Clarify stable schema tokens for localized WATCHLIST entries

.agents/skills/watchlist-md/SKILL.md

@@ -77,11 +77,12 @@ Use this item shape:
 - next_step_on_fail:
 ```
 
-For open items, populate: ID, status, priority, owner, due_at, created_at,
-source, trigger, action, and done_when. Use safe source pointers only; never store
-signed, tokenized, private, or credential-bearing links. Keep last_checked_at and
-result blank until checked. Use `assistant_on_review` for explicit-review help;
-treat legacy `owner: agent` as `assistant_on_review`.
+For open items, keep field keys and enum values in English; populate: ID,
+status, priority, owner, due_at, created_at, source, trigger, action, and
+done_when. Localize only titles and free-text values. Use safe pointers; never
+store signed, tokenized, private, or credential-bearing links. Keep
+last_checked_at and result blank until checked. Use `assistant_on_review` for
+explicit-review help; treat legacy `owner: agent` as `assistant_on_review`.
 
 Generate IDs from the WATCHLIST timezone: WATCHLIST.md `timezone:` field >
 explicit user timezone > environment/user timezone > Asia/Seoul. Re-read

evals/cases/localized-schema-tokens-kr.json

@@ -0,0 +1,65 @@
+{
+  "id": "localized-schema-tokens-kr",
+  "prompt": "한국어로 작성하되 WATCHLIST.md에 추가해줘. 오늘 17:00에 CI 확인.",
+  "locale": "ko",
+  "fixed_now": "2026-05-14T16:30:00+09:00",
+  "fixture": "empty.watchlist.md",
+  "should_trigger_skill": true,
+  "expected": {
+    "operation": "add_item",
+    "status": "open",
+    "due_at": "2026-05-14T17:00:00+09:00",
+    "scheduler": "none",
+    "required_fields": [
+      "source",
+      "trigger",
+      "action",
+      "done_when"
+    ],
+    "forbidden_response_substrings": [
+      "I'll remind you",
+      "I will remind you",
+      "I'll check later",
+      "I will check later",
+      "자동으로 알려드릴게요",
+      "제가 나중에 확인할게요"
+    ],
+    "schema_tokens": {
+      "must_use_field_keys": [
+        "schema_version",
+        "automation",
+        "timezone",
+        "status",
+        "priority",
+        "owner",
+        "due_at",
+        "created_at",
+        "source",
+        "trigger",
+        "action",
+        "done_when",
+        "last_checked_at",
+        "result",
+        "next_step_on_fail"
+      ],
+      "must_use_enum_values": [
+        "open",
+        "P1",
+        "assistant_on_review"
+      ],
+      "must_not_use_localized_schema_tokens": [
+        "상태",
+        "우선순위",
+        "담당자",
+        "기한",
+        "생성일",
+        "출처",
+        "계기",
+        "작업",
+        "완료조건",
+        "열림",
+        "어시스턴트"
+      ]
+    }
+  }
+}

evals/check_policy_markers.py

@@ -41,7 +41,7 @@
         "pending result for later review",
         "safe link",
         "Scope pre-authorized watchlist recording",
-        "For open items, populate",
+        "keep field keys and enum values in English",
         "confirm ID, due_at",
         "WATCHLIST.md `timezone:` field",
         "environment/user timezone",

evals/check_semantic_cases.py

@@ -195,6 +195,7 @@ def validate_add_item(
     case_id: str,
     expected: dict[str, object],
     errors: list[str],
+    locale: object = None,
 ) -> None:
     require_keys(
         expected,
@@ -248,6 +249,92 @@ def validate_add_item(
                     f"{forbidden_operation}"
                 )
 
+    validate_schema_tokens(case_id, locale, expected, errors)
+
+
+def validate_schema_tokens(
+    case_id: str,
+    locale: object,
+    expected: dict[str, object],
+    errors: list[str],
+) -> None:
+    schema_tokens = expected.get("schema_tokens")
+    if schema_tokens is None:
+        return
+    if not isinstance(schema_tokens, dict):
+        errors.append(f"{case_id}: expected.schema_tokens must be an object")
+        return
+
+    require_keys(
+        schema_tokens,
+        {
+            "must_use_field_keys",
+            "must_use_enum_values",
+            "must_not_use_localized_schema_tokens",
+        },
+        case_id,
+        errors,
+        "expected.schema_tokens",
+    )
+    field_keys = require_string_list(
+        schema_tokens,
+        "must_use_field_keys",
+        case_id,
+        errors,
+        "expected.schema_tokens",
+    )
+    enum_values = require_string_list(
+        schema_tokens,
+        "must_use_enum_values",
+        case_id,
+        errors,
+        "expected.schema_tokens",
+    )
+    localized_tokens = require_string_list(
+        schema_tokens,
+        "must_not_use_localized_schema_tokens",
+        case_id,
+        errors,
+        "expected.schema_tokens",
+    )
+
+    required_field_keys = {
+        "schema_version",
+        "automation",
+        "timezone",
+        "status",
+        "priority",
+        "owner",
+        "due_at",
+        "created_at",
+        "source",
+        "trigger",
+        "action",
+        "done_when",
+        "last_checked_at",
+        "result",
+        "next_step_on_fail",
+    }
+    missing_field_keys = sorted(required_field_keys - field_keys)
+    if missing_field_keys:
+        errors.append(
+            f"{case_id}: schema_tokens.must_use_field_keys missing "
+            f"{', '.join(missing_field_keys)}"
+        )
+
+    if not {"open", "P1", "assistant_on_review"}.issubset(enum_values):
+        errors.append(
+            f"{case_id}: schema_tokens.must_use_enum_values must include "
+            "open, P1, and assistant_on_review"
+        )
+    if locale != "ko":
+        return
+    if not {"상태", "우선순위", "담당자", "기한", "열림"}.issubset(localized_tokens):
+        errors.append(
+            f"{case_id}: schema_tokens.must_not_use_localized_schema_tokens "
+            "must include 상태, 우선순위, 담당자, 기한, and 열림"
+        )
+
 
 def validate_storage_contract(
     case_id: str,
@@ -537,7 +624,7 @@ def validate_case(
         return
 
     if operation == "add_item":
-        validate_add_item(case_id, expected, errors)
+        validate_add_item(case_id, expected, errors, case.get("locale"))
     elif operation == "archive_items":
         validate_archive_items(case_id, expected, errors)
     elif operation == "complete_item":

evals/prompts.csv

@@ -1,5 +1,6 @@
 id,should_trigger,prompt,expected
 add-kr-01,true,"WATCHLIST.md에 추가해줘. 오늘 17:00에 GitHub Actions 결과 확인.","creates one open item; due_at is ISO-8601 +09:00 when current time is available; scheduler status says none"
+localized-schema-tokens-kr,true,"한국어로 작성하되 WATCHLIST.md에 추가해줘. 오늘 17:00에 CI 확인.","creates one open item using English schema field keys and enum values while allowing Korean title and free-text values"
 add-kr-02,true,"배포가 방금 시작됐어. 30분 뒤에 에러 로그 확인해야 해.","creates a deferred check with concrete due_at when current time is available, otherwise uses unscheduled and records ambiguity"
 review-kr-01,true,"오늘 확인할 WATCHLIST.md 보여줘.","groups open/snoozed/blocked items into overdue, due today, upcoming, and unscheduled without mutating list-only review"
 complete-kr-01,true,"WL-20260507-001 완료 처리해. CI 모두 pass 했어.","sets status done; fills last_checked_at and result; moves the item under ## Done by default when that section exists; does not delete the item"

evals/self_checks.yaml

@@ -20,6 +20,49 @@ cases:
         - trigger
         - action
         - done_when
+  - id: localized-schema-tokens-kr
+    prompt: "한국어로 작성하되 WATCHLIST.md에 추가해줘. 오늘 17:00에 CI 확인."
+    expected:
+      status: open
+      due_at: "2026-05-14T17:00:00+09:00"
+      required_fields:
+        - source
+        - trigger
+        - action
+        - done_when
+      schema_tokens:
+        must_use_field_keys:
+          - schema_version
+          - automation
+          - timezone
+          - status
+          - priority
+          - owner
+          - due_at
+          - created_at
+          - source
+          - trigger
+          - action
+          - done_when
+          - last_checked_at
+          - result
+          - next_step_on_fail
+        must_use_enum_values:
+          - open
+          - P1
+          - assistant_on_review
+        must_not_use_localized_schema_tokens:
+          - 상태
+          - 우선순위
+          - 담당자
+          - 기한
+          - 생성일
+          - 출처
+          - 계기
+          - 작업
+          - 완료조건
+          - 열림
+          - 어시스턴트
   - id: add-kr-02
     prompt: "배포가 방금 시작됐어. 30분 뒤에 에러 로그 확인해야 해."
     expected:

evals/test_check_watchlist.py

@@ -593,8 +593,9 @@ def test_skill_runtime_polish_markers_stay_precise(self):
             "> explicit user timezone > environment/user timezone > Asia/Seoul."
         )
         required_information = (
-            "For open items, populate: ID, status, priority, owner, due_at, "
-            "created_at, source, trigger, action, and done_when."
+            "For open items, keep field keys and enum values in English; "
+            "populate: ID, status, priority, owner, due_at, created_at, source, "
+            "trigger, action, and done_when."
         )
 
         self.assertIn("pending result for later review", text)
@@ -610,6 +611,7 @@ def test_skill_runtime_polish_markers_stay_precise(self):
         self.assertIn("due_at", text)
         self.assertNotIn("due time", text)
         self.assertIn(required_information, " ".join(text.split()))
+        self.assertIn("Localize only titles and free-text values", " ".join(text.split()))
         self.assertNotIn("done condition", text)
         self.assertIn("confirm ID, due_at, action, done_when, and scheduler status", " ".join(text.split()))
         self.assertIn("watchlist timezone", normalized_text)