feat(corpus): honor A2AT_LANGUAGE for prompts and case files

a2a-t-corpus/README.md

@@ -9,14 +9,14 @@ It currently covers **Task-T**, **Negotiation-T** (RAN energy-saving target / in
 Three things to know before use:
 
 - **Test-only**: lives under `a2a-t-corpus/`, is not packaged into the `a2a_t` wheel, and is deselected in CI with `-m 'not live'`.
-- **A real LLM is required**: when the three required `A2AT_LLM_*` settings are missing it fails with an actionable error; the `.env` keys match the project-root `env.example`.
+- **A real LLM is required**: when the three required `A2AT_LLM_*` settings are missing it fails with an actionable error; the `.env` keys (`A2AT_LANGUAGE`, `A2AT_LLM_*`) reuse the project-root `env.example` names.
 - **The case format is shared with the Java repository**: case files, the schemas under `schemas/`, and the tools under `tools/` are kept identical across both repositories.
 
 ## 2. Module directory
 
 ```
 a2a-t-corpus/
-├── env.example                      configuration template, with the three required A2AT_LLM_* entries
+├── env.example                      configuration template, with A2AT_LANGUAGE and the three required A2AT_LLM_* entries
 ├── .env                             local configuration (secrets, gitignored; copied from env.example)
 ├── README.md                        English guide
 ├── README_zh.md                     Chinese guide
@@ -220,6 +220,7 @@ python a2a-t-corpus/tools/inputCsvToJson.py --reverse --csv a2a-t-corpus/suites/
 
 - **Only `input_case_*.json` is committed**: the CSV design table is for human review, `input_case_*.json` is what the code runs; they hold the same content, but only the JSON is committed.
 - **Adding a case needs no new code**: just create a scenario directory (`suites/<extension>/resources/<scenario>/`) and put the `input_case_*.json` in it; the suite discovers scenarios at collection time. Afterwards run the matching `test_*_self_guard.py` for a structural check.
+- **Non-default languages**: set `A2AT_LANGUAGE` in `a2a-t-corpus/.env` (or the process environment) and place the translated file next to the Chinese one as `input_case_from_text_<language>.json` / `input_case_from_data_<language>.json`, for example `input_case_from_text_en-US.json`. Keep the existing `caseDimension` / `caseCategory` vocabulary (for example `正常` / `异常`) so the report statistics and the self-guard category contract stay consistent. When no case file exists for the configured language, the workflow suites skip with an actionable message; no code change is needed once the file is added.
 
 ### 4.3 Case file format
 
@@ -254,6 +255,8 @@ All commands run from the repository root. First prepare the LLM configuration (
 cp a2a-t-corpus/env.example a2a-t-corpus/.env
 ```
 
+`A2AT_LANGUAGE` selects the prompt-resource language and the case-file variant. It defaults to `zh-CN`; set `en-US` to use the English templates. The corpus reads `a2a-t-corpus/.env` and the process environment (the project-root `.env` is not read). English cases live next to the Chinese ones as `input_case_from_text_en-US.json` / `input_case_from_data_en-US.json`; when no case file exists for the configured language, the workflow suites skip with an actionable message. Non-default languages write `output_result_<flow>_<language>.json` and `summary_report_<flow>_<language>.json`, so the Chinese run results are not overwritten.
+
 **Run every case**:
 
 ```
@@ -288,7 +291,7 @@ uv run pytest a2a-t-corpus/suites/task/test_task_t_from_text_workflow.py -m live
 - For Negotiation-T / Notification-T, just replace the suite path with `a2a-t-corpus/suites/negotiation` or `a2a-t-corpus/suites/notification`; the scenario name supports wildcards too.
 - Every executed case prints to the console as it completes; failed or crashed cases additionally print their full call record. Result files reflect only the current run, and a filtered run overwrites only the matched cases.
 - `--corpus-output-dir=<dir>` redirects the result files and the summary report to a given directory; by default result files write back into each scenario directory and the summary report goes to `a2a-t-corpus/.corpus/`.
-- Each scenario produces `output_result_from_text.json` / `output_result_from_data.json`, aggregated into `summary_report_<flow>.json` and a console table (pass rate overall / by scenario / by dimension / by category / by API step, with the 95% target marked, plus latency p50/p95 and token totals).
+- Each scenario produces `output_result_<flow>.json` (for non-default languages `output_result_<flow>_<language>.json`), aggregated into `summary_report_<flow>.json` and a console table (pass rate overall / by scenario / by dimension / by category / by API step, with the 95% target marked, plus latency p50/p95 and token totals).
 - Without an LLM, run the structural self-guards first (these run in CI):
 
 ```

a2a-t-corpus/README_zh.md

@@ -9,14 +9,14 @@ a2a-t-corpus 是 A2A-T SDK 的数据驱动准确性验证模块。它按用例
 使用前需了解三点：
 
 - **仅用于测试**：位于 `a2a-t-corpus/`，不打包进 `a2a_t`，CI 中以 `-m 'not live'` 排除。
-- **需要真实 LLM**：三个必填配置 `A2AT_LLM_*` 缺失时会直接报错并给出提示；`.env` 的键名与项目根 `env.example` 一致。
+- **需要真实 LLM**：三个必填配置 `A2AT_LLM_*` 缺失时会直接报错并给出提示；`.env` 的键名（`A2AT_LANGUAGE`、`A2AT_LLM_*`）与项目根 `env.example` 一致。
 - **用例格式与 Java 仓库共用**：用例文件、`schemas/` 下的 Schema、`tools/` 下的工具，在两个仓库中保持一致。
 
 ## 2. 模块目录
 
 ```
 a2a-t-corpus/
-├── env.example                      配置模板，含三个必填 A2AT_LLM_* 项
+├── env.example                      配置模板，含 A2AT_LANGUAGE 与三个必填 A2AT_LLM_* 项
 ├── .env                             本地配置（密钥，gitignore；由 env.example 复制）
 ├── README.md                        英文说明
 ├── README_zh.md                     中文说明
@@ -220,6 +220,7 @@ python a2a-t-corpus/tools/inputCsvToJson.py --reverse --csv a2a-t-corpus/suites/
 
 - **代码仓库只提交 `input_case_*.json`**：CSV 设计表供人评审用，`input_case_*.json` 供代码运行用，两者内容一致，但只需要提交 JSON。
 - **新增用例不需要新增代码**：只需新建测试场景目录（`suites/<扩展>/resources/<场景>/`）并放入 `input_case_*.json`，套件在收集阶段会自动发现场景。新增后可用对应扩展的 `test_*_self_guard.py` 做一次结构检查。
+- **非默认语言**：在 `a2a-t-corpus/.env`（或进程环境变量）中设置 `A2AT_LANGUAGE`，并把翻译后的用例文件与中文文件放在同一场景目录，命名为 `input_case_from_text_<语言>.json` / `input_case_from_data_<语言>.json`，例如 `input_case_from_text_en-US.json`。`caseDimension` / `caseCategory` 建议沿用现有词表（如 `正常` / `异常`），以保持报表统计口径和 self-guard 分类契约一致。若配置语言下没有用例文件，workflow 套件会以明确的 skip 信息提示；补齐文件后无需改代码即可运行。
 
 ### 4.3 用例文件格式
 
@@ -254,6 +255,8 @@ python a2a-t-corpus/tools/inputCsvToJson.py --reverse --csv a2a-t-corpus/suites/
 cp a2a-t-corpus/env.example a2a-t-corpus/.env
 ```
 
+`A2AT_LANGUAGE` 同时决定提示词资源语言和用例文件变体，默认 `zh-CN`；设置为 `en-US` 后使用英文模板。corpus 读取的是 `a2a-t-corpus/.env` 和进程环境变量（不会读取项目根目录的 `.env`）。英文用例与中文用例放在同一场景目录，命名为 `input_case_from_text_en-US.json` / `input_case_from_data_en-US.json`；若配置语言下没有对应用例文件，workflow 套件会以明确的 skip 信息提示。非默认语言的结果文件为 `output_result_<flow>_<语言>.json`、汇总为 `summary_report_<flow>_<语言>.json`，不会覆盖中文运行结果。
+
 **执行全部用例**：
 
 ```
@@ -288,7 +291,7 @@ uv run pytest a2a-t-corpus/suites/task/test_task_t_from_text_workflow.py -m live
 - Negotiation-T / Notification-T 只需把套件路径换成 `a2a-t-corpus/suites/negotiation` 或 `a2a-t-corpus/suites/notification`，场景名同样支持通配符。
 - 每条用例执行完成会立即打印到控制台；失败或崩溃的用例会额外打印完整的调用记录。结果文件只反映本次运行，过滤运行时只覆盖命中的用例。
 - `--corpus-output-dir=<目录>` 可把结果文件和汇总报表写到指定目录；默认结果文件写回各场景目录，汇总报表写到 `a2a-t-corpus/.corpus/`。
-- 每个场景生成 `output_result_from_text.json` / `output_result_from_data.json`，并汇总为 `summary_report_<flow>.json` 与控制台表格（总体 / 按场景 / 按维度 / 按分类 / 按 API 步的通过率，标注 95% 目标，含耗时 p50/p95 与 Token 合计）。
+- 每个场景生成 `output_result_<flow>.json`（非默认语言为 `output_result_<flow>_<语言>.json`），并汇总为 `summary_report_<flow>.json` 与控制台表格（总体 / 按场景 / 按维度 / 按分类 / 按 API 步的通过率，标注 95% 目标，含耗时 p50/p95 与 Token 合计）。
 - 未配置 LLM 时，可先运行结构自检（此测试进 CI）：
 
 ```

a2a-t-corpus/conftest.py

@@ -9,6 +9,8 @@
   fails fast with an actionable error exactly when someone actually tries to run the live suites.
 - ``pytest_generate_tests``: parametrizes the workflow suite modules over the discovered scenario
   cases (test id = ``<scenario>/<case id>``), the ``@TestFactory`` counterpart.
+- ``A2AT_LANGUAGE`` (default ``zh-CN``) selects the prompt language and the case-file variant, for
+  example ``input_case_from_text_en-US.json``; a language without any committed case file skips.
 
 Suite modules declare four constants: ``EXTENSION_FOLDER``, ``INPUT_CASE_FILE``, ``FLOW_TYPE`` and
 ``RUNTIME_KIND`` (``"task"``, ``"negotiation"`` or ``"notification"``); ``API_NAMES`` defaults to the
@@ -21,7 +23,8 @@
 from typing import Any
 
 import pytest
-from engine.constants import TASK_API_NAMES
+from engine.config import resolve_language
+from engine.constants import DEFAULT_LANGUAGE, TASK_API_NAMES, localized_case_file
 from engine.discover import ScenarioScanner
 from engine.loader import CaseFileLoader
 
@@ -52,6 +55,12 @@ def _runtime_kind(request: pytest.FixtureRequest) -> str:
     return getattr(request.module, "RUNTIME_KIND", "task")
 
 
+@pytest.fixture(scope="session")
+def corpus_language() -> str:
+    """The configured prompt and case-file language, resolved before the LLM runtime loads."""
+    return resolve_language()
+
+
 @pytest.fixture(scope="session")
 def task_workflow_runtime() -> Any:
     """The shared Task-T SDK runtime, assembled lazily on first use.
@@ -128,14 +137,19 @@ def workflow_engine(request: pytest.FixtureRequest, workflow_runtime: Any) -> An
 
 
 @pytest.fixture(scope="module")
-def workflow_context(request: pytest.FixtureRequest, summary_output_dir: Path | None) -> Any:
+def workflow_context(
+    request: pytest.FixtureRequest,
+    summary_output_dir: Path | None,
+    corpus_language: str,
+) -> Any:
     """The result collector of one suite module; the transcript writes on teardown."""
     from engine.suite import WorkflowRunContext
 
     flow_type = getattr(request.module, "FLOW_TYPE", None)
     if not flow_type:
         pytest.fail("workflow suite modules must declare FLOW_TYPE", pytrace=False)
-    context = WorkflowRunContext(flow_type)
+    run_flow_type = flow_type if corpus_language == DEFAULT_LANGUAGE else f"{flow_type}_{corpus_language}"
+    context = WorkflowRunContext(run_flow_type)
 
     def finalize() -> None:
         context.write_outputs(output_dir_override=summary_output_dir, summary_dir_override=summary_output_dir)
@@ -149,16 +163,28 @@ def pytest_generate_tests(metafunc: pytest.Metafunc) -> None:
     if not _WORKFLOW_FIXTURES.issubset(metafunc.fixturenames):
         return
     extension_folder = getattr(metafunc.module, "EXTENSION_FOLDER", None)
-    flow_file_name = getattr(metafunc.module, "INPUT_CASE_FILE", None)
-    if not extension_folder or not flow_file_name:
+    base_flow_file = getattr(metafunc.module, "INPUT_CASE_FILE", None)
+    if not extension_folder or not base_flow_file:
         return
 
+    language = resolve_language()
+    flow_file_name = localized_case_file(base_flow_file, language)
     scenario_filter = metafunc.config.getoption("--corpus-scenario")
     case_filter = metafunc.config.getoption("--case-filter")
-    scenarios = ScenarioScanner.filter_scenarios(
-        ScenarioScanner.discover(extension_folder, flow_file_name),
-        scenario_filter,
-    )
+    scenarios = ScenarioScanner.discover_optional(extension_folder, flow_file_name)
+    if not scenarios:
+        reason = (
+            f"no {language} corpus case file found: expected "
+            f"suites/{extension_folder}/resources/<scenario>/{flow_file_name}; add the "
+            f"language-specific case file or set A2AT_LANGUAGE={DEFAULT_LANGUAGE}"
+        )
+        metafunc.parametrize(
+            ("scenario", "case"),
+            [pytest.param(None, None, marks=pytest.mark.skip(reason))],
+            ids=["none"],
+        )
+        return
+    scenarios = ScenarioScanner.filter_scenarios(scenarios, scenario_filter)
     loader = CaseFileLoader()
     parameters: list[tuple[Any, Any]] = []
     ids: list[str] = []

a2a-t-corpus/engine/assembler.py

@@ -2,9 +2,9 @@
 
 Every component is produced by the production builders; the sole test seam is the recording client
 injected through the ``llm_client`` parameter. The unified ``A2ATConfig`` is constructed directly
-from the corpus test configuration (packaged zh-CN prompt resources, corpus attempt limit), so the
-module never reads the production ``.env`` loading paths. A separate runtime is provided for each
--T extension the corpus exercises (Task-T / Negotiation-T / …).
+from the corpus test configuration (packaged prompt resources of the configured language, corpus
+attempt limit), so the module never reads the production ``.env`` loading paths. A separate
+runtime is provided for each -T extension the corpus exercises (Task-T / Negotiation-T / …).
 """
 
 from __future__ import annotations
@@ -45,9 +45,6 @@
     "task_runtime",
 ]
 
-#: Prompt language fixed by the corpus (the packaged zh-CN resources serve the business payloads).
-_LANGUAGE = "zh-CN"
-
 #: Fixed LLM client-config entries the corpus bridge does not override.
 _HISTORY_WINDOW = 10
 _SESSION_MAX_TOTAL = 300
@@ -87,7 +84,7 @@ def task_runtime() -> Runtime:
     recorder = _new_recorder(env)
 
     config = A2ATConfig(
-        prompt=PromptRuntimeConfig(language=_LANGUAGE, source_type="packaged"),
+        prompt=PromptRuntimeConfig(language=env.language, source_type="packaged"),
         input_limits=InputLimitConfig(),
         llm=LlmRuntimeConfig(max_attempts=env.max_attempts),
     )
@@ -120,7 +117,7 @@ def negotiation_runtime() -> NegotiationRuntime:
     recorder = _new_recorder(env)
 
     orchestrator = NegotiationGenerationOrchestratorBuilder(
-        language=_LANGUAGE,
+        language=env.language,
         llm_client=recorder,
         max_attempts=env.max_attempts,
     ).build()
@@ -150,7 +147,7 @@ def notification_runtime() -> NotificationRuntime:
     recorder = _new_recorder(env)
 
     config = A2ATConfig(
-        prompt=PromptRuntimeConfig(language=_LANGUAGE, source_type="packaged"),
+        prompt=PromptRuntimeConfig(language=env.language, source_type="packaged"),
         input_limits=InputLimitConfig(),
         llm=LlmRuntimeConfig(max_attempts=env.max_attempts),
     )