Backend-Lag-Fix Phase 1 — entlastet die CPU-Dauerschleife im Mail-Stack:
- delete mail-scan-cron.ts: der 30-Min-Nitro-Cron scannte alle User parallel
(Promise.allSettled) und war redundant zum IMAP-IDLE-Daemon (Single Source
of Truth). Reine Dauerlast ohne Mehrwert.
- imap-idle: In-Flight-Guard (scanInFlight + coalescePending). triggerScan ist
jetzt re-entry-safe — pro Connection max. 1 aktiver + 1 pending Scan statt
bis zu 8 gestapelt pro 2-Min-NOOP-Tick. Gilt für NOOP + exists-Event.
- plan-features: Pro mailAgents 3->2 (+ Math.min-Hack in coach/message aufgeräumt).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
SOS (urge.tsx) uses /api/coach/message as a stateless LLM proxy for game
comments, share drafts and the stream fallback — sending SOS_BOOT +
[INTERN:] prompts. The endpoint persisted the full messages array into
coachSession for pro/legend users, so those internal prompts and the raw
JSON replies leaked into the Coach chat history as visible bubbles.
- Reactivate the sosMode flag (already sent by all three SOS call-sites):
when set, the endpoint skips coachSession persistence, memory extraction
and feedback detection — pure LLM proxy, no shared state.
- Add a defensive filter on /api/coach/history that strips internal
messages (SOS_BOOT, [INTERN:], [SYSTEM-HINT], raw JSON / [[CHIPS]]
replies) so already-contaminated sessions self-heal on next load.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
LLM-Prompt (message.post + sos-stream):
- LANG_INSTRUCTIONS Map raus, ersetzt durch dynamische Instruktion
'Reply in {detectedFromUser} ... fallback: {appLang}'
- Lyra matcht jetzt die Sprache der letzten User-Message (per
detectLang Unicode-Detection); App-Locale ist nur noch Fallback
- Instruktion doppelt eingehängt (Anfang + Ende des System-Prompts)
gegen recency bias bei langen deutschen Prompts
TTS (speak dispatcher + speak-cartesia + speak-elevenlabs):
- Kein 'de'-Default mehr für language. detectLang(text, locale) leitet
Sprache primär aus dem Antwort-Text ab (Arabic/Cyrillic/CJK/Turkish-
Letters), Locale als Fallback
- Cartesia + ElevenLabs: language/language_code nur senden wenn
ableitbar, sonst Provider auto-detect statt erzwungenem 'de'
- speak-cartesia: sonic-2 → sonic-3 (Multi-Lang, war beim Dispatcher-
Fix gestern vergessen worden)
- Google: en-US neutraler Fallback statt de-DE-Bias
Neu: server/utils/detect-lang.ts
Custom-Domain-Slots sind jetzt EIN gemeinsamer Pool für web + mail
(Pro 10 / Legend 20) statt getrennter web/mail-Buckets. Free-Tier ist
entfallen — PLAN_LIMITS hat nur noch pro + legend, getPlanLimits
defaultet auf pro.
Backend:
- plan-features: customDomains ist eine Zahl (CustomDomainLimits weg)
- index.post: Slot-Check gegen Gesamt-Count, Fehler einheitlich LIMIT_REACHED
- index.get: liefert { items, count, limit }
- change-preview + coach/message an die neue Form angepasst
Frontend:
- useCustomDomains: count/limit (Zahlen) statt countsByType/limits
- AddDomainSheet: ein generischer Limit-Hinweis (error_limit_reached)
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
plan-features.customDomains is now { web, mail } per plan instead of a
single number. Free 5+5, Pro 5+5, Legend 10+10 — the user explicitly
chose separate pools so users don't have to trade a website slot for a
mail-pattern slot or vice versa.
- countActiveCustomDomainsSplit(userId) groupBy type → { web, mail }
(mail aggregates mail_domain + mail_display_name). Old single-count
function stays as a deprecated alias for any caller still on it.
- POST /api/custom-domains: body-compat accepts both { pattern, kind }
(current frontend) and { domain, type } (legacy / direct). kind='mail'
is split into mail_domain vs mail_display_name server-side based on
whether the pattern looks like a domain. Slot check is per-bucket;
errors are WEB_LIMIT_REACHED / MAIL_LIMIT_REACHED so the UI can show
the right limit-reached message per tab.
- GET /api/custom-domains: response shape extended to
{ items, counts: { web, mail }, limits: { web, mail } } so the
frontend can drive the per-tab counter without client-side estimation.
- POST /api/custom-domains/:id/submit: hard-blocks mail_display_name
with 400 DISPLAY_NAME_NOT_SUBMITTABLE. Display-name submission to the
global blocklist is deferred to v1.1 — would require a schema split
on BlocklistDomain that's risky pre-TestFlight. mail_domain still
flows through the community-vote pipeline like web entries.
- auth/me.get.ts, plan/change-preview.get.ts, coach/message.post.ts
updated for the new shape (Lyra prompts untouched, only template
variables split web vs mail counts).
24 vitest cases in backend/tests/custom-domains/plan-limits.test.ts
cover the new shape, body compat, bucket logic, and the submit guard;
216/216 total backend tests pass.
LLMs (especially Haiku) keep emitting markdown despite explicit "no markdown"
prompt rule. Mobile app has no markdown renderer — users see raw asterisks.
- New stripMarkdown() util handles **bold**, bullet-lists, headings,
code-fences, links, blockquotes
- /api/coach/message: applies stripMarkdown(text) post-LLM as safety-net
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Per strategist-spec: Lyra-Coach-Mode klarer von SOS-Mode trennen.
- SOS-Mode (urge): crisis-intervention, focused, kurz
- Coach-Mode (lyra): casual, profile-building, philosophy, features
Backend (backend/server/api/coach/message.post.ts):
- COACH_CASUAL_SYSTEM_PROMPT komplett neu strukturiert (~620 tokens)
- Stärkerer Fokus: 3 explicit Aufträge (echtes Gespräch / Profile-Building /
Rebreak sprechen)
- Profile-building-mandate: "wenn du wenig weißt, sag's ehrlich; frag nach
Hobbies/Zielen/Menschen — eingewoben, NICHT als Checkliste"
- Cleanere Mission-Section: Bewegung, Anonymität, kein-pathologisieren,
community-getrieben, DiGA-Listung-Ziel
- Hard-rules klarer: NIE demographics extrahieren (User-Form ist tabu),
kein Sucht-Vokabular, kein medical-advice
- Existing PLAN_DETAILS-template-var bleibt
- Memory-system unverändert (lyra-memories table, extractAndStoreMemories
fire-and-forget — kein schema-change nötig)
Frontend Mode-Badges:
- app/lyra.tsx (Coach-Mode): Header-pill "Coach" in brandOrange-tint neben
Lyra-name
- app/urge.tsx (SOS-Mode): Header-pill "SOS" in error/red-tint neben
Lyra-name (alt: "Lyra · SOS [v2]" inline-text → cleaner badge-style)
i18n:
- coach.modeBadge.coach + coach.modeBadge.sos in DE + EN
Switch-Logic: route-based (lyra.tsx vs urge.tsx → separate persona via
backend endpoint). Kein User-Toggle — User soll nicht entscheiden müssen
"bin ich grade in Krise?".
Implementation Risk: LOW — schema-neutral, prompt-only + 2 small UI badges.
Erste Beta-Testing-Phase: ~1-2 Wochen iterieren bei Feedback.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Coach-Page ist NICHT SOS — User ist nicht in Krise, will small-talk,
Reflexion, Feature-Wuensche, Philosophie. „Lockere Lyra".
Aenderungen:
- Neuer Prompt COACH_CASUAL_SYSTEM_PROMPT (exportiert): warm/neugierig/
manchmal humorvoll, bis 4-5 Saetze, darf eigene Empfehlungen + Mini-
Meinungen formulieren, lädt Feedback aktiv ein. Kein Crisis-Framing.
Sprachregeln (keine Pathologisierung, kein „Sucht") gelten unveraendert.
- Tier-LLM-Routing analog zu sos-stream:
Free/Pro = Groq llama-3.3-70b-versatile (Fallback llama-3.1-8b)
Legend = OpenRouter anthropic/claude-haiku-4.5 (Fallback claude-3.5-haiku)
- max_tokens 280→500 (Coach darf laenger antworten)
- Demographics-Injection (analog sos-stream): birthYear/gender/etc als
USER-DEMOGRAPHIE-Block in Prompt (read-only, kein Extract)
- sosMode-Branch deprecated — Frontend kann den Param noch senden, wird
ignoriert. Folge-TODO: UI-Agent entfernt sosMode aus Coach-Call.
NICHT geändert:
- TTS-Endpoints bleiben plan-agnostisch (Frontend routet by tier)
- sos-stream.get.ts/sos-stream.post.ts unberuehrt (importieren weiter
COACH_SYSTEM_PROMPT, kein Breaking Change)
Memory:
- project_llm_per_plan.md (tier-LLM-Default-Logic)
- feedback_anonymity_nickname.md
- feedback_demographics_user_initiated.md
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
