feat(otel): include per-span token usage metrics in OTel export

[christso]

Context

AgentV captures detailed token usage from both Copilot CLI (via ACP usage_update) and Claude SDK (input_tokens, output_tokens, cache_read_input_tokens) providers. However, the OTel exporter only includes aggregate trace-level counts as root span attributes, not per-message/per-LLM-call token breakdowns.

Current behavior

In otel-exporter.ts, the root span gets:

• agentv.trace.event_count — total tool calls
• agentv.trace.cost_usd — total cost
• agentv.trace.llm_call_count — LLM call count

Child gen_ai.generation spans get model name, duration, and content — but no token usage.

Proposal

Add token usage attributes to each LLM child span using the exact GenAI semantic convention names:

// In exportMessage(), for assistant messages:
if (msg.tokenUsage) {
  span.setAttribute('gen_ai.usage.input_tokens', msg.tokenUsage.inputTokens);
  span.setAttribute('gen_ai.usage.output_tokens', msg.tokenUsage.outputTokens);
  if (msg.tokenUsage.cacheCreationInputTokens != null)
    span.setAttribute('gen_ai.usage.cache_creation.input_tokens', msg.tokenUsage.cacheCreationInputTokens);
  if (msg.tokenUsage.cacheReadInputTokens != null)
    span.setAttribute('gen_ai.usage.cache_read.input_tokens', msg.tokenUsage.cacheReadInputTokens);
}

Note: The GenAI spec has no cost attribute (gen_ai.usage.cost does not exist). Cost should remain as agentv.trace.cost_usd on the root span only.

Data flow prerequisite

The Message type in packages/core/src/evaluation/providers/types.ts needs tokenUsage if not already present. Check what each provider currently captures:

Provider	Token data available	Where captured
Claude SDK	input_tokens, output_tokens, cache_read_input_tokens, cache_creation_input_tokens	claude.ts — from result.usage
Copilot CLI	used (input tokens), no output breakdown	copilot-cli.ts — from ACP usage_update
Copilot SDK	input_tokens, output_tokens	copilot-sdk.ts — from SDK response
Azure/AI SDK	Varies by model	ai-sdk.ts — from Vercel AI SDK response

If Message.tokenUsage doesn't exist, add it as an optional field and propagate from providers.

Files to modify

1. packages/core/src/evaluation/providers/types.ts — Add tokenUsage?: ProviderTokenUsage to Message type (if missing)
2. packages/core/src/evaluation/providers/claude.ts — Attach per-message token usage to Message objects
3. packages/core/src/evaluation/providers/copilot-cli.ts — Same
4. packages/core/src/observability/otel-exporter.ts — Read msg.tokenUsage in exportMessage() and set attributes
5. Tests — Verify token attributes appear on child spans

Acceptance criteria

• LLM child spans include gen_ai.usage.input_tokens and gen_ai.usage.output_tokens when available
• Cache token attributes use correct names: gen_ai.usage.cache_creation.input_tokens, gen_ai.usage.cache_read.input_tokens
• Root span agentv.trace.cost_usd remains unchanged (no gen_ai.usage.cost — doesn't exist in spec)
• Providers that don't report per-message tokens gracefully omit the attributes (no zeros or nulls)
• Unit tests verify attributes on child spans

References

• GenAI token usage attributes — gen_ai.usage.* section
• AgentV OTel exporter: packages/core/src/observability/otel-exporter.ts:178-253
• Depends on feat(otel): adopt OTel GenAI semantic conventions for trace attributes #298 for attribute naming consistency

Testing Approach

Unit Tests (InMemorySpanExporter)

const exporter = new InMemorySpanExporter();
// Run mock eval with known token counts (e.g., mock provider returns tokenUsage: { input: 100, output: 50, cached: 20 })

const spans = exporter.getFinishedSpans();
const genSpan = spans.find(s => s.attributes['gen_ai.operation.name'] === 'chat');
expect(genSpan.attributes['gen_ai.usage.input_tokens']).toBe(100);
expect(genSpan.attributes['gen_ai.usage.output_tokens']).toBe(50);
expect(genSpan.attributes['gen_ai.usage.cache_read.input_tokens']).toBe(20);

What to Assert

• Per-span token attributes present on gen_ai.chat spans (not just root)
• Cache token attributes only present when provider reports them
• Total tokens on root span match sum of child spans
• Token attributes use correct GenAI convention names (not agentv.*)