MentorAgent 1.3.0 release notes
What MentorAgent 1.3.0 adds and fixes: actions turned on and off outside your code, tenants, any vector store, whole-conversation tests, and how to upgrade.
On this page
MentorAgent 1.3.0 was published on NuGet on 11 November 2026 (UTC). It is a minor version: it adds features and fixes defects, with no intended break. It also fixes a security defect of severity S1 in /mentor/chat: see Security.
The five packages require each other at exactly this version: upgrade them together. MentorAgent 1.3.0 runs on .NET 10. To upgrade, see Upgrade from 1.2.0.
Highlights
Section titled “Highlights”- Turn actions on and off from outside your code. An action can be turned off in the code. A governor then decides at run time whether each action is on, which roles it needs, whether the user confirms it, and what the model reads about it.
- One instance for several tenants. Tell MentorAgent which company a request belongs to, and memory, rate limits, conversations, metrics and retrieval stay apart.
- A knowledge base on any vector store. pgvector, Qdrant, Azure AI Search or an in-memory collection, with no source of your own to write.
- An announced action is checked. When a reply says "I'm adding the note" and no tool ran, the agent gets one more chance to call it, or to say that it was not done.
- Whole conversations in your tests. Run multi-turn conversations through the real agent, and check which tools each turn calls.
- A link from another site no longer runs a turn. A security fix for
/mentor/chatwith cookie sign-in.
New features
Section titled “New features”Turn actions on and off from outside your code
Section titled “Turn actions on and off from outside your code”The code decides which methods are actions. Which actions are on for a customer, which roles they need, and whether they ask for a confirmation often change after a release. An administration console decides them. [MentorAction(Enabled = false)] turns an action off in the code. An IMentorActionGovernor, registered with AddMentorActionGovernor<T>(), answers for each action whether it is on, the roles it requires, whether the user confirms it, and the description the model reads. Each answer can be null, or MentorConfirmationChoice.Code, to keep what the code says.
public class CaseArchive
{
// Shipped off: it runs only where the governor turns it on
[MentorAction(Description = "Archives a closed support case",
Enabled = false, RequiresConfirmation = true)]
public string ArchiveCase(string caseNumber) =>
$"Case {caseNumber} archived.";
}
public class ConsoleGovernor(IActionSettings settings)
: IMentorActionGovernor
{
// null: what the code says; false: off; true: on
public ValueTask<bool?> IsEnabledAsync(
MentorActionRequest request, CancellationToken cancellationToken)
=> settings.IsEnabledAsync(request.ToolName, cancellationToken);
// null: the code's roles; an empty list: no role needed
public ValueTask<IReadOnlyList<string>?> GetRequiredRolesAsync(
MentorActionRequest request, CancellationToken cancellationToken)
=> settings.GetRolesAsync(request.ToolName, cancellationToken);
// Code: what the code says; Required: ask; None: do not ask
public ValueTask<MentorConfirmationChoice> GetConfirmationAsync(
MentorActionRequest request, CancellationToken cancellationToken)
=> ValueTask.FromResult(MentorConfirmationChoice.Code);
// null: the code's description
public ValueTask<string?> GetDescriptionAsync(
MentorActionRequest request, CancellationToken cancellationToken)
=> ValueTask.FromResult<string?>(null);
}An action that is off is offered to no agent, no model reads about it, the MCP server does not list it and a call to it does not run. The governor is asked when a session builds its tools, on every tools/list of the MCP server, and before every call. A session keeps the tools that it was built with. Turning an action off applies at once, from its next call. Turning an action on applies from the next session. A confirmation that the governor adds applies from the next call, and a confirmation that it takes away applies from the next session.
A governor can widen what the code allows: turn on an action that ships off, require no role, or take a confirmation away. Whoever implements it holds that power, so check there what may be widened and with whose approval. A governor that throws, or does not answer within ActionPolicyTimeout, stops the action. An application has one governor: a second registration throws. The A2A card follows only the code. The confirmation box keeps the description from the code.
The decisions say where an answer came from: MentorActionDecision.RolesFromOutside and ConfirmationFromOutside, and the new reason codes disabled and governor-failed.
One instance for several tenants
Section titled “One instance for several tenants”An application that serves more than one company from one instance needs MentorAgent to keep each company apart. An IMentorTenantResolver, registered with AddMentorTenantResolver<T>(), names the tenant of each request. Your actions read it from IMentorTenant.Id.
// The tenant is a claim of the signed-in user
public class ClaimTenantResolver : IMentorTenantResolver
{
// null: the request is refused
public ValueTask<string?> ResolveAsync(
MentorTenantRequest request, CancellationToken cancellationToken)
=> ValueTask.FromResult(
request.User?.FindFirst("tenant_id")?.Value);
}
// An action reads the tenant of the request
public class OrderActions(IMentorTenant tenant, IOrderStore store)
{
[MentorAction(Description = "Counts the open orders",
ReadOnly = true)]
public Task<int> CountOpenOrders() =>
store.CountOpenAsync(tenant.Id!);
}With a resolver:
- No request without a tenant. A resolver that throws, or answers
nullor an empty text, refuses the request before the model or the action. A turn whose tenant is not the one its conversation began with is refused too. - Memory, rate limits and conversations are the tenant's own. The key of the memory and of
RateLimitPerUsercarries the tenant, and a session snapshot is restored for its tenant only. - Metrics name the tenant. Every counter and the
mentor.turnspan have thementoragent.tenanttag. The dashboard andGET /mentor/admin/metricsshow the reader's tenant only. - Retrieval and logs know the tenant.
MentorRagRequest.TenantIdgives a knowledge-base source the turn's tenant. MentorAgent filters nothing itself: the source returns that tenant's documents. Every log line of a turn is written in a scope withmentoragent.tenant.
Without a resolver there is one tenant, as before, and IMentorTenant.Id is null.
A knowledge base on any vector store
Section titled “A knowledge base on any vector store”For a collection of Microsoft.Extensions.VectorData, such as pgvector, Qdrant, Azure AI Search or an in-memory one, you no longer write an IMentorRagSource. Register the collection with AddMentorVectorStoreRag, and say which part of a record is the document:
builder.Services.AddMentorVectorStoreRag(
sp => sp.GetRequiredService<VectorStore>()
.GetCollection<Guid, Doc>("docs"),
new MentorVectorStoreRagOptions<Doc>
{
Content = d => d.Text, // required
Title = d => d.Title,
SourceUrl = d => d.Url,
// With a tenant resolver: the tenant's documents only
Filter = request => d => d.Tenant == request.TenantId,
});
builder.Services.AddMentorAgent(options =>
{
options.UseRag = true;
// options.EmbeddingGenerator = ...;
});The options also give the title, the link and the metadata of a record, and Filter says which records a request may see, for example those of its tenant.
The source reads every score as a similarity, whatever the distance function of the collection. With CosineDistance, the Postgres default, it uses 1 minus the distance, so RagMinScore keeps the nearest documents. With a Euclidean, Manhattan or Hamming distance, or with no distance function, the registration fails at startup.
With EmbeddingGenerator set, the message is embedded once per turn, and routing, tool filtering and memory share that vector. The source can also have a generator of its own. Without either, the collection embeds the text itself.
An announced action is checked
Section titled “An announced action is checked”A model can end its reply with "I'm registering the note" and call nothing, or say that something is done when nothing ran (BUG-220). CheckAnnouncedActions is on by default. After a turn in which no tool was called, a model checks whether the reply announces an action or reports one as done. If it does, the agent gets one more request in the same turn, with the same tools. That request tells the agent to call the tool now, or to say that it was not done. The reply already shown is not held back, and the call passes the same checks as any other.
The check costs a few hundred tokens after every reply that called no tool. It goes to ClassifierChatClient, or to ChatClient without one, so a small model there keeps the cost down. It works on Path A with the history kept by MentorAgent, not with UseServiceManagedHistory nor with a pre-built Agent.
Whole conversations in your tests
Section titled “Whole conversations in your tests”MentorEvaluator.RunAsync sends each case to the model alone. RunConversationsAsync runs whole conversations through the agent itself, with its tools, roles, filters, knowledge base, memory and the history from one turn to the next. A turn can require tools, forbid others, set a token budget and a limit on the time to the first token, or expect a refusal.
var report = await evaluator.RunConversationsAsync(
[
new EvalConversation("close a case",
[
new EvalTurn("What is the status of case C-1?")
{
ExpectTools = ["get_case_status"],
},
new EvalTurn("Close it, please.")
{
ExpectTools = ["close_case"],
MaxTimeToFirstToken = TimeSpan.FromSeconds(3),
},
]),
new EvalConversation("out of scope",
[
new EvalTurn("Write me a poem about the sea")
{
ExpectRefusal = EvalRefusal.OutOfScope,
},
]),
],
new MentorEvalOptions
{
// Sign in a test user here: anonymous otherwise
PrepareSession = (services, ct) => Task.CompletedTask,
// Default: every confirmation is answered no
// ApproveConfirmations = true,
MaxTotalTokens = 20_000,
});
foreach (var r in report.Results.Where(r => !r.Passed))
output($"{r.Conversation} turn {r.Turn}: {r.Notes}");
report.ThrowIfFailed();Every confirmation is answered no by default: an evaluation does not write data on its own. MentorEvalOptions.ApproveConfirmations answers yes, and PrepareSession sets the user or the page of the conversation before the first turn, for example to sign in a test user. Each conversation runs in a session of its own. See Test without a model for the tests that need no model.
Other features
Section titled “Other features”- The rules of actions, by name.
ACTION-RULES.md, in the repository, lists the rules by which a method becomes an action, how actions are named, and how roles and confirmations add up. Each rule has a name that does not change and a test. A rule that comes in, changes or is retired is named in the release notes of its version. In 1.3.0 the governor and the agent sources bring eight new rules:enabled.off-by-default,enabled.external-decides,surface.governor-session,roles.external-replace,confirm.external-decides,describe.external-replace,declarative.functions-matchanddeclarative.no-hosted-tools. Four rules now also cover the governor:surface.mcp-withholds-gated,surface.card-ungated-only,roles.checked-before-confirmationandconfirm.predicate-adds. Every rule is in How methods become actions. - An agent source says what it left out.
IMentorAgentSource.LoadAsyncreturns the agents and aMentorAgentRejectionfor each definition left out or tool taken away. AnIMentorAgentSourceObserver, registered withAddMentorAgentSourceObserver<T>(), receives the report each time a coordinator is built. An existing source works unchanged. - Options you can change while the application runs.
MentorRuntimeOptions.Changeablelists them, and says whether a change counts from the next turn, such asLanguageorConfirmationTimeout, or from the next session, such asMentorshipLevel. Every other option is read before the application starts. - "Not found in the documents". With
RagNoResultsNotice, off by default, the model is told when the knowledge-base search kept no document, so it does not answer from memory. - Failed searches are counted. With
EnableObservability, a knowledge-base search that fails is counted onmentoragent.rag_failures, and each search has amentor.rag.searchspan. - A source hears when the conversation changes.
IMentorRagSource.OnConversationChangedAsynctells a source that keeps state that the conversation was reset, restored or replaced. - The coordinator's sampling.
CoordinatorSamplingsets temperature, top-p, top-k, the output limit, the penalties and the seed on every request of the coordinator. A value that you leave unset keeps the provider's default. - A warmer start.
WarmUpChatModel, withWarmUpAtStartup, sends the coordinator's model one small request at startup, so the first user does not wait for the connection or for a local model to load. - A provider that ignores a required tool call is reported. Ollama ignores
tool_choice: "required", soRequireLookupForRecordQuestionsandRequireActionForChangeRequestshad no effect with it, and nothing said so. Now a warning says so, once per provider client. - A call to an action not offered is reported. A call to an action that the agent was not given in the turn runs nothing, as before. An
IMentorActionObservernow receives its decision,Deniedwith the reason codenot-offered. - A project template. The repository has a
dotnet newtemplate,templates/mentoragent-blazor: a Blazor Server helpdesk with three actions, roles, confirmations and a test project that runs whole conversations with a scripted model. It is installed from a clone of the repository, not from NuGet.
Improvements
Section titled “Improvements”- A knowledge-base source that answers
nullfails at once. The message says that a source that finds nothing returns an empty list, instead of aNullReferenceExceptionlater on. The turn goes on without documents, as before. - Declarative agents: functions that name no action are taken away. In
Novior.MentorAgent.Declarative, akind: functionentry that names no action of the session is removed from the agent with a warning. The name can be unknown, or the action can be off. Before, the agent had a function with nothing behind it: the model could call it, nothing ran, and nothing was logged. - A check of the public API. At pack time, every package is checked against the public API of 1.2.0.
- Description and tags on NuGet. The packages say what MentorAgent is: AI agents in a .NET application under your rules. The tags of
Novior.MentorAgent,Novior.MentorAgent.ServerandNovior.MentorAgent.Blazornow includeai-governance,governance,authorization,roles,approvalandhuman-in-the-loop, and thementortag is gone from all five packages.
Each defect has a severity from S1, the most severe, to S4: see Versioning and support.
- BUG-220 (S3): with
HitlModeset toBlockingand a Chat Completions client, the model could announce an action that asks for no confirmation, and end its reply without calling it. NowCheckAnnouncedActionsgives it one more request to call the action, or to say that it was not done. On a sample application, with the check off 18 of 24 replies said a note was written when it was not; with it on, 23 of 24 notes were written. - BUG-110 (S3): when a restored conversation was replaced by a new one, because the service no longer knew it, the knowledge-base search kept the recent messages of the conversation that was gone. Now they are cleared, as on a reset or a restore.
- BUG-229 (S3): a knowledge-base source that embeds the message with a model of its own could be handed the vector of
EmbeddingGenerator, or routing, tool filtering and memory could be handed the vector of the source. It depended on which of them asked first in the turn. Now each embedding model gets its own vector, still computed once per turn.
The security fix is under Security.
Security
Section titled “Security”BUG-230 (S1), in Novior.MentorAgent.Server. GET /mentor/chat ran a turn for any request that the browser sent with the sign-in cookie. With cookie sign-in (SameSite=Lax), a link clicked on another site ran a turn as the signed-in user, with the question that the link's author chose. This happened even with RequireAuthorization.
Now GET and POST /mentor/chat answer 403 before the turn when the browser's Fetch Metadata says that another site started the request. That is the case when Sec-Fetch-Site is cross-site or same-site, and Sec-Fetch-Mode is not cors. A warning chat.cross-site-refused is logged, without the question. These requests work as before: calls from your own page, cors requests, and clients that send no Fetch Metadata (such as servers and mobile apps).
Two cases are left to your application: a SameSite=None sign-in cookie, and browsers that send no Fetch Metadata, such as Safari before 16.4. For them, add a check of your own, such as antiforgery tokens or a custom request header: see Upgrade from 1.2.0.
Breaking changes
Section titled “Breaking changes”This version has no intended breaking change. Three changes of behavior are worth checking when you upgrade: CheckAnnouncedActions is on by default, /mentor/chat refuses requests that another site started, and declarative agents lose the functions that name no action. The steps are in Upgrade from 1.2.0.
Deprecations
Section titled “Deprecations”This version deprecates nothing.
Dependency updates
Section titled “Dependency updates”| Package | From | To | Why |
|---|---|---|---|
Microsoft.Agents.AI and Microsoft.Agents.AI.Workflows | 1.26.0 | 1.27.0 | The Agent Framework 1.27 line. Its two breaking changes are in parts MentorAgent does not use: AddFoundryResponses for hosted Foundry applications, and the redirect callbacks of DefaultHttpRequestHandler. |
Microsoft.Agents.AI.Declarative (in Novior.MentorAgent.Declarative) | 1.26.0-rc1 | 1.27.0-rc1 | The same line. |
Microsoft.Extensions.VectorData.Abstractions | (none) | 10.10.0 | New, for the vector store source. Agent Framework already brought it. |
The other dependencies are unchanged: Microsoft.Extensions.AI 10.11.0, ModelContextProtocol 2.3.0, A2A 1.0.0-preview3.
Known issues
Section titled “Known issues”The defects that are still open in 1.3.0, with what to do about each one, are on the Known issues page, together with the prerelease packages and the package versions to pin. The limits that are not defects are in Known limitations.
Upgrade from 1.2.0
Section titled “Upgrade from 1.2.0”- Upgrade the five packages together. Each package requires the others at exactly its own version.
- Check
/mentor/chatif a front end calls it from another site. A request that another site started now gets403. Your own page andcorsrequests work as before. With aSameSite=Nonesign-in cookie, or for browsers without Fetch Metadata, add a check of your own: see Security. - Decide about
CheckAnnouncedActions. It is on by default, and costs a few hundred tokens after each reply that called no tool. GiveClassifierChatClienta small model, or setCheckAnnouncedActions = false. - Check your declarative agents. A
kind: functionentry that names no action of the session is now removed, and a warning names it. Fix the name, or remove the entry. - Update the links you write by hand. If a page links
MentorAgent.cssandMentorAgent.jsitself, as theindex.htmlof a Blazor WebAssembly client or of a .NET MAUI app does, change?v=1.2.0to?v=1.3.0in both links. - If you use pgvector with Semantic Kernel.
Microsoft.SemanticKernel.Connectors.PgVector1.74.0-preview is built onMicrosoft.Extensions.VectorData10.1. MentorAgent references version 10.10, and with it every search fails withMissingMethodException. UseCommunityToolkit.VectorData.PgVectorinstead.
If you upgrade from 1.1.0, first read Upgrade from 1.1.0 in the 1.2.0 notes.
Full list of changes
Section titled “Full list of changes”Every entry of this version is in the 1.3.0 section of CHANGELOG.md (external site), which is also the release notes on NuGet (external site).
See also
Section titled “See also”- Release notes — Every stable MentorAgent version, newest first, with its date and what it brings; how versions are numbered and supported, and how dependencies are followed.
- MentorAgent 1.2.0 release notes — What MentorAgent 1.2.0 adds and fixes: only permitted actions reach the model, your own policies, who confirmed, the AI notice, and how to upgrade.
- Known issues — The open defects of the current MentorAgent version and how to avoid them, the prerelease packages it depends on, and the package versions to pin.
- Known limitations — What MentorAgent does not do by design, as opposed to a defect: platforms, hosted tools, open endpoints, default checks and A2A chains.
- Troubleshooting — Fix common problems with MentorAgent: startup errors, blocked requests, roles, confirmation, local models. And where to get help or report a bug.
