ax-go-agent-observability

AxAgent Observability For Go

This skill helps an agent write Go code with the generated Ax package github.com/ax-llm/ax/packages/go. Use the generated package API, examples, and manifests; do not import TypeScript-only APIs unless you are editing the TypeScript package.

When To Use

• Inspect agent traces, runtime envelopes, usage, or action logs.
• Register the process-wide usage observer and attribute model calls by tenant, user, request, run, or feature.
• Attach callbacks for model/tool activity and runtime progress.
• Debug agent loops through generated package state and examples.

Package Facts

• Language: Go.
• Package: github.com/ax-llm/ax/packages/go.
• Package API docs: API.md and axir-api.json.
• Capability manifest: axir-capabilities.json.
• Runnable examples: examples/.
• Real network support: yes.
• Scripted no-key transport support: yes.
• Runtime profiles: javascript-goja.

Core Pattern

helper := ax.NewAgent("question:string -> answer:string", nil)
out := helper.Forward(llm, map[string]ax.Value{"question": "How should I proceed?"}, nil)

Centralized Usage Observer

Use the process-wide usage observer for application accounting across many agents, API routes, tenants, and users. Keep per-agent usage accessors for inspecting one agent instance after a run.

axllm.SetUsageObserver(func(event axllm.AxUsageEvent) {
    usageQueue.Enqueue(event)
})
// Later: axllm.SetUsageObserver(nil)

• The observer receives one normalized event for each completed chat or embedding call that reports provider usage. A fully consumed stream emits once; an unconsumed or cancelled stream may not emit.
• Events include the operation, AI/provider name, model, normalized tokens, streaming flag, optional usage context, and available session or remote request IDs.
• Attach usageContext in AI service options for stable application or environment defaults. Attach it in call or agent-forward option maps for tenant, user, request, run, and feature attribution.
• Per-call context overrides service defaults. Nested attributes are shallow-merged.
• The observer is process-wide, best-effort, and fail-open. Registering again replaces the previous observer. Clear it during test teardown or shutdown when appropriate.
• The observer runs on the request path. Production callbacks should synchronously enqueue into a bounded concurrent queue and return immediately, then persist or aggregate out of band. Use a shared durable pipeline across processes or serverless instances.
• Keep identifiers opaque and attributes low-cardinality. Do not attach prompts, responses, secrets, or other sensitive payloads.
• Calculate currency cost downstream against a versioned provider/model pricing table.
• Runnable provider example: src/examples/go/generation/usage_observer.go.