Skip to content

Roadmap

Purpose of this document: Keep contributors focused on the breakthrough — local extensible conversation simulation — not on scope that would compromise that focus. If a proposal fits this document’s MVP criteria, it belongs here. If it belongs in a “not now” section, it should wait.


A build qualifies as MVP when every item below is true:

  1. A player can open a scenario, have a complete text conversation with an NPC, and reach a scored debrief — entirely offline, on a mid-spec laptop.
  2. Scenario packs are plain YAML. A non-programmer can fork and edit one without writing code.
  3. The schema validator catches malformed packs at import time and prints an actionable error.
  4. At least one official scenario pack ships with four distinct, playable scenarios.
  5. Voice input (Whisper STT) and voice output (Kokoro / sherpa-onnx TTS) work as opt-in features on the same local machine.
  6. A workbench UI lets a contributor inspect live state, replay a turn, and validate a pack without touching the CLI.
  7. No outbound network call is made during play. The offline smoke test exits zero.
  8. The README and docs are accurate enough that a new contributor can run the project from git clone to first conversation in under 15 minutes.

Work proceeds in this sequence. Later items depend on earlier items shipping and being stable. Do not start an item before its predecessor passes its acceptance criteria.

#ItemWhy first
1Text conversation loopEverything else is UI on top of this
2Schema / validatorPack authors need a contract before writing content
3One excellent scenarioProves the loop works end-to-end with real content
4Model managerStable download + version-lock before adding more dependencies
5Voice input (Whisper STT)Adds a new input mode without changing the core loop
6Voice output (Kokoro / sherpa-onnx TTS)Closes the full local audio loop
7Workbench UITooling for contributors; requires stable loop and schema
8More official packsContent diversity once the system is proven
9README / demo polishFinal surface for GitHub launch

  • Monorepo skeleton — directory structure, licensing, tooling
  • Developer scripts (setup.sh, dev.sh, PowerShell equivalents)
  • Model registry with checksums and explicit license disclosure flow
  • Local TTS adapter (Kokoro / sherpa-onnx, synthetic voices, no voice cloning)
  • Text-only workbench — chat panel + live state inspector
  • Four official packs — Job Interview Basics, Everyday Negotiation, Language Café, Difficult Conversations (four playable scenarios each)
  • README rewritten for instant GitHub comprehension
  • Core conversation loop: player turn → LLM inference → NPC response → state update
  • Debrief generation from rubric scores
  • Browser UI connected to Python backend via WebSocket
  • JSON Schema covering scenarios, NPCs, rubrics, scenes, safety policies
  • Pack validator CLI (convsim validate-pack <path>)
  • Pack import/export flow in the browser workbench
  • Offline smoke test CI gate
  • Whisper STT integration (voice input, local — requires whisper.cpp runtime)
  • Full TTS playback connected to scenario engine (Kokoro — requires local server)
  • Workbench: state inspector, pack editor, validation, export
  • Layered safety system with global non-overridable rules and per-pack policies
  • Cross-platform CI pipeline with acceptance tests, pack validation, and release smoke matrix
  • Contribution guide, code of conduct, security policy, issue templates
  • Full documentation set (architecture, install, safety, privacy, scenario authoring, troubleshooting)

Remaining polish — before Milestone 1 tag

Section titled “Remaining polish — before Milestone 1 tag”
  • Real UI screenshots (replace SVG placeholders in README)
  • Desktop app with bundled backend (Tauri sidecar for convsim-core)
  • Automated real-model CI smoke test

See the post-alpha issues list for the full triaged list with reasons and milestone assignments.


Future work — outside MVP, not forgotten

Section titled “Future work — outside MVP, not forgotten”

These areas are intentionally out of scope for MVP. They are listed here so contributors know they have been considered and deferred, not overlooked.

Recording them here does not create a commitment to build them. Any future work in these areas needs its own proposal, design, and acceptance criteria.

  • Animated NPC avatars or portrait art
  • Scene backgrounds and atmospheric overlays
  • Custom fonts, themes, or visual design systems beyond the base UI
  • Multi-NPC scenes (more than one NPC active at once)
  • Branching dialogue trees with author-defined decision nodes
  • Long-term memory across sessions (persistent relationship state)
  • Emotion-to-voice prosody mapping (anger, warmth, uncertainty in TTS)
  • In-app pack browser / discovery feed
  • Community ratings and reviews
  • Pack signing and trust tiers
  • Hosted pack registry (CDN or P2P distribution)

See the marketplace architecture for the post-launch marketplace design baseline — entry gate criteria, distribution path comparison, and scope of schema, signing, moderation, payment, and reporting changes required before any paid content ships.

  • Classroom dashboard (instructor sees student session aggregates)
  • LMS integration (SCORM / xAPI export)
  • Analytics and cohort reporting
  • Enterprise SSO or managed deployment

The following are explicitly out of scope and should not be proposed as MVP additions. A future proposal can revisit them, but the bar is high: any yes must not compromise offline-first, local-model, or open-source principles.

TopicWhy not now
VR / ARRequires platform-specific SDKs, hardware, and a whole second rendering stack — incompatible with MVP simplicity
MultiplayerServer infrastructure, latency, identity — all the things MVP deliberately avoids
Cloud inferenceContradicts the local-first promise; would require an account and send user content to a third party
Mobile appsApp store rules, on-device model size, and touch UI are a separate product effort
Marketplace / monetizationPayment rails, refund policy, fraud, curation burden — not a developer tool concern
NSFW contentAbove PG-13 is a hard platform boundary; the safety system enforces this at the validator level
Celebrity or public-figure packsDefamation, right-of-publicity, and likeness risk with no upside for the simulator’s core use case
Complex character animationRequires a game-engine-class renderer; out of proportion for a conversation tool
Therapy / law / medical positioningThese are licensed professions; shipping a product that implies clinical efficacy creates liability without benefit