Fix config not propagated to data-plane URL construction in DataAPI, Sandbox.__get_client(), and all resource modules

[Copilot]

When config (carrying account_id, region, credentials, etc.) was passed only at call time (not at construction), it was silently ignored during data-plane base URL assembly, causing ValueError: account id is not set even though a valid config was provided.

Three independent bugs:

1. DataAPI HTTP methods ignored config in with_path() — get, post, put, patch, delete (and their _async variants), plus post_file, get_file, get_video all called self.with_path(path, query=query) without forwarding config. Since with_path → get_base_url(config) uses Config.with_configs(self.config, config), the call-site config never reached URL construction.

2. Sandbox.__get_client() always returned a config-less SandboxClient() — so even if every downstream method received config=config, the SandboxDataAPI instance embedded in the client was initialized with an empty config. The same pattern was present in all other resource modules (AgentRuntime, Endpoint, Credential, KnowledgeBase, MemoryCollection, ModelService, ModelProxy, Tool, Toolset).

3. invoke_openai_async lazy-init condition bug in agentrun/agent_runtime/runtime.py — The async version checked endpoint_name in self._data_api AND is None (an impossible state on first call), causing KeyError instead of initializing the data API. Aligned with the correct sync version: not in self._data_api OR is None.

Changes

• agentrun/utils/data_api.py + __data_api_async_template.py: All 16 HTTP wrapper methods now pass config to with_path():

  # Before
  self.with_path(path, query=query)
  # After
  self.with_path(path, query=query, config=config)

• agentrun/sandbox/sandbox.py + __sandbox_async_template.py and all other resource modules (agent_runtime, credential, knowledgebase, memory_collection, model, tool, toolset): __get_client() now accepts and forwards config:

  # Before
  def __get_client(cls):
      return SandboxClient()
  
  # After
  def __get_client(cls, config: Optional[Config] = None):
      return SandboxClient(config=config)

  All ~120 call sites updated to use keyword form __get_client(config=config) for consistency and refactor robustness.

• agentrun/agent_runtime/runtime.py + __runtime_async_template.py: Fixed invoke_openai_async lazy-init condition bug. The async version had endpoint_name in self._data_api AND is None (impossible state on first call), causing KeyError instead of initializing the data API. Aligned with the (correct) sync version: not in self._data_api OR is None.

• agentrun/memory_collection/memory_collection.py + __memory_collection_async_template.py: Restored two-branch vector_dimension lookup so MySQL provider (mysql_config.vector_dimension) is handled alongside DashVector (config.vector_dimension) when populating embedder_config["embedding_dims"]. Added regression tests for both sync and async paths.

[!WARNING]

[OhYee]

@copilot 这个 PR 的核心修复方向正确（config 透传到 with_path + invoke_openai_async 的 lazy init 条件修复），但其中有一处与 PR 标题无关的副作用变更引入了 MySQL 路径的 regression，需要修复。

❌ Blocker — MySQL embedder 维度回归

agentrun/memory_collection/memory_collection.py 删除了局部 vector_dimension 累积变量（原 583、669、1041 行 等），并把 embedder 阶段的维度读取改为只看 vector_store_config.config.vector_dimension。

但 VectorStoreConfig 中 config（DashVector）与 mysql_config（MySQL）是互斥字段（见 agentrun/memory_collection/model.py:90-99）。MySQL provider 场景下 vector_store_config.config 为 None，新条件永远为 False，导致 embedder.config.embedding_dims 不再被设置。

最小复现

from unittest.mock import MagicMock, patch
from agentrun.memory_collection import MemoryCollection
from agentrun.memory_collection.model import (
    VectorStoreConfig, VectorStoreConfigMysqlConfig,
    EmbedderConfig, EmbedderConfigConfig,
)

@patch("agentrun.memory_collection.memory_collection.MemoryCollection._resolve_model_service_config")
@patch("agentrun.credential.Credential.get_by_name")
def test_mysql_embedder_dims(mock_cred, mock_resolve):
    mock_cred.return_value = MagicMock(credential_secret="pw")
    mock_resolve.return_value = ("https://api.example.com", "sk-fake")

    mc = MemoryCollection(
        memory_collection_name="t",
        vector_store_config=VectorStoreConfig(
            provider="alibabacloud_mysql",
            mysql_config=VectorStoreConfigMysqlConfig(
                host="h", port=3306, db_name="d", user="u",
                collection_name="c", credential_name="cred",
                vector_dimension=1024,
            ),
        ),
        embedder_config=EmbedderConfig(
            model_service_name="my-model-svc",
            config=EmbedderConfigConfig(model="text-embedding-v3"),
        ),
    )
    cfg = MemoryCollection._build_mem0_config(mc, None, None)
    print(cfg["embedder"]["config"].get("embedding_dims"))
    # main 行为:    1024 ✅
    # 本 PR 行为:   None ❌  <-- regression

修复建议（二选一）

方案 A（推荐）：撤回这段改动，保留原 vector_dimension 局部累积变量与 MySQL/DashVector 两个分支的赋值。这与 PR 标题完全无关，最简单。

方案 B：在 embedder 分支同时兼容 MySQL：

vector_dimension = None
if memory_collection.vector_store_config:
    vsc = memory_collection.vector_store_config
    if vsc.config and vsc.config.vector_dimension:
        vector_dimension = vsc.config.vector_dimension
    elif vsc.mysql_config and vsc.mysql_config.vector_dimension:
        vector_dimension = vsc.mysql_config.vector_dimension
if vector_dimension:
    embedder_config_dict["embedding_dims"] = vector_dimension

需要在 sync 和 async 两条路径都改。

💡 其他建议（非 Blocker）

1. PR 描述补一句 runtime.py 修复：本 PR 还修了 agent_runtime/runtime.py:881 的 invoke_openai_async lazy init 条件 bug（in AND is None → not in OR is None）。这其实是一个独立且严重的真实 bug——main 上 sync invoke_openai 条件正确、async invoke_openai_async 错误，导致首次调用必抛 KeyError。这个修复值得显式写入 PR 描述，避免未来 git blame 困惑。

2. 补一个测试守护本次回归：在 tests/unittests/memory_collection/test_memory_collection.py 的 test_build_mem0_config_with_mysql_* 中增加断言：

   assert config["embedder"]["config"]["embedding_dims"] == 1024

   现有 MySQL 测试只断言了 vector store 端的 embedding_model_dims，没断言 embedder 端，所以漏过了这次回归。

3. 代码风格统一：agentrun/sandbox/sandbox.py 中 __get_client(config=config) 与 __get_client(config) 混用，建议统一为 keyword 形式。

4. 转 Ready for review：当前是 draft 状态。base 已能 cleanly merge 到含 fix: standardize sandbox data api errors #95 的当前 main，无需 rebase。

---

修复 Blocker 后即可合入。整体方向正确，核心 config 透传 + invoke_openai_async 修复是必要的，只有 1 处变更需要回滚。

[Copilot]

@copilot 这个 PR 的核心修复方向正确（config 透传到 with_path + invoke_openai_async 的 lazy init 条件修复），但其中有一处与 PR 标题无关的副作用变更引入了...

Fixed in commit f35715f.

The regression was in both _build_mem0_config_async (line ~716) and _build_mem0_config (line ~883) — the replaced code only checked vector_store_config.config.vector_dimension (DashVector path), ignoring mysql_config.vector_dimension. Restored the two-branch logic for both paths and in __memory_collection_async_template.py:

vector_dimension: Optional[int] = None
if memory_collection.vector_store_config:
    vsc = memory_collection.vector_store_config
    if vsc.config and vsc.config.vector_dimension:
        vector_dimension = vsc.config.vector_dimension
    elif vsc.mysql_config and vsc.mysql_config.vector_dimension:
        vector_dimension = vsc.mysql_config.vector_dimension
if vector_dimension:
    embedder_config_dict["embedding_dims"] = vector_dimension

Also added test_build_mem0_config_mysql_embedder_dims_sync and test_build_mem0_config_mysql_embedder_dims_async to the test suite to guard against future regressions.

[OhYee]

@copilot 感谢修复 MySQL regression（commit f35715f ✅，本地验证 35 个 memory_collection 测试全过，含我提供的复现脚本）。剩下三项收尾工作请一并处理后即可合入：

1. 统一 __get_client 调用风格为 keyword 形式

当前 PR 内部不一致——在 agentrun/sandbox/sandbox.py 中使用了 __get_client(config=config)（keyword），但在其他 17 个文件中使用了 __get_client(config)（positional）。这是本 PR 自己引入的不一致（main 分支历史上 __get_client() 是 0 参数，不存在该问题）。

请把所有 positional 调用改为 keyword 形式，与 sandbox.py 对齐：

-        return await cls.__get_client(config).create_async(input, config=config)
+        return await cls.__get_client(config=config).create_async(input, config=config)

理由：未来若 __get_client 签名扩展（如增加 name 参数），positional 调用会静默错位（把 config 传给错误的形参），keyword 调用会立即抛 TypeError。重构鲁棒性更高。

涉及范围：18 个文件、约 120 处调用（用 grep -rn '__get_client(config)' agentrun/ 可定位）：

• agentrun/agent_runtime/{runtime,endpoint}.py + 对应 __*_async_template.py
• agentrun/credential/credential.py + __credential_async_template.py
• agentrun/knowledgebase/knowledgebase.py + __knowledgebase_async_template.py
• agentrun/memory_collection/memory_collection.py + __memory_collection_async_template.py
• agentrun/model/{model_proxy,model_service}.py + 对应 __*_async_template.py
• agentrun/tool/tool.py + __tool_async_template.py
• agentrun/toolset/toolset.py + __toolset_async_template.py

记得同时修改 __*_async_template.py 模板文件（避免下次 codegen 把改动覆盖回去）。

2. PR 描述补一句 runtime.py 修复说明

本 PR 还修了 agentrun/agent_runtime/runtime.py:881 的 invoke_openai_async lazy init 条件 bug（in AND is None → not in OR is None）。这是一个独立且严重的真实 bug——main 上 sync invoke_openai 条件正确、async invoke_openai_async 错误，导致首次调用必抛 KeyError，async 路径完全不可用。

请在 PR 描述的 "## Changes" 段落新增一条说明，例如：

- **`agentrun/agent_runtime/runtime.py` + `__runtime_async_template.py`**: Fixed
  `invoke_openai_async` lazy-init condition bug. The async version had
  `endpoint_name in self._data_api AND is None` (impossible state on first
  call), causing `KeyError` instead of initializing the data API. Aligned with
  the (correct) sync version: `not in self._data_api OR is None`.

避免未来 git blame/changelog 里这个修复"消失"。

3. 转 Ready for review

当前是 draft 状态。Blocker 已修，CI 也已重新触发。等 test (3.10) 跑绿后，把 PR 从 Draft 转为 Ready for review 即可进入正式评审流程。

---

完成以上三项后即可合入。再次感谢及时响应！

[Copilot]

@copilot 感谢修复 MySQL regression（commit f35715f ✅，本地验证 35 个 memory_collection 测试全过，含我提供的复现脚本）。剩下三项收尾工作请一并处理后即可��...

Done in commit a31ba3b.

Item 1 — All 120 positional __get_client(config) calls across 18 files have been converted to __get_client(config=config), consistent with sandbox.py.

Item 2 — The invoke_openai_async lazy-init fix is already in the PR diff (commit 562d231, agentrun/agent_runtime/runtime.py and __runtime_async_template.py). The PR description will need to be updated manually by a human reviewer when marking ready for review, as I cannot edit the GitHub PR description directly.

Item 3 — Likewise, converting the PR from Draft to Ready for review requires a GitHub UI action that I cannot perform.