What causes "Failed to review code block" in microsoft/autogen?

Use the model/configuration the sample was written for (a strong tool-calling model such as gpt-4o class) so the inner agent reliably emits the ReviewCodeBlock call.

microsoft/autogen · error · Exception

Failed to review code block

Error message

Failed to review code block

What it means

Thrown by the reviewer agent's middleware in the Fibonacci group-chat sample. The middleware expects the inner agent's reply to contain a ReviewCodeBlock tool/function call; it first checks the reply for the function call, and if absent it re-prompts the inner agent to convert the content into ReviewCodeBlock arguments. If that second attempt still does not yield a usable function call, the middleware falls through to throw new Exception("Failed to review code block"). It is a sample-authored parse/validation failure, not an AutoGen core error.

Source

Thrown at dotnet/samples/AgentChat/AutoGen.Basic.Sample/Example07_Dynamic_GroupChat_Calculate_Fibonacci.cs:221

                                From = "code_reviewer",
                            };

                            return msg;
                        }
                    }
                    else
                    {
                        var originalContent = reply.GetContent();
                        var prompt = $@"Please convert the content to ReviewCodeBlock function arguments.

        ## Original Content
        {originalContent}";

                        reply = await innerAgent.SendAsync(prompt, msgs, ct);
                    }
                }

                throw new Exception("Failed to review code block");
            })
            .RegisterPrintMessage();

        return reviewer;
    }
    #endregion create_reviewer

    public static async Task RunWorkflowAsync()
    {
        long the39thFibonacciNumber = 63245986;
        var kernel = DotnetInteractiveKernelBuilder
            .CreateDefaultInProcessKernelBuilder()
            .Build();

        var gpt4o = LLMConfiguration.GetOpenAIGPT4o_mini();

        #region create_workflow
        var reviewer = await CreateReviewerAgentAsync(gpt4o);

View on GitHub (pinned to 027ecf0a37)

Solutions

Use the model/configuration the sample was written for (a strong tool-calling model such as gpt-4o class) so the inner agent reliably emits the ReviewCodeBlock call.
Verify the reviewer's inner agent actually has the ReviewCodeBlock function contract registered (FunctionCallMiddleware with the function definition) before the conversion prompt is sent.
Increase maxTokens / retry on the inner SendAsync so truncated replies don't strip the tool call.
Make the middleware tolerant: on final failure return a TextMessage verdict (e.g. 'need more info') instead of throwing, so the workflow loop can continue.

Example fix

// before
reply = await innerAgent.SendAsync(prompt, msgs, ct);
}

throw new Exception("Failed to review code block");

// after (bounded retry, then degrade gracefully)
reply = await innerAgent.SendAsync(prompt, msgs, ct);
}

var reviewCall = reply.GetToolCalls().FirstOrDefault(c => c.FunctionName == "ReviewCodeBlock");
if (reviewCall is null)
{
    return new TextMessage(Role.Assistant, "ReviewCodeBlock not produced; ask coder to resend the code block.", from: "reviewer");
}

Defensive patterns

Strategy: try-catch

Validate before calling

var reviewCall = reply.GetToolCalls()?.FirstOrDefault(c => c.FunctionName == "ReviewCodeBlock");
if (reviewCall is null)
{
    // don't reach the throw: re-prompt once more or degrade gracefully
    reply = await innerAgent.SendAsync(prompt, msgs, ct);
}

Type guard

static bool HasReviewCodeBlockCall(IMessage reply) =>
    reply.GetToolCalls()?.Any(c => c.FunctionName == "ReviewCodeBlock") == true;

Try / catch

try
{
    return await next(msgs, option, ct);
}
catch (Exception ex) when (ex.Message == "Failed to review code block")
{
    return new TextMessage(Role.Assistant, "Could not produce a structured review; ask coder to resend.", from: "reviewer");
}

Prevention

Use tool-call-capable models for reviewer agents — weak models frequently skip function calls.
Verify the function contract is registered on the inner agent before relying on tool-call extraction.
Bound retries and degrade to a text verdict instead of throwing from middleware.

When it happens

Trigger: Two consecutive failures: (1) the coder's reply does not trigger the model to emit the ReviewCodeBlock function call, and (2) the explicit 'Please convert the content to ReviewCodeBlock function arguments' re-prompt also returns a reply without a ReviewCodeBlock call (or with arguments the connector cannot extract). Typical with weaker models, truncated replies, or an overloaded/incorrect chat client behind the reviewer.

Common situations: Swapping the sample's model for a smaller/cheaper one that rarely emits tool calls; function-call schemas not registered on the inner agent; reply truncated by max tokens so the tool call is missing; rate-limit or auth issues making the inner reply an error message instead of a tool call.

Related errors

AI-assisted analysis of microsoft/autogen@027ecf0a37 (2026-08-15). Data as JSON: /api/errors/d6491870b56e2122. Report an issue: GitHub.