fix(agentscope): support QwenPaw v2 task handoffs

[sipercai]

Description

QwenPaw 2 advances the AgentScope reply async generator through a heartbeat
layer that may call __anext__() and aclose() from different
asyncio.Task contexts. The original AgentScope v2 instrumentation kept
OpenTelemetry and probe-owned ContextVar.Token values across yields, so a
later Task could attempt to reset a token created by another Context.

QwenPaw 2 also closes downstream AgentScope iterators when its ENTRY stream is
closed. Before this follow-up, AgentScope treated the resulting
GeneratorExit as a tool/ReAct failure, and nested LLM/STEP/TOOL spans could
lose the active AGENT parent after a Task handoff.

This change:

• removes the probe-owned ReAct round ContextVar lifecycle and derives the
  1-based round from AgentScope v2's agent.state.cur_iter;
• instruments QwenPaw 2 qwenpaw.runtime.runtime.Runtime.run while preserving
  QwenPaw 1 and CoPaw compatibility;
• keeps ENTRY, AGENT, ReAct, and tool contexts attached only while one
  underlying __anext__() or aclose() is executing, then detaches them in
  the same Task before yielding to the caller;
• treats GeneratorExit from a normal downstream aclose() as successful
  completion while preserving a real error raised by the downstream close;
• restores the QwenPaw ENTRY context while it closes the downstream iterator;
• uses the shared hook_advice fail-open boundary for probe-owned lifecycle
  operations while leaving business iteration and close operations outside
  advice;
• records QwenPaw 2 TTFT only after a non-empty text/tool delta or completed
  assistant output, ignoring response-created, in-progress, and heartbeat
  envelopes;
• preserves the original business result, yielded chunks, cancellation, and
  exception identity when instrumentation callbacks fail;
• adds a combined QwenPaw Runtime plus real AgentScope v2 middleware regression
  that validates ENTRY -> AGENT -> LLM and
  ENTRY -> AGENT -> STEP -> TOOL across fresh Tasks;
• pins mcp<2 only in the AgentScope-oldest and QwenPaw v1/legacy test
  environments because those framework versions use the MCP 1.x API.

This addresses the cross-Context crash, missing QwenPaw 2 request root,
incorrect GeneratorExit error statuses, and broken nested AgentScope
parentage reported in #252. Duplicate instrumentation remains an independent
follow-up and is not claimed as fixed here.

Addresses #252

Type of change

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• This change requires a documentation update

How Has This Been Tested?

• Reproduced the old behavior on the previous PR head: successful tool and
  ReAct spans ended as ERROR / GeneratorExit, the nested child was parented
  directly to ENTRY, and cross-Task token detach was logged.
• Re-ran the same lifecycle reproduction on the follow-up: tool and ReAct
  spans end as UNSET, the child is parented to AGENT, and
  child_parent_is_agent=true.
• AgentScope latest:
  py311/py312/py313-test-loongsuite-instrumentation-agentscope-latest —
  24 passed per environment.
• AgentScope oldest:
  py310/py313-test-loongsuite-instrumentation-agentscope-oldest —
  94 passed per environment.
• QwenPaw 2.0.1:
  py312-test-loongsuite-instrumentation-qwenpaw-latest —
  22 passed, 4 skipped.
• QwenPaw 1 and CoPaw legacy:
  py310-test-loongsuite-instrumentation-qwenpaw-{v1,legacy} —
  14 passed, 1 skipped per environment.
• Fresh-cache tox -e precommit.
• Static LoongSuite PR readiness checker, branch-diff privacy scan,
  cassette privacy scan, and git diff --check.
• Weaver live-check of emitted AgentScope AGENT/LLM/STEP/TOOL spans:
  no violations; only existing development-stability and provider-enum advice.

Validation Evidence

Spec and Scope

• Linked issue/spec: QwenPaw 2.0.0.post4 instrumented is error：loongsuite_agentscope_v2_react_round created in a different Context #252.
• Changed surface: AgentScope v2 async stream Context ownership, normal-close
  semantics, QwenPaw 2 downstream close behavior, compatibility test
  dependencies, focused tests, and changelogs.
• Explicit non-goal: duplicate instrumentation is not changed by this PR.

Local Checks

Check	Command	Result	Notes
Static readiness	python3 "$PIPELINE_SKILL_DIR/scripts/check_loongsuite_pr_readiness.py" --repo .	pass	No readiness errors or warnings.
Precommit	PRE_COMMIT_HOME="$(mktemp -d /tmp/loongsuite-precommit.XXXXXX)" uvx --from 'tox<5' tox -e precommit	pass	Fresh hook cache; no hook-generated changes.
AgentScope latest	uvx --from 'tox<5' tox -c tox-loongsuite.ini -p auto -e py311-test-loongsuite-instrumentation-agentscope-latest,py312-test-loongsuite-instrumentation-agentscope-latest,py313-test-loongsuite-instrumentation-agentscope-latest	pass	24 passed per environment.
AgentScope oldest	uvx --from 'tox<5' tox -c tox-loongsuite.ini -p auto -e py310-test-loongsuite-instrumentation-agentscope-oldest,py313-test-loongsuite-instrumentation-agentscope-oldest	pass	94 passed per environment.
QwenPaw latest	uvx --from 'tox<5' tox -c tox-loongsuite.ini -e py312-test-loongsuite-instrumentation-qwenpaw-latest	pass	22 passed, 4 skipped.
QwenPaw compatibility	uvx --from 'tox<5' tox -c tox-loongsuite.ini -p auto -e py310-test-loongsuite-instrumentation-qwenpaw-v1,py310-test-loongsuite-instrumentation-qwenpaw-legacy	pass	14 passed, 1 skipped per environment.
Plugin lint	uvx --from 'tox<5' tox -c tox-loongsuite.ini -p auto -e lint-loongsuite-instrumentation-agentscope,lint-loongsuite-instrumentation-qwenpaw	pass	Both ruff environments passed.
Privacy scan	git diff origin/main -- . | rg -q '<credential-and-local-path-patterns>'	pass	No credentials, tokens, private paths, or user-name matches.
Independent review	bounded implementation review	partial	Blocking lifecycle and compatibility findings were applied; the final automated refresh timed out and is not counted as a pass.

Business Isolation

Scenario	Status	Evidence
prepare/start fault	pass	Injected Agent start failure preserves the business stream and leaves no current-span leak.
response mapping/stop fault	pass	Finish callback failure does not replace the business result.
fail callback fault	pass	The original business exception object remains observable.
normal early close/GeneratorExit	pass	Fresh-Task aclose() closes downstream once; ENTRY/AGENT/STEP/TOOL remain UNSET.
downstream close error	pass	The original close exception object is re-raised and affected spans are marked ERROR with the real error type.
cancellation	pass	Cancellation identity is preserved and Agent cleanup completes.
cross-task stream	pass	__anext__() and aclose() run in fresh Tasks without an instrumentation detach warning or current-span leak.
clean sibling after fault	pass	Sequential requests finish as separate root traces with no leaked parent.

Functional Matrix

Scenario	Status	Evidence
non-streaming model call	pass	AgentScope v2 focused model-call tests.
streaming reply/tool flow	pass	Cross-Task reply and acting async-generator tests.
agent/tool/ReAct	pass	Combined QwenPaw Runtime plus real AgentScope middleware test emits AGENT, LLM, STEP, and TOOL.
error path	pass	Business iteration errors, cancellation, callback faults, and close errors are covered.
concurrency/context isolation	pass	Fresh Tasks advance and close streams; concurrent tools share the same agent iteration round.
dependency bounds	pass	AgentScope latest/oldest plus QwenPaw latest/v1/legacy environments pass.

Real E2E Matrix

Scenario	Status	Command or Demo	Evidence
non-streaming	N/A	QwenPaw 2 Runtime.run is an async iterator only.	The final-message-only path is covered against installed QwenPaw 2.0.1.
streaming	blocked	LOONGSUITE_PYTHON_SITE_BOOTSTRAP=True OTEL_TRACES_EXPORTER=console OTEL_METRICS_EXPORTER=none qwenpaw app	A disposable provider-backed workspace was not configured for this follow-up.
concurrency	blocked	Same app command, then submit concurrent turns.	Cross-Task isolation is covered deterministically; provider-backed concurrency was not rerun.
agent/tool/ReAct	blocked	Same app command, then submit a tool-using turn.	The real middleware composition is covered locally; provider-backed tool execution was not rerun.
tool-heavy	blocked	Same app command, then submit a bounded multi-tool turn.	Requires a configured tool-capable workspace.
error path	blocked	Same app command with a provider/tool failure fixture.	Business and probe error isolation are covered locally; provider-backed failure was not rerun.

Telemetry and Weaver

Check	Status	Command or Artifact	Notes
Span tree / span kinds	pass	Combined QwenPaw Runtime plus AgentScope middleware regression	Validates ENTRY -> AGENT -> LLM and ENTRY -> AGENT -> STEP -> TOOL.
Normal-close statuses	pass	Cross-Task combined regression	ENTRY, AGENT, LLM, STEP, and TOOL remain UNSET; no error.type.
Error statuses	pass	Downstream close-error regressions	Preserves the real exception and records ValueError.
Context isolation	pass	Fresh-Task __anext__()/aclose() regressions	No instrumentation detach log or current-span leak.
Weaver live-check	pass	weaver registry live-check -r <registry>/model --advice-profile loongsuite-genai --input-format json ...	AgentScope AGENT/LLM/STEP/TOOL sample had no violations. Development-stability and existing agentscope provider-enum advice remain.
QwenPaw custom attributes	follow-up	Full ENTRY sample	Existing qwenpaw.* and copaw.* custom attributes still need registry registration or a compatibility decision.

CI

• GitHub checks: all checks passed on remote head 3ffb1c50.
• The local lifecycle follow-up commit is not yet pushed; pushing it will
  trigger a fresh CI run.
• Known unrelated failures: none.
• Merge state remains Draft / review required.

Does This PR Require a Core Repo Change?

• Yes. - Link to PR:
• No.

Checklist:

See contributing.md for styleguide, changelog guidelines, and more.

• Followed the style guidelines of this project
• Changelogs have been updated
• Unit tests have been added
• Documentation has been updated

[ralf0131]

Summary

Well-structured fix for cross-Task context safety in AgentScope v2 and QwenPaw v2 instrumentation. The refactoring correctly addresses the GeneratorExit handling, context token lifecycle, and nested span parentage issues reported in #252.

Review Highlights

• Correctness: Proper context attach/detach pattern using dataclasses (_AgentState, _ActingState, _EntryState) to track state across async boundaries. The hook_advice fail-open boundary is correctly applied to probe-owned lifecycle operations.
• GeneratorExit handling: Distinguishes between normal downstream aclose() (successful completion) and real errors — critical for QwenPaw v2's heartbeat layer.
• Backward compatibility: Preserves QwenPaw 1 and CoPaw compatibility while adding v2 support.
• Tests: Comprehensive regression tests validate ENTRY -> AGENT -> LLM and ENTRY -> AGENT -> STEP -> TOOL across fresh Tasks.

Minor Note

The large diff is mostly CI workflow configuration (1500+ lines). Core logic changes are focused and well-scoped.

---

Automated review by github-manager-bot