Skip to content

fix(knowledge): repair proper nouns in video-digest ASR transcripts - #8112

Merged
kyle-sexton merged 3 commits into
mainfrom
work-resolve-issue/issue-7746-39105869342-1
Oct 19, 2026
Merged

kyle-sexton merged 3 commits into
mainfrom
work-resolve-issue/issue-7746-39105869342-1

Conversation

@melodic-automation-lanes

@melodic-automation-lanes melodic-automation-lanes Bot commented Oct 18, 2026 •

Copy link
Copy Markdown
Contributor

Refs: #7746

Hold: a human merges this

Summary

Opened by the work-resolve-issue lane. The repair lexicon is now built from the title, description and harvested links for every transcript strategy, and ASR cues go through repairCues, so repairedTermCount counts their repairs. Nothing is passed to Whisper as an initial prompt.

Initial prompt dropped per the owner's decision (Issue Triage Q20); research: .work/issue-triage-course-correction/research/whisper-prompt/RESEARCH.md

Not verified

  • Lowercase title terms such as 'pstack' are not proper-noun-shaped, so they are not extracted; only capitalized or mixed-case title terms enter the lexicon. Not run: the whole extraction directory as one suite (ran the four affected test files).

Fix

  • plugins/knowledge/skills/video-digest/extraction/transcript/proper-noun-repair.js: buildRepairLexicon takes an optional title and extracts terms from it before the description
  • plugins/knowledge/skills/video-digest/extraction/transcript/write-transcript.js: lexicon built for all strategies, ASR cues repaired and counted; runAsr gets no initial prompt
  • plugins/knowledge/skills/video-digest/extraction/transcript/proper-noun-repair.test.js: title-only term and no cross-text phrase tests
  • plugins/knowledge/skills/video-digest/extraction/transcript/write-transcript.test.js: ASR repair/count test
  • plugins/knowledge/skills/video-digest/reference/sources/x.md: lexicon is title, description and links; ASR cues are repaired too
  • .changes/knowledge/fix-video-digest-asr-vocabulary-3d7a9c52.md: patch changelog fragment, repair only

Verification

The run reports:

  • run-test-file write-transcript.test.js: 33 passed
  • run-test-file proper-noun-repair.test.js: 17 passed
  • run-test-file asr-transcribe.test.js: 7 passed
  • run-test-file run-watch-degradation.test.js: 1 passed
  • check-changelog-fragments: all 143 fragments valid
  • Fresh-context subagent review: one finding (uncapped prompt), moot now that the initial prompt is dropped (Q20).
  • The lane run: https://github.com/melodic-software/claude-code-plugins/actions/runs/39105869342

Left out: the items under Not verified.

Related

#7746

🤖 Generated with Claude Code

… video's proper nouns

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
@github-actions

github-actions Bot commented Oct 18, 2026 •

Copy link
Copy Markdown
Contributor

claude-security-review reviewed 2e6b4f87e384430d9a364e1149137498550ae3af: 0 finding(s), 0 as inline comments.

Incremental security review of 2e6b4f8 (changes since f83017e): no security issues.

The three files changed are a changelog fragment, a JSDoc comment in write-transcript.js, and a prose paragraph in reference/sources/x.md. No executable code changed. The wording changes drop the claim that the lexicon is passed to Whisper as an initial_prompt and keep the repair-only description.

No guardrail, hook, allowlist or standing instruction was deleted or softened. The "never auto-installed" statement stays in the doc. The protected-class register was not consulted because nothing was removed that it would apply to.

@github-actions

github-actions Bot commented Oct 18, 2026 •

Copy link
Copy Markdown
Contributor

claude-review reviewed 2e6b4f87e384430d9a364e1149137498550ae3af: 0 finding(s), 0 as inline comments.

No findings in the incremental diff since f83017e.

I read the diff, the surrounding code and the PR context. I did not run the tests.

The change drops the Whisper initial_prompt biasing and keeps the lexicon for proper-noun repair only. The three changed files agree with each other:

  • Changelog fragment: it now says "proper-noun repair" and no longer mentions the Whisper initial prompt.
  • write-transcript.js: the JSDoc now reads "its cues repaired with the same lexicon". The comment at line 377 says the lexicon is never fed to ASR as an initial prompt.
  • reference/sources/x.md: its repair-only paragraph now includes the title and ASR cues, which matches the code.

The existing test at write-transcript.test.js:481 already asserts that runAsr receives no initialPrompt property. asr-transcribe.js still accepts an optional initialPrompt, but nothing in this PR sets it.

Advisory, not blocking: the PR title ("bias and repair ...") and the Summary and Fix sections of the PR body still describe a capped initialPrompt passed to Whisper. After this change the code no longer does that. The title still passes the Conventional Commits check. The body should be updated before a human merges.

No closing-issue criteria applied. The PR uses Refs: #7746, and the context file has no closing-issue section.

@melodic-automation-lanes

Copy link
Copy Markdown
Contributor Author

pr-unblock / address-feedback finished round 1 of 3 on 602bd46.

@melodic-automation-lanes

Copy link
Copy Markdown
Contributor Author

The pr-unblock lane stopped on this pull request and needs a person: a review finding is uncertain or deferred, so a person decides it.

  • Reason: feedback-uncertain
  • Resume key: 602bd46e7ce18b71d2b0793c06ef0cc520718679 (the head the lane stopped on)

To resume, deal with the reason above, remove the needs-human label (and do-not-merge, if set), then dispatch the pr-unblock workflow with pr-number=8112 and activity=address-feedback:

gh workflow run pr-unblock.yml -R melodic-software/claude-code-plugins -f pr-number=8112 -f activity=address-feedback

@melodic-automation-lanes melodic-automation-lanes Bot added the needs-human Human-in-the-loop required; autonomous sessions must not resolve items carrying this. label Oct 18, 2026
…ompt

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
@melodic-automation-lanes

Copy link
Copy Markdown
Contributor Author

pr-unblock / address-feedback finished round 2 of 3 on f83017e.

The owner decided this change repairs transcripts only, with no Whisper
initial prompt. The changelog fragment, the writeEnvelopeTranscriptArtifacts
docblock and the X source reference now say the lexicon (title,
description, harvested links) repairs caption and ASR cues alike.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@kyle-sexton kyle-sexton changed the title fix(knowledge): bias and repair video-digest ASR transcripts with the video's proper nouns fix(knowledge): repair proper nouns in video-digest ASR transcripts Oct 19, 2026
@kyle-sexton

Copy link
Copy Markdown
Member

Decision recorded.

Q5: Repair transcripts only, and drop the change that passes a title-and-description term list to Whisper as initial_prompt?

Yes. Repair only. Basis: the local probe (no gain, more hallucination); Whisper keeps only the prompt tail; OpenAI's archived prompting guide calls prompting "not especially reliable". The owner merges this PR once the docs and title match.

Answered by the owner in the Issue Triage owner interview (tab 5, 2026-10-19) with "accept all Q1 through Q12", after the question was briefed with its recommendation and basis.

@kyle-sexton
kyle-sexton added this pull request to the merge queue Oct 19, 2026
Merged via the queue into main with commit be924dd Oct 19, 2026
57 of 58 checks passed
@kyle-sexton
kyle-sexton deleted the work-resolve-issue/issue-7746-39105869342-1 branch October 19, 2026 03:43
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

needs-human Human-in-the-loop required; autonomous sessions must not resolve items carrying this.

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant