[Infra] Cowork GMS 모델 및 티켓 분기 개선

app/agent.py

@@ -6,6 +6,8 @@
 
 from typing import Any
 
+_STRUCTURAL_LABELS = {"작업 내용:", "목적/배경:", "상세 내용:", "완료 조건:"}
+
 
 class AgentError(RuntimeError):
     pass
@@ -112,7 +114,12 @@ def _parse_plan_payload(payload: dict[str, Any], source_lines: list[str]) -> Pla
             raise AgentError("invalid excluded item")
         covered.add(source_index)
         excluded.append(text.strip())
-    if covered != set(range(len(source_lines))):
+    required = {
+        index
+        for index, line in enumerate(source_lines)
+        if line.strip() not in _STRUCTURAL_LABELS
+    }
+    if not required.issubset(covered):
         raise AgentError("planning response dropped input lines")
     return Plan(tasks=tasks, excluded=excluded)
 
@@ -152,13 +159,19 @@ def plan(self, raw_input: str) -> Plan:
 Rules:
 - Every source_index must appear in at least one task's source_indices or in excluded.
 - Reuse a source_index across tasks only when that single line explicitly contains separate deliverables.
-- Group a task title, its explanatory lines, and its completion criteria into one task.
-- Each genuinely independent deliverable becomes a separate task. Do not turn every detail or completion-criteria bullet into a separate task.
+- First identify deliverable candidates, then attach context and acceptance criteria to each candidate.
+- A single `제목:` block is not proof that all details belong to one task.
+- Split candidates when any one is true: they belong to different repositories or owner teams; each can use a separate PR or merge; one can be postponed while another is still completed; each has independent acceptance criteria or verification; or each can be independently deployed or rolled back.
+- Group candidates only when all are true: they share one owner area; they form one PR or deployment unit; no detail has useful value alone; and one verification result decides completion of the whole group.
+- When a title or detail joins different actions with `및`, `·`, or `/`, re-check whether those actions are independent deliverables. Do not preserve an umbrella title merely because the input used one title block.
+- Mandatory example: Documentation contracts, repository settings, and an AI-agent harness are separate deliverable candidates because each has a useful artifact and independent verification. Input containing those three categories MUST produce three tasks: documentation and contracts; repository configuration; AI-agent harness. A shared umbrella purpose or a possible single PR does not override this rule. Attach each checklist or verification line to its matching task; MUST NOT produce a fourth task from an acceptance criterion.
+- Counterexample: build, push, and digest verification for one image deployment are steps of one task because they form one release unit and have no useful independent completion.
+- Do not turn implementation steps or acceptance-criteria bullets for one deliverable into separate tasks.
 - By default, separate standalone action lines are separate tasks, especially short unindented lines like `cowork 고도화` and `지라 + 깃허브 연동`.
-- Group lines only when their structure clearly marks them as context for a task, such as `작업 내용:`, `목적/배경:`, `상세 내용:`, `완료 조건:`, or bullets and checklist items under those sections.
-- Each repeated `제목:` line starts a new task; attach the following `작업 내용:`, `목적/배경:`, `상세 내용:`, and `완료 조건:` sections to that task until the next `제목:` line.
+- Structured labels such as `작업 내용:`, `목적/배경:`, `상세 내용:`, and `완료 조건:` mark context, not an unconditional instruction to merge every detail under them.
+- Each repeated `제목:` line always starts a new task; a single `제목:` block may still produce multiple tasks when its details meet the independence criteria above.
 - Treat `- [ ]` lines under `완료 조건:` as acceptance criteria, not as separate tasks, and preserve them under the `## 완료 조건` heading in description.
-- Never merge separate standalone action lines merely because they belong to the same project or theme.
+- Never merge separate standalone action lines merely because they belong to the same project, theme, or umbrella title.
 - A task may reference multiple related source lines. Split a single source line only when it clearly contains separate deliverables.
 - Exclude completed items (- [x], strikethrough, 완료), schedules/meetings, and notes/thoughts.
 - Summary is one concise Korean action phrase, maximum 255 characters.
@@ -170,13 +183,26 @@ def plan(self, raw_input: str) -> Plan:
 INPUT_LINES_JSON={encoded}
 """
         payload = self._run(prompt, max_turns=3)
-        try:
-            return _parse_plan_payload(payload, source_lines)
-        except AgentError:
-            correction_prompt = (
-                prompt
-                + "\nschema correction retry: Return the exact top-level wrapper with only "
-                '"tasks" and "excluded" arrays. Never return a single task object.'
-            )
-            corrected_payload = self._run(correction_prompt, max_turns=2)
-            return _parse_plan_payload(corrected_payload, source_lines)
+        required_indices = [
+            index
+            for index, line in enumerate(source_lines)
+            if line.strip() not in _STRUCTURAL_LABELS
+        ]
+        for attempt in range(3):
+            try:
+                return _parse_plan_payload(payload, source_lines)
+            except AgentError as exc:
+                if attempt == 2:
+                    raise
+                correction_prompt = (
+                    prompt
+                    + f"\nValidation failure: {exc}"
+                    + f"\nRequired content source_indices: {required_indices}."
+                    + " Every required index must appear in at least one task or excluded item."
+                    + " Attach missing context and acceptance criteria to the matching task;"
+                    + " do not add a task solely to cover an acceptance criterion."
+                    + " Return the exact top-level wrapper with only "
+                    + '"tasks" and "excluded" arrays. Never return a single task object.'
+                )
+                payload = self._run(correction_prompt, max_turns=2)
+        raise AgentError("planning validation failed")

app/api_agent.py

@@ -36,13 +36,17 @@ def _run(self, prompt: str, *, max_turns: int) -> dict[str, Any]:
                 self.settings.gms_api_url,
                 headers={
                     "Content-Type": "application/json",
-                    "x-api-key": self.settings.gms_api_key,
-                    "anthropic-version": "2023-06-01",
+                    "Authorization": f"Bearer {self.settings.gms_api_key}",
                 },
                 json={
                     "model": self.settings.gms_model,
-                    "max_tokens": 2048,
-                    "messages": [{"role": "user", "content": prompt}],
+                    "messages": [
+                        {
+                            "role": "developer",
+                            "content": "Answer in Korean. Return only the JSON object requested by the user message.",
+                        },
+                        {"role": "user", "content": prompt},
+                    ],
                 },
             )
         except httpx.TimeoutException as exc:
@@ -54,13 +58,10 @@ def _run(self, prompt: str, *, max_turns: int) -> dict[str, Any]:
         if response.status_code != 200:
             raise AgentError(f"GMS request rejected: HTTP {response.status_code}")
         try:
-            content = response.json()["content"]
-            text = "\n".join(
-                item["text"]
-                for item in content
-                if isinstance(item, dict) and item.get("type") == "text"
-            )
-        except (ValueError, KeyError, TypeError) as exc:
+            text = response.json()["choices"][0]["message"]["content"]
+            if not isinstance(text, str) or not text.strip():
+                raise TypeError
+        except (ValueError, KeyError, IndexError, TypeError) as exc:
             raise AgentError("invalid GMS response") from exc
         return _extract_json(text)
 

app/api_smoke.py

@@ -6,7 +6,7 @@
 
 from .agent import _extract_json, _parse_plan_payload
 
-GMS_URL = "https://gms.ssafy.io/gmsapi/api.anthropic.com/v1/messages"
+GMS_URL = "https://gms.ssafy.io/gmsapi/api.openai.com/v1/chat/completions"
 JIRA_BASE_URL = "https://ssafy.atlassian.net"
 PROJECT_KEY = "S15P11A705"
 BOARD_ID = "13941"
@@ -27,14 +27,13 @@ def run_read_only_smoke(
     gms = http.post(
         GMS_URL,
         headers={
-            "x-api-key": gms_key,
-            "anthropic-version": "2023-06-01",
+            "authorization": f"Bearer {gms_key}",
             "content-type": "application/json",
         },
         json={
-            "model": "claude-opus-4-8",
-            "max_tokens": 1024,
+            "model": "gpt-5.4-nano",
             "messages": [
+                {"role": "developer", "content": "Answer in Korean. Return JSON only."},
                 {
                     "role": "user",
                     "content": (
@@ -50,14 +49,11 @@ def run_read_only_smoke(
     if gms.status_code != 200:
         raise RuntimeError(f"GMS request failed with HTTP {gms.status_code}")
     try:
-        body = gms.json()
-        text = next(
-            item["text"]
-            for item in body["content"]
-            if isinstance(item, dict) and item.get("type") == "text"
-        )
+        text = gms.json()["choices"][0]["message"]["content"]
+        if not isinstance(text, str) or not text.strip():
+            raise TypeError
         _parse_plan_payload(_extract_json(text), [source])
-    except (KeyError, TypeError, ValueError, StopIteration) as exc:
+    except (KeyError, IndexError, TypeError, ValueError) as exc:
         raise RuntimeError("GMS returned an invalid structured response") from exc
 
     jira = http.get(

app/config.py

@@ -13,9 +13,9 @@ class Settings:
     cookie_secure: bool
     agent_timeout_seconds: int
     mattermost_webhook_url: str | None
-    gms_api_url: str = "https://gms.ssafy.io/gmsapi/api.anthropic.com/v1/messages"
+    gms_api_url: str = "https://gms.ssafy.io/gmsapi/api.openai.com/v1/chat/completions"
     gms_api_key: str | None = None
-    gms_model: str = "claude-opus-4-8"
+    gms_model: str = "gpt-5.4-nano"
     jira_email: str | None = None
     jira_api_token: str | None = None
     jira_board_id: int = 13941
@@ -55,10 +55,10 @@ def from_env(cls) -> "Settings":
             mattermost_webhook_url=mattermost_webhook_url,
             gms_api_url=os.getenv(
                 "GMS_API_URL",
-                "https://gms.ssafy.io/gmsapi/api.anthropic.com/v1/messages",
+                "https://gms.ssafy.io/gmsapi/api.openai.com/v1/chat/completions",
             ),
             gms_api_key=gms_api_key,
-            gms_model=os.getenv("GMS_MODEL", "claude-opus-4-8"),
+            gms_model=os.getenv("GMS_MODEL", "gpt-5.4-nano"),
             jira_email=jira_email,
             jira_api_token=jira_api_token,
             jira_board_id=int(os.getenv("JIRA_BOARD_ID", "13941")),

app/static/index.html

@@ -92,10 +92,25 @@ <h1 id="input-guide-title">AI와 대화해서 티켓을 구체화하세요</h1>
 
 ## 티켓 분리 규칙
 
-1. 서로 독립적으로 구현, 검증 또는 배포할 수 있는 작업은 각각 별도 티켓으로 분리합니다.
-2. 하나의 작업을 수행하기 위한 세부 단계는 별도 티켓으로 과도하게 나누지 않습니다.
-3. 회의, 일정, 단순 메모, 아이디어만 있는 내용, 이미 완료된 작업은 티켓에서 제외합니다.
-4. 제외 여부가 불명확하면 추측하지 말고 사용자에게 질문합니다.
+1. 최종 결과를 작성하기 전에 독립 산출물 후보를 먼저 나열합니다.
+2. 다음 중 하나라도 해당하면 별도 티켓 후보로 분리합니다.
+   - 서로 다른 저장소 또는 소유 팀이 작업함
+   - 별도의 PR로 구현하거나 병합할 수 있음
+   - 하나를 연기해도 다른 작업을 완료할 수 있음
+   - 각각 독립적인 완료 조건과 검증 방법이 있음
+   - 독립적으로 배포하거나 롤백할 수 있음
+3. 다음 조건을 모두 만족할 때만 하나의 티켓으로 묶습니다.
+   - 동일한 담당 영역에서 수행함
+   - 하나의 PR 또는 하나의 배포 단위임
+   - 세부 항목 하나만 완료해서는 독립적인 가치가 없음
+   - 완료 여부를 하나의 검증 결과로 판단할 수 있음
+4. 제목이나 상세 내용에서 서로 다른 행동이 `및`, `·`, `/`로 연결되면 별도 티켓이 필요한지 다시 검토합니다.
+5. 여러 후보가 있거나 분리 여부가 불명확하면 세부 질문 전에 “현재 N개의 티켓 후보로 보입니다: A, B, C. 이대로 분리할까요?”라고 한 번 확인합니다.
+   - 분리 예시: 문서 계약 작성, 저장소 설정, 에이전트 하네스 구성이 함께 있으면 공통 목적이나 단일 PR 가능성과 관계없이 반드시 3개의 별도 티켓으로 분리합니다. 각 체크리스트는 대응하는 구현 티켓의 완료 조건으로 붙이고, 완료 조건을 네 번째 티켓으로 분리하지 않습니다.
+   - 유지 예시: 이미지 하나를 배포하기 위한 build, push, digest 확인은 하나의 배포 단위를 이루므로 한 티켓의 세부 단계입니다.
+6. 하나의 작업을 수행하기 위한 세부 단계와 그 완료 조건은 별도 티켓으로 과도하게 나누지 않습니다.
+7. 회의, 일정, 단순 메모, 아이디어만 있는 내용, 이미 완료된 작업은 티켓에서 제외합니다.
+8. 제외 여부가 불명확하면 추측하지 말고 사용자에게 질문합니다.
 
 ## 사실 기반 규칙
 
@@ -128,7 +143,7 @@ <h1 id="input-guide-title">AI와 대화해서 티켓을 구체화하세요</h1>
 - [ ] 검증 가능한 완료 기준 1
 - [ ] 검증 가능한 완료 기준 2
 
-여러 티켓이 필요한 경우 위 형식을 반복합니다.
+여러 티켓이 필요한 경우 위 형식을 반복하고, 각 티켓마다 제목 항목을 다시 작성합니다. 하나의 제목 아래에 독립 산출물을 함께 넣지 않습니다.
 
 ## 제목 작성 규칙
 

tests/test_agent.py

@@ -2,7 +2,7 @@
 
 import pytest
 
-from app.agent import AgentError, _parse_plan_payload
+from app.agent import AgentError, PlanningAgent, _parse_plan_payload
 
 
 def test_planner_schema_covers_every_input_line():
@@ -46,3 +46,126 @@ def test_planner_rejects_malformed_source_indices(source_indices):
     }
     with pytest.raises(AgentError, match="invalid planned source indices"):
         _parse_plan_payload(payload, ["로그인 오류 재현"])
+
+
+def test_planner_rechecks_independent_deliverables_inside_one_title_block():
+    class CapturingPlanner(PlanningAgent):
+        prompt = ""
+
+        def _run(self, prompt, *, max_turns):
+            del max_turns
+            self.prompt = prompt
+            return {
+                "tasks": [
+                    {
+                        "source_indices": list(range(8)),
+                        "summary": "프론트엔드 레포 초기 세팅",
+                        "description": "초기 세팅과 에이전트 하네스 구성",
+                    }
+                ],
+                "excluded": [],
+            }
+
+    planner = CapturingPlanner()
+    planner.plan(
+        """제목: 프론트엔드 레포 초기 세팅 및 에이전트 하네스 구성
+작업 내용:
+- AGENTS.md와 API 계약 문서 작성
+- PR 템플릿과 VS Code 설정 추가
+- Claude agents와 commands 구성
+완료 조건:
+- 문서와 설정이 병합됨
+- Claude Code에서 하네스가 인식됨"""
+    )
+
+    for rule in (
+        "A single `제목:` block is not proof that all details belong to one task.",
+        "a separate PR or merge",
+        "postponed while another is still completed",
+        "independent acceptance criteria or verification",
+        "independently deployed or rolled back",
+        "`및`, `·`, or `/`",
+        "Documentation contracts, repository settings, and an AI-agent harness are separate deliverable candidates",
+        "MUST produce three tasks",
+        "documentation and contracts; repository configuration; AI-agent harness",
+        "MUST NOT produce a fourth task from an acceptance criterion",
+        "build, push, and digest verification for one image deployment are steps of one task",
+    ):
+        assert rule in planner.prompt
+
+
+def test_planner_allows_omitted_structural_labels_but_keeps_content_coverage():
+    lines = [
+        "제목: 레포 설정",
+        "작업 내용:",
+        "- PR 템플릿 추가",
+        "완료 조건:",
+        "- PR 체크리스트 표시",
+    ]
+    payload = {
+        "tasks": [
+            {
+                "source_indices": [0, 2, 4],
+                "summary": "PR 템플릿 설정",
+                "description": "PR 체크리스트가 표시되도록 설정한다.",
+            }
+        ],
+        "excluded": [],
+    }
+
+    plan = _parse_plan_payload(payload, lines)
+
+    assert [task.summary for task in plan.tasks] == ["PR 템플릿 설정"]
+
+
+def test_planner_retry_names_validation_failure_and_required_indices():
+    class RetryingPlanner(PlanningAgent):
+        prompts = []
+
+        def _run(self, prompt, *, max_turns):
+            del max_turns
+            self.prompts.append(prompt)
+            indices = [0] if len(self.prompts) == 1 else [0, 1]
+            return {
+                "tasks": [
+                    {
+                        "source_indices": indices,
+                        "summary": "레포 설정",
+                        "description": "PR 템플릿과 VS Code 설정을 추가한다.",
+                    }
+                ],
+                "excluded": [],
+            }
+
+    planner = RetryingPlanner()
+    plan = planner.plan("PR 템플릿 추가\nVS Code 설정 추가")
+
+    assert [task.summary for task in plan.tasks] == ["레포 설정"]
+    assert "Validation failure: planning response dropped input lines" in planner.prompts[1]
+    assert "Required content source_indices: [0, 1]" in planner.prompts[1]
+
+
+def test_planner_allows_two_schema_correction_retries():
+    class ThirdAttemptPlanner(PlanningAgent):
+        calls = 0
+
+        def _run(self, prompt, *, max_turns):
+            del prompt, max_turns
+            self.calls += 1
+            indices = [0] if self.calls < 3 else [0, 1]
+            return {
+                "tasks": [
+                    {
+                        "source_indices": indices,
+                        "summary": "레포 설정",
+                        "description": "레포 설정을 추가한다.",
+                    }
+                ],
+                "excluded": [],
+            }
+
+    planner = ThirdAttemptPlanner()
+    plan = planner.plan("PR 템플릿 추가\nVS Code 설정 추가")
+
+    assert planner.calls == 3
+    assert [task.summary for task in plan.tasks] == ["레포 설정"]