Attention strip: richer source-health items (stale/reject/dup/thin/rate-limit)

Per Codex — make the Overview strip diagnostic without making the operator hunt
through tables. Aggregated (one calm line per condition with a count), volume-
gated, conservative thresholds:

* Stale: active+visible source, last success > 10 days ago (warn).
* High rejection: >=20 ingested, acceptance < 25% (info).
* High duplicate: >=10 accepted, accepted-dup > 50% (info).
* Thin images: >=10 served, per-source image coverage < 25% (info).
* Long rate-limit: retry_after_at more than 12h out (info).

source_health gains a per-source images count + image_coverage. _attention takes
an optional now (for tests). Existing site-wide items (global image coverage,
thin brief, unread feedback) unchanged.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

goodnews/queries.py

@@ -8,6 +8,7 @@ the original source — never stored bodies.
 from __future__ import annotations
 
 import sqlite3
+from datetime import UTC, datetime, timedelta
 
 from .feeds import MAX_BACKOFF_MINUTES
 
@@ -307,6 +308,9 @@ def source_health(conn: sqlite3.Connection) -> list[dict]:
             (SELECT COUNT(*) FROM articles a WHERE a.source_id = s.id AND a.duplicate_of IS NOT NULL) AS duplicates,
             (SELECT COUNT(*) FROM articles a JOIN article_scores sc ON sc.article_id = a.id
              WHERE a.source_id = s.id AND sc.accepted = 1 AND a.duplicate_of IS NULL) AS served,
+            (SELECT COUNT(*) FROM articles a JOIN article_scores sc ON sc.article_id = a.id
+             WHERE a.source_id = s.id AND sc.accepted = 1 AND a.duplicate_of IS NULL
+               AND a.image_url IS NOT NULL AND a.image_url != '') AS images,
             datetime(
                 (SELECT MAX(r.finished_at) FROM ingest_runs r
                  WHERE r.source_id = s.id AND r.finished_at IS NOT NULL),
@@ -327,6 +331,7 @@ def source_health(conn: sqlite3.Connection) -> list[dict]:
         # Curation quality: of what this source got ACCEPTED, how much was a
         # duplicate of content already served (accepted_total − served = accepted dupes).
         d["accepted_dup_rate"] = round(100 * (accepted - d["served"]) / accepted) if accepted else None
+        d["image_coverage"] = round(100 * (d["images"] or 0) / d["served"]) if d["served"] else None
         # Match the REAL scheduler gate: due = the later of the streak-backoff time
         # and any retry_after_at rest (UTC strings sort chronologically).
         due_times = [t for t in (d["next_due_at"], d["retry_after_at"]) if t]
@@ -335,20 +340,66 @@ def source_health(conn: sqlite3.Connection) -> list[dict]:
     return out
 
 
-def _attention(content: dict, sources: list[dict], feedback_unread: int) -> list[dict]:
+# Attention thresholds — conservative + volume-gated, so the strip is a calm
+# operator nudge rather than a noisy scold. Items are aggregated (one line per
+# condition with a count), not one line per offending source.
+_STALE_DAYS = 10
+_REJECT_MIN_INGESTED = 20
+_REJECT_RATE = 25      # acceptance below this % (with enough volume)
+_DUP_MIN_ACCEPTED = 10
+_DUP_RATE = 50         # accepted-duplicate above this %
+_IMG_MIN_SERVED = 10
+_IMG_COVERAGE = 25     # per-source image coverage below this %
+_LONG_REST_HOURS = 12
+
+
+def _attention(content: dict, sources: list[dict], feedback_unread: int, now: datetime | None = None) -> list[dict]:
     """The 'Attention Needed' strip: what an operator should look at, soft-toned
     (warn = act soon, info = worth a glance). Derived from the same data shown
     elsewhere, so it never disagrees with the detail sections."""
     items: list[dict] = []
     n = lambda c: "" if c == 1 else "s"  # noqa: E731 — tiny pluralizer
+    now = now or datetime.now(UTC)
+    active = [s for s in sources if (s.get("status") or ("active" if s.get("active") else "paused")) == "active"]
 
-    resting = [s for s in sources if s.get("active") and (s.get("failures") or 0) > 0]
+    resting = [s for s in active if (s.get("failures") or 0) > 0]
     if resting:
         items.append({"level": "warn", "text": f"{len(resting)} source{n(len(resting))} backing off after failures"})
     flagged = [s for s in sources if s.get("review_flag")]
     if flagged:
         items.append({"level": "warn", "text": f"{len(flagged)} source{n(len(flagged))} flagged for review"})
 
+    # Stale: active, visible feeds whose last success is well in the past.
+    stale_cutoff = (now - timedelta(days=_STALE_DAYS)).strftime("%Y-%m-%d %H:%M:%S")
+    stale = [s for s in active if s.get("content_visible", 1) and s.get("last_success_at") and s["last_success_at"] < stale_cutoff]
+    if stale:
+        items.append({"level": "warn", "text": f"{len(stale)} source{n(len(stale))} haven't updated in over {_STALE_DAYS} days"})
+
+    # High rejection: enough ingested volume, low acceptance.
+    rejecting = [s for s in active if (s.get("total_articles") or 0) >= _REJECT_MIN_INGESTED
+                 and s.get("acceptance_rate") is not None and s["acceptance_rate"] < _REJECT_RATE]
+    if rejecting:
+        items.append({"level": "info", "text": f"{len(rejecting)} source{n(len(rejecting))} accepting under {_REJECT_RATE}% of submissions"})
+
+    # High accepted-duplicate: enough accepted volume, mostly echoing others.
+    duping = [s for s in active if (s.get("accepted_total") or 0) >= _DUP_MIN_ACCEPTED
+              and s.get("accepted_dup_rate") is not None and s["accepted_dup_rate"] > _DUP_RATE]
+    if duping:
+        items.append({"level": "info", "text": f"{len(duping)} source{n(len(duping))} mostly duplicating other feeds"})
+
+    # Low image coverage (info, not a warning).
+    thin = [s for s in active if (s.get("served") or 0) >= _IMG_MIN_SERVED
+            and s.get("image_coverage") is not None and s["image_coverage"] < _IMG_COVERAGE]
+    if thin:
+        items.append({"level": "info", "text": f"{len(thin)} source{n(len(thin))} with thin image coverage (under {_IMG_COVERAGE}%)"})
+
+    # Long rate-limit rest (info).
+    rest_cutoff = (now + timedelta(hours=_LONG_REST_HOURS)).strftime("%Y-%m-%d %H:%M:%S")
+    long_rest = [s for s in sources if s.get("retry_after_at") and s["retry_after_at"] > rest_cutoff]
+    if long_rest:
+        items.append({"level": "info", "text": f"{len(long_rest)} source{n(len(long_rest))} rate-limited for {_LONG_REST_HOURS}h+"})
+
+    # Site-wide signals.
     served = content.get("served") or 0
     with_image = content.get("with_image") or 0
     if served and (with_image / served) < 0.70: