[kcarnold]

Superseded. This PR has been split and rebuilt on current main:

• Drop the reviewed/auto tier from slide translations #182: drop the reviewed/auto tier
• Surface the translator's per-slide notes on the entry #184: surface the per-slide notes (stacked on Drop the reviewed/auto tier from slide translations #182). Notes now live on the slide translation entry (note?, next to provenance) rather than on the conversation.

The Scripture-credit follow-up that was in a comment here is now issue #181. Close this PR once those land.

---

What

Two halves, same idea: stop tracking a quality tier nobody acts on, and start showing the per-slide notes that actually say "look at this one."

Deleted — the reviewed / auto status

SlideTranslationEntry is now { text, provenance }. Gone: SlideStatus, status, reviewedAt, and the viewer's "unreviewed" badge.

• resolveSlideTranslation collapses from a two-pass quality walk to "first entry along the fallback chain."
• slideLibrary.ts no longer stamps status: 'reviewed' / reviewedAt. It rebuilds records on load rather than storing them verbatim, so the dropped fields don't ride along into the next write of the on-disk library.
• slide_translator.py (the Python work-ahead worker) no longer needs to recognize a protected entry. _store_translations now writes only keys that are missing. Keys are content-addressed, so an entry that exists already translates this exact text — whoever wrote it knew at least as much as this run does.
• The review grid's per-cell chip says Saved, not Reviewed. That is what it always meant: the draft matches the library.

Surfaced — the model's per-slide notes

set_translations has always had an optional per-segment note ("a genuine ambiguity or judgement call worth flagging"). applySetTranslations threw it away, and the conversation panel summarized the call as ✍️ set translations: French (5) — so the notes existed and were visible nowhere.

Now they ride on the conversation: SlideConversation.notes, content-addressed by language + source text. Each agent run is seeded with the item's current notes and returns the whole set, so the server replaces rather than merges, and a follow-up touching one slide can't silently drop the notes on the others. A rewrite with no note withdraws the note it had; a revise_translation deliberately leaves notes alone.

The review screen is reorganized around them:

• the service-item picker is now a row of chips instead of a <select>, each badged 📝 n with its note count — the "at a glance, which things have notes" view;
• a flagged row is tinted and marked in the source column;
• a flagged cell leads with its note, above the textarea;
• a count (or "nothing flagged") sits next to the screen title.

The set_translations tool description now tells the model its note is displayed and marks the item — so it has a reason to be sparing.

Behavior change worth a look

A Haitian Creole viewer with no Creole text for a slide now sees the French one, whatever produced it. Previously an unreviewed French text was withheld and the slide read "(not translated)". With the tier gone this is the only coherent reading of the fallback chain, and in practice the gap only opens on a surprise slide the work-ahead worker hasn't reached — but it is a real change, not a refactor.

Smoke test

Touches §3 Proclaim sync — "Slide translations show for the on-air item; review screen loads its conversation." Worth adding to that pass: open an item with notes from the picker and confirm the badge count matches the flagged rows.

Verification

npm run typecheck, npm run lint, npm test (600 passed), uv run pytest — all green. New tests cover the note lifecycle in the agent (attach, withdraw, carry across an untouched follow-up, drop when the source text is gone) and the note rendering in SlideReview.

docs/slide-translations-plan.md gets an amendment at the end recording the removal, in the style of the existing post-build corrections, rather than a rewrite of 385 lines that describe the old design.

🤖 Generated with Claude Code

https://claude.ai/code/session_01RuD8vmbn9ku8wbVQYMH5BH

---

Generated by Claude Code

[kcarnold]

A follow-up plan once 169/170 land:

Credit Scripture versions on translated slides

Context

The slide agent (draftItemTranslations / runSlideTranslationAgent in nlp.ts) grounds
Scripture slides in published Bibles fetched through lookup_bible_passage (bible.ts,
BIBLE_TRANSLATIONS). Nothing records which published text a slide came from, so viewers
never see a credit. Some of these texts carry licenses that require one; for example, hatbsa
is CC BY-SA 4.0 (ebible.org/Scriptures/details.php?id=hatbsa). The goal is a small,
unobtrusive credit under the translated slide, shown only when the slide reproduces Scripture
(at least about one verse). A phrase borrowed for a song gets no credit.

Decisions made: the model declares the credit, and the viewer shows version + license
only (no reference). On the base: it was "stack on PR 152", but 169/170 now move the
pipeline too (see below). You expect 169/170 to land soon, so: build on main after 169/170. If 152 is still
open then, stack on 152 (rebased onto that main), since it owns the entry shape and the notes
UI this extends.

Design

1. Sources registry: src/textSources.ts (new, pure)

• TextSource = { id, kind: 'scripture' | 'creed', name, publisher, license, licenseUrl, sourceUrl }
• TEXT_SOURCES: Record<string, TextSource>, keyed by helloao id (hatbsa, fra_ncl, spa_r09).
  Creeds and confessions are added here later under kind: 'creed', with no schema change.
• formatCredit(source) returns e.g. Bib la — Société Biblique Haïtienne · CC BY-SA 4.0.
• Entries store the id, not the rendered string. If a credit turns out wrong, we fix the
  registry once and every stored slide picks up the fix.
• bible.ts keeps BIBLE_TRANSLATIONS, and a test asserts that every id in it has a
  TEXT_SOURCES entry. Adding a Bible without its credit then fails CI.

2. Data model (on top of PR 152)

• SlideTranslationEntry in src/slideTranslation.ts gets optional sourceId?: string and
  adapted?: boolean; both ride the same path below.
• It is threaded through TranslationBlockResult (nlp.ts) → translateItem
  (src/slideItemTranslation.ts) → the returned entries → slideLibrary.ts (persist, and keep
  it through the record rebuild that PR 152 added) → the /api/slideLibrary upsert body.
• slide_translator.py _store_translations: copy sourceId when present.

3. Model declares it: nlp.ts

• Add two optional fields to the set_translations segment schema, both named generically:
  • source, an enum. It is 'bible' today and gains 'creed' once a creed lookup tool
    exists. The description gives the rule: set only when this slide reproduces at least
    one full verse (or sentence) of text you looked up; omit it for allusions, short phrases,
    and song lyrics.
  • adapted, a boolean: whether the slide alters the published wording. We trust the agent
    on this.
• The model doesn't echo a reference. I checked what a lookup returns:
  { reference: "PSA 23", passages: { French: … } }, where the reference is the requested
  range. That is often wider than one slide (a chapter lookup, with the slide holding verses
  1–3). So echoing it would be loose matching at best. It also isn't needed: the credit
  depends only on which version, and that's fixed by the language (BIBLE_TRANSLATIONS).
• Server-side guard: attach sourceId = BIBLE_TRANSLATIONS[language] only if the
  conversation messages contain a lookup_bible_passage response that returned a passage
  in that language. Scanning messages, rather than a per-run list, covers follow-up runs
  whose lookup happened in an earlier round. The model can't name a license, and it can't
  get a credit for a version it never fetched.
• Store sourceId (and adapted) at write time, so a later change to BIBLE_TRANSLATIONS
  doesn't relabel old slides.
• Working-copy semantics follow PR 152's notes: a rewrite without source withdraws the
  credit, and revise_translation leaves it alone.
• One sentence goes in bibleSection of buildSlideTranslationPrompt pointing at the fields.

4. Viewer: src/SlideTranslationViewer.tsx

• When resolved.entry.sourceId maps to a source, render a footer line through SlideText.
  Add a footer prop next to header if SlideText lacks one. Style it text-xs, low
  opacity, and keep it outside the fit-text measurement so it doesn't shrink the slide.
• The license name links to licenseUrl. This matters for phone viewers; on the projector it
  is just text.
• No label for a verbatim quote: the line itself (version · publisher · license) is
  self-explanatory. There is one localized string, adaptedFrom ("adapted from" / "adapté
  de" / …), in src/strings.ts for all four locales. It is used only when adapted is true.
  It follows the viewer's UI locale like every other string.

5. Review screen: src/SlideReview.tsx

• Show the same credit line in the cell, next to PR 152's note, so the reviewer can see the
  claim. A × to clear it is optional and goes in only if it turns out trivial.

6. Follow-up issue (not built)

• "Require a Scripture credit on BiblePassage items". itemKind already reaches
  proclaimPresentations (yjs_publisher.py) but is not sent to /api/translateItem
  (only itemTitle is, from proclaim_service.py and src/slideTranslationApi.ts). Per your
  answer, this goes to an issue. Also include: the source-language credit (Proclaim's own
  English version on currentSlide), and localized references if that's wanted later.

If PRs 169/170 land first (server-owned slide sync)

Most of the plan is untouched. 169 doesn't change nlp.ts (the agent, its tools, the
prompt), the viewer, the review screen, or slideLibrary.ts. The changes are:

• slide_translator.py is deleted. Its entry write moves to writeSlideTranslations in
  slideSync.ts, so sourceId/adapted get copied there. Python is no longer involved.
• One entry point instead of two. Both /api/translateItem and the server's
  translate-ahead now go through draftItem in server.ts, so the credit is threaded once.
• The BiblePassage rule becomes cheap. The snapshot now carries itemKind to the server
  (slideSync.ts), and proclaimPresentations already stores it for the review screen. So
  passing itemKind into draftItem → buildSlideTranslationPrompt is a few lines: "this
  item is a Bible passage; every slide must declare source: 'bible'". This moves from
  follow-up issue into scope.
• Ordering conflict to watch: 169 still skips status === 'reviewed' entries, and 152
  deletes status. Whichever lands second has to reconcile that. This work is cleaner on
  main after both than stacked on either one.

Things to settle while implementing (flagged, not assumed)

• Registry values (supplied by you; hatbsa's publisher name still gets copied verbatim
  from its ebible page):
  • fra_ncl: Sainte Bible néo-Crampon Libre, © 2022 Fraternité de Tibériade, CC BY-SA 4.0.
    Its terms say an adapted text must indicate the change, which the adapted → "adapted
    from" line covers.
  • spa_r09: Reina-Valera 1909, public domain. No credit is legally required, but we
    still show "Reina-Valera 1909 · dominio público". Crediting only the licensed versions
    would read as inconsistent to a viewer, and the line costs nothing.
  • hatbsa: Bib la (Haitian Bible Society), CC BY-SA 4.0.
• ShareAlike. Under CC BY-SA an adapted slide is itself under CC BY-SA. The credit line
  already names that license, so nothing extra is shown, but it's worth one line in the PR.

Files

src/textSources.ts (new), bible.ts, nlp.ts, src/slideTranslation.ts,
src/slideItemTranslation.ts, slideLibrary.ts, server.ts (upsert body),
slide_translator.py (or, after 169, slideSync.ts + draftItem in server.ts),
src/SlideTranslationViewer.tsx, src/SlideText.tsx,
src/SlideReview.tsx, src/strings.ts, and a short note in docs/slide-translations-plan.md
in PR 152's amendment style.

Verification

• Unit tests: registry covers BIBLE_TRANSLATIONS; source: 'bible' attaches sourceId
  when the history holds a lookup for that language (including one from an earlier round),
  and is dropped when it doesn't; a rewrite without source withdraws it; adapted renders
  "adapted from"; slideLibrary round-trips sourceId; the viewer renders the footer
  only when sourceId resolves; Python _store_translations copies it.
• npm run typecheck, npm run lint, npm test -- --no-color, uv run pytest.
• Smoke §3 (Proclaim sync): replay tests/fixtures/synthetic_service.jsonl and confirm a
  Scripture item shows the credit and a song doesn't.
• Push to claude/vibrant-albattani-lk70bu based on PR 152's head, and open a PR noting it
  stacks on PR 152.