[RichardoMrMu]

What

loongsuite-instrumentation-strands now tolerates the framework-owned OpenTelemetry baggage detach
that Strands runs when a stream is closed from another context. This fixes the
loongsuite-instrumentation-strands-latest failures on Python 3.10-3.13.

Root cause

strands-agents 1.60.0 (published 2026-10-12T19:37Z) added an invocation level baggage context to
Agent.stream_async:

• strands/agent/agent.py:1427-1432 - baggage_ctx = context_api.get_current(), set_baggage(...)
  per configured attribute, then baggage_token = context_api.attach(baggage_ctx)
• strands/agent/agent.py:1479-1480 - finally: context_api.detach(baggage_token)

ContextVar.reset only accepts a token in the context that created it, so when the application
closes the async generator elsewhere the framework's finally runs in a copied context and
opentelemetry.context.detach logs Failed to detach context and raises
ValueError: <Token ...> was created in a different Context.

The plugin wraps Agent.stream_async and closes the wrapped stream, so closing it with
asyncio.create_task(stream.aclose()) (which copies the current context) reaches that finally from
another context. tests/test_instrumentor.py::test_generator_exit_from_early_break_cleans_state_and_next_sibling,
which asserts that stream cleanup leaves no detach failure behind, therefore started to fail with
1.60.0 only.

Observed while checking this branch (base 2026-10-13T03:26:35Z):

• failing: run https://github.com/alibaba/loongsuite-python/actions/runs/38364194877 (head 402984cecafb),
  jobs 115185430508 (3.13), 115185430132 (3.12), 115185429758 (3.11), 115185429600 (3.10) - every
  strands-latest job failed, strands-oldest (1.50.2) stayed green in the same run.
• green control for the same tests: run https://github.com/alibaba/loongsuite-python/actions/runs/38215548191
  (strands-agents 1.59.0).

No existing issue tracks this regression: #54 is the original instrumentation feature request, and the
failure was first noticed while #275's CI ran (unrelated to that PR's changes).

Fix

New src/opentelemetry/instrumentation/strands/_compat.py:

• BaggageContextCompat replaces strands.agent.agent.context_api with a proxy only when the
  installed framework accepts baggage_attributes (so strands-agents < 1.60.0 is left untouched),
  and restores the original binding on uninstrument().
• _TolerantContextApi pairs every framework attach with a private ContextVar probe. detach
  resets the real token when the probe proves the token is still valid in the current context; when
  it is not, the impossible reset is skipped and nothing is logged. Tokens the proxy did not hand out
  are delegated unchanged.

Tolerating the exception after the fact is not enough: opentelemetry.context.detach logs
Failed to detach context before it re-raises, so the error record would still appear. The probe
decides before the call and uses only a private ContextVar, so the OTel logging path is never
entered.

_instrument installs the shim fail-open (a failure logs a warning and instrumentation continues),
_uninstrument restores the binding, and the README/CHANGELOG describe the behaviour.

Verification

Local runs at the same repository revision, only strands-agents differing (-p no:cacheprovider):

environment	before this change	with this change
strands-agents 1.60.0	1 failed, 31 passed (test_generator_exit_from_early_break_cleans_state_and_next_sibling, same ValueError as CI)	37 passed
strands-agents 1.59.0	green	35 passed, 2 skipped
strands-agents 1.50.2 (oldest)	green	35 passed, 2 skipped

RED discriminator: with BaggageContextCompat.install disabled, the same test fails again on 1.60.0
(1 failed), so the shim is what makes it pass.

New tests cover the cross-context detach being tolerated while the value stays attached in the
attaching context, unrelated ValueError still propagating, frameworks without baggage_attributes
never being patched, the instrumentor patching and restoring context_api, and the framework's
invocation baggage attach still happening once per invocation without leaking afterwards.

python -m ruff check instrumentation-loongsuite/loongsuite-instrumentation-strands and
python -m ruff format --check are clean (ruff 0.14.1, as pinned in dev-requirements.txt).

Notes

• The cross-context case replaces the framework's reset with one extra ContextVar.set binding on the
  framework's own binding; the plugin's own tokens keep their exact pairing.
• Upstream could avoid this by detaching inside the context that attached the token, or by skipping
  the detach when the context changed; the shim keeps the instrumentation green until then.
• Unrelated observation while testing: with the instrumentation installed, the framework's invocation
  baggage is current for invoke_agent span start, but the plugin's stream context hooks attach their
  own context, so nested plugin spans and model calls do not see that baggage. That is pre-existing
  behaviour and is unchanged here.

[CLAassistant]

All committers have signed the CLA.

[sipercai]

Validated the exact head with before/after reproduction of cross-task stream closure, the real Strands/OpenAI SDK reasoning and tool loop (two user turns per session, concurrent streaming sessions), cancellation and clean sibling/context checks, and source-to-cloud trace readback. The detach regression is fixed without changing business output, span count or token totals. The original plugin suite also passes against Strands 1.62.0 (37 tests). The separate pre-existing reasoning/Agent-output field issues are tracked in #282. Fresh-cache tox -e precommit passes with a clean worktree.