Ai/issue 2 devloop

AGENTS.md

@@ -27,8 +27,9 @@ work:
    ```
 
 7. Only after approval may AI create a branch and PR.
-8. The MVP implementation creates a PR scaffold only. It must not change service
-   code automatically.
+8. The GitHub Actions implementation creates a PR scaffold only.
+9. A human may then comment `/ai implement` to allow the local mini PC runner to
+   continue that PR branch with Codex CLI.
 
 ## Hard Rules
 
@@ -54,10 +55,23 @@ work:
 ## Approval Rules
 
 - The approval signal is a GitHub Issue comment containing `/ai approve`.
+- The implementation signal is a GitHub Issue comment containing `/ai implement`.
 - The workflow should only act on Issues labeled for DevLoop proposals.
 - Optional repository variable `DEVLOOP_APPROVERS` can restrict who may approve.
   Use a comma-separated list of GitHub usernames.
 
+## Continuing An Approved DevLoop PR
+
+When asked to continue an approved DevLoop PR:
+
+1. Confirm the current branch starts with `ai/issue-`.
+2. Read `docs/ai-ideas/issue-N.md`.
+3. Read `AI_POLICY.md`.
+4. Implement only the approved scope.
+5. Run relevant tests.
+6. Commit to the same branch.
+7. Push the branch so the existing PR updates.
+
 ## Suggested Agent Behavior
 
 - Prefer documentation, tests, and narrow workflow improvements before product

AI_POLICY.md

@@ -18,6 +18,8 @@ Allowed:
 - Create an `ai/` branch after approval.
 - Open a PR scaffold for the approved idea.
 - Add documentation-only planning artifacts for the approved idea.
+- Detect `/ai implement` through a local mini PC runner.
+- Run Codex CLI locally on an approved `ai/issue-*` PR branch.
 - Use manual proposal mode when a human wants to provide a Codex-drafted idea
   instead of spending API credits.
 
@@ -33,6 +35,8 @@ Not allowed:
 - Secret creation, rotation, or disclosure.
 - OpenAI API calls from the approved implementation workflow.
 - More than one AI-generated idea per proposal workflow run.
+- Local runner commits without a human `/ai implement` command.
+- Local runner pushes to `main`.
 
 ## Required GitHub Configuration
 
@@ -78,6 +82,17 @@ The implementation workflow only responds to Issue comments whose trimmed body i
 
 The Issue must also carry the `devloop` and `ai-proposed` labels.
 
+## Implementation Signal
+
+The local mini PC runner only continues implementation after a human comments:
+
+```text
+/ai implement
+```
+
+The runner should work on the existing `ai/issue-*` PR branch and must not merge
+the PR.
+
 ## Human Review Expectations
 
 Before approving an Issue, verify that:

docs/ai-ideas/issue-2.md

@@ -0,0 +1,49 @@
+# DevLoop 아이디어 #2: [DevLoop] 검색 결과 카드에 하이라이트된 키워드 표시 개선
+
+        이 파일은 GitHub Issue에서 사람이 승인한 뒤 생성된 PR scaffold입니다.
+        의도적으로 실제 서비스 코드는 수정하지 않습니다.
+
+        ## 승인된 이슈
+
+        - 이슈: #2
+        - 제목: [DevLoop] 검색 결과 카드에 하이라이트된 키워드 표시 개선
+
+        ## 원본 제안
+
+        ## 문제
+
+현재 검색 결과 카드에서 사용자가 왜 해당 글이 검색되었는지 이해하기 어려운 경우가 있어, 검색 결과의 신뢰도와 탐색 효율성이 떨어질 수 있습니다.
+
+## 지금 필요한 이유
+
+TechCase의 핵심 가치는 실제 기업 적용 사례를 빠르게 이해하는 데 있으므로, 검색 결과 카드에 검색어와 매칭된 키워드를 명확히 하이라이트하여 사용자가 결과의 맥락을 빠르게 파악할 수 있도록 하는 것이 중요합니다. 이는 검색 품질 평가에도 긍정적인 영향을 줄 수 있습니다.
+
+## 제안 범위
+
+프론트엔드 `apps/web` 내 검색 결과 카드 컴포넌트에 하이라이트된 키워드 표시 UI 개선 및 관련 문서(`docs/search-design.md`)에 하이라이트 정책과 구현 방식을 명확히 문서화하는 작업으로 제한합니다. 서비스 코드나 백엔드 검색 로직 변경은 포함하지 않습니다.
+
+## 하지 않을 일
+
+검색 알고리즘 변경, 백엔드 Elasticsearch 쿼리 수정, 키워드 추출 로직 개선, 인증 및 권한 관련 변경, 인프라나 배포 설정 변경은 포함하지 않습니다.
+
+## 위험 요소 / 가드레일
+
+UI 변경에 따른 사용자 혼란 가능성은 낮으며, 문서화 중심 작업이므로 서비스 안정성에 미치는 영향은 거의 없습니다. 다만 하이라이트 표현이 과도하거나 부족하지 않도록 적절한 밸런스 조정이 필요합니다.
+
+## 검증 방법
+
+개선된 검색 결과 카드에서 검색어와 매칭된 키워드가 명확히 하이라이트되어 표시되는지 확인합니다. 여러 검색어 조합에 대해 하이라이트가 일관되게 동작하는지 UI 테스트를 수행하고, 기존 사용자 피드백과 비교해 탐색 편의성이 향상되었는지 검토합니다.
+
+## 승인 방법
+
+이 아이디어를 진행하려면 이슈 댓글에 정확히 `/ai approve`를 남겨주세요.
+승인 후 DevLoop가 `ai/` 브랜치와 PR scaffold를 생성합니다.
+
+DevLoop fingerprint: `4cb0bf0e9a61`
+
+        ## 구현 메모
+
+        - 승인된 범위 안에서만 변경합니다.
+        - 인증, 결제, 데이터베이스 마이그레이션, 인프라는 건드리지 않습니다.
+        - 자동 머지는 활성화하지 않습니다.
+        - 사람 리뷰를 요청하기 전에 검증 메모를 남깁니다.

docs/devloop-runner.md

@@ -0,0 +1,90 @@
+# DevLoop Local Runner
+
+This document describes the mini PC runner that continues an approved DevLoop PR
+with Codex CLI.
+
+## Flow
+
+1. DevLoop creates an idea Issue.
+2. A human comments `/ai approve`.
+3. GitHub Actions creates an `ai/issue-*` scaffold PR.
+4. A human comments `/ai implement` on the Issue.
+5. The mini PC runner detects the command.
+6. The runner checks out the matching PR branch locally.
+7. The runner invokes `codex exec` with `--sandbox workspace-write`.
+8. The runner can commit and push changes back to the same PR branch.
+
+## Required Local Setup
+
+Install and authenticate these on the mini PC:
+
+- `git`
+- `python3`
+- `codex`
+- GitHub credentials that can read Issues, comment on Issues, fetch branches,
+  and push to the repository
+
+Required environment variables:
+
+- `GITHUB_REPOSITORY`: for example `SmileJune/techcase`
+- `GITHUB_TOKEN` or `GH_TOKEN`: token for the local runner
+
+Optional environment variables:
+
+- `DEVLOOP_WORKSPACE`: path to the local clone. Defaults to the current
+  directory.
+- `DEVLOOP_RUNNER_INTERVAL`: polling interval in seconds. Defaults to `300`.
+- `DEVLOOP_APPROVERS`: comma-separated GitHub usernames allowed to use
+  `/ai implement`.
+- `DEVLOOP_CODEX_MODEL`: optional Codex model override.
+- `CODEX_BIN`: optional full path to the Codex CLI executable. Useful for
+  systemd, cron, or SSH non-interactive shells where `~/.local/bin` is not on
+  `PATH`.
+
+Use a dedicated clone for the runner. The runner refuses to start work when the
+workspace is dirty.
+
+## Dry Run
+
+Check what the runner would do without running Codex:
+
+```bash
+python3 scripts/devloop/runner.py \
+  --repo SmileJune/techcase \
+  --once
+```
+
+## Execute Without Commit
+
+Run Codex locally, but leave changes uncommitted for manual inspection:
+
+```bash
+python3 scripts/devloop/runner.py \
+  --repo SmileJune/techcase \
+  --once \
+  --execute-codex
+```
+
+## Execute And Push
+
+Run Codex, commit the resulting allowed changes, and push to the existing PR
+branch:
+
+```bash
+python3 scripts/devloop/runner.py \
+  --repo SmileJune/techcase \
+  --once \
+  --execute-codex \
+  --commit \
+  --push
+```
+
+## Safety Notes
+
+- The runner only responds to `/ai implement`, not `/ai approve`.
+- The runner uses `codex exec --sandbox workspace-write`.
+- The runner refuses to commit changes under `.github/workflows/`, `infra/`,
+  and `apps/backend/alembic/`.
+- The runner does not merge PRs.
+- The runner does not push to `main`; it pushes only to the existing
+  `ai/issue-*` PR branch.

scripts/ai/implement_idea.py

@@ -197,7 +197,7 @@ def create_pr(repo: str, token: str, base: str, branch: str, issue: dict[str, An
                 f"""
                 ## 승인된 아이디어
 
-                Closes #{number}
+                Refs #{number}
 
                 ## 변경 사항
 

scripts/ai/prompts/unified_employee.md

@@ -33,3 +33,7 @@ Idea selection rules:
 - Prefer ideas that can be validated without production access.
 - Prefer ideas that are easy to review and easy to revert.
 - Avoid vague platform rewrites, broad redesigns, or multi-week projects.
+- Do not repeat or closely paraphrase recent DevLoop proposals. If recent
+  proposals focus on result cards, keyword highlights, snippet display, or
+  contributor guidelines, choose a materially different area such as evaluation,
+  source metadata, onboarding, failure handling, or maintainer workflow.

scripts/ai/propose_idea.py

@@ -93,7 +93,7 @@ def github_request(
     path: str,
     token: str,
     payload: dict[str, Any] | None = None,
-) -> dict[str, Any]:
+) -> dict[str, Any] | list[Any]:
     data = None if payload is None else json.dumps(payload).encode("utf-8")
     request = urllib.request.Request(
         f"{GITHUB_API}{path}",
@@ -186,7 +186,46 @@ def manual_idea_from_env() -> dict[str, str]:
     return {field: require_env(env_name) for field, env_name in env_names.items()}
 
 
-def api_idea() -> dict[str, str]:
+def list_recent_devloop_issues(repo: str | None, token: str | None) -> list[dict[str, Any]]:
+    if not repo or not token:
+        return []
+
+    limit = int(env("DEVLOOP_RECENT_ISSUE_LIMIT", "10"))
+    query = urllib.parse.urlencode(
+        {
+            "state": "all",
+            "labels": "devloop,ai-proposed",
+            "per_page": str(limit),
+            "sort": "created",
+            "direction": "desc",
+        }
+    )
+    result = github_request("GET", f"/repos/{repo}/issues?{query}", token)
+    if not isinstance(result, list):
+        raise RuntimeError("Unexpected GitHub issues response.")
+    return [issue for issue in result if "pull_request" not in issue]
+
+
+def recent_issue_context(issues: list[dict[str, Any]]) -> str:
+    if not issues:
+        return "No previous DevLoop proposals were available."
+
+    lines = []
+    for issue in issues:
+        body = str(issue.get("body") or "")
+        excerpt = " ".join(body.split())[:500]
+        lines.append(
+            "\n".join(
+                [
+                    f"- #{issue.get('number')} ({issue.get('state')}): {issue.get('title')}",
+                    f"  excerpt: {excerpt}",
+                ]
+            )
+        )
+    return "\n".join(lines)
+
+
+def api_idea(recent_context: str) -> dict[str, str]:
     prompt = PROMPT_PATH.read_text(encoding="utf-8")
     context = read_repo_context()
     model = env("OPENAI_MODEL", "gpt-4.1-mini")
@@ -213,6 +252,10 @@ def api_idea() -> dict[str, str]:
 
                         {context}
 
+                        Recent DevLoop proposals that must not be repeated or closely paraphrased:
+
+                        {recent_context}
+
                         Return a JSON object with exactly these string fields:
                         title, problem, rationale, scope, non_goals, risks, validation.
                         All field values must be written in Korean. Keep the title concise.
@@ -243,12 +286,12 @@ def validate_idea(idea: dict[str, Any]) -> dict[str, str]:
     return normalized
 
 
-def load_idea() -> dict[str, str]:
+def load_idea(recent_context: str) -> dict[str, str]:
     mode = env("DEVLOOP_PROPOSAL_MODE", "api").lower()
     if mode == "manual":
         return validate_idea(manual_idea_from_env())
     if mode == "api":
-        return api_idea()
+        return api_idea(recent_context)
     raise RuntimeError("DEVLOOP_PROPOSAL_MODE must be 'manual' or 'api'.")
 
 
@@ -285,7 +328,7 @@ def build_issue_body(idea: dict[str, str]) -> str:
 def search_existing_issue(repo: str, token: str, fingerprint: str) -> str | None:
     query = urllib.parse.urlencode(
         {
-            "q": f"repo:{repo} is:issue is:open {fingerprint}",
+            "q": f"repo:{repo} is:issue {fingerprint}",
             "per_page": "1",
         }
     )
@@ -333,7 +376,10 @@ def ensure_label(repo: str, token: str, name: str, color: str, description: str)
 
 def main() -> int:
     dry_run = env("DEVLOOP_DRY_RUN", "false").lower() == "true"
-    idea = load_idea()
+    repo = env("GITHUB_REPOSITORY")
+    token = env("GITHUB_TOKEN") or env("GH_TOKEN")
+    recent_context = recent_issue_context(list_recent_devloop_issues(repo, token))
+    idea = load_idea(recent_context)
     title = f"[DevLoop] {idea['title']}"
     body = build_issue_body(idea)
     fingerprint = body.rsplit("`", 2)[1]
@@ -343,11 +389,13 @@ def main() -> int:
         return 0
 
     repo = require_env("GITHUB_REPOSITORY")
-    token = require_env("GITHUB_TOKEN")
+    token = env("GITHUB_TOKEN") or env("GH_TOKEN")
+    if not token:
+        raise RuntimeError("GITHUB_TOKEN or GH_TOKEN is required.")
 
     existing = search_existing_issue(repo, token, fingerprint)
     if existing:
-        print(f"Matching open DevLoop issue already exists: {existing}")
+        print(f"Matching DevLoop issue already exists: {existing}")
         return 0
 
     ensure_label(repo, token, "devloop", "5319e7", "Managed by the DevLoop MVP workflow.")