[codex] sync skills to named client profiles

docs/skill-management.md

@@ -33,6 +33,15 @@ DNA Memory 将“长期记忆”和“行为 Skill”分开治理：
 Hermes 的目标目录必须链接到同一个 `skill_root` 真源。不要复制三份后分别
 修改，否则召回边界和隐私限制会随客户端漂移。
 
+`platform_skill_roots` 的键可以使用 `platform:instance`，例如
+`hermes:dev`。当注册表目标为 `hermes` 时，`skills sync` 会自动展开到
+所有已配置的 `hermes:*` 实例；新增 Hermes profile 后先登记其 Skill 根目录，
+再运行 `skills sync --apply`。不需要在注册表中重复列出每个 profile。
+
+Memory Loop 必须传真实客户端元数据：Hermes 使用 `client=hermes`。
+只有运行时明确暴露稳定会话 ID 时才传 `session_id`；否则省略，禁止使用
+`session_12345` 等占位值。
+
 对依赖历史的实质任务，Memory Loop 是强制前置步骤；“继续之前的工作”、
 “按上次方案”、已有项目路径、用户长期偏好、已知错误和开放事项都属于触发
 条件。翻译、当前时间、一步格式化等完全自包含任务仍然跳过召回。

scripts/skill_manager.py

@@ -74,7 +74,16 @@ def build_sync_plan(shared_root: Path, platform_roots: Dict[str, Path], registry
     plan = []
     for name, settings in sorted(registry.get("skills", {}).items()):
         source = shared_root / name
-        for platform in settings.get("targets", []):
+        platforms = []
+        for target in settings.get("targets", []):
+            matches = [
+                platform for platform in platform_roots
+                if platform == target or platform.startswith(f"{target}:")
+            ]
+            for platform in matches or [target]:
+                if platform not in platforms:
+                    platforms.append(platform)
+        for platform in platforms:
             root = platform_roots.get(platform)
             target = (root / name) if root else Path("<unconfigured>") / name
             if root is None:

skills/dna-memory-loop/SKILL.md

@@ -16,7 +16,10 @@ names an existing project or path, or depends on a durable preference, known err
 workflow, project state, or open loop.
 
 1. Extract one to four distinctive terms from the request, project, error, or expected result.
-2. Call `memory_recall` separately for each term. Include the real client and session ID when available.
+2. Call `memory_recall` separately for each term. Always pass the real client:
+   use `client=codex`, `client=claude-code`, or `client=hermes` for those runtimes.
+   Never invent client or session metadata. Pass the current stable session ID only
+   when the runtime exposes it; otherwise omit session_id instead of using a placeholder.
 3. Deduplicate by memory ID. Inject at most five memories and about 2,000 tokens total.
 4. Use only relevant results. Current files, processes, remote state, and tests override stale memory.
 

tests/test_bundled_skills.py

@@ -46,3 +46,12 @@ def test_memory_loop_requires_recall_for_history_dependent_work():
     assert "one to four distinctive terms" in body
     assert "simple, self-contained" in body
     assert "Memory failure must not block" in body
+
+
+def test_memory_loop_requires_truthful_client_metadata():
+    body = (ROOT / "skills/dna-memory-loop/SKILL.md").read_text(encoding="utf-8")
+
+    assert "client=hermes" in body
+    assert "Never invent client or session metadata" in body
+    assert "omit session_id" in body
+    assert "placeholder" in body

tests/test_skill_manager.py

@@ -52,3 +52,21 @@ def test_sync_plan_never_overwrites_conflicts(tmp_path):
     assert actions[("shared-ok", "codex")] == "ok"
     assert actions[("shared-ok", "claude")] == "blocked_conflict"
     assert actions[("missing-target", "codex")] == "missing_source"
+
+
+def test_sync_plan_expands_platform_target_to_named_instances(tmp_path):
+    shared = tmp_path / "shared"
+    source = skill(shared, "memory-loop", "---\nname: memory-loop\n---\n")
+    roots = {
+        "hermes": tmp_path / "hermes-default",
+        "hermes:htian": tmp_path / "hermes-htian",
+        "codex": tmp_path / "codex",
+    }
+    registry = {"skills": {"memory-loop": {"targets": ["hermes"]}}}
+
+    plan = build_sync_plan(shared, roots, registry)
+
+    assert [(item.platform, item.action, item.source) for item in plan] == [
+        ("hermes", "create_link", str(source)),
+        ("hermes:htian", "create_link", str(source)),
+    ]