@guandata/decidex 0.1.0 → 0.1.2

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (3) hide show
  1. package/README.md +28 -1
  2. package/package.json +1 -1
  3. package/src/cli.js +219 -17
package/README.md CHANGED
@@ -94,12 +94,14 @@ decidex session send SESSION_ID --message "继续拆解渠道" --wait --follow
94
94
  decidex session stop SESSION_ID
95
95
  decidex session artifacts list SESSION_ID
96
96
  decidex session artifacts download SESSION_ID --all --dir ./artifacts
97
+ decidex session evidence list SESSION_ID
98
+ decidex session evidence download SESSION_ID --all --dir ./evidence
97
99
  decidex run diagnose RUN_ID --session-id SESSION_ID --json
98
100
  ```
99
101
 
100
102
  `session send --wait` 要求消息立即启动;会话正忙时服务端会在入队前返回冲突,避免 CLI 假失败后重试造成重复入队。不带 `--wait` 时结果可能是 `kind=direct` 或 `kind=queued`。
101
103
 
102
- 下载默认拒绝覆盖已有文件,也拒绝通过符号链接写出目标目录;确认覆盖普通文件时加 `--overwrite`。`run diagnose` 未显式传 `--session-id` 时会使用最近 50 条本地 Run 映射;若本地配置不可写,远端 Run 仍会继续,CLI 会先输出 Session/Run ID,之后可用显式 ID 恢复诊断。当前后端未向用户 API 暴露 Runner 的 Pod termination reason 和资源指标,结果会将这两项标记为 `not_exposed`。
104
+ `session evidence` 通过服务端会话文件索引发现实际 `.boe` 文件;证据文件位于 `workspace/analysis/biz-ontology-evidence/`,不会混入普通产物列表。下载默认拒绝覆盖已有文件,也拒绝通过符号链接写出目标目录;确认覆盖普通文件时加 `--overwrite`。`run diagnose` 未显式传 `--session-id` 时会使用最近 50 条本地 Run 映射;若本地配置不可写,远端 Run 仍会继续,CLI 会先输出 Session/Run ID,之后可用显式 ID 恢复诊断。当前后端未向用户 API 暴露 Runner 的 Pod termination reason 和资源指标,结果会将这两项标记为 `not_exposed`。
103
105
 
104
106
  ## Skill
105
107
 
@@ -158,6 +160,31 @@ decidex eval stop AGENT_ID RUN_ID
158
160
 
159
161
  CLI 按最新 `origin/main` 校验该文件:只接受 `caseIds`、`targetModelSelection` 和 `judgeModelSelection`。使用 `--cases all` 时由 CLI 获取并覆盖 `caseIds`;Agent 没有评测案例时会在本地停止。
160
162
 
163
+ 查询、接管或批量监控已经启动的评测:
164
+
165
+ ```bash
166
+ decidex eval runs list AGENT_ID
167
+ decidex eval runs get AGENT_ID RUN_ID
168
+ decidex eval runs wait AGENT_ID RUN_ID --follow --output result.json
169
+ decidex eval runs watch evaluation-runs.json --parallel 3 --follow --output summary.json
170
+ decidex eval evidence download AGENT_ID RUN_ID --all --dir ./evidence
171
+ ```
172
+
173
+ `evaluation-runs.json`:
174
+
175
+ ```json
176
+ {
177
+ "runs": [
178
+ { "label": "A/S07", "agentId": "AGENT_ID", "runId": "RUN_ID" },
179
+ { "label": "B/S07", "agentId": "AGENT_ID", "runId": "RUN_ID" }
180
+ ]
181
+ }
182
+ ```
183
+
184
+ `watch` 只在状态或计数变化时向 stderr 输出进度,最终 JSON 保留每个 Run 的完整 Case 结果;任一 Run 失败、停止或监控出错时退出码为 1。
185
+
186
+ `eval evidence download` 按 Case 分目录下载评测会话中的 `.boe`。该命令要求服务端评测详情返回 `targetSessionId`;旧版服务端会给出明确升级提示。
187
+
161
188
  ## 组织
162
189
 
163
190
  ```bash
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "@guandata/decidex",
3
- "version": "0.1.0",
3
+ "version": "0.1.2",
4
4
  "description": "CLI for building and operating Decidex ontologies, agents, skills, sessions, and evaluations.",
5
5
  "license": "UNLICENSED",
6
6
  "type": "module",
package/src/cli.js CHANGED
@@ -9,6 +9,7 @@ import { parseArgs } from "node:util";
9
9
 
10
10
  const COOKIE_NAME = "decidex_session";
11
11
  const DEFAULT_CONFIG_PATH = join(homedir(), ".config", "decidex", "config.json");
12
+ const BIZ_ONTOLOGY_EVIDENCE_DIR = "workspace/analysis/biz-ontology-evidence/";
12
13
 
13
14
  const HELP = `decidex - Decidex command line client
14
15
 
@@ -40,6 +41,8 @@ Usage:
40
41
  decidex session stop SESSION_ID
41
42
  decidex session artifacts list SESSION_ID
42
43
  decidex session artifacts download SESSION_ID --all --dir DIR [--overwrite]
44
+ decidex session evidence list SESSION_ID
45
+ decidex session evidence download SESSION_ID --all --dir DIR [--overwrite]
43
46
  decidex run diagnose RUN_ID [--session-id ID] [--agent-id ID] --json
44
47
 
45
48
  decidex skill list
@@ -52,7 +55,12 @@ Usage:
52
55
  decidex eval cases export AGENT_ID FILE
53
56
  decidex eval cases plan AGENT_ID FILE [--prune]
54
57
  decidex eval cases apply AGENT_ID FILE [--prune]
55
- decidex eval run AGENT_ID FILE [--cases all] [--wait] [--output FILE] [--timeout SECONDS]
58
+ decidex eval run AGENT_ID FILE [--cases all] [--wait] [--follow] [--output FILE] [--timeout SECONDS]
59
+ decidex eval runs list AGENT_ID [--cursor CURSOR]
60
+ decidex eval runs get AGENT_ID RUN_ID
61
+ decidex eval runs wait AGENT_ID RUN_ID [--follow] [--output FILE] [--timeout SECONDS]
62
+ decidex eval runs watch FILE [--parallel N] [--follow] [--output FILE] [--timeout SECONDS]
63
+ decidex eval evidence download AGENT_ID RUN_ID --all --dir DIR [--overwrite]
56
64
  decidex eval stop AGENT_ID RUN_ID
57
65
 
58
66
  decidex ontology list
@@ -92,12 +100,16 @@ to stderr, so stdout or --output remains machine-readable.`,
92
100
  decidex session send SESSION_ID (--message TEXT | --input FILE) [--wait] [--follow]
93
101
  decidex session stop SESSION_ID
94
102
  decidex session artifacts ...
103
+ decidex session evidence ...
95
104
 
96
105
  With --wait, the message must start immediately; a busy session returns a conflict instead
97
106
  of queuing an untrackable run. Without --wait, the result can be kind=direct or kind=queued.`,
98
107
  "session artifacts": `Usage:
99
108
  decidex session artifacts list SESSION_ID
100
109
  decidex session artifacts download SESSION_ID --all --dir DIR [--overwrite]`,
110
+ "session evidence": `Usage:
111
+ decidex session evidence list SESSION_ID
112
+ decidex session evidence download SESSION_ID --all --dir DIR [--overwrite]`,
101
113
  "run diagnose": `Usage:
102
114
  decidex run diagnose RUN_ID [--session-id ID] [--agent-id ID] --json
103
115
 
@@ -111,13 +123,22 @@ Runtime Pod termination details require backend support.`,
111
123
  decidex skill delete ID`,
112
124
  eval: `Usage:
113
125
  decidex eval cases list|export|plan|apply ...
114
- decidex eval run AGENT_ID FILE [--cases all] [--wait] [--output FILE]
126
+ decidex eval run AGENT_ID FILE [--cases all] [--wait] [--follow] [--output FILE]
127
+ decidex eval runs list|get|wait|watch ...
128
+ decidex eval evidence download AGENT_ID RUN_ID --all --dir DIR [--overwrite]
115
129
  decidex eval stop AGENT_ID RUN_ID`,
116
130
  "eval cases": `Usage:
117
131
  decidex eval cases list AGENT_ID
118
132
  decidex eval cases export AGENT_ID FILE
119
133
  decidex eval cases plan AGENT_ID FILE [--prune]
120
134
  decidex eval cases apply AGENT_ID FILE [--prune]`,
135
+ "eval runs": `Usage:
136
+ decidex eval runs list AGENT_ID [--cursor CURSOR]
137
+ decidex eval runs get AGENT_ID RUN_ID
138
+ decidex eval runs wait AGENT_ID RUN_ID [--follow] [--output FILE] [--timeout SECONDS]
139
+ decidex eval runs watch FILE [--parallel N] [--follow] [--output FILE] [--timeout SECONDS]`,
140
+ "eval evidence": `Usage:
141
+ decidex eval evidence download AGENT_ID RUN_ID --all --dir DIR [--overwrite]`,
121
142
  ontology: `Usage:
122
143
  decidex ontology list|get|export|plan|apply ...`,
123
144
  api: `Usage:
@@ -435,13 +456,25 @@ async function listSessionArtifacts(agentId, sessionId, values) {
435
456
  return (result.files ?? []).filter((file) => file.relativePath?.startsWith("artifacts/"));
436
457
  }
437
458
 
438
- function safeDestination(root, artifactPath) {
439
- const destination = resolve(root, artifactPath.replace(/^artifacts\//, ""));
459
+ async function listSessionEvidence(agentId, sessionId, values) {
460
+ const result = await request(`${sessionPath(agentId, sessionId)}/files?force=1&evidence=1`, { values });
461
+ return (result.files ?? []).filter((file) => {
462
+ const name = file.relativePath?.slice(BIZ_ONTOLOGY_EVIDENCE_DIR.length);
463
+ return file.relativePath?.startsWith(BIZ_ONTOLOGY_EVIDENCE_DIR) && name && !name.includes("/") && name.toLowerCase().endsWith(".boe");
464
+ });
465
+ }
466
+
467
+ function safeRelativeDestination(root, outputPath) {
468
+ const destination = resolve(root, outputPath);
440
469
  const child = relative(root, destination);
441
- if (!child || child.startsWith("..") || isAbsolute(child)) throw new Error(`非法产物路径:${artifactPath}`);
470
+ if (!child || child.startsWith("..") || isAbsolute(child)) throw new Error(`非法输出路径:${outputPath}`);
442
471
  return destination;
443
472
  }
444
473
 
474
+ function safeDestination(root, artifactPath) {
475
+ return safeRelativeDestination(root, artifactPath.replace(/^artifacts\//, ""));
476
+ }
477
+
445
478
  async function assertSafeArtifactTarget(root, destination, overwrite) {
446
479
  let current = root;
447
480
  for (const part of relative(root, dirname(destination)).split(sep).filter((item) => item && item !== ".")) {
@@ -808,6 +841,35 @@ async function sessionCommand(action, args) {
808
841
  if (values.query) query.set("q", values.query);
809
842
  return print(await request(`/api/agent-sessions?${query}`, { values }));
810
843
  }
844
+ if (action === "evidence") {
845
+ const [evidenceAction, sessionIdValue] = positionals;
846
+ expectPositionals(positionals, 2, `session evidence ${evidenceAction || "COMMAND"} SESSION_ID`);
847
+ const sessionId = required(sessionIdValue, "SESSION_ID");
848
+ const session = await resolveSession(sessionId, values);
849
+ const evidence = await listSessionEvidence(session.agentId, sessionId, values);
850
+ if (evidenceAction === "list") {
851
+ return print({ schemaVersion: 1, sessionId, agentId: session.agentId, evidence });
852
+ }
853
+ if (evidenceAction === "download") {
854
+ if (!values.all) throw new Error("evidence download 需要 --all");
855
+ const directory = resolve(required(values.dir, "--dir DIR"));
856
+ await mkdir(directory, { recursive: true });
857
+ const root = await realpath(directory);
858
+ const targets = evidence.map((item) => ({ item, destination: safeRelativeDestination(root, item.name) }));
859
+ for (const { destination } of targets) await assertSafeArtifactTarget(root, destination, values.overwrite);
860
+ const downloaded = [];
861
+ for (const { item, destination } of targets) {
862
+ const bytes = await request(`${sessionPath(session.agentId, sessionId)}/files/content?path=${encodeURIComponent(item.relativePath)}&download=1`, {
863
+ returnBytes: true,
864
+ values,
865
+ });
866
+ await writeArtifact(root, destination, bytes, values.overwrite);
867
+ downloaded.push({ ...item, path: destination });
868
+ }
869
+ return print({ schemaVersion: 1, sessionId, directory, downloaded });
870
+ }
871
+ throw new Error(`未知 session evidence 命令:${evidenceAction || "(空)"}`);
872
+ }
811
873
  if (action !== "artifacts") {
812
874
  expectPositionals(positionals, 1, `session ${action || "COMMAND"} SESSION_ID`);
813
875
  const sessionId = required(positionals[0], "SESSION_ID");
@@ -1035,10 +1097,161 @@ function evaluationRunSpec(value) {
1035
1097
  return value;
1036
1098
  }
1037
1099
 
1100
+ function evaluationRunPath(agentId, runId) {
1101
+ return `/api/personal-agents/${encodeURIComponent(agentId)}/evaluation-runs/${encodeURIComponent(runId)}`;
1102
+ }
1103
+
1104
+ function evaluationProgress(run, label) {
1105
+ const counts = run.counts ?? {};
1106
+ return `[${label}] ${run.status} evaluated=${counts.evaluated ?? 0}/${counts.total ?? run.caseCount ?? "?"} passed=${counts.passed ?? 0} notPassed=${counts.notPassed ?? 0} failed=${counts.evaluationFailed ?? 0}`;
1107
+ }
1108
+
1109
+ async function waitEvaluationRun(agentId, runId, values, label = runId) {
1110
+ let lastProgress;
1111
+ try {
1112
+ return await poll(
1113
+ (deadline) => request(evaluationRunPath(agentId, runId), { deadline, values }),
1114
+ (result) => ["COMPLETED", "STOPPED", "FAILED"].includes(result.run.status),
1115
+ positiveSeconds(values.timeout),
1116
+ values.follow
1117
+ ? (result) => {
1118
+ const progress = evaluationProgress(result.run, label);
1119
+ if (progress !== lastProgress) process.stderr.write(`[decidex] ${progress}\n`);
1120
+ lastProgress = progress;
1121
+ }
1122
+ : undefined,
1123
+ );
1124
+ } catch (error) {
1125
+ error.message = `${error.message}; agentId=${agentId}; runId=${runId}; 可运行 decidex eval stop ${agentId} ${runId}`;
1126
+ throw error;
1127
+ }
1128
+ }
1129
+
1130
+ function evaluationRunTargets(value) {
1131
+ if (!value || typeof value !== "object" || Array.isArray(value) || Object.keys(value).some((key) => key !== "runs") || !Array.isArray(value.runs) || value.runs.length === 0) {
1132
+ throw new Error("watch 文件必须是包含非空 runs 数组的 JSON 对象");
1133
+ }
1134
+ const seen = new Set();
1135
+ return value.runs.map((item, index) => {
1136
+ if (!item || typeof item !== "object" || Array.isArray(item)) throw new Error(`runs[${index}] 必须是对象`);
1137
+ if (Object.keys(item).some((key) => !new Set(["label", "agentId", "runId"]).has(key))) throw new Error(`runs[${index}] 包含不支持的字段`);
1138
+ for (const field of ["agentId", "runId"]) if (typeof item[field] !== "string" || !item[field].trim()) throw new Error(`runs[${index}].${field} 不能为空`);
1139
+ if (item.label !== undefined && (typeof item.label !== "string" || !item.label.trim())) throw new Error(`runs[${index}].label 不能为空`);
1140
+ const key = `${item.agentId}\0${item.runId}`;
1141
+ if (seen.has(key)) throw new Error(`watch 文件包含重复 Run:${item.agentId}/${item.runId}`);
1142
+ seen.add(key);
1143
+ return { label: item.label ?? item.runId, agentId: item.agentId, runId: item.runId };
1144
+ });
1145
+ }
1146
+
1147
+ function positiveInteger(value, name) {
1148
+ const number = Number(value);
1149
+ if (!Number.isInteger(number) || number <= 0) throw new Error(`${name} 必须是正整数`);
1150
+ return number;
1151
+ }
1152
+
1153
+ async function evalRunsCommand(args) {
1154
+ const { values, positionals } = commandArgs(args, {
1155
+ cursor: { type: "string" },
1156
+ follow: { type: "boolean", default: false },
1157
+ output: { type: "string" },
1158
+ parallel: { type: "string", default: "3" },
1159
+ timeout: { type: "string", default: "1800" },
1160
+ });
1161
+ const [action, agentId, runId] = positionals;
1162
+ if (action === "list") {
1163
+ expectPositionals(positionals, 2, "eval runs list AGENT_ID");
1164
+ const query = values.cursor ? `?cursor=${encodeURIComponent(values.cursor)}` : "";
1165
+ return print(await request(`/api/personal-agents/${encodeURIComponent(agentId)}/evaluation-runs${query}`, { values }));
1166
+ }
1167
+ if (action === "get") {
1168
+ expectPositionals(positionals, 3, "eval runs get AGENT_ID RUN_ID");
1169
+ return print(await request(evaluationRunPath(agentId, runId), { values }));
1170
+ }
1171
+ if (action === "wait") {
1172
+ expectPositionals(positionals, 3, "eval runs wait AGENT_ID RUN_ID");
1173
+ const result = await waitEvaluationRun(agentId, runId, values);
1174
+ await emit({ schemaVersion: 1, agentId, runId, run: result.run }, values.output);
1175
+ if (result.run.status !== "COMPLETED") process.exitCode = 1;
1176
+ return;
1177
+ }
1178
+ if (action === "watch") {
1179
+ expectPositionals(positionals, 2, "eval runs watch FILE");
1180
+ const targets = evaluationRunTargets(await readJsonFile(required(agentId, "FILE")));
1181
+ const parallel = Math.min(positiveInteger(values.parallel, "parallel"), targets.length);
1182
+ const results = new Array(targets.length);
1183
+ let nextIndex = 0;
1184
+ async function worker() {
1185
+ while (nextIndex < targets.length) {
1186
+ const index = nextIndex++;
1187
+ const target = targets[index];
1188
+ try {
1189
+ const result = await waitEvaluationRun(target.agentId, target.runId, values, target.label);
1190
+ results[index] = { ...target, run: result.run, monitorError: null };
1191
+ } catch (error) {
1192
+ results[index] = { ...target, run: null, monitorError: error.message };
1193
+ }
1194
+ }
1195
+ }
1196
+ await Promise.all(Array.from({ length: parallel }, worker));
1197
+ const summary = {
1198
+ total: results.length,
1199
+ completed: results.filter((item) => item.run?.status === "COMPLETED").length,
1200
+ stopped: results.filter((item) => item.run?.status === "STOPPED").length,
1201
+ failed: results.filter((item) => item.run?.status === "FAILED").length,
1202
+ errors: results.filter((item) => item.monitorError).length,
1203
+ };
1204
+ await emit({ schemaVersion: 1, summary, runs: results }, values.output);
1205
+ if (summary.completed !== summary.total) process.exitCode = 1;
1206
+ return;
1207
+ }
1208
+ throw new Error(`未知 eval runs 命令:${action || "(空)"}`);
1209
+ }
1210
+
1211
+ async function evalEvidenceCommand(args) {
1212
+ const { values, positionals } = commandArgs(args, {
1213
+ all: { type: "boolean", default: false },
1214
+ dir: { type: "string" },
1215
+ overwrite: { type: "boolean", default: false },
1216
+ });
1217
+ const [action, agentId, runId] = positionals;
1218
+ if (action !== "download") throw new Error(`未知 eval evidence 命令:${action || "(空)"}`);
1219
+ expectPositionals(positionals, 3, "eval evidence download AGENT_ID RUN_ID");
1220
+ if (!values.all) throw new Error("eval evidence download 需要 --all");
1221
+ const result = await request(evaluationRunPath(agentId, runId), { values });
1222
+ const runCases = result.run?.cases ?? [];
1223
+ if (runCases.some((runCase) => !Object.hasOwn(runCase, "targetSessionId"))) {
1224
+ throw new Error("Decidex 服务端未返回评测 Case 的 targetSessionId;请先升级服务端");
1225
+ }
1226
+ const evidence = (await Promise.all(runCases.map(async (runCase, index) => {
1227
+ if (!runCase.targetSessionId) return [];
1228
+ const files = await listSessionEvidence(agentId, runCase.targetSessionId, values);
1229
+ return files.map((file) => ({ ...file, caseId: runCase.id, casePosition: runCase.position, sessionId: runCase.targetSessionId, outputDirectory: `case-${index + 1}` }));
1230
+ }))).flat();
1231
+ const directory = resolve(required(values.dir, "--dir DIR"));
1232
+ await mkdir(directory, { recursive: true });
1233
+ const root = await realpath(directory);
1234
+ const targets = evidence.map((item) => ({ item, destination: safeRelativeDestination(root, join(item.outputDirectory, item.name)) }));
1235
+ for (const { destination } of targets) await assertSafeArtifactTarget(root, destination, values.overwrite);
1236
+ const downloaded = [];
1237
+ for (const { item, destination } of targets) {
1238
+ const bytes = await request(`${sessionPath(agentId, item.sessionId)}/files/content?path=${encodeURIComponent(item.relativePath)}&download=1`, {
1239
+ returnBytes: true,
1240
+ values,
1241
+ });
1242
+ await writeArtifact(root, destination, bytes, values.overwrite);
1243
+ downloaded.push({ caseId: item.caseId, casePosition: item.casePosition, sessionId: item.sessionId, name: item.name, relativePath: item.relativePath, path: destination });
1244
+ }
1245
+ return print({ schemaVersion: 1, agentId, runId, directory, downloaded });
1246
+ }
1247
+
1038
1248
  async function evalCommand(action, args) {
1039
1249
  if (action === "cases") return evalCasesCommand(args);
1250
+ if (action === "runs") return evalRunsCommand(args);
1251
+ if (action === "evidence") return evalEvidenceCommand(args);
1040
1252
  const { values, positionals } = commandArgs(args, {
1041
1253
  cases: { type: "string" },
1254
+ follow: { type: "boolean", default: false },
1042
1255
  output: { type: "string" },
1043
1256
  timeout: { type: "string", default: "1800" },
1044
1257
  wait: { type: "boolean", default: false },
@@ -1058,19 +1271,8 @@ async function evalCommand(action, args) {
1058
1271
  }
1059
1272
  const started = await request(`/api/personal-agents/${agentId}/evaluation-runs`, { method: "POST", body: evaluationRunSpec(body), values });
1060
1273
  if (!values.wait) return emit(started, values.output);
1061
- const runId = encodeURIComponent(started.run.id);
1062
1274
  process.stderr.write(`[decidex] evaluation run=${started.run.id}\n`);
1063
- let finished;
1064
- try {
1065
- finished = await poll(
1066
- (deadline) => request(`/api/personal-agents/${agentId}/evaluation-runs/${runId}`, { deadline, values }),
1067
- (result) => ["COMPLETED", "STOPPED", "FAILED"].includes(result.run.status),
1068
- positiveSeconds(values.timeout),
1069
- );
1070
- } catch (error) {
1071
- error.message = `${error.message}; runId=${started.run.id}; 可运行 decidex eval stop ${positionals[0]} ${started.run.id}`;
1072
- throw error;
1073
- }
1275
+ const finished = await waitEvaluationRun(positionals[0], started.run.id, values);
1074
1276
  await emit(finished, values.output);
1075
1277
  if (finished.run.status !== "COMPLETED") process.exitCode = 1;
1076
1278
  }