@delta-infra/cli 1.1.18 → 1.1.19

package/README.md

@@ -2,6 +2,13 @@
 
 Delta Agent Cloud 的命令行工具，用于管理 Delta Sandbox GPU/CPU 计算任务和调用 Science 工具。
 
+## v1.1.19 Release Notes
+
+- `feat(config)`: `delta-cli config init` 与安装向导自动写入默认 `base_url`，移除首次使用的人工配置步骤。
+- 新增 `delta-protein` skills：蛋白质结构预测/设计/分析等 40+ 工具 reference 与跨工具工作流。
+- 更新 `xrd` science reference（云端 XRD 相识别、artifact、CIF 模拟与精修契约）。
+- 内部：GitHub 公共仓库同步与发布管线优化（sync 规划器及单元测试）。
+
 ## v1.1.18 Release Notes
 
 - Translate the three active agent skills (`delta-shared`, `delta-sandbox`, and `delta-science`) and their maintained references into English.
@@ -20,7 +27,7 @@ npx @delta-infra/cli@latest install
 1. **全局安装** `@delta-infra/cli`（含 Go 二进制）
 2. **选择 AI 工具平台**（可多选）：通用 Agent 目录 `~/.agents/skills/`（Codex/Cursor/OpenCode）、Claude Code `~/.claude/skills/`、OpenCode 原生 `~/.config/opencode/skills/`、Cursor `~/.cursor/skills/`、Memento-S 开发模式 `~/memento_s/skills/`
 3. **安装 AI Skills** 到所选平台的 skills 目录
-4. **初始化配置** `delta-cli config init`（交互式，非 TTY 跳过）
+4. **初始化配置** 自动写入默认服务端地址（无需人工填写 base_url）
 5. **身份认证** `delta-cli auth login`（交互式，非 TTY 跳过）
 
 AI Agent 也可以直接使用以下指令完成安装：
@@ -141,9 +148,9 @@ go install github.com/delta-infra/delta-infra-cli/cmd/delta-cli@latest
 delta-cli config init
 ```
 
-执行时会交互式提示输入服务端地址（默认：`https://delta-infra-nacos.yangtzeailab.com`）以及 Bearer Token（可选，也可后续用 `auth login` 配置）。
+服务端地址无需人工填写：默认使用 `https://delta-infra-nacos.yangtzeailab.com`（已有配置则沿用其中的 base_url），仅交互式提示 Bearer Token（可选，也可后续用 `auth login` 配置）。
 
-非交互式环境（如 CI）会直接使用默认值，也可以通过 flag 或环境变量一次性指定：
+如需指向自建服务器，可通过 flag 或环境变量一次性指定：
 
 ```bash
 # 通过 flag
@@ -456,7 +463,7 @@ make release
 ```bash
 export GH_TOKEN=<your-github-pat>
 export NPM_TOKEN=<your-npm-token>
-./release.sh v1.1.18
+./release.sh v1.1.19
 ```
 
 脚本执行步骤：

package/checksums.txt

@@ -1,5 +1,5 @@
-3fd24fd43adacb420e155e3e636beb43af693112b6b3c70b92de706f74fd85c6 *delta-cli-darwin-amd64.tar.gz
-c134129e704633f45108b03dcdcd40bf3c60eab0b29ef0227ca5b75d11da2a06 *delta-cli-darwin-arm64.tar.gz
-4e1a7f8d4aea6e1c49390ccdba84b9dae8b20d39eababcb7b01a8d1d6789f310 *delta-cli-linux-amd64.tar.gz
-b609bace1814405a3eceb0f298e95056d27630e92b9b489c90bc48890ea52eaa *delta-cli-linux-arm64.tar.gz
-06bddaa9d5733720b61985e2552036f660ebc2076afece1a1674d5c56a4f8b6b *delta-cli-windows-amd64.zip
+a85ffee5fb646cc77fc8d5afe4d34e3b6ed713b1f7abc3987fe3c7f43f5ab1da *delta-cli-darwin-amd64.tar.gz
+22106409c2b10b38fb71cae558e645fc2b1310561613ba8e8a0037eb2655cdaa *delta-cli-darwin-arm64.tar.gz
+9f0656be23adce565aeec0ae3db6bbae4fee484f8b784323ad13041219af7dd3 *delta-cli-linux-amd64.tar.gz
+354c1a28dc0f6d47797276098efbd09bc0ae2394a74ac6eb07ed1e341c3111fc *delta-cli-linux-arm64.tar.gz
+20656d7fbfe7d36d656bbb6a12b09a3b8fe789c2c1b716e4ed8c2096109f1fd0 *delta-cli-windows-amd64.zip

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@delta-infra/cli",
-  "version": "1.1.18",
+  "version": "1.1.19",
   "description": "Delta Agent Cloud CLI for sandbox compute and science tools.",
   "bin": {
     "delta-cli": "scripts/run.js"
@@ -16,7 +16,8 @@
     "skills/manifest.json",
     "skills/delta-sandbox/",
     "skills/delta-shared/",
-    "skills/delta-science/"
+    "skills/delta-science/",
+    "skills/delta-protein/"
   ],
   "os": [
     "darwin",

package/scripts/install-wizard.js

@@ -129,9 +129,7 @@ const messages = {
     step2Skip:      "已安装，跳过",
     step2Fail:      "Skills 安装失败。运行以下命令重试: npx skills add %s -y -g",
     step3:          "正在初始化配置...",
-    step3BaseUrl:   "Delta Agent Cloud 服务器地址",
-    step3Skip:      "跳过配置初始化",
-    step3Done:      "配置已初始化",
+    step3Done:      "配置已初始化 (Server: %s)",
     step3Fail:      "配置初始化失败。运行以下命令重试: delta-cli config init",
     step4:          "身份认证",
     step4Confirm:   "是否进行身份认证？",
@@ -142,7 +140,7 @@ const messages = {
     step4Fail:      "认证失败。运行以下命令重试: delta-cli auth login",
     done:           "安装完成！\n现在可以让你的 AI 工具使用 delta-cli 管理 Delta Sandbox 资源或调用 Science 工具。",
     cancelled:      "安装已取消",
-    nonTtyHint:     "要完成配置，请在终端中运行：\n  delta-cli config init\n  delta-cli auth login\n若还没有 API Key，请先在网页端生成：\n  %s",
+    nonTtyHint:     "要完成配置，请在终端中运行：\n  delta-cli auth login\n若还没有 API Key，请先在网页端生成：\n  %s",
     uninstallConfirm: "将移除全局包、AI Skills 和配置文件。是否继续？",
     uninstallCancelled: "卸载已取消",
     uninstallDone:  "卸载完成",
@@ -164,9 +162,7 @@ const messages = {
     step2Skip:      "Already installed. Skipped",
     step2Fail:      "Failed to install skills. Run manually: npx skills add %s -y -g",
     step3:          "Initializing config...",
-    step3BaseUrl:   "Delta Agent Cloud server URL",
-    step3Skip:      "Skipped config initialization",
-    step3Done:      "Config initialized",
+    step3Done:      "Config initialized (Server: %s)",
     step3Fail:      "Failed to init config. Run manually: delta-cli config init",
     step4:          "Authorization",
     step4Confirm:   "Would you like to authenticate now?",
@@ -177,7 +173,7 @@ const messages = {
     step4Fail:      "Failed to authorize. Run delta-cli auth login to retry",
     done:           "You are all set!\nYour AI tool can now use delta-cli to manage Delta Sandbox resources or invoke Science tools.",
     cancelled:      "Installation cancelled",
-    nonTtyHint:     "To complete setup, run interactively:\n  delta-cli config init\n  delta-cli auth login\nIf you don't have an API key yet, generate one first:\n  %s",
+    nonTtyHint:     "To complete setup, run interactively:\n  delta-cli auth login\nIf you don't have an API key yet, generate one first:\n  %s",
     uninstallConfirm: "This will remove the global package, AI Skills and config. Continue?",
     uninstallCancelled: "Uninstall cancelled",
     uninstallDone:  "Uninstall complete",
@@ -510,32 +506,14 @@ async function stepConfigInit(msg) {
   s.start(msg.step3);
 
   const existingConfig = getExistingConfig();
-  s.stop(msg.step3);
+  // 服务器地址无需人工配置：沿用已有值，否则直接用内置默认值。
+  const baseUrl = (existingConfig && existingConfig.base_url) || DEFAULT_BASE_URL;
 
-  if (existingConfig && existingConfig.base_url) {
-    const reuse = await p.confirm({
-      message: `发现已有配置 (Server: ${existingConfig.base_url})，继续使用？`,
-    });
-    if (p.isCancel(reuse)) {
-      handleCancel(reuse, msg);
-      return;
-    }
-    if (reuse) {
-      p.log.info(msg.step3Skip);
-      return;
-    }
-  }
-
-  const baseUrl = handleCancel(await p.text({
-    message: msg.step3BaseUrl,
-    initialValue: existingConfig && existingConfig.base_url
-      ? existingConfig.base_url
-      : DEFAULT_BASE_URL,
-  }), msg);
+  s.stop(msg.step3);
 
   try {
     writeConfig({ baseUrl, existing: existingConfig });
-    p.log.success(msg.step3Done);
+    p.log.success(fmt(msg.step3Done, baseUrl));
   } catch (e) {
     p.log.error(e.message || String(e));
     p.log.error(msg.step3Fail);
@@ -654,7 +632,16 @@ async function doInstall() {
     const platforms = ["agents"];
     await stepInstallSkills(msg, platforms);
     const existing = getExistingConfig();
-    console.log(fmt(msg.nonTtyHint, dashboardAPIKeysUrl(existing && existing.base_url)));
+    // 非交互安装同样预置默认服务器地址，免去后续人工配置 config init。
+    try {
+      writeConfig({
+        baseUrl: (existing && existing.base_url) || DEFAULT_BASE_URL,
+        existing,
+      });
+    } catch (e) {
+      p.log.error(e.message || String(e));
+    }
+    console.log(fmt(msg.nonTtyHint, dashboardAPIKeysUrl(getExistingConfig() && getExistingConfig().base_url)));
   }
 }
 

package/skills/delta-protein/AGENTS.md

@@ -0,0 +1,73 @@
+# skills/delta-protein — Protein Analysis Skill Package
+
+**Parent:** `../AGENTS.md`
+
+## OVERVIEW
+
+Single public `delta-protein` skill: the unified entry point for Huada protein and
+bioinformatics online analysis tools (mutation prediction, structure prediction, MSA,
+protein design, docking, pocket/active-site analysis, conservation/InterPro,
+enzyme kinetics, solubility/pH prediction, and PubMed/PubChem/UniProt/PISA/PLIP
+lookup). It is host-neutral and routed exclusively through the Delta CLI
+(`science invoke`) to the Science Server's `huada_protein_tools` tool. Per-endpoint
+operation contracts live in `references/` and are loaded on demand; there are no
+separate planner-visible per-tool skills and no workflow endpoints.
+
+## STRUCTURE
+
+```
+delta-protein/
+├── SKILL.md                 # single entry point: routing, CLI-only execution, safety red lines
+├── agents/
+│   └── openai.yaml          # host agent registration
+└── references/
+    └── <endpoint>.md        # one registered endpoint's operation contract
+                             #  (39 endpoints under tool `huada_protein_tools`)
+```
+
+## DATA SOURCE
+
+- Tool/endpoint identity and request fields come from the live Catalog
+  (`delta-cli science list`, `delta-cli science endpoints list <tool>`) plus the
+  authoritative parameter registry in
+  `agent/protein_common/tool_registry/definitions.py` of the `delta-search-backend`
+  project (per-tool `params`, `required_groups`, `default_params`, `timeout`).
+- Response envelope is the Huada contract `{"status","message","tool_type",
+  "resultjson","download_url"}`; per-tool `resultjson` keys are derived from the
+  tool implementations, not invented.
+- Each registered endpoint is a single tool call. Multi-step card workflows
+  (cards/*.json) are NOT registered as endpoints and are not documented here; an
+  agent may chain single endpoints in sequence when the task requires it.
+
+## CONVENTIONS
+
+- JSON/query operations go through `delta-cli science invoke --tool huada_protein_tools --endpoint <name>`; binary artifacts from `download_url` go through `delta-cli science file download`
+- Endpoint names are exact server contracts, read from the Catalog — never invent fields/enums from memory
+- Success is reported only when the subprocess exit is 0 AND top-level JSON `ok` is `true`; `data` holds the service response
+- `references/` contains only registered real endpoints; routing and tool selection belong in `SKILL.md`
+- Host-neutral: `delta-cli` is the sole service boundary; no host-specific tool/env dependencies
+- Discover newly-deployed endpoints via `delta-cli science endpoints list huada_protein_tools`
+
+## MAINTENANCE
+
+- `scripts/gen_references.py` regenerates the `## Output Projection` and
+  `## Output Template` sections in every reference from
+  `scripts/output_metadata.json` (single source of truth for field projection /
+  presentation form / downstream reuse / provenance marker). Stdlib only, idempotent.
+- Workflow: update `output_metadata.json` → run
+  `python scripts/gen_references.py` → run `python scripts/gen_references.py --check`.
+  `--check` fails loudly on missing sections, drifting titles, broken SKILL links or
+  references not covered by metadata.
+- `output_metadata.json` `provenance` is `tested` (verified by a real `delta-cli`
+  invocation) or `derived` (from ToolSpec descriptions / agent templates, marked
+  "pending live calibration"). Newly verified endpoints should be flipped to `tested`.
+- After changing references/SKILL, resync the opencode install dir:
+  `python scripts/gen_references.py --sync "%USERPROFILE%\.config\opencode\skills\delta-protein"`.
+
+## ANTI-PATTERNS
+
+- Do not call `curl`/`requests`/`httpx`/browser/PowerShell web directly against any Huada or Gateway URL
+- Do not bypass delta-cli or attach local MCP tool servers directly
+- Do not fabricate protein sequences, structures, activity data, or algorithm results, and do not pass local-library results off as Science results
+- Do not create extra JSON/CSV/report/image artifacts unless the user explicitly asks
+- Do not break relative links between SKILL.md and `references/`

package/skills/delta-protein/SKILL.md

@@ -0,0 +1,389 @@
+---
+name: delta-protein
+description: "The unified entry point for the company's Huada protein/bioinformatics online analysis capabilities — all calls go through the Delta CLI. Any zero-shot or multi-site mutation prediction, structure prediction (Protenix/ESMFold2), sequence extraction (PDB/FASTA/Foldseek), MSA computation, optimal-pH and solubility prediction for proteins/mutants, enzyme kinetic parameters (kcat/Km), consensus mutation and EVcouplings coevolutionary analysis, protein design (MPNN/RFdiffusion3), molecular docking and Vina scoring, pocket prediction, surface charge, flexible sites, tunnel analysis, PLIP/PISA interactions, active sites, InterPro/ConSurf annotations, single-site mutation primer design, and PubMed/PubChem/UniProt/ProTrek retrieval — all protein analysis tasks must use this Skill; call the huada_protein_tools endpoint via delta-cli, and never bypass the Delta CLI through local MCP servers, legacy gateway scripts, or web search. Simple queries, comparisons or verifications return concise text directly; analysis/multi-step/design/mutation/evaluation tasks produce one structured Markdown report (skeleton and per-task-type chapter structure in SKILL.md) after all endpoints complete. By default, do not create additional JSON/CSV, image or chart files unless the user explicitly asks."
+metadata:
+  requires:
+    bins: ["delta-cli"]
+  cliHelp: "delta-cli science --help"
+---
+
+# Delta Protein
+
+Convert natural-language protein research tasks into the minimal set of verifiable
+Delta CLI Science calls. Expose only this Skill; endpoint selection, parameters,
+cross-endpoint handoffs and result validation are all handled inside this Skill. All
+endpoints belong to the single tool `huada_protein_tools`, and each registered
+endpoint is a single tool call (no workflow endpoints exist).
+
+## Service Call Boundary
+
+- All online protein/bioinformatics operations ultimately go through the Delta CLI:
+  JSON-compatible endpoints use
+  `delta-cli science invoke --tool huada_protein_tools --endpoint <name> --data JSON`;
+  file/binary artifacts use `delta-cli science file download`.
+- Use the CLI's standard authentication and the Science service address derived from
+  `base_url` (`{base_url}/science_tool`). Never override the service route via
+  environment variables, command-line arguments or business URLs.
+- It is forbidden to access Huada, the Gateway, or any business service URL directly
+  with `curl`, `requests`, `httpx`, a browser, or PowerShell web commands; do not
+  assemble/attach a local MCP tool server.
+- Local tools (RDKit/pymatgen/self-run algorithms, etc.) may be used for format
+  conversion, file parsing or display, but their origin must be made explicit and kept
+  separate from CLI results; they must never be passed off as Science results.
+- Before executing, read the corresponding reference and combine required fields and
+  request parameters per that reference; when a reference is missing, stale, or
+  inconsistent with the live catalog, query the real-time catalog with
+  `delta-cli science list` and `delta-cli science endpoints list huada_protein_tools`.
+  Never create fields, enum values or branches beyond the required combinations from memory.
+- Endpoint names are the single authoritative invocation names. The endpoint passed to
+  the CLI must match the catalog exactly; do not alias-map (the `protein_dock` tool's
+  endpoint in the catalog is `dock`), trim prefixes, or rename-and-retry after a
+  `not_found`. When inconsistent with the catalog, terminate and report the
+  configuration problem.
+- A call may only be reported as successful when the CLI subprocess exits 0 AND the
+  top-level `ok` in its JSON output is `true`. The CLI top-level `data` holds the
+  service response; read the server-side return contract per the corresponding
+  reference — do not recursively search for plausible-looking values, and do not
+  misreport business results such as `status:error` as transport failures.
+- Fields marked as "default" in a reference or ToolSpec are fill-in conveniences for
+  the backend agent orchestration layer; they are NOT omissible when calling the CLI
+  directly. As long as the server validates a field as required, it must be passed
+  explicitly (e.g. `DB` for `consurf_predict`); when validation reports a missing
+  field, supply it per the server error and write the live-verified value back into
+  the reference.
+- JSON-bearing arguments must follow the shell rules of the calling environment;
+  never use bare POSIX single quotes on Windows.
+
+### Cross-platform JSON Passing Convention (Mandatory)
+
+| Platform / execution environment | Authoritative way to pass `--data` / `--params` JSON |
+|---|---|
+| Linux / macOS (sh/bash) | `--data '{"k":"v"}'` (POSIX single quotes, verbatim) |
+| Windows cmd.exe | `--data "{\"k\":\"v\"}"` (cmd double quotes + `\"` escaping) |
+| Windows PowerShell 5.1 / opencode(win32) | always parse via cmd: `cmd /c '... --data "{\"k\":\"v\"}"'` (outer PowerShell uses single quotes verbatim, inner layer uses cmd `\"` escaping), or write the full command to a `.bat`/`.cmd` file and run it |
+
+Rules:
+
+- Do not let PowerShell parse `\"` or bare JSON double quotes: PS 5.1 strips embedded
+  double quotes from arguments when invoking native binaries (`{"k":"v"}` → `{k:v}`,
+  reporting `invalid character 'k'`) or breaks apart the outer quoted string.
+- On win32, opencode's Bash tool is PowerShell: every `delta-cli` call carrying JSON
+  must be wrapped with `cmd /c '<cmd syntax>'` or a `.bat` file, with all `--data`
+  JSON keys/values escaped as `\"`.
+- On Linux/macOS do not switch to `\"` for Windows compatibility; standard POSIX
+  single quotes are fine; documentation examples are host-neutral and actual execution
+  follows this table.
+
+Call template:
+
+```text
+delta-cli science invoke --tool huada_protein_tools --endpoint ENDPOINT --data JSON
+```
+
+File artifact download template (when `download_url` points to raw files such as
+PDB/XLSX/FASTA/ZIP):
+
+```text
+delta-cli science file download --tool huada_protein_tools --endpoint ENDPOINT \
+  --params '{"...":"..."}' --output <path>
+```
+
+Only call the endpoints needed to achieve the goal. File-type parameters
+(`structure_file`/`sequence_file`/`msa_file`/`mutant_list_file` etc.) take
+server-accessible URL/path strings, not local upload paths; when a genuine local
+upload is needed, first derive a `download_url` via endpoints such as
+generate_msa/foldseek/structure_to_sequence and backfill it.
+
+## Choosing an Endpoint
+
+| Protein task | endpoint | Detailed rules |
+| --- | --- | --- |
+| Zero-shot single-site mutant candidates (no experimental data) | `zero_shot_mutant_candidates` | [zero_shot_mutant_candidates.md](references/zero_shot_mutant_candidates.md) |
+| n-site/multi-site mutation activity regression (data available, explicit n) | `protein_evolution` | [protein_evolution.md](references/protein_evolution.md) |
+| Multi-site combinatorial mutation activity (training data, n unspecified) | `multievolve` | [multievolve.md](references/multievolve.md) |
+| Consensus mutation analysis (aligned MSA) | `consensus_mutation` | [consensus_mutation.md](references/consensus_mutation.md) |
+| EVcouplings coevolutionary mutation scoring | `evcouplings_mutation` | [evcouplings_mutation.md](references/evcouplings_mutation.md) |
+| Multi-model mutation result summary (final chain step) | `summarize_result` | [summarize_result.md](references/summarize_result.md) |
+| Single-site mutation primer design (REAP Step 7) | `reap_design_primer` | [reap_design_primer.md](references/reap_design_primer.md) |
+| Crop a protein sequence by regions | `crop_sequence` | [crop_sequence.md](references/crop_sequence.md) |
+| Pure protein structure prediction | `protenix` | [protenix.md](references/protenix.md) |
+| Protein/DNA/RNA/ligand complex structure prediction | `esmfold2_predict` | [esmfold2_predict.md](references/esmfold2_predict.md) |
+| Extract protein sequence from PDB/CIF | `structure_to_sequence` | [structure_to_sequence.md](references/structure_to_sequence.md) |
+| Extract protein sequence from FASTA | `fasta_to_sequence` | [fasta_to_sequence.md](references/fasta_to_sequence.md) |
+| Extract Foldseek/SaProt sequence from structure | `foldseek` | [foldseek.md](references/foldseek.md) |
+| Compute homolog MSA (MMseqs2, uniref100) | `generate_msa` | [generate_msa.md](references/generate_msa.md) |
+| Protein optimal pH prediction | `ephod_predict_protein` | [ephod_predict_protein.md](references/ephod_predict_protein.md) |
+| Mutant optimal pH prediction | `ephod_predict_mutation` | [ephod_predict_mutation.md](references/ephod_predict_mutation.md) |
+| Protein solubility and isoelectric point prediction | `protein_sol_predict_protein` | [protein_sol_predict_protein.md](references/protein_sol_predict_protein.md) |
+| Mutant solubility prediction | `protein_sol_predict_mutation` | [protein_sol_predict_mutation.md](references/protein_sol_predict_mutation.md) |
+| Enzyme kinetic parameter prediction (CataPro) | `catapro_predict` | [catapro_predict.md](references/catapro_predict.md) |
+| Enzyme kinetic parameter prediction (CatPred) | `catpred_predict` | [catpred_predict.md](references/catpred_predict.md) |
+| De novo protein design (RFdiffusion3+MPNN) | `protein_design` | [protein_design.md](references/protein_design.md) |
+| Protein scaffold design (RFdiffusion3) | `rfd3_design` | [rfd3_design.md](references/rfd3_design.md) |
+| MPNN sequence design (given structure) | `mpnn_design_sequence` | [mpnn_design_sequence.md](references/mpnn_design_sequence.md) |
+| Protein-ligand molecular docking | `dock` | [dock.md](references/dock.md) |
+| Single-pose scoring of a complex (Vina score_only) | `dock_score` | [dock_score.md](references/dock_score.md) |
+| Pocket prediction (p2rank) | `pocket_prediction` | [pocket_prediction.md](references/pocket_prediction.md) |
+| Surface residue charge analysis (ProCaliper+APBS) | `surface_charge` | [surface_charge.md](references/surface_charge.md) |
+| Flexible-site analysis (B-factor) | `bfitter` | [bfitter.md](references/bfitter.md) |
+| Tunnel/channel analysis (CAVER) | `caver_tunnel_analyze` | [caver_tunnel_analyze.md](references/caver_tunnel_analyze.md) |
+| Protein-ligand interaction analysis | `plip_analyze` | [plip_analyze.md](references/plip_analyze.md) |
+| Protein interface analysis (PDBePISA) | `pisa_query` | [pisa_query.md](references/pisa_query.md) |
+| Residue conservation analysis (ConSurf) | `consurf_predict` | [consurf_predict.md](references/consurf_predict.md) |
+| Enzyme active-site prediction (EasIFA) | `easifa_predict_active_site` | [easifa_predict_active_site.md](references/easifa_predict_active_site.md) |
+| Domain/GO annotation (InterProScan) | `interpro_analyze` | [interpro_analyze.md](references/interpro_analyze.md) |
+| Open-ended protein retrieval (text/sequence/structure) | `protrek_search` | [protrek_search.md](references/protrek_search.md) |
+| Known protein metadata query | `uniprot_query` | [uniprot_query.md](references/uniprot_query.md) |
+| Batch protein metadata query | `uniprot_batch_query` | [uniprot_batch_query.md](references/uniprot_batch_query.md) |
+| PubMed literature search | `pubmed_search` | [pubmed_search.md](references/pubmed_search.md) |
+| PubChem compound search | `pubchem_search` | [pubchem_search.md](references/pubchem_search.md) |
+
+When the tool is clear, read the selected endpoint's reference directly — it already
+contains the required combinations, request fields, defaults and outputs; do not read
+other endpoint docs. Multi-step tasks are decomposed by this Skill's routing rules,
+reading only the references of the endpoints actually involved in each step.
+
+## Routing Boundaries
+
+- Use `protenix` for pure protein structure prediction; use `esmfold2_predict` for
+  complexes involving DNA/RNA/ligands; never mix them.
+- `ephod_predict_protein`/`protein_sol_predict_protein` accept
+  `protein_list`/`protein_list_file` (sequence-file URLs); the `..._mutation`
+  variants accept mutation lists; protein-level and mutation-level inputs cannot be
+  swapped.
+- `uniprot_query` is for exact metadata queries on known accession/gene/protein
+  names; open-ended functional/reaction/family retrieval uses `protrek_search`; the
+  two are not interchangeable.
+- Both `catpred_predict` and `catapro_predict` predict kcat, Km and kcat/Km; call each per
+  its reference's required fields and constraints — the results are not
+  interchangeable in meaning.
+- Without experimental data use `zero_shot_mutant_candidates`; with mutation activity
+  data and an explicit n-site/multi-site mutation specification use
+  `protein_evolution`; when the number of sites is unspecified but training data is
+  provided, use `multievolve`.
+
+This Skill has no workflow endpoints. For multi-step chains, call multiple endpoints
+in dependency order, and for downstream `structure_file`/`msa_file`/box parameters
+only use business fields verified in the previous step:
+
+- `generate_msa`'s output `download_url` becomes the `msa_file` of
+  `consensus_mutation`/`evcouplings_mutation`/`zero_shot_mutant_candidates`;
+- `pocket_prediction`'s output `download_url` (cropped pocket PDB) becomes the
+  `structure_file` of `rfd3_design`;
+- `dock`'s rank-1 `pdb_url` (clean complex PDB) becomes the `structure_file` of
+  `plip_analyze`;
+- structure-prediction/sequence-extraction outputs feed downstream
+  mutation/design/property endpoints; validate success before transitioning.
+- `summarize_result` consumes only a URL list of multiple prediction-result files and
+  is the final step of a mutation-prediction chain.
+
+By default, do not make extra health, catalog or file-operation calls beyond known
+endpoints; the minimal set of read-only calls is acceptable when the user requests it,
+diagnostics require it, or a reference is inconsistent with the live catalog.
+
+## Input Handling
+
+- Parameter source priority: (1) provided verbatim by the user, (2) fields already
+  confirmed in the session (protein sequence/structure/mutation list), (3)
+  `resultjson`/`download_url`/`pdb_url` verified from upstream endpoints. Fabricating
+  missing parameters is forbidden.
+- In-session reuse: each step of a multi-step task may only use business fields
+  verified in the previous step of the current task; do not pull historical workspace
+  results from other sessions to complete the current task. For concrete backfill
+  chains of upstream artifacts see "Downstream reuse" in each reference's Output
+  Projection; parameter-level examples are in "Routing Boundaries".
+- Defaults vs server-required: fields marked as "default" in a reference or ToolSpec
+  are fill-in conveniences for the backend agent orchestration layer; they are NOT
+  omissible when calling the CLI directly. As long as the server validates a field as
+  required, it must be passed explicitly (e.g. `DB` for `consurf_predict`). When
+  validation reports a missing field, supply it per the server error and write the
+  live-verified value back into the reference.
+- File-type parameters (`structure_file`/`sequence_file`/`fasta_file`/`msa_file`/
+  `mutant_list_file`/`training_dataset`/`result_files` etc.) take server-accessible
+  URL/path strings, not local upload paths; when a genuine local upload is needed,
+  first derive a `download_url` via generate_msa/foldseek/structure_to_sequence and
+  backfill it.
+- Pass multi-valued parameters per the type in the corresponding reference: string
+  lists (`model_name`/`cofactors`/`target_columns`), numbers
+  (`topk`/`rfd3_n`/`max_results`), booleans (`fix`/`use_msa`), region lists
+  (`regions` of `crop_sequence`). Enum values (`input_type`/`query_type`/`DB`/
+  `sampling_method` etc.) may only take the values listed in the reference.
+
+## Safety, Retry & Artifacts
+
+- Remote tasks such as submission, docking, and design require explicit user
+  authorization. Do not retry when the outcome is unknown, timed out, or
+  disconnected — the remote computation may already have happened; do not retry
+  `status:error` or empty results with different parameters.
+- For catalog and clearly read-only, idempotent queries, network-class 502/504 or
+  timeout may be retried at most once with identical parameters.
+- After a failed CLI call, return a short summary of the raw error. When the user
+  explicitly asks for troubleshooting, run read-only `delta-cli auth status`,
+  `delta-cli config show` or `delta-cli ... --help`; never log in yourself, modify
+  config or environment variables, forge a UID, token or API key, or retry by
+  changing the endpoint, parameters or call chain.
+- `X-User-Uid`, login, quota or permission errors are terminal authentication errors;
+  forward only the raw error and the login hint. The Skill never logs in by itself,
+  checks credentials, guesses user identity, or attempts to circumvent quota.
+- Heavy calls run serially. A client timeout does not mean the remote computation was
+  cancelled.
+- Do not create CSV, images or reports unless the user asks for files; after
+  downloading binary artifacts, verify them by filename/MIME.
+- Never fabricate protein sequences, structures, mutations, activities, pH,
+  solubility, energies, or infer their mechanisms; prediction results must not be
+  described as measured results; do not present model variance or scores as
+  experimental error.
+- Paths, JSON, charts or reports auto-attached by the outer planner are not equivalent
+  to an explicit user request. If the task is essentially just a query, comparison or
+  verification, return the minimal text result directly.
+- When the user explicitly asks for public background material or literature, web
+  search is allowed, but it must not be presented as a Science service call or
+  substitute for CLI results; label sources separately.
+
+## Output
+
+Compose the final answer from the business response in the current CLI `data` as
+defined by the corresponding reference; preserve strings, numbers, warnings and
+provenance verbatim. Do not add units, mechanisms, citations, links, predicted values
+or memorized knowledge the service did not return. Re-running the CLI to sort, lay
+out, tabulate or summarize is forbidden, as is hand-copying results into new data
+literals; simple filtering, comparison and text organization happen directly in the
+final answer.
+
+### Output Modes (two-mode decision)
+
+- **Report mode (default)**: whenever the task produces prediction/analysis/design
+  results, output a report even for a single endpoint: property predictions
+  (ephod/protein_sol/catpred/catapro), structure prediction and sequence extraction,
+  mutation prediction and design, docking/scoring/pocket/charge/tunnel/conservation/
+  active-site/annotation, primer design, etc. Also report mode when the user
+  explicitly asks for a "report/detailed/summary", or when multiple endpoints/
+  reviews/comparisons are involved. **The whole task produces exactly one report**:
+  no matter how many endpoints are called, run all endpoints in sequence first
+  (validate only `ok:true` at each step and stash key fields), then generate the
+  report once at the end.
+- **Concise mode (value-lookup only)**: use it only when the goal is to "look up a
+  ready-made value" — answer in the most minimal text directly, without the report
+  skeleton — e.g. `pubchem_search` for a single compound's molecular weight/formula,
+  `uniprot_query` for a single entry's sequence/metadata,
+  `protrek_search`/`pubmed_search` as pure information retrieval display. **Whenever
+  prediction/analysis endpoints such as `ephod_predict_*` are called — even just
+  once — report mode is mandatory**; do not downgrade to concise mode because of a
+  "single return".
+
+### Report Mode: one report for the whole task (skeleton)
+
+The report follows fixed numbered sections; numbering is consecutive and never skips.
+If a section genuinely has no content, write "No relevant results in this section"
+and state why; never delete the heading. Generic skeleton:
+
+- `## 1. Analysis Overview`: analysis goal / endpoints & methods used / input data
+  sources.
+- `## 2. Core Results`: `### 2.1 Key findings` (key numbers **bold**); `### 2.2
+  Detailed data` (one paragraph per endpoint, see "Tool result snippets"; tables go in
+  the relevant subsection, preceded by an intro sentence).
+- `## 3. Data & Downloads`: all `download_url`/`pdb_url`/`zip_download_url` listed as
+  `[name](full URL)` with type/use noted; when there are none, write "This analysis
+  generated no downloadable files".
+- `## 4. Notes & Recommendations`: prediction limitations, model differences,
+  experimental validation suggestions.
+
+When intent is clear, replace the skeleton with the matched intent-based section
+structure; for mixed types arrange per "Mixed types / others".
+
+### Report Mode: tool result snippets
+
+Under `### 2.2 Detailed data` (or the matching subsection of an intent-based
+section), one paragraph per endpoint, formatted as:
+
+- `### Tool result: <endpoint display name> (step N)` + `**Status**: succeeded /
+  failed (with error summary)`.
+- Data presented per that endpoint's reference "Output template";
+  `summarize_result` gives the full overview table (≤100 rows, prompt download
+  beyond); `pubmed_search` uses a paper table; key numbers bolded.
+- Structure files (`.pdb`/`.cif`/`.ent`) are presented as plain Markdown links
+  `[structure result (.pdb)](full URL)` in "Data & Downloads"; **do not use
+  `:::structure` custom blocks** (backend/frontend-only; the TUI on
+  Windows/Linux/macOS does not render them and shows raw text).
+- A failed tool states the failure reason in its paragraph; never fabricate results;
+  endpoints not executed do not appear in the report.
+
+### Intent-based report sections (by task type)
+
+| Task type | Report sections |
+|---|---|
+| mutation (mutation prediction with data: protein_evolution/multievolve/consensus/EVcouplings/zero-shot + summary) | 1 Analysis Overview / 2 Recommended mutation sites [2.1 Top candidate mutation table (position/wild type/mutation/model/score/predicted effect); 2.2 Conservation reading (if any, annotation only); 2.3 Multi-model comparison (if any)] / 3 Overall recommendation / 4 Data & Downloads / 5 Notes |
+| structure (protenix/esmfold2/sequence extraction/MSA) | 1 Analysis Overview / 2 Structure information [2.1 Quality assessment (confidence/trusted regions); 2.2 Key structural features; 2.3 Sequence information] / 3 Structure files & downloads / 4 Analysis recommendations / 5 Notes |
+| property (pH/solubility/enzyme kinetics) | 1 Analysis Overview / 2 Prediction results [2.1 Key value summary table (property/predicted value/unit/description); 2.2 Wild type vs mutants, if any] / 3 Result interpretation / 4 Data & Downloads / 5 Notes |
+| design (protein_design/rfd3_design/mpnn_design_sequence) | 1 Design overview / 2 Designed sequences [2.1 Representative sequences (monospace); 2.2 Differences from the original sequence, if any] / 3 Design scores (table) / 4 Design strategy analysis / 5 Data & Downloads / 6 Follow-up suggestions |
+| search (pubmed/pubchem/uniprot/protrek) | 1 Search overview (keywords/databases/hit count) / 2 Result details (literature grouped by PMID; compound/entry lists) / 3 Key findings & summary / 4 Data & Downloads / 5 Notes |
+| analysis (active site/conservation/interface/interaction/pocket/charge/tunnel) | 1 Analysis overview / 2 Active site (table: position/residue/site type/function annotation/confidence) / 3 Conserved residues (table: position/residue/score/rank/conclusion/functional significance) / 4 Functional interpretation / 5 Data & Downloads / 6 Research recommendations |
+| Mixed types / others | 1 Analysis flow overview (actual steps & dependencies) / 2 Per-endpoint results (each `### 2.x` per endpoint using its Output template) / 3 Combined analysis / 4 Data & Downloads / 5 Notes & recommendations |
+
+Conservation reading fixed semantics: higher `conserved_score` = less conserved,
+lower = more conserved; higher `conserved_rank` = more conserved (1,2,3 are usually
+variable residues, 7,8,9 are usually conserved).
+
+### Report Mode: organization rules (aligned with agent content orchestration)
+
+- Multi-endpoint reports are organized by the actual call order, with later results
+  citing earlier ones as evidence (note the source field); cross-tool references are
+  noted at the end of the relevant subsection rather than repeating the full data;
+  only cite fields verified in the current task.
+- Tabulate only when the user needs it or the data suits it; column names restated in
+  natural language, never exposing raw JSON key names; keep original references
+  (mutations as `Y126A/Y206F`, PMIDs in the last table column).
+
+On success, return only bounded text generated from the current CLI business result;
+on failure, return only a short summary of the CLI's raw error.
+
+### Keep Execution-layer Details Internal
+
+- Answers present only business results, the necessary interpretation basis,
+  download links and evidence boundaries; **do not output execution details**,
+  including but not limited to:
+  - parameters/defaults/required values (e.g. `DB=UNIREF100`, `chain=A`,
+    `max_results`) — include them in the command, don't restate them in prose;
+  - invocation details (`cmd /c` wrapping, PowerShell escaping, `.bat`, direct exe
+    call);
+  - working process (retry counts, 503, temp files, quota/credit consumption).
+- Endpoints with defaults/server-required fields still set the parameters when
+  executing, but the display only carries business data (e.g. ConSurf output lists
+  conserved positions, not "called with `DB=UNIREF100`").
+- After a successful result, give the final answer directly; do not proactively offer
+  follow-ups such as "want me to re-run verification / generate a new link"
+  (re-running the CLI for verification is forbidden) unless the user explicitly asks.
+
+Any URL (`download_url`, `pdb_url`, service-returned links, etc.) must be rendered in
+one of the following two **safe forms**; bare text URLs are forbidden in prose (to
+prevent Markdown treating `_..._` as italics and stripping underscores, which breaks
+links):
+
+- **inline code backticks** wrapping the full URL, e.g.
+  `` `https://delta-bucket-public.yangtzeailab.com/toolResultFile/protein-test/protein_00735e8a_20261012_103005.xlsx` ``; or
+- **Markdown link** `[name](full URL)` — underscores in the link target are preserved
+  from italic handling, e.g.
+  `[result file](https://delta-bucket-public.yangtzeailab.com/toolResultFile/protein-test/protein_00735e8a_20261012_103005.xlsx)`.
+
+Also state the file type and use. Whenever a URL is cited, verify it against the raw
+JSON returned by the CLI (fields within `data`), not against rendered text.
+
+Keep after every successful result:
+
+```text
+Evidence boundary: only business results from the current Delta CLI `data`
+Call chain: delta-cli
+No supplementation of unreturned content: yes
+```
+
+## Completion Criteria
+
+- Every online operation is evidenced by a CLI exit code of 0 and top-level
+  `ok:true`;
+- Every scientific number comes from a business field in the current call's `data`,
+  as defined by the reference;
+- Downstream inputs come from upstream results verified in the current task;
+- Failures are correctly classified as config, CLI, parameter, Science Server,
+  business service or timeout;
+- No direct HTTP, local science library, cross-session file or manual result
+  substitutes.

package/skills/delta-protein/agents/openai.yaml

@@ -0,0 +1,4 @@
+interface:
+  display_name: "Delta Protein"
+  short_description: "Unified access to Huada protein analysis tools through the Delta CLI"
+  default_prompt: "Use $delta-protein to analyze protein research goals: pick the right endpoint and complete the task by calling huada_protein_tools through the Delta CLI."