fix(agno): split tool-call loops into react steps

[sipercai]

Description

This fixes Agno 2.x tool-call loop instrumentation so every real model provider request is emitted as its own chat span under a ReAct step, instead of wrapping the whole Agno model response loop as one LLM span.

The change keeps direct Model.response usage covered through a fallback wrapper, but uses Agno's per-request processing hooks when they are available. Streaming agent token totals now fall back to the sum of child LLM spans when Agno does not expose final run metrics.

Fixes no existing issue.

Type of change

• Bug fix (non-breaking change which fixes an issue)

How Has This Been Tested?

Please describe the tests that you ran to verify your changes. Provide instructions so we can reproduce. Please also list any relevant details for your test configuration

• tox -c tox-loongsuite.ini -e py311-test-loongsuite-instrumentation-agno -- -q
• tox -c tox-loongsuite.ini -e py39-test-loongsuite-instrumentation-agno,py313-test-loongsuite-instrumentation-agno -- -q
• tox -c tox-loongsuite.ini -e lint-loongsuite-instrumentation-agno
• tox -e precommit
• python "$PIPELINE_SKILL_DIR/scripts/check_loongsuite_pr_readiness.py" --repo .

Validation Evidence

Spec and Scope

• Linked issue/spec: none
• Approved spec/comment: requested Agno tool-call loop / ReAct span repair
• Changed surface: loongsuite-instrumentation-agno instrumentation hooks, span/token aggregation logic, and focused Agno tests

Local Checks

Check	Command	Result	Notes
Static readiness	python "$PIPELINE_SKILL_DIR/scripts/check_loongsuite_pr_readiness.py" --repo .	pass
Precommit	tox -e precommit	pass
Focused tests	tox -c tox-loongsuite.ini -e py311-test-loongsuite-instrumentation-agno -- -q	pass	36 tests
Version edge tests	tox -c tox-loongsuite.ini -e py39-test-loongsuite-instrumentation-agno,py313-test-loongsuite-instrumentation-agno -- -q	pass	36 tests on each env
Focused lint	tox -c tox-loongsuite.ini -e lint-loongsuite-instrumentation-agno	pass
Privacy scan	changed-file scan for secrets and private trace material	pass	no hits

Real E2E Matrix

Scenario	Status	Command or Demo	Evidence
non-streaming	pass	deterministic Agno 2.x latency tool-call smoke with OTLP export	ARMS/CMS cloud readback verified; LLM/tool durations match injected provider/tool latency
streaming	pass	deterministic Agno 2.x streaming latency tool-call smoke with OTLP export	agent and child LLM token totals verified in ARMS/CMS; streaming LLM span covers chunk interval
concurrency	pass	concurrent Agent.arun(stream=True) smoke	no round/token/context leakage observed
agent/tool/ReAct	pass	deterministic two-round tool-call loop	tree verified as Agent -> Step -> LLM/Tool -> Step -> LLM
tool-heavy	N/A	not part of this bug surface	focused fix covers one tool-call batch per round
error path	pass	focused unit coverage for stream close/cancel cleanup	ContextVar cleanup verified

Telemetry and Weaver

Check	Status	Command or Artifact	Notes
Span tree / span kinds	pass	local in-memory spans plus ARMS/CMS cloud readback	tool spans are siblings of LLM spans under ReAct step
Span timing	pass	local in-memory spans plus ARMS/CMS cloud readback	LLM spans cover provider latency; tool spans cover tool execution latency
Token totals	pass	local in-memory spans plus ARMS/CMS cloud readback	agent total equals sum of child LLM totals
Concurrency isolation	pass	concurrent async streaming smoke	per-run ReAct rounds and token aggregation remain isolated
Weaver live-check	pass	weaver registry live-check --advice-profile loongsuite-genai ...	no violations

CI

• GitHub checks: not run yet
• Known unrelated failures: none known
• Follow-up needed: inspect GitHub Actions after PR creation

Does This PR Require a Core Repo Change?

• 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

Refactors Agno instrumentation to split tool-call loops into individual ReAct step spans (one LLM span per provider request), with contextvar-based run-state tracking and token-usage fallback from child LLM spans. Well-structured, follows existing patterns, and includes comprehensive test coverage (sync/async, stream/non-stream, direct model call).

Findings

• [Info] _wrapper.py:643 — Dead code in _response_for_llm_finish: the if condition is redundant since both branches return assistant_message. See inline comment.

Suggestions

The dead conditional in _response_for_llm_finish can be simplified. If the original intent was to validate assistant_message attributes before returning it, consider returning None in the else branch or removing the conditional entirely.

---

Automated review by github-manager-bot

[ralf0131]

Summary

Re-review after new commit fbc32a18 (latency test coverage). The added latency tests correctly validate that LLM spans cover provider latency (≥25ms for 30ms delay) and that direct Model.response calls without agent context do not emit ReAct step spans. The 5-layer span tree assertion (session > turn > step > {llm, tool}) is well-structured.

The previously noted [Info] finding in _response_for_llm_finish (redundant conditional — both branches return assistant_message) is still present. This is non-blocking; recommend cleaning up in a follow-up.

Overall the implementation is solid: ContextVar-based state coordination is clean, the fallback to outer Model.response wrappers for older Agno versions is well-handled, and test coverage is comprehensive.

---

Automated review by github-manager-bot

[ralf0131]

Summary

Refactors Agno 2.x tool-call loop instrumentation to emit one chat span per provider request under a ReAct step, instead of wrapping the entire model response loop as a single LLM span. Uses Agno's per-request processing hooks (_process_model_response, process_response_stream, etc.) when available, with a clean fallback to outer Model.response wrappers for older versions. Streaming agent token totals now fall back to summing child LLM spans when Agno doesn't expose final run metrics.

Findings

No issues found. The implementation is solid:

• _wrap_model_methods feature detection cleanly prefers inner hooks, falling back with a warning when unavailable
• _AgnoRunState + ContextVar properly tracks react round, active step, and accumulated token usage per run
• State lifecycle is correctly managed in try/except/finally across sync, async, and streaming paths
• _is_stream_close check correctly handles the edge case where stream closure triggers finally without a real error
• Token fallback (_apply_agent_token_fallback) only activates when Agno metrics are None, avoiding double-counting
• _uninstrument properly unwraps both inner and fallback wrappers via _unwrap_if_present
• Excellent test coverage: sync/async/stream/async-stream tool-call loops, latency coverage verification, direct model call (no react step), token aggregation from child spans

CI: all checks passing. Changelog updated. CLA: signed.

---

Automated review by github-manager-bot