feat(ai): 设置里加「AI 生成队列」，看得见后台在算什么

队列本来是完全不可见的：页面上一句「排队生成中」说不出自己是下一个、第二十
个，还是已经放弃了——网关挂掉的时候，「还在生成」和「永远不会好」长得一模
一样。今天排查就是这么排的。

- GET /analysis/insight/queue 返回队列（running 在前，其次按优先级和年龄，
  和 worker 实际取任务的顺序一致）、已生成的解读、scope 名到中文标签的映射
  （前端不必再抄一份），以及消费者的限流配置
- POST /analysis/insight/queue/retry：手动把「已放弃」的重新排队，不等冷却。
  自动重试要等冷却是为了不去捶一个正在抽风的上游；人按下重试是他自己判断值得
  再试一次
- 页面在 设置 → AI 生成队列。插队的任务标「插队」——这是整个界面最想让人看见
  的一件事：为什么是它排在最前面
- 「已生成」单独列：队列空了意味着「没有待办」，不是「什么都没生成过」，
  没有这一节这两件事在界面上没法区分

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

backend/routes/analysis.py

@@ -168,8 +168,23 @@ def insight():
 @bp.route("/insight/queue", methods=["GET"])
 @require_auth
 def insight_queue():
-    """What the coach still has to generate — for a progress indicator."""
+    """The coach's queue: what is running, what is waiting, what gave up.
+
+    Also lists what has already been generated, because "nothing queued" and
+    "nothing generated" look the same from the queue alone and mean opposite
+    things.
+    """
     return jsonify({
         "pending": ai_jobs.pending_count(g.user_id),
-        "enabled": ai_jobs.ENABLED,
+        "jobs": ai_jobs.list_jobs(g.user_id),
+        "insights": analysis_svc.list_insights(g.user_id),
+        "scopes": {name: scope.label for name, scope in scopes.SCOPES.items()},
+        "settings": ai_jobs.settings(),
     })
+
+
+@bp.route("/insight/queue/retry", methods=["POST"])
+@require_auth
+def insight_queue_retry():
+    """Re-queue everything that gave up, without waiting out the cooldown."""
+    return jsonify({"pending": ai_jobs.retry_failed(g.user_id)})

backend/services/analysis.py

@@ -571,6 +571,22 @@ def _run_job(user_id, kind, subject):
 jobs.set_runner(_run_job)
 
 
+def list_insights(user_id, limit=60):
+    """Stored answers, newest first — what the coach has actually produced."""
+    rows = query_all(
+        "SELECT kind, subject, model, upstream, created_at FROM ai_insights "
+        "WHERE user_id = ? ORDER BY created_at DESC",
+        [user_id],
+    )
+    return [{
+        "kind": r["kind"],
+        "subject": r["subject"],
+        "model": r["model"],
+        "upstream": r["upstream"],
+        "generatedAt": r["created_at"],
+    } for r in rows[:limit]]
+
+
 def clear_insight_cache(user_id, kind=None):
     if kind:
         execute(

backend/services/jobs.py

@@ -197,6 +197,60 @@ def status_of(user_id, kind, subject):
     }
 
 
+def list_jobs(user_id, limit=60):
+    """The queue as it stands, for the 设置 screen.
+
+    Ordered the way the worker will actually take them — priority, then age —
+    so the list reads as "what happens next" rather than as a table of rows.
+    Finished jobs come last: they are history, not queue.
+    """
+    rows = query_all(
+        "SELECT * FROM ai_jobs WHERE user_id = ? "
+        "ORDER BY CASE status WHEN 'running' THEN 0 WHEN 'pending' THEN 1 "
+        "WHEN 'failed' THEN 2 ELSE 3 END, priority ASC, created_at ASC",
+        [user_id],
+    )
+    return [{
+        "kind": r["kind"],
+        "subject": r["subject"],
+        "status": r["status"],
+        "priority": r["priority"],
+        # The queue only distinguishes "the user is looking at this" from
+        # "backfill"; showing the raw number would mean explaining the scale.
+        "interactive": r["priority"] <= PRIORITY_INTERACTIVE,
+        "attempts": r["attempts"],
+        "maxAttempts": MAX_ATTEMPTS,
+        "error": r.get("error"),
+        "updatedAt": r.get("updated_at"),
+    } for r in rows[:limit]]
+
+
+def settings():
+    """What the consumer is configured to do, for the same screen."""
+    return {
+        "enabled": ENABLED,
+        "concurrency": MAX_CONCURRENT,
+        "gapSeconds": GAP_SECONDS,
+        "maxAttempts": MAX_ATTEMPTS,
+        "retryAfterSeconds": FAILED_RETRY_SECONDS,
+    }
+
+
+def retry_failed(user_id):
+    """Put every given-up job back in the queue, now, at the user's request.
+
+    The automatic retry waits out a cooldown so a flapping upstream is not
+    hammered; a person pressing 重试 has decided it is worth trying again.
+    """
+    execute(
+        "UPDATE ai_jobs SET status = 'pending', attempts = 0, error = NULL, "
+        "holder = NULL, claimed_at = NULL, updated_at = ? "
+        "WHERE user_id = ? AND status = 'failed'",
+        [_iso(_now()), user_id],
+    )
+    return pending_count(user_id)
+
+
 def pending_count(user_id=None):
     sql = "SELECT COUNT(*) AS n FROM ai_jobs WHERE status IN ('pending', 'running')"
     params = []

backend/tests/test_coach.py

@@ -1105,3 +1105,64 @@ class TestGatewayCourtesy:
     def test_there_is_a_gap_between_jobs(self):
         """Back-to-back is what saturates a four-thread box."""
         assert jobs.GAP_SECONDS > 0
+
+
+class TestQueueScreen:
+    def test_queue_endpoint_requires_auth(self, client):
+        assert client.get("/api/analysis/insight/queue").status_code == 401
+        assert client.post("/api/analysis/insight/queue/retry").status_code == 401
+
+    def test_lists_jobs_labels_and_limits(self, client, auth, month):
+        client.get("/api/analysis/insight?scope=sleep", headers=auth)
+        body = client.get("/api/analysis/insight/queue", headers=auth).get_json()
+        assert body["pending"] >= 1
+        assert any(j["kind"] == "sleep" for j in body["jobs"])
+        assert body["scopes"]["sleep"] == "睡眠", "the UI must not restate the list"
+        assert body["settings"]["concurrency"] == jobs.MAX_CONCURRENT
+
+    def test_an_open_screen_is_marked_as_having_jumped_the_queue(
+        self, client, auth, month
+    ):
+        jobs.enqueue(month["id"], "trends", "2026-08-30", priority=jobs.PRIORITY_PREFETCH)
+        client.get("/api/analysis/insight?scope=sleep", headers=auth)
+        by_kind = {j["kind"]: j for j in
+                   client.get("/api/analysis/insight/queue", headers=auth).get_json()["jobs"]}
+        assert by_kind["sleep"]["interactive"] is True
+        assert by_kind["trends"]["interactive"] is False
+
+    def test_running_jobs_are_listed_before_waiting_ones(self, client, auth, month):
+        jobs.enqueue(month["id"], "health", "a")
+        jobs.enqueue(month["id"], "sleep", "b")
+        jobs._claim_next()
+        listed = client.get("/api/analysis/insight/queue", headers=auth).get_json()["jobs"]
+        assert listed[0]["status"] == "running"
+
+    def test_generated_answers_are_listed_separately(
+        self, client, auth, month, gateway, monkeypatch
+    ):
+        """An empty queue means "nothing left to do", not "nothing was done" —
+        the two look the same without this."""
+        answer(monkeypatch, json.dumps({
+            "headline": "h", "points": [], "actions": [], "confidence": "low",
+        }, ensure_ascii=False))
+        analysis_svc.generate_scope_insight(month["id"], "sleep")
+        body = client.get("/api/analysis/insight/queue", headers=auth).get_json()
+        assert any(i["kind"] == "sleep" for i in body["insights"])
+
+    def test_retry_puts_given_up_jobs_back(self, client, auth, month):
+        jobs.enqueue(month["id"], "sleep", "s")
+        row = jobs._claim_next()
+        for _ in range(jobs.MAX_ATTEMPTS):
+            jobs._finish(row["id"], "boom")
+        assert jobs.status_of(month["id"], "sleep", "s")["status"] == "failed"
+
+        resp = client.post("/api/analysis/insight/queue/retry", headers=auth)
+        assert resp.status_code == 200
+        assert jobs.status_of(month["id"], "sleep", "s")["status"] == "pending"
+        assert jobs.status_of(month["id"], "sleep", "s")["attempts"] == 0
+
+    def test_the_queue_is_per_account(self, client, auth, month, make_user):
+        other = make_user("other@example.com")
+        jobs.enqueue(other["id"], "sleep", "s")
+        body = client.get("/api/analysis/insight/queue", headers=auth).get_json()
+        assert body["jobs"] == []