Inside Aspire
The Architecture Behind the AI Loop
Ground decisions. Observe outcomes. Ship intent.
Start presentation
Wisconsin .NET UG
November 17, 2026 at 6pm Central
Make browser behavior observable.
WithBrowserLogs
Console output, exceptions, network events, and screenshots become evidence.
Configure the tracked browser at runtime. Follow the reviewed guide into the CLI and the developer's next decision.
Experimental ASPIREBROWSERLOGS001 in Aspire.Hosting.Browsers. A documented extra capability, not wired into this weather app.
Browser evidence has a visible trail.
- microsoft/aspire#16531Configure tracked browser command
- pr-docs-check run 25092695241Successful existing workflow run
- microsoft/aspire.dev#775Bot draft, maintainer refinement, merge
- browser-logs.mdPublished guide, retrieved by the CLI
Different skills. Visible human review.
doc-writer is a distinct skill.gh aw status "pr-docs-check" \
--repo "microsoft/aspire" --jsondoc-writer guides documentation authoring. aspire-orchestration and aspire-monitoring guide precise app operations.
Actual refinement commit: moved the guide, corrected scope, configuration precedence, and browser sharing. Nothing is dispatched on stage.
Proposals have explicit write boundaries.
pr-docs-check.mdgh aw compilepr-docs-check.lock.yml
on:
pull_request:
types: [closed]
branches:
- main
- release/*
safe-outputs:
create-pull-request:
title-prefix: "[docs] "
draft: true
target-repo:
"microsoft/aspire.dev"
protected-files: blocked{
"number": 775,
"state": "closed",
"merged": true,
"merged_by": {
"login": "davidfowl"
}
}Start with a healthy app.
An exact member, not a plausible guess.
{
"id": "csharp/aspire.hosting.browsers/browserlogsbuilderextensions/methods#withbrowserlogs",
"name": "WithBrowserLogs(IResourceBuilder<T>, string?, string?, BrowserUserDataMode?)",
"language": "csharp",
"kind": "member"
}return markdownResponse(
renderCSharpMemberKindMarkdown(
routeProps.pkg,
routeProps.type,
routeProps.memberKind,
routeProps.allTypes,
base
)
);The response identifies the member and returns its reference content. The renderer preserves that identity in published Markdown.
Actual returned member / reference renderer. Selected fields, not fabricated output.
Guidance chooses. The CLI executes.
Skill
File-based instructions that guide workflow and tool choice.
aspire-monitoring/SKILL.mdA skill selects the observation; it does not replace the tool.
CLI tool
Executable operations that return current evidence.
aspire wait apiaspire describeaspire otel traces api
Public CLI 13.6.1. Exact operations and evidence.
Skills choose the next move.
Route and create
aspireaspire-initaspireifyaspire-project-v2-migrationRun and inspect
aspire-orchestrationaspire-monitoringaspire-deploymentaspire-orchestration selects the exact start/wait target. aspire-monitoring selects the logs and traces that ground the next decision.
Fetch current text. Reuse unchanged work.
using var request = new HttpRequestMessage(
HttpMethod.Get, url);
if (!string.IsNullOrEmpty(cachedETag))
{
request.Headers.IfNoneMatch.Add(
new EntityTagHeaderValue(cachedETag));
}var currentFingerprint =
SourceContentFingerprint.Compute(
content, IndexSchemaVersion);
await _docsCache.SetIndexAsync(
[.. documents], cancellationToken)
.ConfigureAwait(false);The narrative-docs path obtains the cached ETag, reuses matching fingerprints, and parses only when needed. Default corpus: llms-full.txt.
ETag source / cache/index source. Selected lines, wrapped for display.
The parser knows fences.
var span = content.AsSpan();
var codeBlocks = FindCodeBlockRegions(span);
// Find all document boundaries (line indices where H1 headers start)
var docBoundaries = FindDocumentBoundaries(span, codeBlocks);
// Find all section headings (H2+) using the pre-computed fence regions
// passed down from ParseAsync.
var sectionStarts = FindSectionHeadings(docSpan, codeBlocks);Headings inside code fences must not split the document.
Inspect the actual C# parser excerpt in docs/code. The rehearsal uses Aspire's parser, not a clone.
Find the relevant narrative section.
private const float TitleWeight = 10.0f;
private const float SummaryWeight = 8.0f;
private const float HeadingWeight = 6.0f;
private const float CodeWeight = 5.0f;
private const float BodyWeight = 1.0f;Configure browser logs
Rank prose and sections, then retrieve the content. This is not the separate compiler-derived API index.
Pinned narrative weights. Local lexical ranking, not embeddings or model training.
Wire an executable, not a fallback.
var api = builder.AddExecutable("api", executable, nativeDirectory)
.WithHttpEndpoint(env: "PORT")
.WithHttpHealthCheck("/health")
.WithExternalHttpEndpoints()
.WithOtlpExporter(OtlpProtocol.HttpProtobuf);
builder.AddViteApp("frontend",
Path.Combine(builder.AppHostDirectory, "..", "frontend"))
.WithNpm(install: false)
.WithReference(api.GetEndpoint("http")).WaitFor(api)
.WithEnvironment("API_HTTP", api.GetEndpoint("http"))
.WithEnvironment("API_HTTPS", "");Actual AppHost: excerpt omits surrounding configuration. Dynamic ports, readiness, references, and OTLP are explicit.
Observe a running AppHost.
OTLP adds structured logs, traces and metrics.
Official dashboard overview and image source. Starting a C# or TypeScript AppHost launches the integrated dashboard; emitting telemetry still requires instrumentation.
Same viewer. No AppHost required.
aspire dashboard run --non-interactivedocker run --rm `
-p "127.0.0.1:18888:18888" `
-p "127.0.0.1:4317:18889" `
-p "127.0.0.1:4318:18890" `
mcr.microsoft.com/dotnet/aspire-dashboard:latestAny instrumented OTLP emitter can send telemetry. A standalone viewer does not orchestrate your processes.
Documented startup options, not recorded executions. Container ports map UI / gRPC / HTTP; loopback binding is a safety adaptation. Browser authentication remains enabled. The CLI command is foreground/blocking.
Instrumentation makes the evidence.
What happened, with structured context.
Where this request went across boundaries.
How behavior changes over time.
Instrument code, export OTLP, then inspect actual results.
Integrated resource service / independent telemetry service. Designed for development and short-term diagnostics, not unlimited production retention.
Start from deployable intent.
AppHost modelTarget integrationEnvironment resource
aspire add docker --non-interactivevar environment = builder
.AddDockerComposeEnvironment("env")
.WithDashboard();Declare deployable application resources in that model.
Explicitly labeled Docker Compose example, not a user-chosen target or the weather AppHost. AddDockerComposeEnvironment / WithDashboard. No integration or graph was changed here.
Publish describes the target.
aspire publish --apphost 'path\to\deployment\AppHost.cs' `
--output-path 'publish-output' --non-interactiveCompose publishing generates docker-compose.yaml and .env.
It does not build images or apply the deployment.
Documented publish pipeline. Public CLI 13.6.1 syntax; substitute a configured, deployment-ready AppHost. No generated output is presented as a capture.
Deploy applies the model.
aspire deploy --apphost 'path\to\deployment\AppHost.cs' `
--environment '<environment>' --non-interactiveDeploy works from the AppHost model, not an old publish folder.
Compose fills parameters, builds images, and runs Compose up.
Documented deploy pipeline. It deploys declared resources, not the AppHost process. No provisioning or deployment was performed for this presentation.
Same intent. Explicit boundaries.
Choose the supported integration and environment.
Resolve parameters, secrets, authentication and allowed side effects.
Inspect the deployed resources and observed health.
aspire-deployment guides the workflow; the CLI and target provider perform it.
Documented target setup / unchanged shipped deployment skill. Same model does not mean identical target behavior.
Aspire makes AI
more efficient.
AI loops
Aspire dashboard
Aspire deployment
Get the code
Earlier architecture.
Current surface.
This rehearsal
llms-full.txt, CLI docs commands, and a separate API-reference index.
Optional appendix. Reusable fetching, parsing, indexing, and retrieval outlive a particular tool surface.
Reported documentation outcomes
82feature-docs PRs merged
44.8hmedian docs time-to-merge
Reported window: May 3-June 2, 2026.
Historical human-reviewed data, not a fresh SLA.
Optional appendix. Source and measurement context