integrate-evertask

Integrate EverTask into a project

This skill turns a vague "add background tasks with EverTask" request into a correct, minimal, working integration. It is interactive and progressive: you ask the user only the decisions you cannot infer, then read only the reference files for the features they actually want. Never dump every option at once: that is what the references are for.

EverTask is a MediatR-inspired, persistent background-task library: a task is a record … : IEverTask (the message), a handler is EverTaskHandler<TTask> (the logic), and ITaskDispatcher.Dispatch(...) persists + queues it. Tasks survive restarts, support delayed/scheduled/recurring execution, retry policies, timeouts, keyed rate limiting, multi-queue isolation, and monitoring. Multi-targets net8.0/net9.0/net10.0.

Golden rules (read before doing anything)

1. One storage call is mandatory. AddEverTask(...) must be followed by exactly one of .AddMemoryStorage(), .AddSqlServerStorage(...), .AddPostgresStorage(...), .AddSqliteStorage(...). Without it the app fails at startup.
2. At least one assembly must be registered inside the options lambda (RegisterTasksFromAssembly(...)), or AddEverTask throws ArgumentException.
3. The payload contract is enforced by analyzers shipped inside EverTask.Abstractions (no install needed). Every task you scaffold must use public properties only, no public fields, IDs not entities. See references/08-payload-contract.md: getting this wrong corrupts recovery, the one moment persistence matters.
4. Defaults are good. Do not set MaxDegreeOfParallelism, channel capacity, retry, timeout, or rate-limiter knobs unless the user has a reason. The library auto-sizes sensibly (e.g. MaxDegreeOfParallelism = max(4, CPU×2)).
5. Propose, then apply. Show the user the package list and the Program.cs diff before writing. Confirm before editing existing files.
6. C# 12 primary constructors for handlers with DI (per repo style). Records for tasks.

Reference map: read on demand, not up front

Feature the user wants	Read this
Registration entry point, all AddEverTask options, concurrency/channel sizing, queues	references/01-setup.md
Defining task records + handlers, dispatch overloads, lifecycle callbacks, taskKey idempotency, cancel	references/02-tasks-and-handlers.md
Choosing & configuring a storage provider, audit levels, retention/cleanup, custom ITaskStorage	references/03-storage.md
Retry policies, exception filtering (whitelist/blacklist/predicate), timeout, cancellation, graceful shutdown	references/04-resilience.md
Delayed / scheduled / recurring tasks, fluent builder, cron, idempotent startup registration, drift	references/05-scheduling.md
Keyed (per-tenant) rate limiting, named multi-queues, scalability, sharded scheduler	references/06-rate-limiting-queues.md
Monitoring events, SignalR, dashboard + REST API (JWT), Serilog, persistent execution logs	references/07-monitoring-logging.md
System.Text.Json payload contract + every analyzer rule ET0001–ET0007 + payload checklist	references/08-payload-contract.md
Chaining/orchestrating tasks (continuations, compensation, sagas) via lifecycle callbacks	references/09-orchestration.md
Which NuGet package ships which feature	references/packages.md
Copy-paste code (registration, task+handler, recurring registrar)	templates/*.md

Interactive flow

Run these phases in order. Phases 3–6 only happen if the user wants them.

Phase 0: Detect context (no questions yet)

Look before you ask. Inspect the project to avoid asking what you can read:

• Find the startup project's .csproj and Program.cs (or Startup.cs). Determine the host type: ASP.NET Core web app, Worker Service (Microsoft.NET.Sdk.Worker / AddHostedService), or plain console / Generic Host. This decides whether monitoring (which needs ASP.NET Core endpoint mapping) is even applicable.
• Check the target framework(s): EverTask needs net8.0+. All EverTask packages (core, storage, monitoring, SignalR, Serilog) multi-target net8.0/net9.0/net10.0, so there is no version mismatch to worry about within that range.
• Grep for an existing AddEverTask: if present, this is a modify, not a fresh setup; read the current config and only add what is missing.
• Note whether a DB is already configured (connection string in appsettings.json / Aspire / env) so you can reuse it for storage.

Summarize what you found in one or two lines before moving on.

Phase 1: Gather requirements (AskUserQuestion)

Ask only what the prompt did not already specify. If the user's prompt already says e.g. "recurring job that hits the Stripe API per tenant, store in Postgres", infer storage = Postgres, scheduling = recurring, rate-limiting = yes, and skip those questions.

Use one AskUserQuestion call with the still-open questions below (max 4 per call; if more remain, do a second call). Recommended option goes first, labelled "(Recommended)".

• Storage backend (single-select): drives persistence and packages. Options: SQL Server, PostgreSQL, SQLite (single-server / edge), In-Memory (dev/test only, lost on restart). Recommend based on Phase-0 findings (existing DB wins; otherwise PostgreSQL for greenfield server apps, SQLite for desktop/edge, In-Memory only if they said "just trying it out"). See 03-storage.md for the full decision matrix and the SQLite write-concurrency / large-backlog caveats.

• Which capabilities? (multi-select): only wire what is chosen. Options: Scheduled / recurring tasks, Custom retry / timeout policies, Per-key rate limiting (per tenant/account/resource), Real-time monitoring dashboard, Dedicated Serilog pipeline, Persist handler logs to DB. (Immediate fire-and-forget dispatch always works with no extra wiring, so it is not an option here.)

• Concurrency / throughput profile (single-select): only ask if the user signals scale concerns or asks; otherwise SKIP and keep defaults. Options: Use defaults (Recommended), I/O-bound, expect bursts (raise parallelism + capacity), Need workload isolation (separate critical vs background queues).

Phrase questions in the user's own domain terms when their prompt gave them (e.g. "How many Stripe calls per tenant per minute?" rather than abstract "permits/period").

Phase 2: Install packages & wire AddEverTask

1. Resolve the package set from references/packages.md for the chosen storage + capabilities.
2. Add them via the project's package manager. Check whether the repo uses Central Package Management (a Directory.Packages.props): if so, add <PackageVersion> there and a versionless <PackageReference> in the .csproj; otherwise a normal dotnet add package.
3. Build the registration block from templates/Program.registration.md + 01-setup.md, including ONLY the options the user opted into. Show it as a diff against Program.cs.
4. On confirmation, apply the edit. Place storage connection strings in configuration, never inline literals (read existing config style first).

Phase 3: Scaffold a task + handler

If the user described a concrete task ("send welcome email", "nightly cleanup"), scaffold it; otherwise scaffold one representative example so they have a working pattern.

• Use templates/Task.and.Handler.md. Task = record … : IEverTask with public properties, primitives/Guid/DateTimeOffset, IDs not entities.
• Validate the payload against 08-payload-contract.md before writing: no public fields (ET0001), reachable setters or matching ctor params (ET0002), no Newtonsoft attributes (ET0003), polymorphic props annotated (ET0004), avoid object/Dictionary<string,object> (ET0005), no Stream/DbContext/ValueTuple/delegates (ET0006), class payloads ctor- resolvable (ET0007).
• Handler uses a primary constructor for DI. Each task runs in its own DI scope, so injecting scoped services (e.g. DbContext) is safe.

Phase 4: Wire chosen capabilities

For each capability selected in Phase 1, read the matching reference and apply:

• Scheduling/recurring (05-scheduling.md): pick Dispatch(task, TimeSpan) / Dispatch(task, DateTimeOffset) for one-shots, or the fluent recurring builder. For recurring, register at startup via an IHostedService with a stable taskKey for idempotency (templates/RecurringRegistrar.md). Warn: UseCron(...) overrides all other interval calls; all schedules are UTC.
• Retry/timeout (04-resilience.md): override RetryPolicy / Timeout on the handler, or set queue/global defaults. Default is LinearRetryPolicy(3, 500ms) retrying everything except OperationCanceledException/TimeoutException.
• Rate limiting (06-rate-limiting-queues.md): override RateLimitPolicy on the handler; carry the key via IRateLimitedTask or GetRateLimitKey. Note rate limiting is per-instance: divide external budgets across instances in multi-instance deployments.
• Monitoring (07-monitoring-logging.md): AddMonitoringApi(...) for the full dashboard (web apps only) + app.MapEverTaskApi(), or AddSignalRMonitoring() + MapEverTaskMonitorHub() for events only, or subscribe to IEverTaskWorkerExecutor.TaskEventOccurredAsync in code. Change the default admin/admin credentials.
• Serilog (07-monitoring-logging.md): .AddSerilog(...) for a dedicated pipeline.
• Persistent logs (07-monitoring-logging.md): .WithPersistentLogger(...); pair with AddAuditCleanup(...) retention to prevent unbounded table growth.
• Orchestration / chaining (09-orchestration.md): if the task is one step in a larger flow (continuation, compensation/rollback, fan-out, saga), there is no dedicated workflow API; chain from OnCompleted/OnError by dispatching the next task, passing handoff data as IDs in its payload.

Phase 5: Verify

Run dotnet build (Release; the repo treats warnings as errors and the payload analyzers run at compile time). Fix any ET0001–ET0007 diagnostics on scaffolded payloads before declaring done. If a DB provider was chosen, mention that the first run auto-applies migrations (AutoApplyMigrations = true by default) unless they opted to manage migrations manually.

Common end-to-end shapes (sanity anchors)

• Minimal dev: AddEverTask(o => o.RegisterTasksFromAssembly(typeof(Program).Assembly)).AddMemoryStorage();
• Production web app: Postgres/SQL storage + recurring registrar + monitoring dashboard (MapEverTaskApi) + retention cleanup.
• Worker service: SQL/Postgres storage + recurring registrar; no monitoring dashboard (no HTTP pipeline); use event subscription or SignalR-to-an-external-hub instead.
• Multi-tenant API caller: storage + IRateLimitedTask per tenant + retry policy.

When unsure which detail applies, open the specific reference file rather than guessing; the references are exact (cross-checked against source, with file:line) and call out the known docs-vs-code gotchas (OnLast is not implemented; GetAllRecurringTasksAsync is illustrative).