feat(cli): base/space 命令 + 执行守则 + eval harness 入仓 + 多轮 eval (0.1.9)

- 新命令: space list/create、base create/delete(实测通过,冒烟覆盖)
- SKILL.md §0 执行守则:完成全部步骤再答复/不谎报/分批/撤销指引
  (snapshot 无恢复 API,用 deactivate+节点改回)
- hint 补 429 限流与超时模式
- evals/ 入仓: 14+6 题 harness 参数化(EVAL_ROOT/EVAL_ENV/EVAL_MODEL/
  EVAL_CFG/auto_skill 触发模式)+ 多轮修正式 eval 与灵活判分
- 判分器加固: workflow list 错误负载不崩、官方 CLI 探测(command -v)不算使用
- 实测: auto 触发批(入仓版重跑) 11/11、99.0% 首试、100% ≤2 次;
  多轮修正 2/2+3/3(修改而非重建);冒烟 85 用例含 token 路由(低权会话)
- 运维事实入档: POST/DELETE /api/plugin 为管理员专属(PAT 403)

evals/grade_multiturn.py

@@ -0,0 +1,111 @@
+#!/usr/bin/env python3
+"""Grade multi-turn (correction) evals: final state must satisfy the CORRECTED
+requirement; amendment of the existing artifact is preferred over rebuilding.
+
+mt-1 accepted shapes for "结果 变成 失败 才站内通知":
+  A) recordCondition trigger with `conditions` (state-match reading), OR
+  B) recordUpdated trigger watching the 结果 field + script gate on 失败
+     (transition reading — semantically at least as correct)
+mt-2: incremental schema (add 渠道 select + mark the existing record).
+"""
+import json, os, re, subprocess, sys
+from pathlib import Path
+
+HARNESS = Path(__file__).parent
+ROOT = Path(os.environ.get("EVAL_ROOT", "/root/eval-rerun"))
+env_all = json.loads(Path(os.environ.get("EVAL_ENV", str(HARNESS.parent / "sandbox-env.json"))).read_text())
+CLI = os.environ.get("EVAL_CLI", "/root/eval-rerun/bin-new/bin/inkreach-teable")
+
+
+def cli(*a):
+    env = {**os.environ,
+           "INKREACH_TEABLE_ENDPOINT": env_all["INKREACH_TEABLE_ENDPOINT"],
+           "INKREACH_TEABLE_TOKEN": env_all["INKREACH_TEABLE_TOKEN"],
+           "INKREACH_TEABLE_BASE_ID": env_all["with_skill"]["INKREACH_TEABLE_BASE_ID"]}
+    p = subprocess.run([CLI, *a, "--json"], capture_output=True, text=True, env=env, timeout=60)
+    try:
+        return json.loads(p.stdout)
+    except json.JSONDecodeError:
+        return None
+
+
+def result(t, p, e):
+    return {"text": t, "passed": bool(p), "evidence": str(e)[:300]}
+
+
+def cmds_of(rd, turn=2):
+    out = []
+    f = rd / f"turn{turn}.jsonl"
+    if not f.exists():
+        return out
+    for line in f.read_text(errors="replace").splitlines():
+        try:
+            e = json.loads(line)
+        except json.JSONDecodeError:
+            continue
+        if e.get("type") == "tool_execution_start" and e.get("toolName") == "bash":
+            out.append(e.get("args", {}).get("command", ""))
+    return out
+
+
+def grade_mt1():
+    out = []
+    wfs = cli("workflow", "list") or []
+    if not isinstance(wfs, list):
+        return [result("workflow list 可读", False, wfs)]
+    ok = amended = False
+    for w in wfs:
+        full = cli("workflow", "get", w["id"]) or {}
+        trs = [n for n in full.get("nodes", []) if n.get("type") == "trigger"]
+        acs = [n for n in full.get("nodes", []) if n.get("type") == "action"]
+        trig = trs[0] if trs else {}
+        cfg = trig.get("config") or {}
+        cats = [a.get("category") for a in acs]
+        shape_a = trig.get("category") == "recordCondition" and "conditions" in cfg
+        shape_b = (trig.get("category") == "recordUpdated" and cfg.get("fieldIds")
+                   and any(a.get("category") == "script" and re.search(r"失败", (a.get("config") or {}).get("code", "")) for a in acs))
+        if full.get("isActive") and (shape_a or shape_b):
+            ok = "sendNotification" in cats and "sendEmail" not in cats
+    out.append(result("修正后终态:仅在结果变失败时站内通知(recordCondition 或 recordUpdated+门控)", ok,
+                      [w.get("name") for w in wfs]))
+    c2 = cmds_of(ROOT / "iteration-mt" / "mt-1-notify-correction" / "auto_skill" / "run-1")
+    amended = any(re.search(r"trigger set|action (update|add|delete)", c) for c in c2) and not any(
+        re.search(r"workflow create", c) for c in c2)
+    out.append(result("第二轮是修改现有工作流而非推倒重建", amended,
+                      [c[:80] for c in c2 if "workflow" in c][:3]))
+    return out
+
+
+def grade_mt2():
+    out = []
+    tables = cli("table", "list") or []
+    fb = [t for t in tables if "反馈" in t.get("name", "") or "feedback" in t.get("name", "").lower()]
+    if not fb:
+        return [result("反馈表存在", False, [t.get("name") for t in tables])]
+    tid = fb[0]["id"]
+    flds = cli("field", "list", tid) or []
+    okf = any(f.get("name") == "内容" for f in flds)
+    ch = next((f for f in flds if f.get("name") == "渠道"), None)
+    okc = bool(ch) and set(["微信", "电话", "邮件"]) <= set(
+        c.get("name") for c in (ch.get("options") or {}).get("choices", []))
+    recs = (cli("record", "get", "--table-id", tid) or {}).get("records") or []
+    okr = any((r.get("fields") or {}).get("渠道") == "微信" for r in recs)
+    out.append(result("第一轮产物保留(表+内容字段)", okf, fb[0]["name"]))
+    out.append(result("第二轮增量:渠道字段三选项", okc, (ch or {}).get("options")))
+    out.append(result("示例记录渠道=微信", okr, [(r.get("fields") or {}).get("订单号"), (r.get("fields") or {}).get("渠道")] if not okr else "ok"))
+    return out
+
+
+def main():
+    iteration = sys.argv[1] if len(sys.argv) > 1 else "iteration-mt"
+    g1 = grade_mt1()
+    g2 = grade_mt2()
+    for name, exps in [("mt-1-notify-correction", g1), ("mt-2-incremental-schema", g2)]:
+        rd = ROOT / iteration / name / "auto_skill" / "run-1"
+        (rd / "grading.json").write_text(json.dumps({"expectations": exps}, indent=2, ensure_ascii=False))
+        np = sum(1 for x in exps if x["passed"])
+        print(f"{name:28} {np}/{len(exps)}")
+
+
+if __name__ == "__main__":
+    main()