fix: hide unassessed jobs and cache gate rejections

BUGFIX_LOG.md

@@ -9,6 +9,20 @@
 ---
 ## 2026-08-11
 
+### BUG · 默认职位列表混入大量当前 CV 未评分卡片
+
+**错误**
+修正 0 分显示后，页面仍出现大量灰色长横线。最新 200 条缓存中有 169 条没有当前 CV 的 `job_matches`，其中包含历史搜索结果和已经被相关性闸门拒绝的明显无关职位。
+
+**原因**
+`/api/jobs` 先读取全局最近 200 条原始缓存，再把没有 `match_score` 的职位默认视为可见。相关性闸门的拒绝只写运行审计事件，不按 CV 持久化，导致同一缓存职位在后续搜索中反复进入大模型闸门。
+
+**解决方案**
+默认 Web 列表改为直接查询最新 CV 下已有且未被 `skip` 的真实匹配；新增独立的 `job_relevance_rejections` 缓存，按职位、CV、JD 内容和 prompt 版本复用闸门拒绝。换 CV、JD 更新或 prompt 升级仍会重新判断，且不会为拒绝结果伪造 0 分评分。
+
+**结果**
+新增回归测试覆盖当前 CV 列表、无 CV 兼容、全新/缓存 JD 的拒绝落库、拒绝缓存作用域、重复搜索跳过和缓存清理；完整 pytest（255 tests）、Ruff、Python 编译和前端语法检查均通过。
+
 ### BUG · 未评分职位卡被错误显示为 0 分
 
 **错误**

CHANGELOG.md

@@ -4,6 +4,9 @@
 
 ### Bug Fixes
 
+- **Current-CV job lists no longer mix in unassessed historical cards** (`server.py` / `cache.py` / `search_prefilter.py` / `search_assessment_stage.py`)
+  The default Web list now reads visible `job_matches` for the latest CV across the cache instead of taking the newest 200 raw cache rows and treating missing matches as relevant. JD relevance-gate rejections are cached separately by job, CV, JD content, language, and prompt version, so the same rejected cached job does not consume another model call on the next search. A different CV, changed JD, or updated gate prompt still triggers reassessment; no synthetic zero-score `MatchScore` is created.
+
 - **Unassessed job cards no longer appear as zero-score matches** (`templates/index.html`)
   Job cards whose current-CV score is `null` now show a gray em dash instead of being coerced to `0` by JavaScript. Genuine numeric zero scores still display as zero, preserving the distinction between "not assessed" and "assessed with no match".
 

README.md

@@ -132,6 +132,7 @@ DEFAULT_MODEL=gemini-3.5-flash-lite
 - **Dynamic title seniority gate**: uses CV eligible/stretch/blocked levels to remove obvious title-level mismatches before JD matching
 - **Experience-gap hard reject**: prefilter records `skip_exp`, and `jd_assessment` directly rejects roles whose explicit years-required exceeds the candidate's experience by more than 3 years
 - **Explainable matching**: job details include score breakdown, risks, skill matches, and recommendation tiers
+- **Current-CV result list**: the default Web list shows only real, non-skipped matches for the latest CV; CV-scoped relevance rejections are reused until the CV, JD content, language, or gate prompt changes
 - **Risk-only relocation / office attendance handling**: same-country relocation and office-attendance requirements are recorded in `risks / risk_penalty` but are not treated as `location_score` penalties
 - **Artifact hub**: generate and reuse Interview Prep, Cover Letter, and CV Optimization from the job detail panel
 - **Log panel**: level filtering, keyword highlight, auto-refresh

docs/RUNBOOK.md

@@ -294,7 +294,7 @@ jobradar_email_sync_runs_total{status="failed",reason="auth"}
 | Indeed 抓取结果为 0 | JobSpy 被 Indeed 限流（常见，非故障） | 等几小时重试；查日志中 JobSpy WARNING |
 | Adzuna 返回 429 | 速率限制 | 调大 `_MIN_INTERVAL`（当前 1.2s）；核对 `.env` 中 `ADZUNA_APP_ID` / `ADZUNA_APP_KEY` |
 | LLM 评估全部拒绝 | `cv_summary` / `cv_skills` 提取失败 | 检查 CV 解析结果；`uv run jobradar assess` 补跑评估 |
-| 职位无模型评分 | 当前 `cv_hash` 下尚无 `job_matches` 记录（换 CV 后常见） | `uv run jobradar assess` 为当前 CV 补算 |
+| 职位无模型评分 | 当前 `cv_hash` 下尚无 `job_matches` 记录（换 CV 后常见）；默认 Web 列表不会展示这类历史卡片 | 仅在需要主动补算历史 JD 时运行 `uv run jobradar assess` |
 
 ### 安全注意
 

jobradar/assessment.py

@@ -17,6 +17,7 @@
 
 BATCH_SIZE = 8  # 每批 JD 数量，兼顾 context 长度与 token 节省
 TITLE_RELEVANCE_PROMPT_VERSION = "title_relevance_v4"
+JD_ASSESSMENT_PROMPT_VERSION = "jd_assessment_v1"
 
 _LANGUAGE_NAMES = {"zh": "中文", "en": "English", "es": "Español"}
 _TITLE_KEYWORD_STOPWORDS = {
@@ -27,6 +28,10 @@
 }
 
 
+def jd_assessment_prompt_version(language: str = "zh") -> str:
+    return f"{JD_ASSESSMENT_PROMPT_VERSION}:{language}"
+
+
 def gate_worker_count(provider: str) -> int:
     """Bound independent gate calls without overloading local providers."""
     return 1 if provider in {"ollama", "local"} else 2

jobradar/cache.py

@@ -165,6 +165,19 @@
     PRIMARY KEY (job_id, cv_hash)
 );
 
+CREATE TABLE IF NOT EXISTS job_relevance_rejections (
+    job_id            TEXT NOT NULL,
+    cv_hash           TEXT NOT NULL,
+    description_hash  TEXT NOT NULL,
+    reason            TEXT NOT NULL DEFAULT '',
+    score             INTEGER NOT NULL DEFAULT 0,
+    model_name        TEXT NOT NULL DEFAULT '',
+    prompt_version    TEXT NOT NULL DEFAULT '',
+    created_at        TEXT NOT NULL,
+    updated_at        TEXT NOT NULL,
+    PRIMARY KEY (job_id, cv_hash)
+);
+
 CREATE TABLE IF NOT EXISTS interview_preps (
     job_id            TEXT NOT NULL,
     cv_hash           TEXT NOT NULL,
@@ -465,12 +478,35 @@ def merge_job_raw_source(dedup_key: str, source_entry: dict) -> None:
         )
 
 
-def get_recent_jobs(limit: int = 50, language: str = "zh") -> list[JobResult]:
-    """按抓取时间倒序返回最近 limit 条未过期职位。"""
+def get_recent_jobs(
+    limit: int = 50,
+    language: str = "zh",
+    *,
+    require_match: bool = False,
+) -> list[JobResult]:
+    """按抓取时间倒序返回最近职位，可限定为当前 CV 下已有可见匹配。"""
     with _conn() as con:
-        rows = con.execute(
-            "SELECT * FROM job_cache ORDER BY fetched_at DESC LIMIT ?", (limit,)
-        ).fetchall()
+        if require_match:
+            cv_hash = get_latest_cv_hash()
+            if not cv_hash:
+                return []
+            rows = con.execute(
+                """
+                SELECT job_cache.*
+                FROM job_cache
+                JOIN job_matches
+                  ON job_matches.job_id = job_cache.dedup_key
+                 AND job_matches.cv_hash = ?
+                 AND job_matches.recommendation != 'skip'
+                ORDER BY job_cache.fetched_at DESC
+                LIMIT ?
+                """,
+                (cv_hash, limit),
+            ).fetchall()
+        else:
+            rows = con.execute(
+                "SELECT * FROM job_cache ORDER BY fetched_at DESC LIMIT ?", (limit,)
+            ).fetchall()
     jobs = [_row_to_job(r) for r in rows]
     from jobradar.jd_profile import jd_profile_prompt_version
 
@@ -702,6 +738,75 @@ def get_job_match(job_id: str, cv_hash: str, description: str = "", prompt_versi
     return MatchScore.model_validate_json(row["score_json"])
 
 
+def get_job_relevance_rejection(
+    job_id: str,
+    cv_hash: str,
+    description: str = "",
+    prompt_version: str = "",
+) -> dict | None:
+    with _conn() as con:
+        row = con.execute(
+            """
+            SELECT reason, score, model_name, description_hash, prompt_version
+            FROM job_relevance_rejections
+            WHERE job_id = ? AND cv_hash = ?
+            """,
+            (job_id, cv_hash),
+        ).fetchone()
+    if row is None:
+        return None
+    if description and row["description_hash"] != _description_hash(description):
+        return None
+    if prompt_version and row["prompt_version"] != prompt_version:
+        return None
+    return {
+        "reason": row["reason"],
+        "score": row["score"],
+        "model_name": row["model_name"],
+        "prompt_version": row["prompt_version"],
+    }
+
+
+def save_job_relevance_rejection(
+    *,
+    job_id: str,
+    cv_hash: str,
+    description: str,
+    reason: str,
+    score: int,
+    model_name: str = "",
+    prompt_version: str = "",
+) -> None:
+    now = datetime.utcnow().isoformat()
+    with _conn() as con:
+        con.execute(
+            """
+            INSERT INTO job_relevance_rejections
+              (job_id, cv_hash, description_hash, reason, score, model_name,
+               prompt_version, created_at, updated_at)
+            VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?)
+            ON CONFLICT(job_id, cv_hash) DO UPDATE SET
+              description_hash = excluded.description_hash,
+              reason = excluded.reason,
+              score = excluded.score,
+              model_name = excluded.model_name,
+              prompt_version = excluded.prompt_version,
+              updated_at = excluded.updated_at
+            """,
+            (
+                job_id,
+                cv_hash,
+                _description_hash(description),
+                reason,
+                score,
+                model_name,
+                prompt_version,
+                now,
+                now,
+            ),
+        )
+
+
 def _attach_latest_match(job: JobResult, language: str = "zh") -> None:
     latest_cv_hash = get_latest_cv_hash()
     if not latest_cv_hash:
@@ -1159,6 +1264,7 @@ def clear_all() -> None:
         con.execute("DELETE FROM jd_profiles")
         con.execute("DELETE FROM job_summaries")
         con.execute("DELETE FROM job_matches")
+        con.execute("DELETE FROM job_relevance_rejections")
         con.execute("DELETE FROM interview_preps")
         con.execute("DELETE FROM cover_letters")
         con.execute("DELETE FROM cv_optimizations")
@@ -1189,6 +1295,10 @@ def delete_jobs(dedup_keys: list[str]) -> int:
             f"DELETE FROM job_matches WHERE job_id IN ({placeholders})",
             dedup_keys,
         )
+        con.execute(
+            f"DELETE FROM job_relevance_rejections WHERE job_id IN ({placeholders})",
+            dedup_keys,
+        )
         con.execute(
             f"DELETE FROM interview_preps WHERE job_id IN ({placeholders})",
             dedup_keys,
@@ -1619,6 +1729,10 @@ def clean_expired() -> int:
                 f"DELETE FROM job_summaries WHERE job_id IN ({placeholders})",
                 expired_job_ids,
             )
+            con.execute(
+                f"DELETE FROM job_relevance_rejections WHERE job_id IN ({placeholders})",
+                expired_job_ids,
+            )
         # 删除有明确截止日期且已过期的 JD
         r1 = con.execute(
             "DELETE FROM job_cache WHERE expires_at IS NOT NULL AND expires_at < ?",

jobradar/search_assessment_stage.py

@@ -7,7 +7,11 @@
 from typing import Callable
 
 from jobradar import cache
-from jobradar.assessment import JDAssessment, batch_assess_jds
+from jobradar.assessment import (
+    JDAssessment,
+    batch_assess_jds,
+    jd_assessment_prompt_version,
+)
 from jobradar.jd_profile import jd_profile_prompt_version
 from jobradar.job_evaluation import evaluate_job_once, job_evaluation_prompt_version
 from jobradar.logger import get_logger
@@ -322,6 +326,15 @@ def _is_visible_job(job_obj) -> bool:
                     details={"score": assessment.score, "cached": True},
                 )
                 llm_rejected += 1
+                cache.save_job_relevance_rejection(
+                    job_id=cached_job.dedup_key,
+                    cv_hash=cv_hash,
+                    description=cached_job.description_snippet,
+                    reason=assessment.reason,
+                    score=assessment.score,
+                    model_name=f"{llm.provider}/{llm.model}",
+                    prompt_version=jd_assessment_prompt_version(language),
+                )
             write_cache(
                 {
                     "title": cached_job.title,
@@ -443,6 +456,16 @@ def _is_visible_job(job_obj) -> bool:
                     "raw_sources": raw_sources,
                 }
             )
+            if not assessment.relevant and cv_hash and llm is not None:
+                cache.save_job_relevance_rejection(
+                    job_id=key,
+                    cv_hash=cv_hash,
+                    description=content,
+                    reason=assessment.reason,
+                    score=assessment.score,
+                    model_name=f"{llm.provider}/{llm.model}",
+                    prompt_version=jd_assessment_prompt_version(language),
+                )
             if assessment.relevant:
                 summary_job = JobResult(
                     title=title,

jobradar/search_prefilter.py

@@ -5,6 +5,7 @@
 from typing import Callable
 
 from jobradar import cache
+from jobradar.assessment import jd_assessment_prompt_version
 from jobradar.filters import infer_title_seniority, is_title_seniority_ok
 from jobradar.logger import get_logger
 from jobradar.matching import match_prompt_version
@@ -95,7 +96,13 @@ def classify_cache_hit(cached_job: JobResult, cv_hash: str, language: str = "zh"
         prompt_version=match_prompt_version(language),
     )
     if match is None:
-        return "reassess"
+        rejection = cache.get_job_relevance_rejection(
+            cached_job.dedup_key,
+            cv_hash,
+            cached_job.description_snippet,
+            prompt_version=jd_assessment_prompt_version(language),
+        )
+        return "skip" if rejection is not None else "reassess"
     return "skip" if match.recommendation == "skip" else "reuse"
 
 

jobradar/server.py

@@ -603,7 +603,10 @@ def delete_application(application_id: int) -> dict:
 
 @app.get("/api/jobs")
 def get_jobs(limit: int = 200, language: str = "zh") -> list[dict]:
-    jobs = cache.get_recent_jobs(limit, language=language)
+    require_match = bool(cache.get_latest_cv_hash())
+    jobs = cache.get_recent_jobs(limit, language=language, require_match=require_match)
+    if require_match:
+        jobs = [j for j in jobs if j.match_score is not None]
     jobs = [j for j in jobs if j.is_effectively_relevant]
     jobs.sort(key=lambda j: (j.effective_score if j.effective_score is not None else -1), reverse=True)
     return [_job_to_dict(j) for j in jobs]

tests/test_application_tracking.py

@@ -226,6 +226,40 @@ def test_jobs_route_remains_bound_to_get_jobs():
     assert route.endpoint.__name__ == "get_jobs"
 
 
+def test_jobs_api_requests_only_current_cv_matches(monkeypatch):
+    from jobradar import server
+
+    calls = {}
+
+    def fake_recent_jobs(limit, language="zh", require_match=False):
+        calls["limit"] = limit
+        calls["language"] = language
+        calls["require_match"] = require_match
+        return []
+
+    monkeypatch.setattr(server.cache, "get_latest_cv_hash", lambda: "current-cv")
+    monkeypatch.setattr(server.cache, "get_recent_jobs", fake_recent_jobs)
+
+    assert server.get_jobs() == []
+    assert calls == {"limit": 200, "language": "zh", "require_match": True}
+
+
+def test_jobs_api_preserves_cache_visibility_without_a_cv(monkeypatch):
+    from jobradar import server
+
+    calls = {}
+
+    def fake_recent_jobs(limit, language="zh", require_match=False):
+        calls["require_match"] = require_match
+        return []
+
+    monkeypatch.setattr(server.cache, "get_latest_cv_hash", lambda: "")
+    monkeypatch.setattr(server.cache, "get_recent_jobs", fake_recent_jobs)
+
+    assert server.get_jobs() == []
+    assert calls["require_match"] is False
+
+
 def test_application_nav_button_is_not_nested_in_config_button():
     from pathlib import Path