@delta-infra/cli 1.0.72 → 1.0.74

package/skills/rdkit/references/operations.md

@@ -0,0 +1,64 @@
+# RDKit operation 契约
+
+所有示例都在 `SKILL.md` 的 wrapper 模板中设置 `endpoint` 和 `payload`。业务结果直接
+位于 `r["native"]`。只做用户请求所需的一次调用，成功后调用 `step_finish`。
+
+## health
+
+- 情景：明确检查 RDKit 服务状态。
+- 代码：`endpoint="health"; payload=None`，调用时省略 `--data-json`。
+- 投影：status、可用性、metadata、limits、warnings 中存在的字段。
+
+## parse
+
+- 情景：只校验并规范化一个 SMILES/InChI。
+- 代码：`endpoint="parse"; payload={"input":"CCO","format":"smiles"}`
+- 投影：valid、canonical_smiles、formula、inchi_key、warnings/error。
+
+## descriptors
+
+- 情景：只计算一个已知 SMILES 的描述符。
+- 代码：`endpoint="descriptors"; payload={"smiles":"CCO","descriptor_set":"basic","sanitize":True}`
+- 投影：valid 与 descriptors 中实际返回的键。
+
+## batch-descriptors
+
+- 情景：对多个已知且无需再次解析的 SMILES 计算同一组描述符。
+- 代码：`endpoint="batch-descriptors"; payload={"smiles":["CCO","c1ccccc1"],"descriptor_set":"basic","sanitize":True}`
+- 投影：results 中每项的输入 SMILES、valid、descriptors、warnings/error。
+
+## batch-parse-describe
+
+- 情景：同时校验、规范化并计算一个或多个分子的描述符。
+- 代码：`endpoint="batch-parse-describe"; payload={"molecules":[{"id":"aspirin","smiles":"CC(=O)OC1=CC=CC=C1C(=O)O"}],"descriptor_set":"basic","sanitize":True}`
+- 投影：results 中 id、valid、canonical_smiles、descriptors、warnings/error。
+
+## render
+
+- 情景：用户明确要求分子结构图。
+- 代码：`endpoint="render"; payload={"smiles":"CCO","image_format":"png","width":500,"height":400}`
+- 投影：只解码文档指定的 base64 字段；验证 PNG magic bytes 后再保存。
+
+## fingerprint
+
+- 情景：明确要求一个分子的 Morgan 指纹。
+- 代码：`endpoint="fingerprint"; payload={"smiles":"CCO","radius":2,"n_bits":2048}`
+- 投影：valid、fingerprint_metadata、on_bits、warnings/error。
+
+## similarity
+
+- 情景：比较两个已知 SMILES。
+- 代码：`endpoint="similarity"; payload={"query_smiles":"CCO","target_smiles":"CCN","radius":2,"n_bits":2048}`
+- 投影：native 实际返回的 similarity、metric、fingerprint metadata。
+
+## similarity-matrix
+
+- 情景：比较三个及以上分子的所有两两相似度。
+- 代码：`endpoint="similarity-matrix"; payload={"molecules":[{"id":"a","smiles":"CCO"},{"id":"b","smiles":"CCN"}],"fingerprint":{"type":"morgan","radius":2,"n_bits":2048},"sanitize":True}`
+- 投影：matrix 与 `ranked_pairs[{a,b,similarity}]`；禁止 i/j/score。
+
+## substructure
+
+- 情景：用 SMARTS 在一个或多个目标 SMILES 中筛选子结构。
+- 代码：`endpoint="substructure"; payload={"query_smarts":"C(=O)O","target_smiles":["CC(=O)O","CCO"],"sanitize":True}`
+- 投影：query、每个目标的 valid/match/matches 与 warnings/error。

package/skills/synbo-service/SKILL.md

@@ -0,0 +1,31 @@
+---
+name: synbo-service
+description: 当用户需要对分类反应条件空间做初始采样，或基于真实历史反应结果使用 SynBO 推荐下一批催化剂、溶剂、碱等条件时使用。支持 health、initialize、optimize；所有调用只走 Delta CLI，重型 optimize 串行执行且超时不自动重试。
+---
+
+# SynBO 服务
+
+## 强制规则
+
+- 只用 `python_repl` 执行 `$SKILLS_ROOT/delta-science/scripts/invoke.py`，
+  tool=`synbo`；禁止 Bash、路径探测、字面 CLI、直接 HTTP 和本地 SynBO/GP。
+- `python_repl` 工具参数只能包含 `code`，禁止 `deps/dependencies/packages`。
+- 业务结果直接位于 `r["native"]`，禁止 `native.data`。
+- `health` 无 body；`initialize` 和 `optimize` 使用 `--data-json`。
+- 重型调用串行执行；timeout/断连结果未知，不自动重试。
+
+## operation 选择
+
+- `initialize`：没有 measured previous results 时使用。提供 `condition_dict`、
+  `opt_metrics`、方向/range/weight 设置、batch_size、sampling_method 和 seed。
+- `optimize`：只有存在真实历史时使用。每一行必须包含全部 condition 列和全部数值
+  metric 列；有界 CPU 测试用 `accuracy:"tiny"`、`device:"cpu"`。
+- descriptors 如提供，键必须与 condition_dict 完全一致，index 列存在且其余为数值。
+
+完整成功 payload 见 `references/endpoints.md`。若 optimize 返回数组维度错误、timeout、
+空 recommendations 或后端失败，报告“未产生推荐”，不得用本地 GP、UCB、响应面、
+随机采样或人工排序替代。
+
+通过 wrapper 调用一次，显式打印 recommendations/metadata/warnings 或有界错误；成功
+后立即 `step_finish(status="done", summary=<同一 JSON>)`，失败则
+`step_finish(status="failed", summary=<原始错误短摘要>)`。禁止后续工具。

package/skills/synbo-service/agents/openai.yaml

@@ -0,0 +1,4 @@
+interface:
+  display_name: "SynBO ??????"
+  short_description: "?? Delta CLI ????????????????"
+  default_prompt: "?? $synbo-service ??????????????????"

package/skills/synbo-service/references/endpoints.md

@@ -0,0 +1,157 @@
+# SynBO API Endpoint Reference
+
+All paths are relative to `http://111.2.199.31:52317/api/v1`.
+
+This reference is for the exposed HTTP gateway. It is not a local SynBO project
+runner.
+
+## Health
+
+`GET /chem/synbo/health`
+
+Reports SynBO import status, Torch/CUDA runtime availability, service limits,
+default device, and visible CUDA devices.
+
+Important fields inside gateway `data`:
+
+```json
+{
+  "status": "ok",
+  "synbo_available": true,
+  "metadata": {
+    "service": "synbo-service",
+    "engine": "synbo",
+    "version": "0.1.0",
+    "synbo_version": "0.1.0",
+    "default_device": "cpu",
+    "cuda_visible_devices": ""
+  },
+  "limits": {
+    "max_batch_size": 128,
+    "max_reaction_space_size": 200000,
+    "max_previous_rows": 10000
+  },
+  "runtime": {
+    "torch_available": true,
+    "cuda_available": false,
+    "cuda_device_count": 0,
+    "cuda_device_names": []
+  },
+  "warnings": []
+}
+```
+
+## Initialize
+
+`POST /chem/synbo/initialize`
+
+Use for initial condition sampling when no measured previous results exist.
+
+Request body:
+
+```json
+{
+  "condition_dict": {
+    "catalyst": ["Pd(OAc)2", "Pd(PPh3)4", "Pd2(dba)3"],
+    "solvent": ["THF", "Dioxane", "Toluene"],
+    "base": ["K2CO3", "NaOEt", "DBU"]
+  },
+  "opt_metrics": ["yield"],
+  "opt_metric_settings": [
+    {"opt_direct": "max", "opt_range": [0, 100], "metric_weight": 1.0}
+  ],
+  "batch_size": 3,
+  "sampling_method": "random",
+  "desc_normalize": "minmax",
+  "random_seed": 42
+}
+```
+
+Supported `sampling_method` values: `sobol`, `random`, `lhs`, `kmeans`.
+
+## Optimize
+
+`POST /chem/synbo/optimize`
+
+Use after measured experiment rows are available. Each previous row must contain
+all condition columns and all objective metric columns.
+
+Request body:
+
+```json
+{
+  "condition_dict": {
+    "catalyst": ["Pd(OAc)2", "Pd(PPh3)4", "Pd2(dba)3"],
+    "solvent": ["THF", "Dioxane", "Toluene"],
+    "base": ["K2CO3", "NaOEt", "DBU"]
+  },
+  "opt_metrics": ["yield"],
+  "previous_results": [
+    {"batch": 0, "catalyst": "Pd(OAc)2", "solvent": "THF", "base": "K2CO3", "yield": 42.0},
+    {"batch": 0, "catalyst": "Pd(PPh3)4", "solvent": "Dioxane", "base": "NaOEt", "yield": 55.0},
+    {"batch": 0, "catalyst": "Pd2(dba)3", "solvent": "Toluene", "base": "DBU", "yield": 63.0}
+  ],
+  "batch_size": 2,
+  "accuracy": "tiny",
+  "device": "cpu",
+  "random_seed": 42
+}
+```
+
+Useful optional fields:
+
+- `accuracy`: `tiny`, `low`, `medium`, `high`, or `ultra`.
+- `acq_func`: `EHVI`, `UCB`, `ParEGO`, or `NEI`.
+- `surrogate_model`: `GP`, `RF`, `BNN`, or `BayesianLinear`.
+- `device`: `cpu`, `auto`, `cuda`, or `cuda:0`.
+- `cuda_visible_devices`: request-level CUDA visibility hint. Use only when the
+  user explicitly asks for it and `/health` indicates a CUDA-capable runtime.
+- `constraints`: map condition type to prohibited values.
+
+## Descriptors
+
+If omitted, SynBO uses generated OneHot descriptors. To provide descriptors:
+
+```json
+{
+  "descriptors": {
+    "solvent": [
+      {"name": "THF", "polarity": 0.207, "boiling_point_C": 66.0},
+      {"name": "Toluene", "polarity": 0.099, "boiling_point_C": 110.6}
+    ]
+  },
+  "descriptor_index_col": "name"
+}
+```
+
+`descriptors` must contain exactly the same keys as `condition_dict`. Descriptor
+rows must include the configured index column and numeric descriptor columns.
+
+## Error Handling
+
+Gateway errors have the standard envelope:
+
+```json
+{
+  "code": 1,
+  "message": "SynBO service returned HTTP 400",
+  "data": {
+    "endpoint": "/synbo/optimize",
+    "service_url": "http://synbo-service:8897",
+    "synbo_response": {
+      "detail": "previous_results is missing metric columns: ['yield']"
+    }
+  }
+}
+```
+
+Common causes are oversized reaction spaces, duplicate condition values,
+non-numeric descriptors, missing previous-result columns, missing numeric metric
+values, and requesting CUDA when no GPU is available.
+
+The service also enforces its own per-request time limit that is independent of
+the client HTTP `timeout`. A slow call returns a timeout error such as
+`"SynBO service timed out"`. This is expected for heavy optimizer settings
+(GP + EHVI, higher accuracy) or when a concurrent request is occupying the
+shared GPU. Retry once the service is idle, or lower `accuracy`, narrow the
+reaction space, or use a GPU device.