.NET: add experimental session-scoped skill providers

[adirh3]

Proposal and API alignment

Related to #2651.

This implementation is ready for maintainer review, not a claim that the feature or a .NET-only rollout has been agreed. The contribution guide asks for maintainer alignment and cross-language consistency; the issue describes the use case, and this PR offers a concrete .NET implementation for that discussion.

Motivation

SDK hosts may store reusable skills in application-managed storage rather than project files. Today, such a host must materialize skill files or expose a separate custom loading tool. A session-scoped provider lets these skills participate in the existing native skill tool without either workaround.

The published CLI 1.0.83 protocol already contains internal/experimental skillProvider.list, skillProvider.read, and hasSkillProvider support. This change exposes an idiomatic experimental .NET facade over that existing mechanism. It does not make the raw transport handlers public or change generated protocol code.

Changes

• Add [Experimental] SkillProvider with asynchronous metadata listing and lazy Markdown reads, plus a typed descriptor.
• Add non-serialized SessionConfigBase.SkillProvider and preserve it in create/resume config clones.
• Register typed reverse-RPC handlers before listening and bind each provider to its owning session before the legacy create/resume requests.
• Send hasSkillProvider only when a provider is configured.
• Resolve callbacks by the existing session registry, clear provider references during cleanup, and use the existing RPC error/cancellation and source-generated JSON infrastructure.
• Add focused public-API unit tests and a .NET usage example.

Example API:

#pragma warning disable GHCP001
await using var session = await client.CreateSessionAsync(new SessionConfig
{
    SkillProvider = new ApplicationSkills(),
    EnableSkills = true,
});

// ApplicationSkills implements:
// Task<IReadOnlyList<SkillProviderDescriptor>> ListAsync(CancellationToken)
// Task<string> ReadAsync(string name, CancellationToken)

Scope and behavior

• Opt-in: sessions without a provider keep their existing wire configuration.
• Text-only and pathless; no supporting-file/asset provider, virtual filesystem, or storage policy.
• Native activation is unchanged: no extra loading tool, tool override, or transport proxy.
• The binding is ephemeral and must be supplied again on resume.
• EnableSkills = false keeps the provider bound but dormant. CopilotClientMode.Empty callers should explicitly enable skills.
• Catalog metadata and Markdown frontmatter must agree; the native runtime retains validation responsibility.
• This does not add cloud-session support; all cloud configurations are rejected before connecting, including callers supplying an explicit session ID.
• The public contract documents the native limits: at most 1,024 catalog descriptors, 1 MiB of aggregate metadata, and 1 MiB of UTF-8 text per Markdown response.
• No CLI version bumps, dependency changes, workflow modifications, or application-specific behavior.

Validation

On the current SDK main base (a675b55):

• 296 targeted .NET 8 tests passed across provider/lifecycle, cloning, JSON-RPC, serialization, public DTO, and forward-compatibility coverage, including both cloud session-ID cases.
• The existing .NET 8 test project disables reflection-based System.Text.Json serialization.
• Release builds passed for net8.0, net10.0, and netstandard2.0, with zero warnings or errors.
• git diff --check passed.

Commands (from dotnet):

dotnet test test/GitHub.Copilot.SDK.Test.csproj -f net8.0 -c Release -p:CopilotSkipCliDownload=true --filter "FullyQualifiedName~ClientSessionLifetimeTests|FullyQualifiedName~CloneTests|FullyQualifiedName~JsonRpcTests|FullyQualifiedName~SerializationTests|FullyQualifiedName~PublicDtoTests|FullyQualifiedName~ForwardCompatibilityTests"
dotnet build src/GitHub.Copilot.SDK.csproj -c Release --no-restore -p:CopilotSkipCliDownload=true

These committed tests use the SDK's existing fake server and public APIs. Separately, a host prototype against the real CLI 1.0.83 runtime exercised native tool loading, file-skill coexistence, reload, resume, and subagent skill preloading. That prior prototype evidence is not a claim that runtime E2E coverage is included in this PR.

A further isolated probe of this exact PR against CLI 1.0.84-8 passed genuinely asynchronous list/read callbacks, native/file skill coexistence, session isolation, reload/resume rebinding, read-error recovery, and disabled/no-provider controls. It used locally mocked inference; the probe is separate from this PR's committed tests. Callback cancellation follows the existing RPC connection lifetime, not turn-abort lifetime.

Maintainer questions

• Is exposing this native provider capability aligned with the intended public SDK roadmap?
• Is an abstract provider with list/read methods the preferred .NET shape, or should it use a different facade?
• Should corresponding language implementations land together before the API is offered, even experimentally?

[Copilot]

🟡 Changes recommended

Explicit-ID cloud sessions bypass validation even though the protocol rejects skill providers for all cloud creation flows.

Get a fresh assessment by requesting another Copilot review.

Pull request overview

Adds an experimental .NET facade for session-scoped, in-memory skill providers proposed in #2651.

Changes:

• Adds provider APIs, descriptors, configuration, and documentation.
• Implements session-scoped reverse-RPC routing and cleanup.
• Adds lifecycle, serialization, cloning, and isolation tests.

File summaries

File	Description
dotnet/src/SkillProvider.cs	Defines the experimental provider API.
dotnet/src/Types.cs	Adds provider configuration and cloning.
dotnet/src/Session.cs	Stores and invokes session providers.
dotnet/src/Client.cs	Registers callbacks and wire flags.
dotnet/README.md	Documents provider usage.
dotnet/test/Unit/CloneTests.cs	Tests configuration cloning.
dotnet/test/Unit/ClientSessionLifetimeTests.cs	Enables split lifetime tests.
dotnet/test/Unit/ClientSessionLifetimeTests.SkillProvider.cs	Tests provider lifecycle and RPC behavior.

Review details

Suppressed comments (1)

dotnet/src/SkillProvider.cs:44

• The protocol enforces a 1 MiB UTF-8 byte limit on each Markdown response (nodejs/src/generated/rpc.ts:21508-21512). Include that bound in this public contract so implementations can enforce it before returning content that the runtime will reject.

    /// The complete Markdown, including YAML frontmatter matching the metadata returned by
    /// <see cref="ListAsync"/>. The native runtime rejects inconsistent metadata.

• Files reviewed: 8/8 changed files
• Comments generated: 3
• Review effort level: Balanced (auto)

Note

Copilot is running an experiment and ran this review at Balanced.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[adirh3]

Addressed in 82b8aa8: every Cloud + SkillProvider configuration now fails before connecting, including explicit session IDs. The parameterized regression test covers both generated and explicit IDs. The 296 targeted tests and all three SDK target-framework builds pass.

[adirh3]

Addressed in 82b8aa8: the configuration remarks, README, and PR description now consistently state that cloud sessions do not support skill providers regardless of who supplies the session ID.

[adirh3]

Addressed in 82b8aa8: ListAsync documents the 1,024-descriptor and 1 MiB aggregate metadata limits; ReadAsync documents the 1 MiB UTF-8 response limit. The README includes the same bounds. Validation remains with the native runtime.

[adirh3]

Closing as superseded by the upstream native skill-provider API, now available in the official GitHub.Copilot.SDK 1.0.17-preview.7 package.

Lumi has been migrated from the temporary custom SDK to the official ISkillProvider contract. Validation covered lazy native activation, updated skill bodies, provider rebinding on resume, delegated agents, and coexistence with file-based skills. Windows and Linux builds, targeted/native integration tests, fresh-chat live checks, and single-file publishing/runtime startup with the full CLI sign-in contract all passed. Observed full-suite flakiness also reproduced on unchanged Lumi main and passed in focused isolation; no migration regression was found.

This PR no longer needs to merge. Thanks!