feat(mastra): add Openlayer observability exporter for Mastra

[viniciusdsmello]

Adds OpenlayerExporter, a first-class Mastra observability exporter, so Mastra
applications can send agent, workflow, model, and tool traces to Openlayer with a
small configuration snippet.

Closes OPEN-12306.

import { Mastra } from '@mastra/core';
import { Observability } from '@mastra/observability';
import { OpenlayerExporter } from 'openlayer/lib/integrations/mastra';

export const mastra = new Mastra({
  observability: new Observability({
    configs: {
      openlayer: {
        serviceName: 'my-service',
        exporters: [new OpenlayerExporter()],
      },
    },
  }),
});

With OPENLAYER_API_KEY and OPENLAYER_INFERENCE_PIPELINE_ID set, that is the whole
integration.

Why a rewrite shim was needed

@mastra/otel-exporter already emits OTel GenAI semconv v1.38 — the convention
Openlayer's OTLP ingest reads. But it only sets gen_ai.input.messages /
gen_ai.output.messages on MODEL_GENERATION spans. Openlayer builds a row's input and
output from the root span, which instead carries mastra.<type>.input / .output
and is ignored.

Posting a Mastra-shaped trace unmodified produced correct hierarchy, cost, tokens,
latency and error status — and openlayer_output: {} with openlayer_inputs: [].

So this exporter is a gap-filler, not a translator. @mastra/arize converts gen_ai →
OpenInference and @mastra/langfuse converts gen_ai → langfuse.*; we convert nothing
and only fill in what Mastra omits. The mapping rules were established by probing the
live endpoint, not inferred:

Root-span shape	Row input / output
gen_ai.input.messages + gen_ai.output.messages (1.38 parts)	✅ populated
gen_ai.prompt / gen_ai.completion	❌ empty
gen_ai span events	❌ empty
OpenInference input.value / output.value	❌ empty
Traceloop traceloop.entity.input / .output	❌ empty

OpenInference is deliberately not targeted — the issue lists it as a reference, but
it was measured to yield empty rows against Openlayer.

What ships

File	Purpose
src/lib/integrations/mastra/genaiMessages.ts	Pure: coerce any Mastra value into semconv 1.38 parts messages
src/lib/integrations/mastra/spanRewriter.ts	Pure: every attribute-mapping rule
src/lib/integrations/mastra/otlpExporter.ts	OTLPTraceExporter subclass applying the rewrite in export()
src/lib/integrations/mastra/index.ts	OpenlayerExporter, config resolution
src/lib/integrations/mastra.ts	Flat facade so the module is importable (see below)
examples/mastra-tracing.ts	Runnable agent + workflow example

Dependencies are optional peers only — nothing added to dependencies, so consumers
who don't use Mastra pay nothing.

Notable decisions

• Import path is openlayer/lib/integrations/mastra. scripts/utils/postprocess-files.cjs
  regenerates dist/package.json's exports from a directory scan and discards
  hand-authored subpaths, so openlayer/integrations/mastra would 404 on a real install.
  A flat facade module makes the module reachable via the generated ./lib/* wildcard,
  matching the convention every existing example already uses.
  This also means the shipped ./integrations/claude-agent-sdk export is dead today —
  latent only because it is undocumented. Worth a separate issue.
• Provider-slug normalization. Mastra reports gen_ai.provider.name = openai.responses,
  which has no Openlayer cost-table entry, so every user following the docs would have
  gotten silent $0 cost. PROVIDER_SLUG_ALIASES maps verified dotted slugs to real
  ones; unknown values pass through untouched, and a debug log fires on an unmapped dotted
  slug so a future miss is visible rather than silent.
• MODEL_CHUNK dropped by default, via dropSpanTypes — Mastra emits one span per
  streaming chunk. Deliberately not named excludeSpanTypes, which already exists one
  layer up in Mastra's own config.
• MODEL_STEP spans are kept even though they are ~57% of steps: dropping a span type
  does not reparent its children, and dropping these silently lost the tool-call span.

Testing

• 48 unit tests across three suites — message coercion, every rewrite rule (including the
  negative cases), and config resolution.
• A live end-to-end test, skipped without credentials, that publishes a real trace and
  asserts on the row fetched back: non-empty input content, non-empty output text,
  cost > 0, tokens, model, provider, session id, and the tool call nested in the steps tree.

The live test is the only guard against the empty-I/O defect this integration exists to
fix — every other test uses synthetic data and cannot see it. Five structural-vs-content
false positives were found and fixed while building it, including assertions that would
have passed against an empty row.

Running it live requires NODE_OPTIONS=--experimental-vm-modules (@ai-sdk/openai is
ESM-only); its imports are dynamic so the suite still skips cleanly under a plain
yarn test.

Pre-existing issues found, not fixed here

• ./integrations/claude-agent-sdk is unreachable in the published package (same
  exports codegen cause as above).
• tests/openai-tracer.test.ts fails on main — a function-call output assertion in
  openAiTracer.ts, untouched by this branch.
• yarn test cannot start its mock server in a fresh checkout: .stats.yml carries no
  OpenAPI spec URL, so the tests/api-resources/* suites cannot run.

🤖 Generated with Claude Code

https://claude.ai/code/session_01AX7bmWUeckGwETCQMyt8sp