.NET: Add Redis checkpoint store for workflows - #9253

Motivation & Context

Workflows that run across processes, pods or machines need a durable, shared checkpoint store. On .NET the built-in stores are in-memory and file system (process-exclusive), plus Cosmos DB. Redis is a common choice for this kind of short-lived, shared state (Azure Managed Redis, ElastiCache, self-hosted, Kubernetes), and #2401 asks for a Redis-backed checkpoint store. The earlier .NET attempt, #2799, was closed as outdated with its Copilot review comments still open; this PR is a fresh implementation on current main and addresses all of them (details below).

Description & Review Guide

• What are the major changes?

  • New package Microsoft.Agents.AI.Redis (VersionSuffix=alpha, IsPackable=false until maintainers want to publish it, as was done for Valkey) with RedisCheckpointStore : JsonCheckpointStore and RedisCheckpointStoreOptions (KeyPrefix, Database, TimeToLive). It is built on StackExchange.Redis, which Add Redis-backed checkpoint storage for distributed workflow persistence #2401 names. Types live in Microsoft.Agents.AI.Workflows.Checkpointing, like CosmosCheckpointStore.
  • Storage layout per session: a sorted set of checkpoint ids scored by a per-session commit counter (INCR), a hash with the checkpoint JSON (stored verbatim, so $type discriminators stay first) and a hash with parent ids. One Lua script writes all of it atomically.
  • RetrieveIndexAsync returns checkpoints in commit order, as ICheckpointStore requires and CheckpointManager.GetLatestCheckpointAsync relies on. It stays correct for many checkpoints in the same second and for several processes writing to the same session.
  • All keys of a session share one hash tag built from the session id ({, } and % are escaped, and KeyPrefix may not contain braces), so the store works on Redis Cluster. The optional TimeToLive is reset on every new checkpoint and covers all keys of the session, so a session expires as a whole.
  • The store takes an IConnectionMultiplexer the application owns. It holds no other resources, so it is not IDisposable and is safe to share between threads.
  • A missing checkpoint throws KeyNotFoundException, like FileSystemJsonCheckpointStore.
  • Tests in Microsoft.Agents.AI.Redis.UnitTests: 40 unit tests against a mocked IDatabase (always run), and 12 tests against a real Redis server that follow the Cosmos emulator pattern: REDIS_CONNECTION_STRING (default localhost:6379), skipped when no server is reachable, failing instead when REDIS_AVAILABLE=true. The server tests include a real workflow that is resumed from Redis with a new connection, store and workflow instance.
  • New sample samples/03-workflows/Checkpoint/CheckpointWithRedis: run, stop part way, then resume from the latest checkpoint found in Redis. Added to the solution, the 03-workflows README and eng/verify-samples (skipped there because it needs a Redis server).

• What is the impact of these changes? Additive only: a new package, test project and sample. No existing public API changes. New central package version: StackExchange.Redis 3.4.0.

• What do you want reviewers to focus on?

  • Package location. I added a separate Microsoft.Agents.AI.Redis package instead of extending Microsoft.Agents.AI.Valkey, because: Add Redis-backed checkpoint storage for distributed workflow persistence #2401 asks for StackExchange.Redis; Python already ships agent-framework-redis, so the name matches across languages; and in the Valkey PR (.NET: Adds Valkey to chat message history - issue 5445 #5542) a reviewer asked whether that package should be called Redis, and the Valkey maintainers preferred to keep the Valkey package Valkey-specific (Valkey.Glide only guarantees Redis 6.2–7.2). The Valkey package also has no dependency on Workflows today. I'm happy to move the store into the Valkey package if you prefer.
  • Feature usage telemetry. I did not add a FeatureIndex bit, since docs/specs/feature-usage-bit-registry.md treats new bits as a maintainer decision. I can add a redis row (next free .NET index is 75) if you want one.
  • StackExchange.Redis version. 3.4.0 is the latest stable release and fits the current central pins. I can lower it to the 2.x line if you'd rather not push 3.x onto applications.
  • Scope: no DI helpers or connection-string constructors, to keep the first version small. Applications usually already register one IConnectionMultiplexer.

Related Issue

Fixes #2401

Supersedes #2799 (closed as outdated). There is no other open PR for this issue. #2401 also covers Python; this PR is .NET only.

AI Assistance

• No material AI assistance was used.
• This is an AI-assisted contribution. I reviewed, understood, and verified all submitted content and accept responsibility for it.

AI assistance details: implementation, tests, sample and PR description were drafted with an AI coding assistant (Claude), starting from a Redis checkpoint store I already maintain in my own project. I reviewed the design and code and ran the builds and tests listed above against a real Redis server.

Contribution Checklist

• The code builds clean without any errors or warnings
• All unit tests pass, and I have added new tests where possible
• The PR follows the Contribution Guidelines
• This PR links to an agreed issue with no competing open PR, or the Related Issue section documents a trivial-change or repository-automation exception.
• This is not a breaking change. If it is a breaking change, add the breaking change label (or add "[BREAKING]" to the title prefix, before or after any language prefix) — a workflow keeps the label and title prefix in sync automatically.