#485)

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

…r span (#484)

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

…ies (1:1)

Previously _begin_obs skipped the obs wrapper for ANY Temporal activity (Option
A) and only stamped the ambient RunActivity span, so all business spans in a turn
collapsed onto ONE obs span (52:1). But inside a *business* activity, start_span
and end_span run in the SAME process, so a wrapper is safe there. Option A is
only required for the SDK's own dispatched START_SPAN/END_SPAN activities (the
in_temporal_workflow path), where start and end are separate activities on
possibly different workers.

Discriminate on activity type: _in_tracing_dispatch_activity() is true only for
the "start-span"/"end-span" activities. For everything else (sync, or a business
activity) open a real per-step wrapper — it nests under the interceptor's ambient
RunActivity span and closes in-process, giving each business span its own obs
span (1:1), matching the sync path. The bounded _OBS_HANDLES registry backstops
any mis-discrimination.

…P startup

Replace the SDK's inline observability edge (#484/#485/#490/#491) with thin
shims that delegate to the shared sgp-obs library:

- obs_ids.py / obs_span.py / temporal.py / trace.py: delegate correlation, the
  per-step wrapper span, the begin-obs decision, and the Temporal interceptors to
  sgp_obs.traces backends + Correlator.
- adk/pyproject.toml: add sgp-obs==0.2.0rc1; root pyproject: add the CodeArtifact
  `scale` index + explicit source for sgp-obs.

Also wire sgp_obs.traces.init_tracing() into the BaseACPServer lifespan. Without
it, get_tracer() resolves to the API-default ProxyTracerProvider: obs wrapper
spans never record or export, so no span carrying the <source>.business_trace_id
reverse anchor reaches the collector. init_tracing adopts an app-installed
provider if present, else installs one with the OTLP exporter (fail-open).

Validated end-to-end in sgp-dev (rocket-mock + audit): agent registers, 20/20
business spans correlate (obs ids in span metadata), per-step wrapper spans
record with the reverse anchor, and force_flush to the collector succeeds.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

_acp_dispatch_span was per-service boundary glue living in the SDK. The
ACP->Temporal dispatch is a transport boundary, not service logic, so it now
lives once in the library (sgp_obs.traces.dispatch_span). Replace the 45-line
function + its sys/contextlib imports with two `with dispatch_span(...)` calls,
passing the business id as an attribute (agentex.task_id) so the generic library
stays source-agnostic.

Bumps the sgp-obs pin to 0.3.0 (adds dispatch_span + the ALWAYS_ON sampler fix).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Now that init_tracing auto-instruments outbound httpx, the span-export calls get
traced themselves and nest back into the trace being exported — PUT /v5/spans/batch
(SGP) and POST /spans (Agentex), plus the egp auth/db subtree a propagated
traceparent drags in. Wrap both processors' async exports in
sgp_obs.traces.suppress_instrumentation so the export makes no span and injects no
traceparent. Verified in sgp-dev: export-pollution spans went 51 -> 0 while each
business step still shows its real work (e.g. run_agent_streamed -> the LLM call).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

…on by default

Every agentex agent makes outbound HTTP (LLM/egp), so httpx instrumentation is
universal — pull the extra in the SDK so init_tracing's egress instrumentation
(and the export suppression that pairs with it) actually functions without each
agent opting in. DB instrumentation (sgp-obs[db]) stays per-agent opt-in.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Wrap both completion_stream (sync) and acompletion_stream (async) with
sgp_obs.traces.instrument_stream so every litellm streaming call emits one
gen_ai.chat span with TTFT/TTAT events + the decode-window timing (tps, tpot,
output tokens) — filling the "span for TTFT but nothing for the streaming" gap
for non-Temporal / litellm agents. Provider extractors count content-bearing
deltas; fail-open, chunks pass through unchanged.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Wire SGPLLMGateway.{completion_stream,acompletion_stream} through
sgp_obs.traces.instrument_stream so scale-gp streaming gets the same gen_ai.chat
span (TTFT/TTAT + decode-window timing) as litellm. Both gateways yield the same
OpenAI-shaped Completion chunk, so the extractors + attribute builder move to a
shared _stream_obs module (litellm refactored to use it; scale-gp tags
gen_ai.system=scale-gp). Streaming instrumentation is now uniform across adapters.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

After init_tracing, register sgp_obs's openai-agents tracing bridge so a Runner
turn's internal phases (generation/tool/handoff) show up in Tempo under the
business step instead of a dark gap. Fail-open no-op for non-openai-agents agents.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Every agent's message/send and task/create handler is now wrapped with sgp-obs
`business_trace(business_trace_id=lambda p: p.task.id)` at registration, so an
agent gets the full business<->obs correlation edge — the turn anchor plus the
process-local correlation context the openai-agents bridge and auto-egress read
for their reverse tag — with NO per-agent code. Proven manually on analyst-agent
(trace 9ad202d0b0bbdc0a78d8e5719c830811); this bakes it into the base server so
no agent has to decorate.

The decorator preserves the handler's shape (sync / async / async-generator), so
streaming message/send handlers keep streaming.

Fail-open + version-tolerant: `_with_business_trace` no-ops when the installed
sgp-obs predates `business_trace` (defensive import, same stance as init_tracing),
so this is inert until the sgp-obs pin is bumped to the release that ships it
(0.5.0) — no code change needed here at that point.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

…struction

Pass the flag into the sgp_obs Correlator that _begin_obs builds for every
adk business span, so the opt-out actually takes effect on the primary
correlation path. Behavior lives in sgp_obs (scaleapi#158071); this is the
construction pass-through.

Version-resilient (_build_correlator): an installed sgp_obs that predates the
`correlate_business` parameter raises TypeError on the kwarg, so we fall back to
the no-arg form and correlation stays always-on rather than dying. The flag
takes effect once the sgp-obs pin is bumped to the release that wired it
(0.5.0). Confirmed against the currently-pinned 0.3.0: fallback, no crash.

Test: env parsing + no-crash fallback + (when supported) the flag reaching the
Correlator instance.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

… on)

Match sgp_obs env_off: disable ONLY on an explicit {0,false,no,off}; unset,
empty, or an unrecognized value stays ON. Adds "off" to the token set and stops
treating an unrecognized value as True-by-coincidence — now it's True by rule
(opt-out), so a typo can't silently drop correlation. Test covers off-tokens +
empty/typo staying on.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>