XBag 发布面向 Microsoft.Extensions.AI 与 Agent Framework 的独立 .NET 包

Small, independent .NET packages for Microsoft.Extensions.AI and Microsoft Agent Framework, by Murat Ay, the author of XBag, a .NET process-orchestration and decisioning platform. Each package solves one problem, plugs into a stable Microsoft extension point and is useful on its own. Community packages, not affiliated with Microsoft.

Packages

Package	What it does	Status
`XBag.AI.Testing`	Scripted, recording `IChatClient` and assertions: test tools, middleware and agents with no API key	0.1.0-preview.1
`XBag.AI.Testing.Agents`	Scripted, recording `AIAgent` for testing agent decorators and pipelines	0.1.0-preview.1
`XBag.Agents.Stores.EntityFrameworkCore`	Durable chat history and agent sessions in SQL Server, PostgreSQL and SQLite	0.1.0-preview.1
`XBag.Agents.Stores.EntityFrameworkCore.Workflows`	Workflow checkpoints in SQL Server, PostgreSQL and SQLite	0.1.0-preview.1
`XBag.Agents.Stores.Redis`	Chat history and agent sessions in Redis, with time to live	0.1.0-preview.1
`XBag.Agents.Stores.Redis.Workflows`	Workflow checkpoints in Redis (sorted set in commit order)	0.1.0-preview.1
`XBag.AI.SpendCaps`	Token and cost caps per run, user and tenant: reserve before, settle after every model call (streaming too)	0.1.0-preview.1
`XBag.AI.SpendCaps.Agents`	`AIAgentBuilder.UseSpendCaps()`: one budget scope per agent run	0.1.0-preview.1
`XBag.AI.Transcripts`	Validate and repair stored conversations (orphan tool calls and results, unanswered approvals) and a guard above the tool loop	0.1.0-preview.1
`XBag.AI.Analyzers`	Roslyn analyzers and code fixes for Microsoft.Extensions.AI / Agent Framework footguns that compile and fail silently (XBAG001-006)	0.1.0-preview.1
`XBag.AI.Prompts`	Prompt templates without Semantic Kernel: variables, `if`, `each`, role-structured messages; values escaped and never able to open a message	0.1.0-preview.1
`XBag.AI.Prompts.Agents`	`PromptInstructionsProvider`: renders a template into an agent's instructions on every run	0.1.0-preview.1
`XBag.AI.Quota`	Tokens-per-minute, requests-per-minute and concurrency limits per model, a FIFO queue instead of 429s, `Retry-After`-aware retry that never retries a started stream	0.1.0-preview.1
`XBag.AI.Quota.ClientModel`	429 / 5xx and `Retry-After` from the OpenAI and Azure OpenAI SDKs (`ClientResultException`) for `XBag.AI.Quota`	0.1.0-preview.1
`XBag.AI.ToolSelection`	Send only the relevant tools per model call (BM25, embeddings or both), plus `search_tools` / `unload_tools` for large catalogs	0.1.0-preview.1
`XBag.AI.RunContext`	Typed per-request context (user, tenant, token) for tools: hidden tool parameters the model cannot fill, a DI scope per tool call, a bearer-token handler for allowed origins	0.1.0-preview.1
`XBag.AI.RunContext.Agents`	`AgentRunOptions.WithRunContext`, `AIAgentBuilder.UseRunContext<T>()`: the same context for every agent run, streaming included	0.1.0-preview.1
`XBag.AI.ToolGuard`	Tool-argument validation with corrective feedback, result caps	planned
`XBag.AI.Adaptive`	Option and structured-output adaptation across models (Qwen, vLLM, OpenAI-compatible endpoints)	planned

Related, published separately: NousToolCalling, prompt-based tool calling for models without native tool calls.

Conventions shared by every package

One package, one problem. No shared runtime "core" package; packages integrate only through Microsoft's own types.
Microsoft.Extensions.AI first. Agent Framework glue lives in separate sub-packages, so the base package stays light.
Preview dependencies are quarantined in prerelease sub-packages; a stable package never pulls a preview one.
Independent versions. A fix to one package never forces a release of the others.
Targets: .NET 8 and .NET 10, nullable on; trimming/AOT analyzers on wherever the dependencies allow it (see below).
Every package README states its pipeline slot (where it sits in a ChatClientBuilder / AIAgentBuilder chain).

Trimming and NativeAOT

Package	Trim / AOT compatible	Why
`XBag.AI.Testing`, `XBag.AI.Testing.Agents`	Yes (analyzers clean)	Only Microsoft.Extensions.AI / Agent Framework abstractions, which are trimmable
`XBag.AI.SpendCaps`, `XBag.AI.SpendCaps.Agents`	Yes (analyzers clean)	Only Microsoft.Extensions.AI, Microsoft.Extensions.Options and Microsoft.Agents.AI; our code uses no reflection
`XBag.AI.Transcripts`	Yes (analyzers clean)	Only Microsoft.Extensions.AI; our code uses no reflection
`XBag.AI.Analyzers`	Not applicable	Runs inside the compiler at build time; nothing is deployed with your app
`XBag.AI.Prompts`, `XBag.AI.Prompts.Agents`	Yes (analyzers clean)	Own template engine, no third-party dependency; `PromptArguments.From<T>` reads properties with trim annotations
`XBag.AI.Quota`, `XBag.AI.Quota.ClientModel`	Yes (analyzers clean)	Only Microsoft.Extensions.AI and System.ClientModel; our code uses no reflection
`XBag.AI.ToolSelection`	Yes (analyzers clean)	Only Microsoft.Extensions.AI; the `search_tools` / `unload_tools` functions build their JSON schema and results by hand, without reflection
`XBag.AI.RunContext`, `XBag.AI.RunContext.Agents`	Yes (analyzers clean)	Only Microsoft.Extensions.AI and Microsoft.Agents.AI; parameter binding goes through `AIFunctionFactoryOptions.ConfigureParameterBinding`, no reflection of our own
`XBag.Agents.Stores.EntityFrameworkCore`, `.EntityFrameworkCore.Workflows`	No	EF Core does not support trimming or NativeAOT, so these packages do not claim it
`XBag.Agents.Stores.Redis`, `.Redis.Workflows`	Not verified	Our code passes the trim/AOT analyzers, but StackExchange.Redis is not annotated as trimmable and no NativeAOT publish was tested

Build and test

dotnet build -c Release

dotnet test -c Release

Requires the .NET 8 and .NET 10 SDKs. Tests run against the real Microsoft pipeline (FunctionInvokingChatClient, ChatClientAgent, the workflow runtime) using the fakes from XBag.AI.Testing; no API keys are needed. The EF Core store tests use SQLite in memory; the Redis store tests start a redis:7-alpine container with Testcontainers, so they need Docker (CI on ubuntu-latest has it).

License

MIT