Software Updates

transcript defaults to the main spoken language

• Product decision: the after-meeting transcript opens in the meeting's MAIN
• spoken language only. The translation column defaults on solely when the
• meeting was CLEARLY multilingual — a second language with real presence
• (>=3 lines AND >=10% of the transcript), which also stops phantom
• languages from occasional STT mislabels from forcing a two-column view.
• The "Show translation" control is now available for EVERY meeting
• (previously hidden for monolingual transcripts), so a viewer can request
• any target language on demand; pair mode still requires a real second
• language.

bilingual language-flip fixes — block jump, merge race, chunk echoes

• Live bilingual test (EN/PT alternating) exposed three defects:
• 1. Cross-language block jump: SubtitlePanel merged live text into the last
• block by SPEAKER only, but confirmed blocks group per-language — so a
• speaker's PT live text streamed into their EN block, then tore out into
• a new block on final. Live-merge (and the standalone-live condition) now
• require the language to match too; a flipped-language live starts its
• own block. Live captions carry sourceLanguage on both clients now
• (pushSubtitle previously dropped it, so the panel couldn't tell).
• 2. Final-merge language race: the merge buffer's flush read
• detectedLangRef 150ms AFTER the final — if the next sentence was in the
• other language, the merged text was translated under the WRONG source
• language, or skipped entirely (tLang === newly-detected), leaving the
• translation column on "…" forever. The language is now captured
• per-final, and a mid-buffer flip flushes the previous language's buffer
• immediately instead of fusing two languages into one line.
• 3. Committed-chunk echoes: routing incremental chunks through the
• streaming path bypassed the untranslated-echo retry guard
• (wordOverlapRatio) — an English chunk echoed back as English PERSISTED
• in the display. Chunks are atomic/guarded again; only the volatile tail
• streams (errors there self-correct on the next interim or final).
• Diagnostics (asked for: "identify when the system tries to switch
• languages and why"): every detection flip now logs [lang-switch]
• {from, to, deepgramDetected, constraintMode} to the console and records a
• zero-cost lang-switch marker in the latency report — rendered as slate
• ticks along the timeline chart, counted in the copy-paste export,
• excluded from all latency percentiles.

DEEPL_API_URL host override + shorter auth-failure breaker hold

• Learned during the Free->Pro upgrade (2026-08-17): the key suffix is NOT a
• reliable host indicator across a plan change. The upgraded account's old
• ":fx" key was rejected by api-free.deepl.com with 403 "Wrong endpoint. Use
• https://api.deepl.com" while ALSO lacking translate permission on the Pro
• host — suffix-based routing would have silently sent every request to the
• wrong place. DEEPL_API_URL now pins the host explicitly when set; suffix
• inference remains the default.
• Breaker auth hold cut 30min -> 5min: a revoked/misscoped key costs one
• wasted probe per window per instance, but a long hold means a key you JUST
• fixed stays locked out — exactly wrong during a key rotation, which is when
• auth errors actually happen.
• Deployed alongside the rotated DEEPL_API_KEY (new Pro key verified with
• three consecutive 200s from api.deepl.com before rotation).

contact-identity safety + live-send compliance gates

• A business name is not a person's name. Discovery output is untrusted, so personalisation now requires a full human name WITH a cited source (or an explicit admin attestation). Stops "Rand Jeweler" being turned into a "Hi Rand" salutation.
• normalizeFullPersonName() requires first + last, unicode-letter parts only; citedContactIdentity() drops any identity the model supplied without a real http(s) source.
• Recipient must be a source-verified contact email.
• PROSPECTING_POSTAL_ADDRESS must contain a real address (CAN-SPAM footer was otherwise shipping the literal "[SET PROSPECTING_POSTAL_ADDRESS]").
• Live provider sending stays closed until PROSPECTING_PROVIDER_OUTREACH_APPROVED=true is set following provider / compliance sign-off, per Brevo's consent-based recipient policy.
• First-touch emails are validated send-ready before dispatch.
• Every send now writes a PROSPECTING_EVENT audit row; proposal status only flips to SENT when the attachment was actually allowed.

fix duplicate captions, whole-sentence pops, and the DeepL double-hop

• translate.ts: DeepL circuit breaker. 456/429 holds 10min, 401/403 holds 30min, 3 consecutive soft failures hold 30s; half-open on expiry, closes on first success. New tryDeepL() helper centralises the bookkeeping. Spend is now attributed to the provider that actually served (everything was booked as "deepl", including every OpenAI fallback).
• All 4 final sites now retire the live caption. New translationSinkRef routes onFinal()'s provisional tail translation to the confirmed line instead of the retired overlay, so the translation column fills immediately rather than sitting on "..." until the whole-sentence request returns.
• Streaming was dead code: stream:true was only sent on the NON-incremental path, but the incremental flag is on in prod. Wired streaming through the incremental path via a new optional onPartial on IncrementalTranslator's translate contract (backwards-compatible; 7 existing tests still pass).
• Both translate routes dropped `context` on the streaming branch, and the streaming DeepL call ignored it — incremental seams lost grammatical agreement whenever streaming engaged.
• Streaming path called lruSet() even for context-dependent results, which the non-streaming path correctly refuses to cache (cache poisoning).
• subtitle-panel.tsx had a JSX block spliced into its header doc-comment by a bad search-replace in an earlier session (committed in HEAD). It compiled, but it had swallowed the autoGrow render logic — the prop was accepted and silently ignored. Comment restored, logic reinstated.
• New src/lib/build-version.ts + next.config env block so each meeting's latency report records WHICH build produced it (commit SHA, branch, env, immutable deployment URL) — previously unanswerable outside Sentry.
• Copy button on the latency report emitting a paste-ready text report with build identity, per-kind and per-role breakdowns, and raw samples.
• Guests recorded no latency at all; samples now carry role: host|guest.
• Latency chart time axis floored to whole minutes, so every label read "0m" on any call under 2 minutes. Now m:ss.
• X-Translate-Provider / X-DeepL-Breaker response headers.

minutes transcript: original-always-left view + resilient translation loading

• The transcript view assumed exactly 2 languages, but Smart-Mode STT can
• emit 3-4 (mislabels + code-switching). langA/langB were picked by first-
• appearance and both columns loaded via Promise.all — ONE hard target
• failing (e.g. whole-transcript Hindi) blanked BOTH columns, leaving every
• line as an empty 'TRANSLATED …' placeholder forever.

strip speaker label from live translated text + in-place revision morph

• 1. Fix: translated live captions were pushed with 'SpeakerName: ' baked in
• (legacy of the old raw overlay), so the name appeared INSIDE the
• translation column under the separately-rendered speaker header. Strip
• it (strict match on the known speaker; promote a bare name prefix to
• the label when no source arrived yet).
• 2. Revision handling in TypewriterText is now an in-place morph: when the
• engine rewrites already-shown text (conjugation fix, word swap once
• 'certain'), the changed word repaints in a single frame while the
• reveal frontier stays put — no more visible erase-and-retype.

smooth adaptive typewriter for BOTH columns + live streaming text

• Two perceived-latency fixes in the subtitle panel (shared by host + guest):
• 1. TypewriterText rewritten as an adaptive smooth-reveal: text never pops
• in as a whole phrase — even a big translation-chunk burst flows out
• char-by-char. Baseline ~55 cps, accelerating so any backlog drains in
• ≤450ms (smooth but never lagging speech). Revisions rewind only to the
• common prefix; the old 'jump to full text after 2s' pop is gone —
• instant render now only applies to lines that MOUNT old (scroll-back).
• 2. The live/interim section now renders the actual streaming source text
• AND the incremental translation chunks (previously it only showed
• '…is speaking' dots, making all the Task-4/5 incremental work invisible
• until the final landed as one phrase-pop). Two-column live layout with
• pulsing dots inline in the speaker label, '…' placeholder anchoring the
• eye where translation will appear, and pulsing caret at the reveal
• frontier of both columns.

latency Task 4 — incremental interim translation (flagged)

• REDUCES cost. Deepgram interims grow ('I' → 'I think' → 'I think we
• should…') and we previously re-translated the WHOLE growing string on
• every interim — O(n²) total characters per sentence sent to DeepL/OpenAI.
• New src/lib/meetings/incremental-translate.ts (pure logic, client-safe,
• 5 unit tests): words stable across ≥2 consecutive interims are committed
• as chunks and translated exactly once; only the short changing tail
• retranslates per interim (O(n) total). Chunk translations are serialized
• so they always land in source order; state resets automatically when
• Deepgram revises already-committed words, and explicitly on every final —
• finals still get one clean whole-sentence translation via the existing
• path, so interim chunk seams never persist.
• Seam quality: preceding committed source words are passed as context —
• DeepL's native 'context' parameter (influences translation, not billed as
• output) and OpenAI's prompt. translateSegment/callDeepL/callOpenAI gain an
• optional context arg; context-dependent results skip ALL caches (LRU, DB,
• client phrase cache) since the same words translate differently in
• different contexts.
• Feature-flagged: NEXT_PUBLIC_MEETINGS_INCREMENTAL_TRANSLATE=1 (off by
• default in prod until tested). Wired into both host and guest room
• clients with per-(speaker,src,tgt) instances.

latency Task 5 — streaming OpenAI fallback for live subtitles

• translate.ts: new translateSegmentStreaming() — same decision tree as translateSegment (LRU hit / same-lang / DeepL success return atomic 'complete' outcomes; only the OpenAI fallback streams). Parses OpenAI SSE frames, forwards raw token text, LRU-caches + records spend on finish.
• Both translate routes: opt-in body.stream flag; streaming responses are marked X-Translate-Streaming: 1, everything else stays JSON (all existing callers unaffected).
• Both room clients: interim path sends stream:true, reads the body progressively, updates the caption per chunk, respects the interimTranslateTokenRef race (cancels the reader mid-stream when a newer interim wins), and feeds the Task-6 phrase cache on completion.
• Finals keep the non-streaming path (with its wordOverlapRatio retry guard) — only interims stream.

latency Tasks 2/3/6 — pre-warm, region pinning, client phrase cache

• Task 2 (pre-warm): both room clients fire one throwaway translate request
• at mount so the Vercel function + DNS/TLS to DeepL/OpenAI are already warm
• and pooled before the first real utterance — kills the visible first-
• translation cold-start tax.
• Task 3 (region pinning): both translate routes pinned preferredRegion=iad1.
• Participants are predominantly US-East + Brazil; iad1 minimizes the
• browser→function leg for both (GRU→iad1 ≈115ms is Brazil's best US region,
• US-East ≈15ms) and sits near OpenAI US inference. Realtime path skips the
• us-west-2 Supabase reads entirely so DB distance is irrelevant here.
• Task 6 (client-side hot-phrase cache): repeated short utterances (≤60
• chars — 'yes', 'okay', greetings) now resolve from an in-tab LRU Map
• (≤500 entries) with ZERO network round-trip after first occurrence.
• Negative cost delta: fewer DeepL/OpenAI calls.
• Tasks 1 (persistent channel), 4 (incremental translation), 5 (streaming
• fallback) remain per docs/meetings/LATENCY-ABSOLUTE-LIMIT-HANDOFF.md.

latency 'absolute limit' handoff brief for Fable 5

• Six-task implementation plan to drive mic-to-translated-subtitle latency
• toward the ~300-500ms physics-adjacent floor at zero/negative cost delta:
• persistent translation channel (LiveKit data channel), pre-warm on room
• mount, region pinning audit, incremental/speculative translation (cuts
• DeepL/OpenAI spend), streaming OpenAI fallback (cost-neutral), and a
• client-side hot-phrase cache. Includes current provider status (DeepL 456
• quota block, new MEETINGS_OPENAI_API_KEY tier limits), cost constraints,
• implementation order, and regression guardrails.

reject unsourced 'estimated' evidence claims outright

• Rand Jeweler regression case: a label:"estimated" claim ("No Google
• reviews found") with NO source URL and NO captured text still passed
• as "supported" because the local-category check accepted a bare
• structuredMetrics object ({rating: 0, reviewCount: 0}) as sufficient
• proof on its own — even though that object was itself just the
• analyst's fabricated placeholder, not an observation. This let a pure
• guess survive to generation, where it then got flagged as contradicted
• by a real, well-sourced Birdeye capture showing Google actually has 51
• reviews for the same prospect.
• Fix: validateEvidenceDeterministic() now rejects any "estimated"-label
• claim with neither a sourceUrl nor captured excerpt as "unsupported",
• before category-specific checks run. generate.ts also excludes anything
• already deterministically unsupported/stale (not just superseded
• duplicates) from what's shown to the LLM validator, since it can only
• ever downgrade already-rejected evidence, never rescue it -- so there's
• no reason to let it re-flag them as noisy 'contradictions'.
• Deleted the specific fabricated Rand Jeweler evidence row and added a
• regression test. All 66 prospecting tests pass.

auto-supersede stale duplicate evidence to stop false contradictions

• Previously, re-scraping a source (e.g. re-checking a prospect's Instagram
• cadence) left the OLD bare-assertion evidence row in place alongside the
• NEW, better-sourced one. Both rows fed into evidence validation, and the
• LLM auditor correctly (but unhelpfully) flagged them as contradicting each
• other every time the prospect was regenerated — even though generation
• still succeeded, this produced a scary, noisy 'Contradictions detected'
• toast on every retry (Koorosh Fine Jewelry case).
• Fix: validateEvidenceDeterministic() now groups evidence by
• (category, normalised source URL) and auto-marks all but the
• most-recently-observed row in each group as 'stale'/superseded, with a
• clear note, before any other checks run. generate.ts also excludes
• superseded rows from what's shown to the LLM validator, so it never even
• sees the old duplicate to flag it. Deleted the specific stale Koorosh
• evidence row and added a regression test mirroring this exact case.

fix stuck whole-transcript translation when DeepL quota exhausted

• Batch the OpenAI fallback (callOpenAIBatch/translateManyViaOpenAI) instead of one HTTP call per transcript line — a 1000+ line transcript translated one-string-at-a-time (even at concurrency 3) blew past the minutes-translation route's 180s timeout whenever DeepL was unavailable, leaving the client stuck on 'TRANSLATED ...' placeholders forever.
• Extract detectLanguages/classifyLanguageOpenAI into new src/lib/meetings/language-detect.ts (no @/lib/prisma import) so standalone scripts run via tsx can use them without tripping the server-only guard. translate.ts now re-exports from it.
• Add scripts/repair-transcript-languages.ts: one-off/reusable tool to re-detect and correct mislabeled MeetingTranscript.sourceLanguage rows (Smart Mode's old snap-back bug mislabeled some genuinely-English lines as Portuguese) and invalidate stale MeetingMinutesTranslation cache rows so they regenerate with corrected labels.
• Ran the repair tool against mtg_1786734013755_vvfuzyyz (Felipe's meeting), fixing 20 mislabeled rows so far; a few more remain pending OpenAI/DeepL rate-limit windows resetting (safe to re-run, idempotent).

animated join loading screen, subtitle overlay shifts for CAD viewer, invoice checkout width fix, stale minutes-translation cache repair

• Fixed the persisted minutes-translation cache never invalidating when the underlying transcript grows after the first translation run — a cached row with fewer transcript entries than the meeting now has was served as-is, leaving every line added since untranslated ('…' in the UI). GET and POST both now compare cached transcript length against the live final-transcript count and treat a mismatch as stale, forcing a fresh translate + overwrite instead of trusting the short array.
• Animated the join/token-generation loading screen (host room).
• Subtitle overlay now shifts left when a 3D CAD viewer / diamond / invoice tile is open, instead of relying solely on pointer-events-none.
• Widened the invoice tile during Stripe embedded checkout so it isn't cramped into the ~38% side-column width that was causing a blank/white screen for guests after clicking pay.

rose gold shininess fix (metalness 1.0) + subtitle overlay no longer blocks tile buttons

• Rose gold metalness was 0.92 vs yellow gold's 1.0 — the lower metalness let diffuse/base-color leak through the PBR shading model, making it read as plastic instead of a shiny polished metal. Bumped to 1.0 to match yellow gold's shininess.
• Subtitle overlay wrapper (both host + guest rooms) is now pointer-events-none with pointer-events-auto only on its resize handle, so the floating subtitle box never blocks clicks on tiles underneath it (e.g. the 3D CAD viewer's close button when the subtitle box floats over the top-right corner).

fix host own-speech translation regression (Smart/Auto/Manual STT constraint modes)

• Host's own speech translated column stopped populating because it was gated solely on translationLangRef (viewer's own reading-translation choice), which is often unset. Added remoteSpokenLangRef fallback that tracks the other participant's spoken language from broadcast captions and uses it as the translation target for MY OWN speech too.
• Restored an unrestricted 'Auto' detection mode (no meeting-language restriction) alongside the existing constrained 'Smart' default.
• Added 'Manual' mode to pick 2+ specific languages to constrain Deepgram's per-utterance detection to.
• None of this was previously committed/deployed, which is why production was still exhibiting the pre-fix behaviour.

remove accidentally-committed temp diagnostic script

bigger transcript PDF typography, translation-quality heuristic, CAD storageKey sync

• Bumped body/text font sizes further (13.5px -> 15.5px body text, 14px -> 16px base, headers/letterhead scaled up to match) and widened the page (760px -> 820px, 24mm->20mm/16mm margins) -- previous pass wasn't enough, still read as small/cramped at real print scale.
• New wordOverlapRatio() heuristic in translate.ts: DeepL/OpenAI can return success but silently pass rambling/disfluent transcript lines through nearly unchanged instead of translating them. Every DeepL batch result and every OpenAI call is now checked for suspicious similarity to the source (>55% word overlap, source != target); DeepL results that look untranslated get a second-opinion OpenAI pass, OpenAI itself gets one retry with a more forceful prompt before giving up.
• transcript-view.tsx looksUntranslated() upgraded from exact-match to the same fuzzy word-overlap check; client-side auto-retranslate trip-wire lowered from 60% exact-duplicate to 25% fuzzy-untranslated.
• rename route now keeps CadFile.storageKey/folder/fileName in sync with Supabase Storage after a move (previously only the storage object moved, breaking DB-driven consumers like the in-meeting 3D display tray)
• scripts/fix-cad-storage-keys.ts: repair script for files broken by past moves; ran with --apply, fixed the Rosser/Stella engagement ring row

fix duplicate-language display, speaker color collisions, and PDF typography

• Speaker colors now assigned by order of first appearance (guaranteed distinct for <=10 speakers) instead of hash-based lookup, which was colliding two different speakers onto the same color
• New looksUntranslated() helper: when a 'translated' column is actually identical text to the original (already same language, or STT mislabeled source), stop rendering it as a highlighted duplicate translation box in both the live view and the transcript PDF export -- this fixed Felipe's English lines showing up highlighted-English twice
• Both PDF exports (minutes summary + transcript) now render inside a page-like white card with generous margins/padding on a neutral background, bumped base font sizes and line-height throughout for a properly 'typed document' look instead of edge-to-edge tiny text

real letterhead PDF export for minutes summary + tidy transcript PDF

• New GET /api/admin/meetings/[id]/minutes-pdf: dedicated print-ready document (Katura letterhead, Summary, Action Items table, Decisions, Open Questions, Risks, Follow-ups, Participants) covering everything above the transcript section, auto-triggers window.print()
• minutes-actions.tsx 'Download PDF' now opens this endpoint in a new tab instead of window.print()-ing the raw app page (which printed sidebar/nav/buttons and looked like a screenshot)
• transcript-view.tsx printPdf: tightened spacing, alternating row shading, ellipsis-truncated speaker labels, em-dash placeholder for empty translation cells, print-color-adjust for zebra striping

fix transcript translation timeout causing 'Retranslate' SyntaxError

• translateBatch OpenAI fallback now uses bounded concurrency (mapWithConcurrency, limit=6) instead of sequential per-line calls, cutting wall-clock time drastically on long transcripts
• bump minutes-translation route maxDuration 60 -> 180 for safety margin
• transcript-view.tsx: fetchTranslation now safely parses responses (reads as text first, handles non-JSON error bodies like timeout/HTML pages) instead of blindly calling res.json(), which threw an opaque SyntaxError
• surface actual error message in the retranslate failure toast when short enough

redesign transcript PDF/txt/copy export to match on-screen grouped 2-column transcript view with letterhead, color-coded speakers, green translation highlight

fix DeepL-outage translation fallback (root cause of duplicate/untranslated transcript columns), auto-repair stale cached translations, green highlight, color-coded + truncated speaker names

fix transcript/translation index misalignment + fixed 2-col language layout; Prospecting: self-healing evidence enrichment retry

session log for 2026-08-14 meetings work + prospecting evidence-validation triage

group transcript by speaker turn, default 2-column translation, Smart Mode STT (constrained language auto-detect)

subtitles borderless-by-default; add missing prospecting SQL migration

• SubtitlePanel: new 'plain' prop renders text with just a subtle blur behind it (no visible box/border/tint). Wired into every subtitle render site in both the host and guest rooms, and flipped the default to plain=true so the boxed container is now OFF by default per request.
• supabase/migrations/20260814_jeweler_prospecting_v3_sync.sql: the original 20260805 migration only had V1 columns. The Prisma schema has since grown V2/V3 columns (contactEmailVerification, closeProbability, nextStep, validationStatus, proposalV2, messageType, etc.) plus 3 whole tables (ProspectNote, ProspectingAiUsage, ProspectingEvent) that were never applied to the DB -- this is exactly why /admin/saas/prospecting's GET /api/admin/prospecting/prospects (which selects evidence.validationStatus and the meetingNotes/ProspectNote relation) was failing with 'did you run the SQL migration?'. New migration is additive/idempotent (IF NOT EXISTS) and safe to run on top of the existing V1 tables.

fix translation silently stopping + subtitle defaults

• translate.ts: add OpenAI gpt-4o-mini fallback when DeepL is missing or fails (quota/expired key/network), so live translation no longer silently shows the untranslated source language.
• meeting-room-client + guest-room-client: persist the viewer's per-meeting translation-language choice in localStorage so leaving+rejoining or flipping the Classic/Railgun engine toggle no longer resets it to the browser locale (which disabled translation for a pt-BR-browser viewer in a pt meeting).
• Subtitle defaults: XL font on desktop; overlay box 50% wider (640->960px).