fix(evolution): clarify learned capabilities and memory capture

desktop/scripts/prepare-bundle.mjs

@@ -30,11 +30,16 @@ const bundleFlavor = process.env.HUGAGENT_DESKTOP_BUNDLE || "full";
 if (!["full", "thin"].includes(bundleFlavor)) {
   throw new Error(`Unknown HUGAGENT_DESKTOP_BUNDLE flavor: ${bundleFlavor}`);
 }
-validateDesktopBuildTarget(
-  process.env.TAURI_ENV_TARGET_TRIPLE
-    ? ["--target", process.env.TAURI_ENV_TARGET_TRIPLE]
-    : [],
-);
+if (bundleFlavor === "full") {
+  // The offline runtime is a native payload, so a cross-architecture target
+  // cannot be packaged with it. Thin bundles carry no runtime — any target,
+  // including universal-apple-darwin, is buildable on any host.
+  validateDesktopBuildTarget(
+    process.env.TAURI_ENV_TARGET_TRIPLE
+      ? ["--target", process.env.TAURI_ENV_TARGET_TRIPLE]
+      : [],
+  );
+}
 const desktopTarget = currentDesktopTarget();
 const dependencyFingerprint = desktopDependencyFingerprint(repoRoot, desktopTarget);
 const python = bundleFlavor === "full" ? findPython() : null;

src/backend/core/evolution/user_settings.py

@@ -258,6 +258,23 @@ def pending_for_user(user_id: str, *, limit: int = 20) -> List[Dict[str, Any]]:
                 action = personal_action(candidate)
                 if action is None:
                     continue
+                change_preview = _change_preview(candidate)
+                if str(candidate.target_kind or "") == "skill":
+                    # Legacy personal cycles could promote a single repeated
+                    # tool call as a "sequence".  Such a candidate has neither
+                    # an order to preserve nor an applicability condition, and
+                    # approving it would install a tautological skill.  Keep
+                    # the evidence intact for audit, but do not present noise as
+                    # a decision the user ought to make.
+                    if change_preview.get("type") == "skill_sequence":
+                        steps = change_preview.get("steps") or []
+                        rules = change_preview.get("ordering_constraints") or []
+                        if len(steps) < 2 and not rules:
+                            continue
+                    elif change_preview.get("type") != "skill_document":
+                        # A skill with no inspectable content is not informed
+                        # consent and would also fail materialisation later.
+                        continue
                 out.append(
                     {
                         "candidate_id": candidate.candidate_id,
@@ -276,7 +293,7 @@ def pending_for_user(user_id: str, *, limit: int = 20) -> List[Dict[str, Any]]:
                         "action": action["action"],
                         "action_label": action["label"],
                         "action_effect": action["effect"],
-                        "change": _change_preview(candidate),
+                        "change": change_preview,
                         "created_at": candidate.created_at.isoformat()
                         if candidate.created_at
                         else None,
@@ -300,7 +317,8 @@ def _change_preview(candidate) -> Dict[str, Any]:
     """
     ir = candidate.ir or {}
     kind = str(candidate.target_kind or "")
-    for change in ir.get("changes") or []:
+    changes = ir.get("changes") or []
+    for change in changes:
         if kind == "skill":
             document = change.get("document")
             if isinstance(document, dict) and document.get("content"):
@@ -327,6 +345,40 @@ def _change_preview(candidate) -> Dict[str, Any]:
                         for op in operations
                     ],
                 }
+    if kind == "skill":
+        # Sequence-only candidates are the common output of the personal
+        # evolution path.  They do not carry an authored SKILL.md, but approval
+        # still materialises a concrete name, description and ordered set of
+        # steps.  Surface that exact structure instead of returning {}, which
+        # made the frontend silently remove the detail button.
+        tools: List[str] = []
+        constraints: List[Dict[str, Any]] = []
+        for change in changes:
+            sequence = change.get("tool_sequence")
+            if isinstance(sequence, list) and sequence and not tools:
+                tools = [str(tool) for tool in sequence]
+            rules = change.get("ordering_constraints")
+            if isinstance(rules, list) and rules:
+                constraints = [dict(rule) for rule in rules if isinstance(rule, dict)]
+
+        rules_only = not tools and bool(constraints)
+        if rules_only:
+            from core.evolution.activation import _tools_from_constraints
+
+            tools = _tools_from_constraints(constraints)
+        if tools:
+            # Reuse the materialiser's wording so the preview and the installed
+            # skill never disagree about what the user accepted.
+            from core.evolution.activation import _skill_description, _skill_label
+
+            return {
+                "type": "skill_sequence",
+                "display_name": _skill_label(tools, rules_only=rules_only),
+                "description": _skill_description(tools, constraints, rules_only=rules_only),
+                "allowed_tools": tools,
+                "steps": tools,
+                "ordering_constraints": constraints,
+            }
     return {}
 
 

src/backend/core/memory/extractors/gate.py

@@ -39,14 +39,28 @@
 - graph: 稳定的实体关系（隶属、负责、依赖、使用、组成、别名、分类）
 - task: 本轮明确提出的多步任务目标（仅本会话内使用）
 
-【判定标准 —— 宁缺毋滥】
-绝大多数对话轮次**没有**值得长期记住的内容。以下都不算：
+【判定标准】
+以下内容都不算：
 - 一次性的问答、查询、闲聊
 - 只对当前这一次请求有效的指令（"这次用表格""先看第三页"）
 - 助手自己说的内容（除非用户明确认可为约定）
 - 模型本来就会的通用常识
 
-只有当**下次对话不知道这条信息就会做错或重复问**时，才算值得记住。
+以下内容应当记住：
+- 用户明确要求“记住”“记录一下”“以后”“下次”“默认”“一律”“不要再”，并给出需要延续的信息；明确说明仅本次有效的除外
+- 用户直接陈述或确认的身份、岗位、部门、单位、职责、联系方式等相对稳定的信息 → identity
+- 用户反复适用的表达与交付偏好，例如语言、详略、格式、称呼、禁忌和默认输出方式 → preference
+- 用户所在组织、团队或项目特有的可复用规则，例如术语定义、统计口径、计算方法、步骤顺序、命名规范、目录约定、必做校验、交付标准和风险红线 → procedural
+- 用户直接陈述或确认的稳定实体关系，例如谁负责什么、哪个部门隶属哪里、系统依赖什么、产品由哪些模块组成、名称与别名的对应关系 → graph
+- 用户明确提出的多步任务，以及尚未完成的目标、步骤、进度、截止时间和交付物 → task；此类信息只在本会话内保留，不作为长期事实
+- 用户对已有信息的纠正、替换或失效声明，例如“不是 A，是 B”“我现在改负责 X”“以后不用旧口径”；应记住新信息并用于更新旧记忆
+- 助手提出的方案被用户明确确认为今后的规则或偏好，例如“对，以后都按这个执行”；按对应类别记住
+
+总判断原则：
+- 只有用户明确说出或明确确认的内容才能作为事实写入，不要把助手的推测当成用户记忆
+- 判断它在未来是否仍然有用：如果下次对话不知道这条信息会导致做错、重复询问或违反用户约定，就应该记住
+- 用户明确要求记住时应优先保留；若内容同时符合多个类别，可以返回多个类别
+- 不要因为信息只出现一次就拒绝；稳定性由语义和用户意图判断，而不是由出现次数判断
 
 【输出格式（严格 JSON，无代码块包裹）】
 {{"classes": ["identity", "procedural"]}}

src/backend/tests/evolution/test_personal_approval.py

@@ -200,6 +200,34 @@ def test_the_tool_sequence_is_surfaced_so_approval_is_informed(db):
     assert pending[0]["tool_sequence"] == ["a", "b"]
 
 
+def test_sequence_only_skill_has_an_inspectable_change_preview(db):
+    for i in range(5):
+        _episode(db, f"e{i}", "alice")
+    _candidate(db, "c1", [f"e{i}" for i in range(5)])
+    db.commit()
+
+    change = US.pending_for_user("alice")[0]["change"]
+    assert change["type"] == "skill_sequence"
+    assert change["display_name"] == "固定调用顺序：a → b"
+    assert change["steps"] == ["a", "b"]
+    assert "`a`、`b`" in change["description"]
+
+
+def test_legacy_single_tool_candidate_is_hidden_without_deleting_evidence(db):
+    for i in range(5):
+        _episode(db, f"e{i}", "alice")
+    _candidate(
+        db,
+        "single",
+        [f"e{i}" for i in range(5)],
+        ir={"changes": [{"tool_sequence": ["view_text_file"]}]},
+    )
+    db.commit()
+
+    assert US.pending_for_user("alice") == []
+    assert db.get(EvolutionCandidate, "single") is not None
+
+
 # ── The queue lists only what it can carry out ───────────────────────────────
 #
 # Every case below used to appear with a 「为我启用」 button whose sole possible

src/backend/tests/evolution/test_promotion_chain.py

@@ -101,6 +101,24 @@ def test_recurring_pattern_is_compiled_into_a_skill_proposal():
     assert "重新规划" not in proposal.rationale
 
 
+def test_one_repeated_tool_call_is_not_mislabelled_as_a_sequence_skill():
+    pattern = P.Pattern(
+        kind=P.PATTERN_SUCCESS_SUBSEQUENCE,
+        signature="call_subagent",
+        support=9,
+        success_rate=1.0,
+        tool_sequence=["call_subagent"],
+        episode_ids=[f"ep-{i}" for i in range(9)],
+    )
+
+    assert (
+        P.promote_tool_sequence_to_skill(
+            pattern, skill_credit=0.9, workflow_credit=0.1
+        )
+        is None
+    )
+
+
 def test_promotion_is_refused_when_orchestration_explains_it_better():
     proposal = P.promote_tool_sequence_to_skill(
         _success_pattern(),

src/frontend/src/api.ts

@@ -3966,6 +3966,14 @@ export type EvolutionChangePreview =
       allowed_tools: string[];
       content: string;
     }
+  | {
+      type: 'skill_sequence';
+      display_name: string;
+      description: string;
+      allowed_tools: string[];
+      steps: string[];
+      ordering_constraints: Array<Record<string, unknown>>;
+    }
   | {
       type: 'memory_ops';
       operations: Array<{