Project

Changelog

Release notes, versions, and updates.

Edit this page on GitHub

Changelog

All notable changes to Marvox are documented here.

Format: Keep a Changelog · Versioning: Semantic Versioning

[1.5.0] — Production Studio, Casting Intelligence & External APIs (June 2026)

Added

Robotics Persona API — Commercial surface that lets external hardware, robots, and agents drive a Marvox character outside story context. POST /api/v2/personas/{character_id}/execute, persona profile and behavior-policy endpoints, and GET /api/v2/personas/group/{group_id}/consensus-state for fleet synchronization. Authenticated with an X-Marvox-API-Key header.
Character App Store — Instantiate personas from pre-built templates via /api/v2/templates and POST /api/v2/personas/from-template/{template_id}.
Transcription (speech-to-text) — POST /api/speech-to-text/projects/{project_id}/transcribe and a standalone POST /api/speech-to-text/transcribe for voice-driven prompts and dictation.
Voice Recipes — Save, update, share, and reuse a voice configuration across characters. CRUD under /api/characteros/projects/{project_id}/voice-recipes, plus public share-token resolution.
Production-grade exports — 19 export formats via /api/projects/{project_id}/export/{export_type}, each with an export audit log: manuscript/text (PDF, Word .docx, EPUB, Final Draft .fdx, Fountain screenplay, Markdown, RTF), audio (MP3, WAV, FLAC), structured data (character JSON, VoiceDNA JSON, canon-graph CSV, dialogue CSV, dialogue XLSX, HTML), and interchange (XLIFF 1.2, SRT subtitles, Yarn game dialogue). Audio, polished-manuscript, and structured-character formats are Pro-gated; lightweight handoff formats are Free.
Casting intelligence — Gender- and age-aware voice casting for every speaker, a per-book casting summary report, and an auto-produce path for full-book audiobook synthesis.

Changed

Performance Director — Audio synthesis now applies a scene-level dramatic pace envelope and emotion-reactive emphasis with micro-pauses for more cinematic delivery.
Agentic voice-casting — A learning loop continuously refines voice-casting decisions from production outcomes, layered into the live production pipeline alongside health monitoring.
Manuscript Studio — New Studio tab with refined speaker-attribution for dialogue-heavy manuscripts.
Billing — Subscription tiers consolidated to Free, Pro, Studio, and Enterprise, with checkout and lifecycle billing handled in-app via Stripe.
Vector storage — RAG embeddings migrated to a native PostgreSQL vector(1536) column.

[1.4.0] — Phase 4: API Key Management & Character Reflection (2026-03-19)

Added

API Key Management — Programmatic access to Marvox without user credentials.
- POST /api/billing/api-keys — Create long-lived API keys (SHA256-hashed, one-time display).
- GET /api/billing/api-keys — List keys with prefix, last-used audit trail.
- DELETE /api/billing/api-keys/{key_id} — Soft-revoke a key immediately.
- Keys use Authorization: Bearer mrvx_... — identical to JWT bearer flow.
- Schema: migration 0011 — api_keys table with indexes on project_id, key_hash, is_active.
Character Reflection — Automatic memory-driven personality evolution.
- POST /api/characteros/projects/{id}/characters/{char_id}/reflect — Manual trigger.
- Nightly automatic reflections at 02:00 UTC via background scheduler.
- Produces: reflection text, emotional arc trend (improving / stable / declining), saved memory ID.
- Error codes: REFLECT_FAILED, REFLECT_NO_MEMORIES.
Inference Cost Tracking — OPENAI_API_USAGE_WARNING_THRESHOLD env var alerts when costs exceed the configured percentage (default: 80%).

Changed

DEPLOYMENT.md — Updated required Railway env vars to include Phase 4 additions.

[1.3.0] — Phase 3: Collaboration & Character Participants (2026-02-20)

Added

CharacterParticipantManager — Characters react to SCENE_COMPLETE events in real-time via WebSocket broadcasts (event type: CHARACTER_REACTION).
Reactions are generated in parallel (semaphore = 3) using CharacterAgent.chat(mode=WRITER_ROOM).
Memory persistence for character reactions via MemoryBridge (non-blocking — failures do not block reactions).
EmotionalArcEvolverAgent — Evolves character personality from accumulated reactions.
ConsensusAnalyzerAgent — Detects conflicting character portrayals across multiple writers.
ReactionAudioStreamAgent — Live audio streaming for character reactions in collaboration sessions.

Fixed

Circular import in collab_routes.py — now uses lazy initialization for CharacterParticipantManager.

[1.2.0] — Phase 2: Audio Pipeline (2026-02-01)

Added

8-step audio production pipeline (AudioPipelineOrchestrator):
1. Parse scene dialogue blocks.
2. Resolve character voices via VoiceSelectionAgent.
3. Configure per-character TTS params via VoiceConfigurationAgent (300+ configs; speed 0.25×–4.0×).
4. Per-line emotion detection (AudioSceneAgent).
5. TTS generation (parallel with asyncio.Semaphore(5) to prevent OOM).
6. Audio composition and mixing via pydub.
7. 5-layer audio QA via AudioContinuityAgent.
8. Store to Vercel Blob; fall back to local file storage in dev.
VoiceCloningAgent — Neural voice DNA cloning.
DNALearningEngine — Evolves voice DNA when overall quality score ≥ 85 %.
VoiceVault — IP-locks Voice DNA to project/org boundaries (signed binding).
AudiobookOrchestrator — Full audiobook production from project canon.

[1.1.0] — Phase 1: CharacterOS Agent Network (2026-01-25)

Added

19-agent CharacterOS network orchestrated by AgentRuntime singleton.
Content generation agents: CharacterAgent, WriterAgent, ReaderAgent, NarratorAgent.
Quality agents: ContinuityAgent (5-layer validation), DialogueQualityAgent, EmotionalBeatAnalyzer, ScenePacingOptimizer.
Supporting agents: DirectorAgent, SummarizationAgent, AtmosphereAgent, QualityAnalysisAgent.
3-pass scene generation: WriterAgent → ContinuityAgent (fast) → optional 1 revision → ContinuityAgent (full) → NarratorAgent → ContinuityAgent (warn-only).
4-mode system: CANON, CANON+INFER, BRANCH, WRITER_ROOM.
RAG pipeline: CanonIndexer with semantic search, chapter/character filters, and CHAROS_RETRIEVAL_DEFAULT_TOP_K tuning.
MemoryBridge — Cross-agent character memory persistence with 90-day auto-pruning.
Prompt safety — sanitize_story_excerpts() + RAG_SAFETY_RULE blocks prompt injection from story content.
Typed agent contracts — all agent boundaries use dump_validated() / validate_retrieved_chunks().
@traced telemetry — structured span logs on all public agent methods.
EvalRecorder — JSONL eval event log for agent decisions and latency.
RuntimePolicy.from_env() — all CHAROS_* tuning knobs centralized.

Infrastructure

Note (historical): Vector storage entries below reflect the stack at the time of this release. The production stack has since migrated to PostgreSQL pgvector — no separate vector DB is required.

~~Upstash Vector as production vector backend (VECTOR_DB_BACKEND=upstash).~~ (replaced by pgvector)
~~ChromaDB for local development.~~ (replaced by pgvector)
BatchEmbeddingService — batched + cached embedding calls (4× faster than per-call).
IntegratedSecurityMiddleware — CSRF, rate limiting, XSS sanitization.
Stripe billing integration with webhook deduplication.

[1.0.0] — MVP Foundation (2026-01-15)

Added

FastAPI backend on Railway (PostgreSQL + Redis).
Next.js 14 frontend on Vercel. (upgraded to Next.js 16.x — see current package.json)
Manuscript upload and analysis pipeline (.txt, .epub, .pdf).
Character extraction — up to 14+ characters per manuscript.
AI-powered story analysis: summary, themes, dialogue patterns, publication readiness.
9-tab analysis dashboard with charts.
JWT authentication, email verification, Google OAuth.
Alembic-managed schema migrations (auto-run at Railway startup).
Docker Compose for local development.

Phase 3: Audio Pipeline (Target: 8 days)

Integrate neural TTS engine
Build multi-voice audio combining
Add voice binding to characters
Create /audio endpoint

Phase 4: Frontend (Target: 13 days)

Build CharacterOS Studio page
Add Scene Simulator UI
Implement mode selector (CANON/CANON+INFER/BRANCH)
Add audio playback controls

Changelog

Release notes, versions, and updates.

Edit this page on GitHub

Changelog

All notable changes to Marvox are documented here.

Format: Keep a Changelog · Versioning: Semantic Versioning

[1.5.0] — Production Studio, Casting Intelligence & External APIs (June 2026)

Added

Robotics Persona API — Commercial surface that lets external hardware, robots, and agents drive a Marvox character outside story context. POST /api/v2/personas/{character_id}/execute, persona profile and behavior-policy endpoints, and GET /api/v2/personas/group/{group_id}/consensus-state for fleet synchronization. Authenticated with an X-Marvox-API-Key header.
Character App Store — Instantiate personas from pre-built templates via /api/v2/templates and POST /api/v2/personas/from-template/{template_id}.
Transcription (speech-to-text) — POST /api/speech-to-text/projects/{project_id}/transcribe and a standalone POST /api/speech-to-text/transcribe for voice-driven prompts and dictation.
Voice Recipes — Save, update, share, and reuse a voice configuration across characters. CRUD under /api/characteros/projects/{project_id}/voice-recipes, plus public share-token resolution.
Production-grade exports — 19 export formats via /api/projects/{project_id}/export/{export_type}, each with an export audit log: manuscript/text (PDF, Word .docx, EPUB, Final Draft .fdx, Fountain screenplay, Markdown, RTF), audio (MP3, WAV, FLAC), structured data (character JSON, VoiceDNA JSON, canon-graph CSV, dialogue CSV, dialogue XLSX, HTML), and interchange (XLIFF 1.2, SRT subtitles, Yarn game dialogue). Audio, polished-manuscript, and structured-character formats are Pro-gated; lightweight handoff formats are Free.
Casting intelligence — Gender- and age-aware voice casting for every speaker, a per-book casting summary report, and an auto-produce path for full-book audiobook synthesis.

Changed

Performance Director — Audio synthesis now applies a scene-level dramatic pace envelope and emotion-reactive emphasis with micro-pauses for more cinematic delivery.
Agentic voice-casting — A learning loop continuously refines voice-casting decisions from production outcomes, layered into the live production pipeline alongside health monitoring.
Manuscript Studio — New Studio tab with refined speaker-attribution for dialogue-heavy manuscripts.
Billing — Subscription tiers consolidated to Free, Pro, Studio, and Enterprise, with checkout and lifecycle billing handled in-app via Stripe.
Vector storage — RAG embeddings migrated to a native PostgreSQL vector(1536) column.

[1.4.0] — Phase 4: API Key Management & Character Reflection (2026-03-19)

Added

API Key Management — Programmatic access to Marvox without user credentials.
- POST /api/billing/api-keys — Create long-lived API keys (SHA256-hashed, one-time display).
- GET /api/billing/api-keys — List keys with prefix, last-used audit trail.
- DELETE /api/billing/api-keys/{key_id} — Soft-revoke a key immediately.
- Keys use Authorization: Bearer mrvx_... — identical to JWT bearer flow.
- Schema: migration 0011 — api_keys table with indexes on project_id, key_hash, is_active.
Character Reflection — Automatic memory-driven personality evolution.
- POST /api/characteros/projects/{id}/characters/{char_id}/reflect — Manual trigger.
- Nightly automatic reflections at 02:00 UTC via background scheduler.
- Produces: reflection text, emotional arc trend (improving / stable / declining), saved memory ID.
- Error codes: REFLECT_FAILED, REFLECT_NO_MEMORIES.
Inference Cost Tracking — OPENAI_API_USAGE_WARNING_THRESHOLD env var alerts when costs exceed the configured percentage (default: 80%).

Changed

DEPLOYMENT.md — Updated required Railway env vars to include Phase 4 additions.

[1.3.0] — Phase 3: Collaboration & Character Participants (2026-02-20)

Added

CharacterParticipantManager — Characters react to SCENE_COMPLETE events in real-time via WebSocket broadcasts (event type: CHARACTER_REACTION).
Reactions are generated in parallel (semaphore = 3) using CharacterAgent.chat(mode=WRITER_ROOM).
Memory persistence for character reactions via MemoryBridge (non-blocking — failures do not block reactions).
EmotionalArcEvolverAgent — Evolves character personality from accumulated reactions.
ConsensusAnalyzerAgent — Detects conflicting character portrayals across multiple writers.
ReactionAudioStreamAgent — Live audio streaming for character reactions in collaboration sessions.

Fixed

Circular import in collab_routes.py — now uses lazy initialization for CharacterParticipantManager.

[1.2.0] — Phase 2: Audio Pipeline (2026-02-01)

Added

8-step audio production pipeline (AudioPipelineOrchestrator):
1. Parse scene dialogue blocks.
2. Resolve character voices via VoiceSelectionAgent.
3. Configure per-character TTS params via VoiceConfigurationAgent (300+ configs; speed 0.25×–4.0×).
4. Per-line emotion detection (AudioSceneAgent).
5. TTS generation (parallel with asyncio.Semaphore(5) to prevent OOM).
6. Audio composition and mixing via pydub.
7. 5-layer audio QA via AudioContinuityAgent.
8. Store to Vercel Blob; fall back to local file storage in dev.
VoiceCloningAgent — Neural voice DNA cloning.
DNALearningEngine — Evolves voice DNA when overall quality score ≥ 85 %.
VoiceVault — IP-locks Voice DNA to project/org boundaries (signed binding).
AudiobookOrchestrator — Full audiobook production from project canon.

[1.1.0] — Phase 1: CharacterOS Agent Network (2026-01-25)

Added

19-agent CharacterOS network orchestrated by AgentRuntime singleton.
Content generation agents: CharacterAgent, WriterAgent, ReaderAgent, NarratorAgent.
Quality agents: ContinuityAgent (5-layer validation), DialogueQualityAgent, EmotionalBeatAnalyzer, ScenePacingOptimizer.
Supporting agents: DirectorAgent, SummarizationAgent, AtmosphereAgent, QualityAnalysisAgent.
3-pass scene generation: WriterAgent → ContinuityAgent (fast) → optional 1 revision → ContinuityAgent (full) → NarratorAgent → ContinuityAgent (warn-only).
4-mode system: CANON, CANON+INFER, BRANCH, WRITER_ROOM.
RAG pipeline: CanonIndexer with semantic search, chapter/character filters, and CHAROS_RETRIEVAL_DEFAULT_TOP_K tuning.
MemoryBridge — Cross-agent character memory persistence with 90-day auto-pruning.
Prompt safety — sanitize_story_excerpts() + RAG_SAFETY_RULE blocks prompt injection from story content.
Typed agent contracts — all agent boundaries use dump_validated() / validate_retrieved_chunks().
@traced telemetry — structured span logs on all public agent methods.
EvalRecorder — JSONL eval event log for agent decisions and latency.
RuntimePolicy.from_env() — all CHAROS_* tuning knobs centralized.

Infrastructure

Note (historical): Vector storage entries below reflect the stack at the time of this release. The production stack has since migrated to PostgreSQL pgvector — no separate vector DB is required.

~~Upstash Vector as production vector backend (VECTOR_DB_BACKEND=upstash).~~ (replaced by pgvector)
~~ChromaDB for local development.~~ (replaced by pgvector)
BatchEmbeddingService — batched + cached embedding calls (4× faster than per-call).
IntegratedSecurityMiddleware — CSRF, rate limiting, XSS sanitization.
Stripe billing integration with webhook deduplication.

[1.0.0] — MVP Foundation (2026-01-15)

Added

FastAPI backend on Railway (PostgreSQL + Redis).
Next.js 14 frontend on Vercel. (upgraded to Next.js 16.x — see current package.json)
Manuscript upload and analysis pipeline (.txt, .epub, .pdf).
Character extraction — up to 14+ characters per manuscript.
AI-powered story analysis: summary, themes, dialogue patterns, publication readiness.
9-tab analysis dashboard with charts.
JWT authentication, email verification, Google OAuth.
Alembic-managed schema migrations (auto-run at Railway startup).
Docker Compose for local development.

Phase 3: Audio Pipeline (Target: 8 days)

Integrate neural TTS engine
Build multi-voice audio combining
Add voice binding to characters
Create /audio endpoint

Phase 4: Frontend (Target: 13 days)

Build CharacterOS Studio page
Add Scene Simulator UI
Implement mode selector (CANON/CANON+INFER/BRANCH)
Add audio playback controls

Changelog

Changelog

[1.5.0] — Production Studio, Casting Intelligence & External APIs (June 2026)

Added

Changed

[1.4.0] — Phase 4: API Key Management & Character Reflection (2026-03-19)

Added

Changed

[1.3.0] — Phase 3: Collaboration & Character Participants (2026-02-20)

Added

Fixed

[1.2.0] — Phase 2: Audio Pipeline (2026-02-01)

Added

[1.1.0] — Phase 1: CharacterOS Agent Network (2026-01-25)

Added

Infrastructure

[1.0.0] — MVP Foundation (2026-01-15)

Added

Phase 3: Audio Pipeline (Target: 8 days)

Phase 4: Frontend (Target: 13 days)

Archive

Changelog

Changelog

[1.5.0] — Production Studio, Casting Intelligence & External APIs (June 2026)

Added

Changed

[1.4.0] — Phase 4: API Key Management & Character Reflection (2026-03-19)

Added

Changed

[1.3.0] — Phase 3: Collaboration & Character Participants (2026-02-20)

Added

Fixed

[1.2.0] — Phase 2: Audio Pipeline (2026-02-01)

Added

[1.1.0] — Phase 1: CharacterOS Agent Network (2026-01-25)

Added

Infrastructure

[1.0.0] — MVP Foundation (2026-01-15)

Added

Phase 3: Audio Pipeline (Target: 8 days)

Phase 4: Frontend (Target: 13 days)

Archive