ap-client 0.4.2.dev0__tar.gz → 0.5.0.dev1__tar.gz

{ap_client-0.4.2.dev0 → ap_client-0.5.0.dev1}/PKG-INFO

@@ -1,18 +1,18 @@
 Metadata-Version: 2.5
 Name: ap-client
-Version: 0.4.2.dev0
+Version: 0.5.0.dev1
 Summary: Agent Platform API Client & CLI
 Requires-Python: >=3.10
-Requires-Dist: irepo-artifact==0.1.2
+Requires-Dist: irepo-artifact==0.2.0
 Requires-Dist: pyyaml>=6.0
 Requires-Dist: requests>=2.28.0
 Requires-Dist: rich>=13.0.0
 Requires-Dist: typer>=0.9.0
 Requires-Dist: websockets>=13.0
 Provides-Extra: all
-Requires-Dist: instance-repo[oss]<2.0.0,>=1.1.2; extra == 'all'
+Requires-Dist: instance-repo[oss]<2.0.0,>=1.1.4; extra == 'all'
 Provides-Extra: dataset
-Requires-Dist: instance-repo[oss]<2.0.0,>=1.1.2; extra == 'dataset'
+Requires-Dist: instance-repo[oss]<2.0.0,>=1.1.4; extra == 'dataset'
 Description-Content-Type: text/markdown
 
 A lightweight Python SDK and command line interface for Agent Platform. It provides helpers for configuring API access and managing templates, datasets, jobs, and groups.

{ap_client-0.4.2.dev0 → ap_client-0.5.0.dev1}/ap_client/akpool_commands.py

@@ -63,7 +63,9 @@ def _fmt(output_format: Optional[str]) -> str:
 def _check_time_range(value: str) -> str:
     normalized = str(value or "").strip().lower()
     if not _TIME_RANGE_RE.fullmatch(normalized):
-        raise typer.BadParameter(f"--time-range must be a duration like 30s/5m/1h, got {value!r}")
+        raise typer.BadParameter(
+            f"--time-range must be a duration like 30s/5m/1h, got {value!r}"
+        )
     return normalized
 
 

{ap_client-0.4.2.dev0 → ap_client-0.5.0.dev1}/ap_client/api.py

@@ -13,6 +13,7 @@ from typing import Any, Literal, Optional, Union
 from urllib.parse import parse_qsl, quote, urlencode, urlsplit, urlunsplit
 
 import requests
+from urllib3.util import Timeout as HTTPTimeout
 
 from .config import Config, ConfigurationError, get_config, has_cluster_header
 from .managed_ak import require_managed_ak_ack, validate_ak_id, validate_ak_selection
@@ -126,7 +127,6 @@ EXTRA_AGG_OPS = ("avg", "sum")
 _OTHERS_MODEL = {
     "name": "Others",
     "series_name": "",
-    "is_default": False,
     "source_id": "",
     "series_id": "",
 }
@@ -171,7 +171,9 @@ def _merge_meta_tags(
     conflicts = [k for k, v in provided.items() if v and str(v).strip() and k in existing_keys]
     if conflicts:
         raise ValueError(
-            "tags 中已包含 " + ", ".join(conflicts) + "，与 meta 参数冲突，请只保留一种方式"
+            "tags already contains "
+            + ", ".join(conflicts)
+            + "; it conflicts with the meta parameter, keep only one"
         )
     extra = [
         f"{key}:{str(value).strip()}"
@@ -343,6 +345,11 @@ def _split_paged(payload: Any) -> tuple[list[dict], dict]:
     )
 
 
+# apiserver meta-tags 端点单页上限（handler.metaTagPaging 截断到 200；
+# 缺省仅 20，目录查询必须显式带上并翻页）。
+_META_TAG_PAGE_SIZE = 200
+
+
 class APIClient:
     """Agent Platform API client."""
 
@@ -401,6 +408,8 @@ class APIClient:
         params: Optional[dict] = None,
         json_body: Any = None,
         timeout: TimeoutType = None,
+        deadline: Optional[float] = None,
+        _return_response: bool = False,
     ) -> Any:
         """Send an HTTP request with X-Request-ID and optional verbose logging."""
         self._require_cluster_config()
@@ -412,6 +421,7 @@ class APIClient:
         verbose = self.config.verbose
         max_attempts = self._max_attempts_for(method)
         for attempt in range(1, max_attempts + 1):
+            attempt_timeout = _deadline_timeout(timeout, deadline)
             request_id = str(uuid.uuid4())
             headers = {"X-Request-ID": request_id}
             if verbose:
@@ -435,9 +445,11 @@ class APIClient:
                     params=params,
                     json=json_body,
                     headers=headers,
-                    timeout=timeout,
+                    timeout=attempt_timeout,
                 )
             except requests.RequestException as exc:
+                if deadline is not None:
+                    _manifest_budget_remaining(deadline)
                 duration_ms = (time.perf_counter() - start) * 1000
                 if verbose:
                     self._log_exception(method, url, request_id, duration_ms, exc)
@@ -453,10 +465,12 @@ class APIClient:
                             delay,
                             error=f"{type(exc).__name__}: {exc}",
                         )
-                    time.sleep(delay)
+                    _deadline_sleep(delay, deadline)
                     continue
                 raise APIError(None, str(exc), request_id, None) from exc
             duration_ms = (time.perf_counter() - start) * 1000
+            if deadline is not None:
+                _manifest_budget_remaining(deadline)
             if verbose:
                 self._log_response(method, url, request_id, resp, duration_ms)
             if not resp.ok:
@@ -472,10 +486,12 @@ class APIClient:
                             delay,
                             status=resp.status_code,
                         )
-                    time.sleep(delay)
+                    _deadline_sleep(delay, deadline)
                     continue
                 detail = _response_error_detail(resp)
                 raise APIError(resp.status_code, detail, request_id, resp)
+            if _return_response:
+                return resp, request_id
             if not resp.content:
                 return None
             try:
@@ -977,21 +993,103 @@ class APIClient:
                 return item
         raise APIError(404, f"meta model {name!r} not found", "client-filter", None)
 
+    def _get_meta_tag_catalog(
+        self, path: str, params: Optional[dict] = None
+    ) -> tuple[list[dict], dict]:
+        """按 benchmarks 同口径自动翻页取全量目录记录，返回 ``(records, pagination)``。
+
+        apiserver ``meta-tags`` 端点分页参数为 ``page`` / ``page_size``（缺省
+        20、上限 200，见 handler.metaTagPaging）；单页取用会把值目录静默截断
+        （``meta-job-type`` 漏掉 ``Eval-Eval`` 时 ``default`` 会被算成空串），
+        故显式传上限并翻页。
+        """
+        records: list[dict] = []
+        pagination: dict = {}
+        page = 1
+        # 与 list_benchmarks 同口径的 100 页上限，防分页异常导致死循环
+        max_pages = 100
+        while page <= max_pages:
+            query = dict(params or {})
+            query["page"] = page
+            query["page_size"] = _META_TAG_PAGE_SIZE
+            page_records, pagination = _split_paged(self._get_central(path, params=query))
+            records.extend(page_records)
+            if not page_records or not pagination:
+                break
+            try:
+                total = int(pagination.get("total"))
+            except (TypeError, ValueError):
+                total = None
+            if total is not None:
+                if len(records) >= total:
+                    break
+            elif len(page_records) < _META_TAG_PAGE_SIZE:
+                break
+            page += 1
+        return records, pagination
+
+    def list_meta_tags(self) -> dict:
+        """meta 标签键目录（apiserver GET /apis/v1/meta-tags/keys，返回 {items:[...]}）。
+
+        apiserver 用 PagedSuccess envelope（{code,data,pagination}）返回，
+        按基准 benchmarks 客户端同口径经 _split_paged 拆包并自动翻页；
+        裸列表等旧形态兜底为空分页。
+        """
+        records, pagination = self._get_meta_tag_catalog("/meta-tags/keys")
+        result = {"items": records}
+        if pagination:
+            result["pagination"] = pagination
+        return result
+
+    def get_meta_tags(self, key: str) -> dict:
+        """单个 meta 键的值目录（GET /apis/v1/meta-tags?key=，返回 {items:[...]}）。
+
+        与 ``list_meta_tags`` 同样自动翻页，避免值目录被单页默认大小截断。
+        """
+        records, pagination = self._get_meta_tag_catalog("/meta-tags", {"key": key})
+        result = {"items": records}
+        if pagination:
+            result["pagination"] = pagination
+        return result
+
     def list_meta_job_types(self) -> dict:
-        """任务类型词表（本地常量，不走接口）。
+        """任务类型词表（目录驱动：GET /apis/v1/meta-tags?key=meta-job-type）。
 
-        job-type 词表固定为 Eval-Eval（默认）+ Others（兜底手动输入），
-        不依赖服务端可用性。default 固定 ``Eval-Eval``（job-type 词表
-        不再走集群配置，``AP_META_DEFAULT_JOB_TYPE`` env 已删）。
+        仅取目录中 ``is_enabled`` 的值（缺失按启用解析）+ 末尾 ``Others``
+        兜底行；``default`` 为 ``Eval-Eval``（在目录中时，否则空串）。
+        目录请求失败或为空 → 回退本地常量（Eval-Eval + Others，fail-open），
+        并在 stderr 打一行告警，词表降级可见但不阻塞命令。
         """
-        default = "Eval-Eval"
+        error: Exception | None = None
+        try:
+            items = self.get_meta_tags("meta-job-type").get("items") or []
+        except Exception as exc:  # fail-open：目录不可用不阻塞词表命令
+            error = exc
+            items = []
+        if items:
+            options = []
+            for item in items:
+                value = str(item.get("value") or "").strip()
+                if value and item.get("is_enabled", True):
+                    options.append({"value": value})
+            if options:
+                default = "Eval-Eval" if any(o["value"] == "Eval-Eval" for o in options) else ""
+                return {
+                    "meta_job_type": {
+                        "default": default,
+                        "options": [*options, {"value": "Others"}],
+                    }
+                }
+        _write_stderr(
+            [
+                "warning: meta-job-type catalog unavailable or empty, "
+                f"falling back to local constants (reason={error or 'empty catalog'})"
+            ]
+        )
         return {
             "meta_job_type": {
-                "default": default,
-                "options": [
-                    {"value": default, "is_default": True},
-                    {"value": "Others", "is_default": False},
-                ],
+                "default": "Eval-Eval",
+                "options": [{"value": "Eval-Eval"}, {"value": "Others"}],
             }
         }
 
@@ -1185,6 +1283,7 @@ class APIClient:
         benchmark: Optional[str] = None,
         experiment: Optional[str] = None,
         ak_id: Optional[str] = None,
+        ak_mode: str = "auto",
         instance_range: Optional[str] = None,
         profile: Optional[str] = None,
     ) -> dict:
@@ -1236,6 +1335,7 @@ class APIClient:
             benchmark=benchmark,
             experiment=experiment,
             ak_id=ak_id,
+            ak_mode=ak_mode,
         )
 
         result = self._post("/jobs", body, timeout=timeout)
@@ -1297,6 +1397,7 @@ class APIClient:
         benchmark: Optional[str] = None,
         experiment: Optional[str] = None,
         ak_id: Optional[str] = None,
+        ak_mode: str = "auto",
         instance_range: Optional[str] = None,
         profile: Optional[str] = None,
     ) -> dict:
@@ -1312,11 +1413,15 @@ class APIClient:
         if not template and not profile_id and profile is None:
             raise ValueError("Either template, profile or profile_id is required")
         validate_ak_selection(ak_id, params, params_list, overrides)
+        if ak_mode not in {"auto", "manual", "synced"}:
+            raise ValueError("ak_mode must be auto, manual or synced")
         body: dict = {}
         if template is not None:
             body["template"] = template
         if ak_id is not None:
             body["ak_id"] = ak_id
+        if ak_mode != "auto":
+            body["ak_mode"] = ak_mode
 
         if params is not None:
             body["params"] = params
@@ -2067,9 +2172,14 @@ class APIClient:
         return self._delete(f"/concurrency-policies/{quote(policy_id)}")
 
     # ==================== Logs and artifacts ====================
-    def cancel_group(self, group_id: str) -> dict:
+    def preview_group_cancel(self, group_id: str) -> dict:
+        """Preview the current Group cancel scope and obtain confirmation if required."""
+        return self._post(f"/groups/{quote(group_id)}/cancel-preview", {})
+
+    def cancel_group(self, group_id: str, confirmation_token: Optional[str] = None) -> dict:
         """Cancel all unfinished jobs in a group."""
-        return self._post(f"/groups/{quote(group_id)}/cancel", {})
+        body = {"confirmation_token": confirmation_token} if confirmation_token else {}
+        return self._post(f"/groups/{quote(group_id)}/cancel", body)
 
     def get_job_logs(
         self,
@@ -2362,6 +2472,52 @@ class APIClient:
             params=params,
         )
 
+    def wait_for_job_artifacts_manifest(self, job_id: str) -> dict:
+        """Poll the new API within one per-Job budget; HTTP failures keep normal retries."""
+        budget = self.config.artifact_manifest_wait_timeout
+        if not math.isfinite(budget) or budget <= 0:
+            raise ConfigurationError(
+                "AP_ARTIFACT_MANIFEST_WAIT_TIMEOUT must be a finite positive number"
+            )
+        deadline = time.monotonic() + budget
+        while True:
+            response, request_id = self._request(
+                "GET",
+                f"/jobs/{quote(job_id, safe='')}/artifacts/manifest",
+                deadline=deadline,
+                _return_response=True,
+            )
+            try:
+                payload = response.json()
+            except ValueError as exc:
+                raise APIError(
+                    response.status_code, "Invalid artifact manifest response", request_id, response
+                ) from exc
+            if not isinstance(payload, dict) or payload.get("job_id") != job_id:
+                raise APIError(
+                    response.status_code, "Invalid artifact manifest response", request_id, response
+                )
+            if response.status_code == 200:
+                if "manifest" not in payload or (
+                    payload["manifest"] is not None and not isinstance(payload["manifest"], list)
+                ):
+                    raise APIError(200, "Invalid artifact manifest response", request_id, response)
+                return payload
+            if response.status_code != 202 or payload.get("status") != "preparing":
+                raise APIError(
+                    response.status_code,
+                    "Invalid artifact preparation response",
+                    request_id,
+                    response,
+                )
+            try:
+                delay = float(response.headers.get("Retry-After", "2"))
+            except ValueError:
+                delay = 2.0
+            if not math.isfinite(delay) or delay <= 0:
+                delay = 2.0
+            _deadline_sleep(delay, deadline)
+
     def get_artifact_file_download_urls(self, job_id: str, paths: list[str]) -> dict:
         """Get download URLs for multiple artifact files."""
         endpoint = f"/jobs/{quote(job_id, safe='')}/artifacts/files/download"
@@ -3275,3 +3431,34 @@ def _summarize_json_value(data: Any, limit: int) -> str:
         summary = {"_count": len(data), "_sample": sample}
         return _truncate_text(_safe_json(summary), limit)
     return _truncate_text(_safe_json(data), limit)
+
+
+def _manifest_budget_remaining(deadline: float) -> float:
+    remaining = deadline - time.monotonic()
+    if remaining <= 0:
+        raise APIError(None, "Artifact manifest wait timed out", None, None)
+    return remaining
+
+
+def _deadline_timeout(
+    timeout: TimeoutType, deadline: Optional[float]
+) -> Union[TimeoutType, HTTPTimeout]:
+    if deadline is None:
+        return timeout
+    remaining = _manifest_budget_remaining(deadline)
+
+    def bounded(value):
+        return min(value, remaining) if value is not None else remaining
+
+    if isinstance(timeout, tuple):
+        connect, read = (bounded(value) for value in timeout)
+    else:
+        connect = read = bounded(timeout)
+    # The connect phase must consume the same remaining budget as the read.
+    return HTTPTimeout(total=remaining, connect=connect, read=read)
+
+
+def _deadline_sleep(delay: float, deadline: Optional[float]) -> None:
+    if deadline is not None:
+        delay = min(delay, _manifest_budget_remaining(deadline))
+    time.sleep(delay)