broader-vids 0.1.15 → 0.1.17

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "broader-vids",
3
- "version": "0.1.15",
3
+ "version": "0.1.17",
4
4
  "description": "Curated YouTube and Bilibili recommendations for Claude Code",
5
5
  "author": {
6
6
  "name": "Broader Vids contributors"
@@ -261,6 +261,7 @@ The dashboard polls its status without blocking navigation or rating drafts.
261
261
  After verification, a file journal and database receipt coordinate session activation.
262
262
  Recovery restores the previous files before a database commit or removes the old session after a commit.
263
263
  Completed and cancelled workers remove their temporary profiles and staged blacklist database.
264
+ Account verification and blacklist requests release each Playwright response body after JSON parsing, including parsing failures.
264
265
  Failed cleanup retains the journal and blocks browser use until recovery succeeds.
265
266
  New and resumed recommendation commands record automated-check time.
266
267
  That time counts toward an explicit total allowance when one exists.
@@ -312,6 +313,9 @@ A successful import returns the saved profile and updates the editor without ano
312
313
  The profile stores disabled built-in tags separately from manual tags.
313
314
  Preference imports preserve these local exclusion choices.
314
315
  Tag saves update rules without fitting a learning model.
316
+ Saved rounds compare interests rather than the profile revision.
317
+ Tag edits do not invalidate saved planning, previews, or publication.
318
+ Each round retains its saved tag rules. New rounds use later edits.
315
319
  The dashboard serializes tag writes per platform and rejects older tag snapshots.
316
320
  An interest save preserves a draft when only exclusion metadata changed.
317
321
  Profile revisions prevent one window from replacing a newer interest edit.
@@ -350,9 +354,18 @@ Rounds with alignment version 1 retain their original instructions and cache ide
350
354
  Older builds reject new alignment-version-2 rounds during recovery.
351
355
  The mapping cache identity includes policy 2 when applicable.
352
356
  A policy change can require text-only mapping without repeating visual assessment.
353
- New rounds freeze ranking policy 4 for short-video and portrait deductions.
357
+ New rounds freeze ranking policy 5 for short-video, portrait, and recorded-course deductions.
354
358
  Recovery uses the ranking policy saved in each round.
355
- Adjustment records use format 2 with portrait deduction and verified dimensions for policy 4.
359
+ Adjustment records retain format 2. Policy 4 adds portrait deduction and verified dimensions.
360
+ Policy 5 adds the course-series deduction and its verified metadata and classification.
361
+ `recommendations/series.py` checks complete Bilibili multipart metadata under one BV ID.
362
+ It rejects missing durations, duplicate parts, mismatched identities, and incomplete part counts.
363
+ The existing text-only alignment batch classifies recorded courses after quality assessment.
364
+ Its cache identity includes the eligible series metadata. Quality evidence remains separate.
365
+ The batch requests course records only for eligible multipart archives.
366
+ Interest matches and course classifications have separate validation. Retries preserve valid results from either task.
367
+ Classification uses the same worker limits, retries, temporary-directory cleanup, and token accounting as content alignment.
368
+ Saved policies 1–4 retain their previous ranking behavior.
356
369
  New media verification reads decoded FFmpeg output dimensions after automatic rotation.
357
370
  Quality-cache results carry saved dimensions into ranking, including after full-media cleanup.
358
371
  A retained, verified frame supplies the orientation for older cached dimensions.
package/docs/REFERENCE.md CHANGED
@@ -327,9 +327,17 @@ A busy browser or recommendation operation requires a later retry.
327
327
  Do not remove another operation's lock.
328
328
 
329
329
  Any blank rating dimension makes a recommended video incomplete.
330
- Counts include all previous recommendation exposures for the requested platforms, with duplicate videos counted once.
330
+ Counts cover only the recommendation list currently shown in the dashboard for each requested platform.
331
+ Each video counts once. An empty dashboard list has no reminder.
332
+ When a new list becomes current, the reminder follows that list. Older ratings and history remain saved.
331
333
  Dismissed videos and calibration-only examples do not count.
332
334
  The report separates videos without ratings from partially rated videos.
335
+ Each platform's `videos` array provides the title, URL, and missing dimension IDs for every incomplete video.
336
+ The `dimension_labels` map provides the visible rating-group names.
337
+ The assistant lists these missing groups beside each video in the reminder.
338
+ In the dashboard, **Reviewed** means all four groups have saved feedback.
339
+ A partial count such as **2/4 rated** shows incomplete feedback. The review panel names the missing groups.
340
+ Unsaved edits do not change the saved count.
333
341
  Choose **Open dashboard and stop** to review feedback.
334
342
  Request recommendations again after review.
335
343
  Alternatively, explicitly choose **Continue this round**.
@@ -445,9 +453,13 @@ A central suspected-crime investigation can match crime while its allegations re
445
453
  Mapping confidence describes relevance, not factual proof. A brief incidental mention need not produce a match.
446
454
  This policy does not rewrite published lists, quality evidence, or saved ratings.
447
455
 
448
- New ranking policy 4 adds a five-point final-score deduction for verified portrait videos.
456
+ Ranking policy 4 adds a five-point final-score deduction for verified portrait videos.
449
457
  Portrait means height greater than width. Unknown dimensions give zero portrait deduction.
450
- Open Score details to inspect the dimensions, portrait deduction, short-video deduction, and final score.
458
+ New rounds use ranking policy 5. It also applies the Bilibili course-series rule.
459
+ For multipart videos under one BV ID, it subtracts ten final-score points when content analysis confirms a recorded course.
460
+ The archive must contain at least two parts. Their verified total duration must exceed five hours.
461
+ Unknown course classifications receive zero deduction and appear in the round report.
462
+ Open Score details to inspect the dimensions, deductions, course-series evidence, and final score.
451
463
 
452
464
  ## Review and rate videos
453
465
 
@@ -629,6 +641,7 @@ It shows input tokens, output tokens, cached input, and each step's share of rep
629
641
 
630
642
  Broad-interest alignment appears as its own model-usage stage.
631
643
  The stage reports alignment calls separately from quality assessment.
644
+ It also includes recorded-course classification when a long Bilibili multipart archive needs that check.
632
645
  Reported tokens equal input plus output. Input totals include cached input.
633
646
  Local search, screening, evidence preparation, and scoring do not use model tokens.
634
647
  Quality assessment combines caption and frame input in the same model call.
@@ -1021,10 +1034,15 @@ It requests the system browser and returns the exact URL.
1021
1034
  If the system cannot start the browser, the command returns `manual_open_required` and keeps the server available.
1022
1035
  Open the returned URL in your system browser if an embedded browser stays blank.
1023
1036
  `--no-browser` returns its URL without opening the system browser.
1024
- The dashboard skill presents that URL in chat.
1037
+ The dashboard skill verifies status in a separate command before it presents the URL in chat.
1038
+ On Windows, startup requests an independent process so the dashboard can outlive its launcher.
1039
+ If the host denies this request, approve a persistent local dashboard through the host's permission prompt and retry.
1040
+ A denied request returns an error before the command reports a running server.
1025
1041
  A busy preferred port causes the server to select a free port.
1026
1042
  Stop authenticates the registered server before requesting shutdown.
1027
1043
  It does not stop a process based only on a PID or port.
1044
+ On Windows, Stop waits up to five seconds for process exit before reporting `stopped`.
1045
+ It returns `stopping` if process exit remains unverified.
1028
1046
  Private control records and bounded logs reside in the profile's `.dashboard/` directory.
1029
1047
  The foreground `serve` command remains available for terminal use.
1030
1048
 
package/docs/SCORING.md CHANGED
@@ -29,11 +29,21 @@ For new assessments without a usable transcript, subtract ten points from qualit
29
29
  The same deduction applies to the learned quality component during ranking.
30
30
  After the learned final score, subtract 20 × clamp((300 − t) / 240, 0, 1), where t is seconds.
31
31
  This deduction is 15 at two minutes, 10 at three, 5 at four, and 0 at five minutes.
32
- New rounds use ranking policy 4, with a maximum short-video deduction of 20 points.
32
+ New rounds use ranking policy 5, with a maximum short-video deduction of 20 points.
33
33
  Saved policy-2 rounds retain the previous ten-point maximum. Policy-1 rounds retain no duration deduction.
34
34
  Policy 4 also subtracts five final-score points for portrait videos with verified height greater than width.
35
35
  The engine adds this deduction to the short-video deduction after learned scoring.
36
36
  A square or landscape video receives zero portrait deduction. Unknown dimensions also give zero deduction.
37
+ Policy 5 also subtracts ten final-score points for a recorded Bilibili course series under one BV ID.
38
+ The archive must contain at least two parts, and their verified total duration must exceed 18,000 seconds.
39
+ Exactly five hours gives zero course-series deduction.
40
+ The content-analysis batch classifies recorded course lessons from the assessment summary and series metadata.
41
+ Apply the deduction only to a recorded-course classification with medium or high confidence.
42
+ An unknown classification gives zero deduction and a warning in the round report.
43
+ Other multipart content gives zero deduction.
44
+ YouTube playlists and Bilibili collections across separate BV IDs are outside this rule.
45
+ The engine adds all final-score deductions after learned scoring, then clamps the result at zero.
46
+ Saved policies 1–4 retain their previous deductions.
37
47
  Eligibility uses evidence quality before these final-score deductions.
38
48
  Policy 3 applies the twenty-point short-video rule.
39
49
  Clamp the adjusted score at zero.
@@ -269,7 +279,16 @@ Published historical lists retain their saved scores.
269
279
  Later runs use the current model.
270
280
 
271
281
  Provisional previews use the round's saved ranking policy and validated interest mappings.
272
- The short-video deduction applies once when the saved policy requires it.
282
+ For policy 5, let $L$ be the final score after learned calibration.
283
+ The published score is
284
+
285
+ $$
286
+ S=\max(0,L-P_{\mathrm{short}}-P_{\mathrm{portrait}}-P_{\mathrm{course}}).
287
+ $$
288
+
289
+ The course deduction is $P_{\mathrm{course}}=10$ when both course-series conditions hold and classification confidence is medium or high.
290
+ Otherwise, $P_{\mathrm{course}}=0$.
291
+ Each deduction applies once when the saved policy requires it.
273
292
  Previews create no recommendation history.
274
293
 
275
294
  ## 6. Structured feedback
package/docs/WORKFLOW.md CHANGED
@@ -45,11 +45,15 @@ Reset keeps tools and Bilibili sign-in while removing learned data and histories
45
45
  Clean removes the complete managed data directory while keeping the plugin installed.
46
46
  Both destructive actions start with a preview and require an explicit request before application.
47
47
 
48
- Before a new round, the engine counts incomplete feedback across earlier recommendations for the requested platforms.
48
+ Before a new round, the engine checks the current dashboard recommendation list for each requested platform.
49
+ An empty list has no feedback reminder. A new current list replaces the earlier list as the reminder target.
50
+ The reminder lists each incomplete video and names its missing rating groups.
49
51
  It counts each canonical video once and excludes dismissed videos and calibration-only examples.
50
52
  Any blank topic, subject, depth, or quality rating makes a video incomplete.
51
53
  Watched status does not change this count.
52
54
  The report separates unrated videos from partially rated videos.
55
+ The dashboard shows a saved count such as **2/4 rated** for partial feedback.
56
+ **Reviewed** means all four rating groups have saved feedback. The review panel lists missing groups.
53
57
  Choose **Open dashboard and stop** or **Continue this round**.
54
58
  The dashboard choice ends the request until you invoke recommendations again.
55
59
  Continuing leaves blank ratings unchanged and permits only this new round.
@@ -101,7 +105,7 @@ A shared batch deduplicates new topic proposals across platforms at its baseline
101
105
  It does not force the platforms to choose different topics.
102
106
 
103
107
  Legacy rounds retain their saved policies when they resume.
104
- New rounds use rotation version 2 and ranking policy 4.
108
+ New rounds use rotation version 2 and ranking policy 5.
105
109
  The language rule changes search text, not interest labels or hierarchy levels.
106
110
  Existing topic labels and aliases stay stable as the hierarchy grows.
107
111
  The engine converts Traditional Chinese keywords to Simplified Chinese and rejects incompatible scripts before a keyword search.
@@ -352,10 +356,18 @@ It then uses exact matches and reports the fallback.
352
356
 
353
357
  The engine applies a duration deduction after the learned final score: 20 × clamp((300 − t) / 240, 0, 1).
354
358
  The deduction is 15 at two minutes, 10 at three, 5 at four, and 0 at five minutes.
355
- New rounds freeze ranking policy 4. Each round stores its policy for continuation.
359
+ New rounds freeze ranking policy 5. Each round stores its policy for continuation.
356
360
  Verified portrait dimensions, where height exceeds width, add a five-point final-score deduction.
357
361
  The engine uses saved media dimensions and reports both deductions in Score details.
358
362
  Missing dimensions give zero portrait deduction.
363
+ For Bilibili, the existing content-analysis batch also classifies recorded courses in verified multipart archives.
364
+ A recorded course with at least two parts and a total duration above five hours receives a ten-point final-score deduction.
365
+ The engine requires medium or high confidence in the course classification.
366
+ It uses the existing archive metadata and content-analysis request. It adds no separate request.
367
+ Retries preserve valid interest matches and course classifications independently.
368
+ The response requires course records only for eligible multipart archives.
369
+ Uncertain classifications receive zero deduction and appear in the round report.
370
+ Score details shows the deduction, part count, series duration, and classification reason.
359
371
  The adjusted score has a floor of zero.
360
372
  The existing evidence quality gate and screening rules remain unchanged.
361
373
  It applies no unrelated-topic quota.
@@ -418,6 +430,7 @@ It identifies manual rules, imported Bilibili bans, and built-in rules.
418
430
  Remove manual rules there when needed.
419
431
  The built-in 视频翻译 tag has a Remove button. Add the tag again to restore that exclusion.
420
432
  Tag removals remain saved after a restart or preference import.
433
+ Tag edits apply to new rounds. Active and resumed rounds retain their saved tag rules.
421
434
  Unblock imported creators on Bilibili, then sync the blacklist.
422
435
 
423
436
  Bilibili sign-in and each new Bilibili round attempt a complete blacklist sync.
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "broader-vids",
3
- "version": "0.1.15",
3
+ "version": "0.1.17",
4
4
  "description": "Codex and experimental Claude Code skills for local YouTube/Bilibili recommendations",
5
5
  "license": "MIT",
6
6
  "engines": {
@@ -11,7 +11,7 @@
11
11
  },
12
12
  "files": [
13
13
  "bin/",
14
- "python/broader_vids-0.1.15-*.whl",
14
+ "python/broader_vids-0.1.17-*.whl",
15
15
  "python/core-requirements.txt",
16
16
  "python/evidence-requirements.txt",
17
17
  "python/asr-requirements.txt",
@@ -48,6 +48,6 @@
48
48
  "tag": "latest"
49
49
  },
50
50
  "broaderVidsRelease": {
51
- "sourceCommit": "3c8ef20b931af015dfa5f788c00e114d3e3c8b31"
51
+ "sourceCommit": "0973f7d06fdbfe25bbe358a27a6fa511a93f0411"
52
52
  }
53
53
  }
package/plugin.json CHANGED
@@ -1,7 +1,7 @@
1
1
  {
2
2
  "$schema": "https://agent-plugins.org/schemas/1.0.0/plugin.schema.json",
3
3
  "name": "broader-vids",
4
- "version": "0.1.15",
4
+ "version": "0.1.17",
5
5
  "description": "Curated YouTube and Bilibili recommendations with bounded quality evidence and local watched/feedback state.",
6
6
  "license": "MIT",
7
7
  "extensions": {
@@ -58,8 +58,12 @@ Optional warnings do not block recommendations.
58
58
  Routine checks do not repeat the first-time introduction.
59
59
 
60
60
  For `feedback_required`, show platform counts and distinguish unrated videos from partially rated videos.
61
+ List every video in each platform's `feedback.platforms[platform].videos` array by title and returned URL.
62
+ For each video, translate its `missing_dimensions` with `feedback.dimension_labels` and state the missing rating groups.
61
63
  Any blank rating dimension makes a video incomplete.
62
- The count includes all prior, non-dismissed recommendations for the requested platforms.
64
+ The reminder covers only the recommendation list currently shown in the dashboard for each requested platform.
65
+ An empty dashboard list has no reminder. A new current list replaces the earlier list as the reminder target.
66
+ Dismissed videos do not count. Earlier pending videos outside the current list do not count.
63
67
  Calibration-only videos do not count.
64
68
  Offer **Open dashboard and stop** or **Continue this round**.
65
69
  For the dashboard choice, open the dashboard and end this recommendation request.
@@ -119,6 +123,8 @@ Report fewer than 100 eligible entries as incomplete without starting quality as
119
123
  A complete final list contains ten videos per platform.
120
124
  Report a partial result as partial.
121
125
  Show all requested platform lists before the final candidate-failure summary.
126
+ Report any `course_series_warnings` after the lists.
127
+ These warnings mean recorded-course classification was uncertain, so the course-series deduction was zero.
122
128
  Use each report's `candidate_failures` to identify skipped videos and downloads recovered by retries.
123
129
  Include video links, actual attempt counts, and brief reasons. Do not claim four attempts when a saved limit stopped retries earlier.
124
130
  Check `download_outcome` before describing a download as failed. A completed download can still have an unsupported assessment.
@@ -20,7 +20,12 @@ If the runtime is missing, use the packaged [setup skill](../broader-vids-setup/
20
20
 
21
21
  For an open request, run `dashboard open` with an explicit absolute `--data-dir`.
22
22
  The command starts or reuses that profile's server and requests the system browser.
23
- Present the returned URL as a clickable link.
23
+ After open finishes, run `dashboard status` in a separate tool call.
24
+ Use the same data directory.
25
+ Present the verified URL as a clickable link only when status reports running.
26
+ If the server stopped or Windows denied process creation, request host permission for a persistent local dashboard.
27
+ Retry once through the host's supported approval mechanism, then verify status again.
28
+ If permission is denied or the server stops again, report the failure and end the request.
24
29
  `browser_opened: true` means the system accepted the browser request. It does not prove that the page rendered.
25
30
  If `browser_opened` is false, show the returned fallback message and link.
26
31
  Use `--no-browser` when the user requests only a link or server startup.
@@ -31,11 +31,20 @@ For a successful explicit setup request, return the full guide produced by `onbo
31
31
 
32
32
  Use the recommendation skill's engine workflow, evidence rules, saved profile, and feedback gate. Run one recommendation command with `--host claude`, the verified `--model` and `--effort`, and the requested platform options.
33
33
 
34
+ For `feedback_required`, list each current dashboard video by title and URL, with its missing rating groups.
35
+ Use `feedback.platforms[platform].videos` and `feedback.dimension_labels` from the report.
36
+ The reminder covers only the current dashboard list for each requested platform. An empty list has no reminder.
37
+ A new current list replaces the earlier list as the reminder target.
34
38
  Do not bypass incomplete feedback without the user's explicit choice. Do not reduce requested platform coverage. Report incomplete results as partial and keep candidate failure details accurate. If either platform pauses, report the original batch cause and end the request. Continue a paused run only when the user asks.
35
39
 
36
40
  ## Dashboard
37
41
 
38
- Open the dashboard by running the engine's `dashboard` command with `--host claude`. Use the returned local URL. Do not start another server when one is already active. Stop only the verified Broader Vids dashboard for the selected profile.
42
+ Open the dashboard with the engine's `dashboard` command and `--host claude`.
43
+ After open finishes, verify `dashboard status` in a separate tool call before presenting the returned local URL.
44
+ If Windows denies persistent startup, request host permission through its supported approval mechanism.
45
+ Verify status after one retry. Report the failure if startup remains blocked.
46
+ Do not start another server when one is already active.
47
+ Stop only the verified Broader Vids dashboard for the selected profile.
39
48
 
40
49
  The dashboard manages ratings, exclusions, interests, imports, and token reports. Do not create ratings or change preferences without the user's instruction. Dashboard access does not require a recommendation round.
41
50