speaks this API, and it enforces strict mode, so the definition alone decides.
+    /// </remarks>
     public static ResponsesFunctionTool ToResponsesTool(ToolDefinition definition) => new()
     {
         Name = definition.Function.Name,
         Description = definition.Function.DescriptionForLLM,
-        Parameters = ToOpenAIParameters(definition),
+        Parameters = ToOpenAIParameters(definition, definition.Function.Strict),
         Strict = definition.Function.Strict,
     };
 
@@ -59,7 +73,7 @@ public static class ProviderToolAdapters
     /// <summary>
     /// The parameter schema for the OpenAI APIs, converted only when strict mode asks for it.
     /// </summary>
-    private static System.Text.Json.JsonElement ToOpenAIParameters(ToolDefinition definition) => definition.Function.Strict
+    private static System.Text.Json.JsonElement ToOpenAIParameters(ToolDefinition definition, bool isStrict) => isStrict
         ? OpenAIStrictToolSchema.FromToolParameters(definition.Function.Parameters)
         : definition.Function.Parameters;
 }