Enable Phoenix tracing for Prompt Agent (hosted) targets

[tangym]

Problem

target.trace: { backend: phoenix } was silently ignored for Prompt Agent (model-only) targets. The trace config was only consumed inside the if target.is_callable: branch in rollout.py, so HostedSession never set up instrumentation.

Fix

When target.trace is set for a hosted target, call phoenix.otel.register(auto_instrument=True) before building the session. This detects the installed openinference-instrumentation-litellm package and monkey-patches litellm so every acompletion call emits spans to Phoenix automatically.

• Idempotent: module-level _hosted_trace_registered flag ensures it runs once per process
• Graceful degradation: logs a warning if arize-phoenix-otel is not installed
• No new dependencies required (uses existing optional deps)

Testing

Ran a smoke config with azure/gpt-5.4-nano + trace: { backend: phoenix }. Confirmed:

• Log: Enabled Phoenix auto-instrumentation for hosted session tracing
• 33 traces landed in Phoenix default project
• All rollouts completed successfully

[Jake Present (jakepresent)]

This looks directionally good to me. I ran a local compile plus targeted runtime/import tests: 140 passed, 1 skipped.

I’d wait to approve/merge until #23 lands and this is rebased, since this currently touches p2m/stages/rollout.py and #23 moves that surface to inference.py / test-case terminology. After the rebase, the only thing I’d want to sanity-check is that the hosted-target trace path still registers Phoenix instrumentation in the renamed inference stage. A tiny unit test around “hosted target + target.trace registers once” would be nice, but I don’t think it needs to block if the smoke result holds.

[Jake Present (jakepresent)]

Found one issue while doing a partial review before the rebase.

The existing targeted tests still pass locally:

• uv run python -m compileall -q p2m
• uv run pytest tests/test_runtime_modes.py tests/test_framework_agnostic.py tests/test_stage_runner_smoke.py -q
• 137 passed, 1 skipped

But directly exercising _ensure_hosted_trace_instrumentation() crashes in my env. phoenix.otel.register(auto_instrument=True) attempts to auto-instrument all installed OpenInference instrumentors, not just LiteLLM. In this checkout it imported the CrewAI/ChromaDB stack and failed with a Pydantic config error before the hosted session was built.

That means target.trace on a hosted target can currently turn into a hard runtime crash, even though the intended behavior is graceful degradation. Also, _hosted_trace_registered = True is set before registration succeeds, so a failed registration attempt can leave the process in a “registered” state without instrumentation.

I think the fix is either:

1. instrument only LiteLLM explicitly, rather than auto_instrument=True across every installed OpenInference integration, or
2. catch broader registration failures, log a warning, and only mark the registration as successful after it actually succeeds.

After that plus the post-#23 rebase from rollout.py to inference.py, I’m good with the direction.

[tangym]

Thanks for the review! Updated with all the fixes:

1. auto_instrument=True switched to explicit register() + LiteLLMInstrumentor().instrument() so we control exactly what gets instrumented and avoid the crash you flagged.

2. Flag timing fixed: _hosted_trace_registered is now set after both register() and instrument() succeed, so a failure mid-setup correctly leaves the flag False for retry on next call.

3. Broader exception handling: added a general except Exception with exc_info=True alongside the ImportError catch, so unexpected failures from Phoenix/OTel config issues get logged instead of bubbling up.

4. Rebased onto main: resolved the renaming from refactor(p2m): terminology migration to maximize broad-developer recognition #23.

Also added unit tests (tests/test_hosted_trace_registration.py) covering the happy path (register-once idempotency), ImportError (missing deps), and RuntimeError (config failure).