Insert-beat is a pure in-scene micro-interaction — adding choices that
lead to change-scene contradicted its purpose. Now insert-beat generates
1-3 richer beats then loops back to the original options, which is the
natural UX for "you glanced at something decorative."
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
User feedback: custom interactions rarely produce new story content because
the classifier heavily biased toward insert-beat (single reaction, no scene
change). Three changes to fix this:
1. Freeform text input now always triggers a full scene generation (skips
the classify step entirely) — users who type expect the story to advance.
2. Vision (background click) classifier de-biased: prompt now favors
change-scene when uncertain, and the code fallback flipped from
insert-beat to change-scene. insert-beat narrowed to pure observation.
3. Insert-beat enhanced: generates 1-3 beats (was 1) with follow-up
choices (was: loop back to original beat). Even when vision classifies
as insert-beat, the player gets richer content and new options.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Squash-merge the cloudflare-migration branch (7 commits by Kai ki) into
staging with conflict resolution, feature integration, and bug fixes.
Engine:
- Paradigm D: single-stream Writer replacing dual-phase Plan/Beats
- Delete Architect agent; story bible generated via Writer <plan> tag
- Modular prompt architecture (segments/registry/builder)
- StreamRouter for tagged stream splitting (<plan>/<story>/<choices>)
Infrastructure:
- Cloudflare Workers deployment (wrangler.jsonc, OpenNext adapter)
- D1 database schema + Drizzle ORM (scaffolded, not yet active)
- R2 storage helpers (scaffolded, not yet active)
- Story persistence API routes + client-side persistence
BYOK (Bring Your Own Key):
- /api/llm/user-proxy with SSRF-protected LLM proxy (+ requireUser auth)
- CORS-aware fetch in ai-client: auto-detect CORS failure, fallback to
server proxy transparently via OpenAI SDK custom fetch
- BYO config support added to classify-freeform and vision routes
- SettingsModal CORS privacy notice (keys never logged/stored)
SSE streaming:
- engineClient.ts: fetchSSE helper for progressive scene events
- startSession/requestScene accept optional emit callback
- Fix SSE error event field name (error → message) in scene/start routes
i18n integration:
- Wire buildLanguageDirective into paradigm D's prompt builder
- Update corsNotice i18n keys (zh-CN/en/ja) with CORS proxy privacy text
- Preserve Session.language + LanguageSwitcher from i18n commit
Co-authored-by: Kai ki <155355644+zbf1009@users.noreply.github.com>
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Two follow-ups from pr-agent review of #79:
1. director.ts voicePromises built a Character WITHOUT stepfunVoiceId, so
on a StepFun server the client (which omits the voice payload to save
FOT) echoed back only voiceDescription — and the server re-scored via
pickStepfunVoiceId every beat instead of honoring the LLM pick. The
whole "CharacterDesigner picks a preset id" mechanism was effectively
bypassed on live StepFun sessions (it only worked for prebaked cards,
which carry stepfunVoiceId in their JSON). Persist stepfunVoiceId onto
the Character so the client→server round-trip keeps the LLM selection.
2. fetchBeatAudio's null-provider branch (probe pending) required
speaker.voice and silently dropped a stepfun-only speaker. Accept any
synthesizable source (voice | stepfunVoiceId | voiceDescription) so a
slow getTtsProvider probe can't drop audio during the first scene's
fetch window. The server resolveVoice normalizes regardless of which
fields arrive.
Make homepage cards and live sessions produce sound when the server is
configured for StepFun TTS, instead of silently failing (the prebaked
Xiaomi voice was useless on a StepFun server, and wasted ~220KB/beat in
Fast Origin Transfer).
Three coordinated changes:
1. CharacterDesigner now picks a StepFun preset voice id directly from the
32-entry catalog in the SAME LLM call that designs the character — zero
extra latency, LLM-grade match quality. The Xiaomi prompt path is
byte-identical to history (verified programmatically) so cache hit rate
and voice quality are preserved. pickStepfunVoiceId (keyword scorer)
remains the fallback for orphan speakers / invalid LLM picks.
2. The 32-preset catalog moves to lib/tts-client/stepfun-voices.json as the
single source of truth, shared by the scorer, the CharacterDesigner
prompt, /api/tts-provider, and the offline enrich script.
3. A new GET /api/tts-provider endpoint lets the client probe the server's
TTS provider at /play mount. fetchBeatAudio then shapes its request body:
on a StepFun server it sends the lightweight stepfunVoiceId /
voiceDescription and omits the ~220KB Xiaomi reference audio (FOT saving
~13MB per protagonist per session on prebaked cards). requestBeatAudio
re-provisions on a provider mismatch before synth, so audio never goes
silent on a cross-provider replay or mid-session provider flip.
New type fields are all optional and backward-compatible: Character.stepfunVoiceId,
BeatAudioRequest.voiceDescription/characterName/stepfunVoiceId, voice made
optional. AGENTS.md updated for the new route, type fields, dependency map,
and StepFun voice-selection flow.
Eliminate the dual code path (raw fetch vs AI SDK) for text and vision.
All providers now go through createLanguageModel() + generateText(),
removing chatOpenAiCompatible/analyzeOpenAiCompatible, the manual Usage
type, summarizeUsage, and responseFormat plumbing from 8 call sites.
Key fix: @ai-sdk/openai v3 defaults to the Responses API (/responses);
DeepSeek only supports Chat Completions, so we use .chat() explicitly.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Thread orientation (portrait|landscape) from client through API, engine,
and image gen. Portrait devices render 1024x1792 (9:16) full-bleed scenes;
desktop/landscape keeps 1792x1024 (16:9). Adds cover-aware click→image
coordinate mapping, session-locked orientation, a shared coerceOrientation
helper, and a choices overflow cap in portrait.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
The Writer was the serial long pole: a single LLM call wrote the scene
skeleton AND the full beats[] graph before anything downstream could
start, so variable-length beat generation blew up tail latency.
Split it into two calls:
- Phase A (runWriterPlan): minimal skeleton the image pipeline needs
(sceneSummary, sceneKey, entryBeatId, cast, entry roster, entry speaker).
Serial, on the critical path, kept lightweight.
- Phase B (runWriterBeats): full beats[] + storyStatePatch, written to
honor the plan. Launched immediately, overlaps the ENTIRE image pipeline
(cards / cinematographer / portraits / painter), awaited last.
Critical path becomes PhaseA + max(imagePipeline, PhaseB), so the long
beat-writing is hidden behind image gen. A Phase B failure degrades to a
single playable beat synthesized from the plan.
Paired distinct-payload A/B (6 content-matched stories, baseline vs split):
- median end-to-end 42.6s -> 32.2s (-24%)
- mean 46.4s -> 33.1s (-29%)
- worst case 74.7s -> 37.6s (halved)
- no content regression: total Writer output tokens 12858 -> 13699
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Add a `tag` option to chat() and have it print one `[cache] <tag>
hit=X miss=Y rate=Z%` line per call. Three Usage-shape variants are
probed in order so the same logger works across providers:
- DeepSeek (v3+): usage.prompt_cache_hit_tokens / *_miss_tokens
- OpenAI / o-series: usage.prompt_tokens_details.cached_tokens
- Anthropic: usage.cache_read_input_tokens / *_creation_*
When none of them are present (MiMo / local Ollama / others) we still
print prompt + completion totals so the cost baseline is visible.
Tag every callsite so the log is greppable:
architect / writer / character-designer / cinematographer / insert-beat
This is the prerequisite for the prefix-cache reordering work that
follows — without per-agent visibility there's no way to tell if a
prompt rearrangement actually moved the needle.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Flatten the pnpm monorepo (apps/web + packages/*) into a single web package at the repo root.
- Move app/lib/components/scripts/public to root; drop apps/web and packages/* wrappers
- Rewrite tsconfig paths (@infiplot/*) to ./lib/*; turbopack.root = __dirname
- Update Vercel (no root-directory) and Cloudflare (pnpm build:cf at root) deploy paths
- Regenerate pnpm-lock.yaml to drop stale workspace importers
- Bump engines.node to >=22 to match wrangler
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
yuanzonghao
Zonghao Yuan
DESKTOP-I1T6TF3\Q