"""Source grounding script unit tests — Codex APPROVED Q1-Q5. dry-run no API, no DB write. 사용자 요구사항 #7 그대로: - source evidence extractor unit - Gemini response schema validation helper - HTML contains expected sections - API key/secret 로그 금지 """ from __future__ import annotations import json import sys from pathlib import Path _REPO_ROOT = Path(__file__).resolve().parents[3] _SCRIPTS = _REPO_ROOT / "backend" / "scripts" if str(_SCRIPTS) not in sys.path: sys.path.insert(0, str(_SCRIPTS)) import experiment_rooftop_source_grounding as exp # noqa: E402 def _bundle(episode: str = "", planning: str = "") -> exp.SourceBundle: return exp.SourceBundle( episode_fulltext=episode, planning_doc_text=planning, episode_meta={"fulltext_length": len(episode)}, project_meta={"planning_doc_length": len(planning)}, ) # --------------------------------------------------------------------------- # Deterministic evidence extractor # --------------------------------------------------------------------------- def test_evidence_hits_sub_space_keywords() -> None: b = _bundle(episode="햇살이 비치는 옥탑방 거실에서 침대 위 어두운 벽면에 붉은 원이 보였다.") rows = exp.extract_evidence(b) keywords = {r.keyword for r in rows} assert "옥탑방" in keywords assert "거실" in keywords assert "침대" in keywords assert "벽면" in keywords def test_evidence_field_hint_resolved() -> None: b = _bundle(episode="식탁 위에 사진이 놓여 있다.") rows = exp.extract_evidence(b) table_rows = [r for r in rows if r.keyword == "식탁"] assert table_rows assert table_rows[0].candidate_field == "furniture.dining_table" assert table_rows[0].source == "episode" def test_evidence_multiple_sources_separate_rows() -> None: b = _bundle( episode="옥탑방 안에서 사건이 일어난다.", planning="이야기는 옥탑방을 배경으로 한다.", ) rows = exp.extract_evidence(b) sources = {r.source for r in rows if r.keyword == "옥탑방"} assert sources == {"episode", "planning_doc"} def test_evidence_quote_window_contains_keyword() -> None: b = _bundle(episode="앞 문장. 햇살이 비치는 옥탑방 거실. 다음 문장.") rows = exp.extract_evidence(b) for r in rows: if r.keyword == "옥탑방": assert "옥탑방" in r.quote # --------------------------------------------------------------------------- # Schema validation helper — strict JSON 구조 sanity # --------------------------------------------------------------------------- def test_rooftop_bible_schema_required_fields() -> None: schema = exp.ROOFTOP_BIBLE_SCHEMA assert schema["type"] == "object" required = set(schema["required"]) expected = { "sub_spaces", "layout_relations", "doors_windows", "kitchen", "bathroom", "furniture", "materials", "condition_age", "socioeconomic_tone", "scale_density", "state_variants", "required_visual_cues", "forbidden_luxury_cues", "unknowns", "inference_limits", } assert expected.issubset(required), f"missing fields: {expected - required}" def test_gemini_instruction_no_truncation_directive() -> None: # Gemini 입력 자르지 마라 (CLAUDE.md 절대 규칙). instr = exp.GEMINI_INSTRUCTION_TEMPLATE # 새 marker: <> / <> (Codex Blocking — str.format placeholder 회피) assert "<>" in instr assert "<>" in instr # 자르기/truncate/[:400] 등 substring 없음. for token in (":400]", ":500]", "truncate", "잘라"): assert token not in instr, f"GEMINI_INSTRUCTION must not contain {token!r}" def test_build_gemini_prompt_no_format_keyerror() -> None: """Codex Blocking: {trusted, weak, unknown} literal brace 가 .format() 와 충돌 → build_gemini_prompt 는 replace 방식이라 안전.""" prompt = exp.build_gemini_prompt( episode_text="옥탑방 거실 본문", planning_text="기획안 본문", ) assert "옥탑방 거실 본문" in prompt assert "기획안 본문" in prompt # literal brace 가 그대로 유지됨 (str.format 안 거침). assert "{trusted, weak, unknown}" in prompt # marker 토큰 (replace 처리됐어야 — 잔존하면 fail). assert "<>" not in prompt assert "<>" not in prompt def test_build_gemini_prompt_empty_inputs() -> None: prompt = exp.build_gemini_prompt(episode_text="", planning_text="") # marker 는 사라지고, instruction 본문은 남음 assert "<>" not in prompt assert "<>" not in prompt assert "EPISODE 1 FULLTEXT" in prompt def test_gemini_instruction_evidence_only_directive() -> None: instr = exp.GEMINI_INSTRUCTION # Codex Q3: Gemini 가 판정자 아님. assert "evidence" in instr.lower() or "evidence" in instr assert "unknowns" in instr # metric 금지 명시 (Codex 권고). assert "metric" in instr or "qualitative" in instr # --------------------------------------------------------------------------- # Luxury cue list + severity enum (Codex Q3/Q4) # --------------------------------------------------------------------------- def test_luxury_cue_list_has_candidates() -> None: for kw in ("marble", "chandelier", "walk-in closet", "luxury bathroom"): assert kw in exp.LUXURY_CUE_CANDIDATES def test_severity_enum_three_levels() -> None: assert exp.SEVERITY_CANDIDATES == ["info", "suspicious", "likely_mismatch"] # --------------------------------------------------------------------------- # Comparison — plate grouping (Codex Q5) # --------------------------------------------------------------------------- def test_comparison_returns_empty_when_base_run_missing() -> None: b = _bundle(episode="옥탑방 거실") evidence = exp.extract_evidence(b) rows = exp.build_comparison_report( base_run_dir=Path("/tmp/nonexistent_base_run"), bundle=b, evidence_rows=evidence, gemini_bible=None, ) assert rows == [] # --------------------------------------------------------------------------- # HTML sections — 사용자 요구사항 #7 # --------------------------------------------------------------------------- def test_render_html_contains_expected_sections(tmp_path) -> None: b = _bundle(episode="옥탑방 거실에 식탁이 있다.") evidence = exp.extract_evidence(b) run_meta = { "run_id": "test_run", "created_at": "2026-05-23T20:00:00", "plan_version": "v1", "project_id": exp.PROJECT_ID, "episode_id": exp.EPISODE_ID, "evidence_count": len(evidence), "comparison_count": 0, "gemini_status": "skipped", "args": {}, "serve_info": None, } exp.render_html( run_dir=tmp_path, bundle=b, evidence_rows=evidence, base_run_dir=None, comparison_rows=[], gemini_bible=None, run_meta=run_meta, ) html = (tmp_path / "index.html").read_text(encoding="utf-8") for header in ( "1. Source bundle", "2. Deterministic source evidence", "3. Gemini rooftop bible", "4. Current base plate", "5. Comparison candidates", "6. Run meta", ): assert header in html, f"HTML missing section header {header!r}" def test_render_html_quote_limit_applied(tmp_path) -> None: # 매우 긴 quote 가 HTML 에 truncate 되는지 (Codex 추가 권고). long_quote = "옥탑방 " + ("거실 " * 200) b = _bundle(episode=long_quote) evidence = exp.extract_evidence(b) run_meta = {"run_id": "qtest", "created_at": "x", "plan_version": "v1", "project_id": exp.PROJECT_ID, "episode_id": exp.EPISODE_ID, "evidence_count": len(evidence), "comparison_count": 0, "gemini_status": "skipped", "args": {}, "serve_info": None} exp.render_html(run_dir=tmp_path, bundle=b, evidence_rows=evidence, base_run_dir=None, comparison_rows=[], gemini_bible=None, run_meta=run_meta) html = (tmp_path / "index.html").read_text(encoding="utf-8") # 정확한 quote 가 잘렸으면 "..." 포함. assert "..." in html # --------------------------------------------------------------------------- # API key/secret 로그 금지 (사용자 요구사항 #7) # --------------------------------------------------------------------------- def test_jsonl_append_does_not_carry_api_key() -> None: payload = { "stage": "gemini_bible", "model": "gemini-3.5-flash", "status": "ok", "latency_ms": 1234, } # 우리 _append_jsonl_if 함수가 payload 키 외의 secret 을 주입하지 않음을 확인. # (단순히 dumps 만 함 — secret 키 추가 없음을 명시 검증) s = json.dumps(payload) for forbidden in ("api_key", "Authorization", "Bearer", "GEMINI_API_KEY"): assert forbidden not in s def test_call_gemini_returns_failed_without_api_key(monkeypatch) -> None: # settings.gemini_api_key (단수) + env GEMINI_API_KEY 둘 다 비움 → failed. from app.core.config import settings # noqa: PLC0415 monkeypatch.delenv("GEMINI_API_KEY", raising=False) monkeypatch.setattr(settings, "gemini_api_key", "") result = exp.call_gemini_bible( episode_text="test", planning_text="test", gemini_model="gemini-3.5-flash", max_attempts=1, ) assert result["status"] == "failed" assert "GEMINI_API_KEY" in result.get("error", "") # --------------------------------------------------------------------------- # Source bundle structure # --------------------------------------------------------------------------- def test_source_bundle_dataclass_defaults() -> None: b = exp.SourceBundle(episode_fulltext="", planning_doc_text="") assert b.missing_inputs == [] assert b.episode_meta == {} assert b.project_meta == {} # --------------------------------------------------------------------------- # Codex Blocking 1 — base plate assets copy # --------------------------------------------------------------------------- def test_copy_base_plate_assets_copies_png_and_json(tmp_path) -> None: # 가짜 base_run_dir 만들기 base_run = tmp_path / "fake_base_run" (base_run / "base_plates").mkdir(parents=True) (base_run / "realized_spatial_cards").mkdir(parents=True) (base_run / "base_plates" / "main_room__eye_level_wide.png").write_bytes(b"PNGDATA") (base_run / "realized_spatial_cards" / "base__main_room__eye_level_wide.json").write_text( '{"plate_id": "base__main_room__eye_level_wide"}', encoding="utf-8" ) dest = exp.copy_base_plate_assets(tmp_path, base_run) assert dest is not None assert (dest / "main_room__eye_level_wide.png").exists() assert (dest / "base__main_room__eye_level_wide.json").exists() # 상대경로 확인 rel = dest.relative_to(tmp_path) assert str(rel).startswith("base_plate_assets/fake_base_run") def test_copy_base_plate_assets_none_for_missing_base() -> None: out = exp.copy_base_plate_assets(Path("/tmp/x"), None) assert out is None def test_html_img_src_uses_local_assets_path(tmp_path) -> None: # 가짜 base_run + asset copy 후 render_html → img src 가 base_plate_assets/... 로 base_run = tmp_path / "fake_base_run" (base_run / "base_plates").mkdir(parents=True) (base_run / "base_plates" / "main_room__eye_level_wide.png").write_bytes(b"PNG") run_dir = tmp_path / "run" run_dir.mkdir() exp.copy_base_plate_assets(run_dir, base_run) b = _bundle(episode="x") run_meta = { "run_id": "asset_test", "created_at": "x", "plan_version": "v1", "project_id": exp.PROJECT_ID, "episode_id": exp.EPISODE_ID, "evidence_count": 0, "comparison_count": 0, "gemini_status": "skipped", "args": {}, "serve_info": None, "base_plate_assets_dir": f"base_plate_assets/{base_run.name}", } exp.render_html(run_dir=run_dir, bundle=b, evidence_rows=[], base_run_dir=base_run, comparison_rows=[], gemini_bible=None, run_meta=run_meta) html = (run_dir / "index.html").read_text(encoding="utf-8") assert 'src="base_plate_assets/' in html # parent path (..) 사용 금지 assert 'src="../' not in html # --------------------------------------------------------------------------- # Codex Blocking 2 — local schema validation # --------------------------------------------------------------------------- def test_gemini_call_uses_jsonschema_validate(monkeypatch) -> None: """call_gemini_bible 에 jsonschema.validate import 가 들어가 있고 schema 위반 응답을 RuntimeError 로 잡는지 — soft test (간접 검증).""" import experiment_rooftop_source_grounding as src_mod # noqa: PLC0415 src = Path(src_mod.__file__).read_text(encoding="utf-8") assert "jsonschema.validate" in src, "Gemini wrapper must call jsonschema.validate" assert "ROOFTOP_BIBLE_SCHEMA" in src