Add Open Telemetry instrumentation

[atheriel]

This commit wraps all LLM model calls in an Open Telemetry span that abides by the (still nascent) semantic conventions for Generative AI clients.

It's very similar in approach to what was done for httr2, and in fact the two of them complement one another nicely:
r-lib/httr2#729.

For example:

library(otelsdk)

Sys.setenv(OTEL_TRACES_EXPORTER = "stderr")

chat <- ellmer::chat_databricks(model = "databricks-claude-3-7-sonnet")
chat$chat("Tell me a joke in the form of an SQL query.")

[atheriel]

Traces that mix LLM model call spans with httr2 spans:

[jcheng5]

cc @cpsievert @schloerke @icarusz

[hadley]

Do we want to (optionally?) also include user and assistant messages, a la https://opentelemetry.io/docs/specs/semconv/gen-ai/gen-ai-events/ ?

[atheriel]

@hadley I do. But there's a ton of disagreement in the OTel LLM community about how to do that, and none of the existing instrumentation libraries work in the same way 😞. Plus the whole "structured body" mechanism the current spec proposes (1) isn't supported by the span API; and (2) is formally deprecated.

So I kind of think we need to noodle on what to do there, and I suggest pushing it into follow-up work. I'm planning on writing up an issue describing what options we have.

I also think we should have first-class support for tool call spans, because that's something that ellmer focuses on specifically. This PR is really the "basic" bit that the title implies.

[hadley]

@atheriel ok, that makes sense. I'm sure there will be a lot of learning as we figure out exactly what is most useful to instrument across packages.

[atheriel]

Moving this back to draft because it has known issues (i.e. the concurrency does not work correctly).

[atheriel]

This has been updated to support async operations and for changes in otel and otelsdk. It now also includes pretty extensive unit tests and support for agent and tool call spans.

[hadley]

One other question: do we want to log something about auth here (i.e. in particular, which auth was automatically picked?)

[shikokuchuo]

Just adding a comment that I've adopted @schloerke's suggestions from #848 (with a couple of extensions as I noted) in 8080dc6, and this is a common approach to fixing the tracer across the packages we're instrumenting.

[shikokuchuo]

@schloerke one of the live-api test failures concerns a test on this block:

p1 <- chat$chat_async("What's the current date in Y-M-D format?") |>
  promises::then(function(result) {
    chat$chat_async("What date will it be 47 days from now?")
  })

The test expectation is for both agent spans to be top level.

I'm thinking this was set up pre-otel promise domains, and want to confirm with you that one should indeed be a child span of the other given it executes within a $then() context. Thanks!

[schloerke]

The test expectation is for both agent spans to be top level.

I'm thinking this was set up pre-otel promise domains, and want to confirm with you that one should indeed be a child span of the other given it executes within a $then() context. Thanks!

I'd still expect both chat$chat_async() calls to be top level.

**Investigating

[schloerke]

The error was introduced in eeb0b84 (#526) 🥸

But!.. It's fixed now. Thank goodness for unit tests

---

Restrictions on the other earlier failing test have been relaxed and now pass

[shikokuchuo]

Thanks @schloerke. I've made a couple more tidy-ups, the only consequential one to move otel to 'suggests', consistent with the other packages we're instrumenting.

Apart from the unit tests themsleves, I've also been testing using your shinychat demo (Shiny, ellmer, httr2 and mirai spans) and the traces all look good.

[shikokuchuo]

@hadley this PR is now ready for review. Thanks!