Logaffe.Client 0.2.2

Logaffe.Client

Delivery of log entries to a logaffe installation — a self-hostable, central logging tool for a single operator and their AI agent.

This is the layer underneath the two logging-framework packages, and it is the one to reach for when the application logs through neither of them:

• Logaffe.Serilog — a Serilog sink
• Logaffe.Extensions.Logging — an ILoggerProvider

Use

using Logaffe.Client;

await using var delivery = new EntryDelivery(new EntryDeliveryOptions
{
    Installation = new Uri("https://logs.example.com"),
    IngestToken  = "logaffe_ingest_…",
});

delivery.Send("""
    {"@t":"2026-08-13T14:12:03.417Z","@mt":"Order {OrderId} shipped","OrderId":4711}
    """);

Send takes one CLEF line and returns immediately. It never throws and never blocks.

A delivery never names a project: the token is the project. The address is scheme and host as the operator reaches the installation — the ingest path is appended and is not a setting.

What it promises

Fire-and-forget, and it means it. Entries go into a bounded in-memory queue and a background loop delivers them. The queue drops the oldest entries when it is full rather than growing, because an unbounded queue turns a logging outage into an outage of the application, which is the one thing this exists to prevent.

Nothing is guaranteed to arrive. There is no durable client-side buffer and no retry that outlives the process. This is affordable because logaffe is additive: the application keeps its own file logging, so a delivery that is lost costs a convenience and never the record.

Failures are reported to you, not swallowed. OnFailure is a delegate rather than an ILogger, so this package asks nothing of the application's logging stack. An installation that could not be reached cannot be told that it could not be reached, so the report belongs in the local log that already exists.

Disposal flushes, with a deadline. Dispose/DisposeAsync spends up to FlushTimeout delivering what is still queued. What does not go in that time is lost — which is what fire-and-forget means.

Settings

The batch limits — a thousand entries, five mebibytes — are product values rather than sender settings, and are applied for you.

An overload takes your own HttpClient when the application manages its own.

The format

One JSON object per line. @t (an instant with an offset or Z) and @mt (a message template) are required; @l defaults to Information; @x carries an exception as text. Everything not beginning with @ is a property.

The server renders the template, so @m is refused rather than ignored — there is one place where rendering happens. Placeholders are substituted only where a property of the same name was delivered; everything else stays character for character, because log content is untrusted and an application logging a raw request body will eventually log braces.

instance, SourceContext, TraceId and SpanId are ordinary properties that the installation promotes to indexed fields when present. Supplying none of them is fully supported.

Full detail: docs/ingestion.md.

Requirements

.NET 8 or later. MIT licensed.