Files
2026-07-13 12:58:18 +08:00

205 lines
8.9 KiB
C#

// STOPGAP: This integration-level header propagation replaces once copilotkit-sdk-dotnet
// ships (Microsoft contribution, ETA mid-2026). When that SDK lands, delete this code
// and use the SDK's built-in header propagation.
// See: https://www.notion.so/copilotkit/3543aa3818528150b6acc5b872ad7fe5
using System.ClientModel.Primitives;
using Microsoft.AspNetCore.Http;
using OpenAI;
// TODO(copilotkit-sdk-dotnet): migrate to SDK-level header propagation
public class AimockHeaderPolicy : PipelinePolicy
{
// Seeded once at startup from Program.cs (where the DI container exists).
// The policy is created statically via CreateOpenAIClientOptions at
// agent-factory construction time and has no DI access, so it reads the
// request's HttpContext through this seeded singleton accessor — mirroring
// the CvDiag.Logger static-seed pattern. IHttpContextAccessor is a singleton
// that resolves the *current* request's HttpContext via a holder the server
// seeds at request entry; that holder flows across the AG-UI SSE-pump
// ExecutionContext boundary, so the headers the middleware stashed on
// HttpContext.Items are visible here at outbound-call time.
public static IHttpContextAccessor? HttpContextAccessor { get; set; }
public override void Process(PipelineMessage message, IReadOnlyList<PipelinePolicy> pipeline, int currentIndex)
{
ApplyHeadersAndDiag(message);
var (backend, ctx, provider, model) = CvdiagLlmContext(message);
backend?.EmitLlmCallStart(ctx!, provider, model, EstimatePromptTokens(message));
var sw = System.Diagnostics.Stopwatch.StartNew();
string? errorClass = null;
try
{
ProcessNext(message, pipeline, currentIndex);
}
catch (Exception ex)
{
errorClass = ex.GetType().Name;
throw;
}
finally
{
sw.Stop();
backend?.EmitLlmCallResponse(ctx!, provider, model, null, sw.ElapsedMilliseconds, errorClass);
}
}
public override async ValueTask ProcessAsync(PipelineMessage message, IReadOnlyList<PipelinePolicy> pipeline, int currentIndex)
{
ApplyHeadersAndDiag(message);
var (backend, ctx, provider, model) = CvdiagLlmContext(message);
backend?.EmitLlmCallStart(ctx!, provider, model, EstimatePromptTokens(message));
var sw = System.Diagnostics.Stopwatch.StartNew();
string? errorClass = null;
// Heartbeat: emit backend.llm.call.heartbeat every 10s while the outbound
// call is outstanding (spec §3; verbose-tier-and-above). The loop is a
// no-op when CVDIAG is off (backend null) — we skip starting it entirely.
using var heartbeatCts = new CancellationTokenSource();
Task? heartbeat = backend is null ? null : HeartbeatLoop(backend, ctx!, sw, heartbeatCts.Token);
try
{
await ProcessNextAsync(message, pipeline, currentIndex);
}
catch (Exception ex)
{
errorClass = ex.GetType().Name;
throw;
}
finally
{
sw.Stop();
heartbeatCts.Cancel();
if (heartbeat is not null)
{
try { await heartbeat; } catch (OperationCanceledException) { /* expected */ }
}
backend?.EmitLlmCallResponse(ctx!, provider, model, null, sw.ElapsedMilliseconds, errorClass);
}
}
private static async Task HeartbeatLoop(CvdiagBackend backend, CvdiagBackend.RequestContext ctx,
System.Diagnostics.Stopwatch sw, CancellationToken token)
{
try
{
while (!token.IsCancellationRequested)
{
await Task.Delay(TimeSpan.FromSeconds(10), token);
backend.EmitLlmCallHeartbeat(ctx, sw.ElapsedMilliseconds);
}
}
catch (OperationCanceledException)
{
// Outbound call completed; stop heartbeating.
}
}
// Resolve the CVDIAG backend + per-request context + outbound provider/model
// at LLM-call time. Returns a null backend when CVDIAG is off so callers
// skip every emit. The request context flows on the request async tree
// (AsyncLocal), seeded by CvdiagInstrumentationMiddleware at ingress.
private static (CvdiagBackend? Backend, CvdiagBackend.RequestContext? Ctx, string Provider, string Model)
CvdiagLlmContext(PipelineMessage message)
{
var backend = CvdiagBackend.Instance;
if (backend is null || !backend.IsEnabled) return (null, null, "openai", "unknown");
var ctx = CvdiagBackend.CurrentRequestContext;
if (ctx is null) return (null, null, "openai", "unknown");
var host = message.Request.Uri?.Host ?? "";
var provider = host.Contains("openai", StringComparison.OrdinalIgnoreCase) ? "openai"
: host.Contains("azure", StringComparison.OrdinalIgnoreCase) ? "azure"
: "openai";
var model = ExtractModel(message) ?? "unknown";
return (backend, ctx, provider, model);
}
// Best-effort: pull "model":"..." out of the outbound chat-completions body
// without fully parsing it (the body is a BinaryContent we must not consume).
private static string? ExtractModel(PipelineMessage message)
{
try
{
var content = message.Request.Content;
if (content is null) return null;
using var ms = new MemoryStream();
content.WriteTo(ms, default);
var json = System.Text.Encoding.UTF8.GetString(ms.ToArray());
var marker = "\"model\":\"";
var i = json.IndexOf(marker, StringComparison.Ordinal);
if (i < 0) return null;
var start = i + marker.Length;
var end = json.IndexOf('"', start);
return end > start ? json[start..end] : null;
}
catch
{
return null;
}
}
// Rough prompt-token estimate (~4 chars/token) over the outbound body size.
private static int EstimatePromptTokens(PipelineMessage message)
{
try
{
var content = message.Request.Content;
if (content is null) return 0;
using var ms = new MemoryStream();
content.WriteTo(ms, default);
return (int)(ms.Length / 4);
}
catch
{
return 0;
}
}
// Forwards the captured x-* headers onto the outbound LLM request and emits
// the CVDIAG outbound breadcrumb. The headers are read from the current
// request's HttpContext.Items via IHttpContextAccessor — HttpContext flows
// across the AG-UI SSE-pump ExecutionContext boundary, so the value the
// middleware stashed is still visible here at outbound-call time. This layer
// appends its hop tag to x-diag-hops on the outbound call.
private static void ApplyHeadersAndDiag(PipelineMessage message)
{
var headers = AimockHeaderContext.Get(HttpContextAccessor?.HttpContext);
foreach (var header in headers)
{
if (string.Equals(header.Key, CvDiag.HeaderDiagHops, StringComparison.OrdinalIgnoreCase))
continue; // set below with this layer's hop appended
// Add-if-absent: preserve correlation IDs and any headers set by prior policies/SDK.
if (!message.Request.Headers.TryGetValue(header.Key, out _))
message.Request.Headers.Set(header.Key, header.Value);
}
// GATING RULE: only deviate from original control flow (append the
// x-diag-hops breadcrumb, emit the per-outbound CVDIAG log) when a
// diagnostic header is actually present. On non-diagnostic traffic the
// outbound request stays byte-identical to pre-instrumentation behavior
// (the inbound x-* forward loop above is original behavior).
bool diagnosticPresent = headers.ContainsKey(CvDiag.HeaderDiagRunId)
|| headers.ContainsKey(CvDiag.HeaderAimockContext);
if (diagnosticPresent)
{
headers.TryGetValue(CvDiag.HeaderDiagHops, out var existingHops);
message.Request.Headers.Set(CvDiag.HeaderDiagHops, CvDiag.AppendHop(existingHops, "backend-ms-agent-harness-dotnet"));
CvDiag.LogOutbound("backend-ms-agent-harness-dotnet", headers, CvDiag.HopCount(existingHops));
}
}
/// <summary>
/// Creates an <see cref="OpenAIClientOptions"/> with the header forwarding policy
/// pre-configured. All OpenAI client instantiations should use this to ensure
/// x-* prefixed headers propagate to outgoing calls.
/// </summary>
// TODO(copilotkit-sdk-dotnet): migrate to SDK-level header propagation
public static OpenAIClientOptions CreateOpenAIClientOptions(string endpoint)
{
var options = new OpenAIClientOptions
{
Endpoint = new Uri(endpoint),
};
options.AddPolicy(new AimockHeaderPolicy(), PipelinePosition.PerCall);
return options;
}
}