feat(examples): add Chinese text ecommerce customer service demo

[sipercai]

Description

This PR adds a text-only Chinese-language LangGraph e-commerce customer-service example. It targets LangGraph 1.2+ and uses the supported LangChain 1.x create_agent API. An LLM intent router selects a presales or aftersales ReAct agent, each specialist uses a separate synthetic tool set, and a final LLM reviews the response before it is returned.

All customer-facing runtime content is Simplified Chinese, including prompts, fictional product/order/policy values, tool descriptions and errors, CLI output, example questions, and fail-open responses. Stable Python identifiers, graph node names, route enum values, and JSON schema keys remain in English. The Chinese README includes a ready-to-use training walkthrough for the scenario and trace structure.

The example is CLI-only and contains no image input, file upload, browser UI, customer data, or production commerce integration. All products, orders, policies, and tool results are fictional fixtures.

Fixes # (N/A)

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?

• cd examples/ecommerce-customer-service && uv run --isolated --no-project --with-requirements requirements-dev.txt pytest tests -q
• cd examples/ecommerce-customer-service && uv run --isolated --no-project --with-requirements requirements-dev.txt --with 'langgraph==1.2.10' --with 'langchain==1.3.14' --with 'langchain-openai==1.4.1' pytest tests -q
• tox -e precommit
• python3 .github/scripts/detect_loongsuite_changes.py with the pull-request changed-file set

Does This PR Require a Core Repo Change?

• Yes. - Link to PR:
• No.

Checklist:

• Followed the style guidelines of this project
• Changelogs have been updated (not required for an example-only change; use Skip Changelog)
• Unit tests have been added
• Documentation has been updated

Validation Evidence

Spec and Scope

• Linked issue/spec: direct feature request; no public issue
• Approved spec/comment: implement a text-only Chinese e-commerce support example without customer data
• Changed surface: bilingual example documentation, Chinese CLI/runtime content, LangGraph workflow, synthetic tools, training walkthrough, and offline tests under examples/ecommerce-customer-service

Local Checks

Check	Command	Result	Notes
Static readiness	pipeline readiness checker and changed-file detector	pass	Example-only diff; detector reports no LoongSuite package changes
Precommit	tox -e precommit	pass	Fresh pre-commit cache; clean worktree afterward
Focused tests	commands listed above	pass	10 tests pass with the declared LangGraph 1.2+ requirements and the pinned current stable 1.2.10 stack
Test-first red	N/A	N/A	No instrumentation behavior or bug fix is introduced
VCR replay-only	N/A	N/A	Offline tests use deterministic injected model responses and synthetic tools; no cassette is committed
External review	earlier two-round diff review plus LangGraph v1 migration review	partial	The Chinese localization diff passed two-round review. The v1 delta was checked against the official migration contract and full tests; the external reviewer could not run because its login had expired.
Privacy scan	changed example files	pass	No customer identifiers, credentials, private endpoints, or local paths

Business Isolation

Scenario	Status	Evidence
prepare/start fault	N/A	No instrumentation wrapper or callback lifecycle changed
response mapping/stop fault	pass	Specialist and final-review failures return bounded Chinese fallback responses
fail callback fault	N/A	No instrumentation callback changed
stream chunk/final callback fault	N/A	The example uses non-streaming invocation
early close/cancellation/GeneratorExit	N/A	The example does not expose a stream API
cross-task or cross-Context stream	N/A	The example does not expose a stream API
clean sibling after fault	pass	Concurrent workflow test proves route and answer state remains isolated

Real E2E Matrix

Scenario	Status	Command or Demo	Evidence
non-streaming	pass	Chinese text-only deployed smoke on LangGraph 1.2.10 / LangChain 1.3.14	Eight real Chinese questions returned HTTP 200: presales 3, aftersales 3, clarify 2
streaming	N/A	N/A	The example intentionally exposes a text question/answer invocation only
concurrency	pass	test_concurrent_invocations_keep_routes_and_answers_isolated	Concurrent states retain separate routes and answers
agent/tool/ReAct	pass	Chinese presales and aftersales deployed smoke	Presales and aftersales questions selected different named agents, executed LangGraph v1 model/tool loops, and reached the final reviewer
tool-heavy	pass	Chinese product/inventory/knowledge flow	Product search, product knowledge, and inventory tools returned synthetic Chinese evidence before the reviewed answer
error path	pass	test_specialist_and_reviewer_fail_open and test_cli_reports_a_chinese_error_for_a_blank_question	Specialist/reviewer and invalid-input failures return stable Chinese responses without escaping

Telemetry and Weaver

Check	Status	Command or Artifact	Notes
Span tree / span kinds	pass	Cloud trace readback from the final LangGraph 1.2.10 Pod	Eight fresh business traces were attributed to the final Pod; presales and aftersales samples contain router LLM, selected named AGENT, repeated LLM/TOOL children, and response-review LLM spans
Content capture modes	N/A	N/A	This example does not change instrumentation or content-capture behavior
Concurrency isolation	N/A	N/A	No instrumentation concurrency behavior changed
Weaver live-check	blocked	Registry live-check	The current registry advisor reports a pre-existing duplicated local-variable definition

CI

• GitHub checks: the previous revision passed all executable CodeQL, LoongSuite, docs, precommit, typecheck, and CLA checks
• Known unrelated failures: one stale changelog event was created before the Skip Changelog label; subsequent PR events see the label
• Follow-up needed: inspect the new revision checks after push

[ralf0131]

Summary

This PR adds a well-structured text-only e-commerce customer service example using LangGraph. The workflow correctly implements intent routing, specialist ReAct agents, and a final review step. Tests provide good coverage with deterministic mock models for offline execution.

Verdict: APPROVED — Clean example code, good test coverage, proper error handling.

Observations (non-blocking)

• workflow.py:_route_intent — The bare except Exception is intentional (fail-open), which is appropriate for an example. In production code, consider logging the exception for debugging.
• tools.py — Clean separation of PRESALES_TOOLS and AFTERSALES_TOOLS. The synthetic data is well-structured and clearly fictional.
• tests/ — The DeterministicToolModel and FixedClassifier patterns are excellent for deterministic testing without external dependencies.

---

Automated review by github-manager-bot

[ralf0131]

Re-approved after new commit (localization to Chinese). The changes look good:

• Clean localization of prompts, tool descriptions, fictional data, and CLI output to Simplified Chinese
• Python identifiers, graph node names, and route enums remain in English (good practice)
• All CI checks passing, CLA signed

---

Automated review by github-manager-bot

[ralf0131]

Summary

This PR adds a well-structured Chinese e-commerce customer service demo using LangGraph, showcasing intent routing, specialist ReAct agents with isolated tool sets, and a review step before responding. Code quality is high with solid error handling, comprehensive tests, and clear documentation.

LGTM — nice example that demonstrates the LoongSuite instrumentation capabilities end-to-end.

---

Automated review by github-manager-bot