Skip to content

AI agent skill that researches any topic across Reddit, X, YouTube, HN, Polymarket, and the web - then synthesizes a grounded summary

fix(youtube): spend transcript budget on in-window videos first; gate stale-yt-dlp nudge on actual fetch failures (#535)

Hammad Khan92c104b911ed
diff --git a/skills/last30days/scripts/last30days.py b/skills/last30days/scripts/last30days.py
index cc0cf66e..0c1a1bb4 100644
--- a/skills/last30days/scripts/last30days.py
+++ b/skills/last30days/scripts/last30days.py
@@ -965,10 +965,12 @@ def main() -> int:
if not args.hiring_signals:
try:
from lib import quality_nudge
+ from lib import youtube_yt as _youtube_yt
# Populate transcript-fetch ratio so quality_nudge can detect the
# degraded-YouTube failure mode (videos returned but transcripts
# silently failed - typically a stale yt-dlp binary).
youtube_items = report.items_by_source.get("youtube") or []
+ _yt_fetch_stats = _youtube_yt.get_transcript_fetch_stats()
instagram_items = report.items_by_source.get("instagram") or []
research_results = {
"youtube_videos_count": len(youtube_items),
@@ -985,6 +987,13 @@ def main() -> int:
"youtube_captions_disabled_count": sum(
1 for it in youtube_items if it.metadata.get("captions_disabled")
),
+ # Actual yt-dlp fetch outcomes for this run. The counts above are
+ # computed from post-pruning items, so they can't tell "fetches
+ # failed (stale binary)" from "fetches succeeded but the videos
+ # were pruned downstream"; the latter was producing false
+ # stale-yt-dlp nudges (#531).
+ "youtube_transcript_fetch_attempts": _yt_fetch_stats["attempts"],
+ "youtube_transcript_fetch_failures": _yt_fetch_stats["failures"],
# Track Instagram returned-zero-items so quality_nudge can detect
# the silent-failure case (SC configured but the v2 reels endpoint
# 500'd through both the original query and the hashtag retry).
diff --git a/skills/last30days/scripts/lib/quality_nudge.py b/skills/last30days/scripts/lib/quality_nudge.py
index dcd5aa76..e1b6545f 100644
--- a/skills/last30days/scripts/lib/quality_nudge.py
+++ b/skills/last30days/scripts/lib/quality_nudge.py
@@ -63,12 +63,23 @@ def _is_youtube_degraded(research_results: dict, threshold: float) -> bool:
who turned off captions can never produce a transcript, so counting that
video toward "fetch failures" produces false positives. A single
captions-disabled video in a small result set was tripping the nudge.
+
+ When actual fetch outcomes are available, they take precedence over the
+ post-pruning ratio: the report counts only see items that survived
+ freshness/relevance pruning, so a run where every transcript fetch
+ succeeded but the fetched videos were later pruned looks identical to a
+ stale-binary run (#531). Zero failures across attempted fetches proves
+ the binary works - don't flag.
"""
videos = int(research_results.get("youtube_videos_count") or 0)
transcripts = int(research_results.get("youtube_transcripts_count") or 0)
captions_disabled = int(research_results.get("youtube_captions_disabled_count") or 0)
if videos <= 0:
return False
+ fetch_attempts = int(research_results.get("youtube_transcript_fetch_attempts") or 0)
+ fetch_failures = int(research_results.get("youtube_transcript_fetch_failures") or 0)
+ if fetch_attempts > 0 and fetch_failures == 0:
+ return False
eligible = videos - captions_disabled
if eligible <= 0:
# Every returned video had captions disabled - upstream content fact,
diff --git a/skills/last30days/scripts/lib/youtube_yt.py b/skills/last30days/scripts/lib/youtube_yt.py
index 5e7380c2..0a4380f0 100644
--- a/skills/last30days/scripts/lib/youtube_yt.py
+++ b/skills/last30days/scripts/lib/youtube_yt.py
@@ -33,6 +33,25 @@ TRANSCRIPT_LIMITS = {
"deep": 8,
}
+# Cumulative yt-dlp transcript-fetch stats for the current process. The final
+# report only sees post-pruning items, so it can't distinguish "fetches failed
+# (stale binary)" from "fetches succeeded but the videos were pruned later".
+# quality_nudge reads these via last30days.py to suppress the stale-yt-dlp
+# nudge when every attempted fetch actually succeeded. yt-dlp path only: the
+# nudge diagnoses the local binary, not the ScrapeCreators API.
+_TRANSCRIPT_FETCH_STATS = {"attempts": 0, "failures": 0}
+
+
+def get_transcript_fetch_stats() -> Dict[str, int]:
+ """Return cumulative transcript-fetch stats for this process."""
+ return dict(_TRANSCRIPT_FETCH_STATS)
+
+
+def reset_transcript_fetch_stats() -> None:
+ """Reset cumulative transcript-fetch stats (used by tests)."""
+ _TRANSCRIPT_FETCH_STATS["attempts"] = 0
+ _TRANSCRIPT_FETCH_STATS["failures"] = 0
+
# Max words to keep from each transcript
TRANSCRIPT_MAX_WORDS = 5000
@@ -712,6 +731,13 @@ def search_and_transcribe(
transcripts = fetch_transcripts_parallel(
candidate_ids, out_captions_disabled=captions_disabled_ids,
)
+ # Record fetch outcomes (captions-disabled videos can never succeed,
+ # so they don't count as failures) for the stale-yt-dlp nudge.
+ _TRANSCRIPT_FETCH_STATS["attempts"] += len(candidate_ids)
+ _TRANSCRIPT_FETCH_STATS["failures"] += sum(
+ 1 for vid in candidate_ids
+ if not transcripts.get(vid) and vid not in captions_disabled_ids
+ )
else:
_log(f"Transcript limit is 0 for depth={depth}, skipping transcript fetch")
@@ -966,8 +992,13 @@ def search_youtube_sc(
transcript_limit = TRANSCRIPT_LIMITS.get(depth, TRANSCRIPT_LIMITS["default"])
if transcript_limit > 0 and items:
attempt_count = min(len(items), transcript_limit * 3)
+ # Same in-window-first ordering as search_and_transcribe(): don't let
+ # an out-of-window back-catalog (kept by the soft date filter above)
+ # consume the transcript budget of videos the freshness scorer keeps.
+ in_window = [i for i in items if i.get("date") and i["date"] >= from_date]
+ out_of_window = [i for i in items if not (i.get("date") and i["date"] >= from_date)]
_log(f"Fetching SC transcripts for up to {attempt_count} videos (target: {transcript_limit})")
- for item in items[:attempt_count]:
+ for item in (in_window + out_of_window)[:attempt_count]:
vid = item["video_id"]
if not vid:
continue
diff --git a/tests/test_quality_nudge.py b/tests/test_quality_nudge.py
index 9802ba47..8f367b31 100644
--- a/tests/test_quality_nudge.py
+++ b/tests/test_quality_nudge.py
@@ -377,6 +377,72 @@ class TestYouTubeCaptionsDisabledDoesNotFalseFlag:
assert "youtube" in q["core_degraded"]
+class TestStaleNudgeRequiresActualFetchFailures:
+ """Zero failed fetches must suppress the stale-yt-dlp nudge (#531).
+
+ The report counts (youtube_videos_count / youtube_transcripts_count) are
+ computed from post-pruning items. A run where every transcript fetch
+ succeeded but the fetched videos were later pruned by freshness scoring
+ looks identical to a stale-binary run from those counts alone, producing
+ a false "stale yt-dlp binary" nudge. When actual fetch outcomes are
+ available and show zero failures, the binary demonstrably works.
+ """
+
+ def test_zero_failures_does_not_flag(self):
+ # The #531 repro: 2 in-report videos, 0 transcripts among them, but
+ # all 6 attempted fetches succeeded (on videos pruned later).
+ q = _compute(
+ ytdlp_installed=True,
+ result_overrides={
+ "youtube_videos_count": 2,
+ "youtube_transcripts_count": 0,
+ "youtube_captions_disabled_count": 0,
+ "youtube_transcript_fetch_attempts": 6,
+ "youtube_transcript_fetch_failures": 0,
+ },
+ )
+ assert "youtube" not in q["core_degraded"]
+
+ def test_actual_failures_still_flag(self):
+ q = _compute(
+ ytdlp_installed=True,
+ result_overrides={
+ "youtube_videos_count": 6,
+ "youtube_transcripts_count": 0,
+ "youtube_captions_disabled_count": 0,
+ "youtube_transcript_fetch_attempts": 6,
+ "youtube_transcript_fetch_failures": 6,
+ },
+ )
+ assert "youtube" in q["core_degraded"]
+
+ def test_missing_fetch_stats_preserves_existing_behavior(self):
+ # Callers that don't pass fetch stats (or the SC path, which doesn't
+ # use the local yt-dlp binary) fall back to the ratio heuristic.
+ q = _compute(
+ ytdlp_installed=True,
+ result_overrides={
+ "youtube_videos_count": 6,
+ "youtube_transcripts_count": 0,
+ "youtube_captions_disabled_count": 0,
+ },
+ )
+ assert "youtube" in q["core_degraded"]
+
+ def test_zero_attempts_preserves_existing_behavior(self):
+ q = _compute(
+ ytdlp_installed=True,
+ result_overrides={
+ "youtube_videos_count": 6,
+ "youtube_transcripts_count": 0,
+ "youtube_captions_disabled_count": 0,
+ "youtube_transcript_fetch_attempts": 0,
+ "youtube_transcript_fetch_failures": 0,
+ },
+ )
+ assert "youtube" in q["core_degraded"]
+
+
class TestInstagramSilentFailure:
"""Instagram is a `bonus` source via SC. Silent-failure detection: if SC
is configured but the source returned zero items, surface a nudge so the
diff --git a/tests/test_youtube_yt.py b/tests/test_youtube_yt.py
index 1414e10f..b1b741a8 100644
--- a/tests/test_youtube_yt.py
+++ b/tests/test_youtube_yt.py
@@ -482,6 +482,79 @@ class TestSearchAndTranscribe(unittest.TestCase):
ft_mock.assert_not_called()
+class TestTranscriptFetchStats(unittest.TestCase):
+ """Track yt-dlp fetch outcomes for quality_nudge (#531 false stale-yt-dlp nudge)."""
+
+ FROM_DATE = "2026-03-01"
+ TO_DATE = "2026-03-31"
+
+ def setUp(self):
+ youtube_yt.reset_transcript_fetch_stats()
+
+ def _make_item(self, video_id, views, date):
+ return {
+ "video_id": video_id,
+ "title": f"Video {video_id}",
+ "url": f"https://www.youtube.com/watch?v={video_id}",
+ "channel_name": "TestChannel",
+ "date": date,
+ "engagement": {"views": views, "likes": 10, "comments": 5},
+ "relevance": 0.8,
+ "why_relevant": "test",
+ "description": "test desc",
+ "duration": 600,
+ }
+
+ def _run(self, items, fake_parallel=None):
+ if fake_parallel is None:
+ def fake_parallel(video_ids, max_workers=5, out_captions_disabled=None):
+ return {vid: "A detailed transcript about the topic." for vid in video_ids}
+ with mock.patch.object(youtube_yt, "search_youtube", return_value={"items": items}), \
+ mock.patch.object(youtube_yt, "fetch_transcripts_parallel", side_effect=fake_parallel):
+ return youtube_yt.search_and_transcribe(
+ "test topic", self.FROM_DATE, self.TO_DATE, depth="default",
+ )
+
+ def test_fetch_stats_track_attempts_and_failures(self):
+ items = [
+ self._make_item("ok1", 3_000, "2026-03-20"),
+ self._make_item("fail1", 2_000, "2026-03-15"),
+ self._make_item("nocap1", 1_000, "2026-03-10"),
+ ]
+
+ def fake_parallel(video_ids, max_workers=5, out_captions_disabled=None):
+ result = {}
+ for vid in video_ids:
+ if vid.startswith("nocap"):
+ result[vid] = None
+ if out_captions_disabled is not None:
+ out_captions_disabled.add(vid)
+ elif vid.startswith("fail"):
+ result[vid] = None
+ else:
+ result[vid] = "A detailed transcript about the topic."
+ return result
+
+ self._run(items, fake_parallel)
+
+ stats = youtube_yt.get_transcript_fetch_stats()
+ self.assertEqual(stats["attempts"], 3)
+ # Captions-disabled videos can never succeed; they are not failures.
+ self.assertEqual(stats["failures"], 1)
+
+ def test_fetch_stats_zero_failures_when_all_succeed(self):
+ # The #531 scenario: every fetch succeeds (on videos later pruned by
+ # freshness scoring). failures must be 0 so quality_nudge does not
+ # blame a stale yt-dlp binary.
+ items = [self._make_item(f"v{i}", 1_000 * (i + 1), "2024-01-15") for i in range(4)]
+
+ self._run(items)
+
+ stats = youtube_yt.get_transcript_fetch_stats()
+ self.assertEqual(stats["attempts"], 4)
+ self.assertEqual(stats["failures"], 0)
+
+
class TestYtdlpSSHRouting(unittest.TestCase):
"""LAST30DAYS_YOUTUBE_SSH_HOST routes yt-dlp invocations through SSH for residential IP."""