Skip to main content

Roadmap

What's shipped, what's in progress, and what's planned.

Shipped​

FeatureWhat landed
Web FrontendNext.js chat UI with streaming, thread management, and memory inspection. Lives in apps/web/.
API LayerFastAPI with REST (POST /api/chat) and WebSocket (/api/chat/stream) endpoints. Thread management, memory status, session end. Lives in apps/backend/api/.
Voice ChatExperimental OpenAI Realtime speech preview with a FastAPI websocket bridge, standalone test harness, and Next.js voice UI. Current path is speech-only and non-agentic. Lives in apps/backend/voice/.
Session FeedbackEnd-of-session thumbs rating captured at /end, /exit, and POST /threads/{id}/end. SQLite-backed, incognito-safe.
Session Trajectory EvalUnified runner for short (inline) and long (checkpoint) trajectory datasets. 25 long-trajectory cases covering modality, boundary enforcement, crisis arcs, closing, venting, and mode transitions. Concurrent hybrid execution with --concurrency, --case, --verbose.
Crisis Gate — LLM-primaryLLM is the primary crisis classifier; regex is fallback only. Override precedence fix, shadow monitoring, prompt hardening (conversation fencing, anti-injection, adversarial examples), strict truth table enforcement.
Dispatcher — LLM-primaryLLM handles all mode + modality classification. Context-blind regex fast paths removed. LLM-based mid-exercise exit detection. Exercise modality persistence.
Knowledge Overhaulsoul.md expanded with therapeutic grounding, cultural sensitivity, repair patterns, boundary-setting voice. identity.md rewritten with product philosophy. boundaries.md expanded with redirection patterns and dependency framing.
OpenAI Embeddingstext-embedding-3-large as default provider, Gemini as fallback. Hybrid RRF retrieval achieves 14/17 recall@5 vs 6/17 token-only.

In progress​

FeatureStatusWhat's left
Response quality rubricDesigned, not implementedLLM-as-judge eval runner to test empathy, tone, banned phrases, question stacking, conciseness. Needs rubric dataset + grading runner.
Memory integration evalDesigned, not implementedTest whether retrieved memory shapes responses. Cross-session continuity, procedural rule enforcement, appropriate recall.
Session feedback — closing modeDesigned, not wiredClosing detection is now LLM-primary; feedback prompt needs to fire on natural closings, not just CLI/API end commands.
Session feedback — voiceDesigned, not wiredVoice disconnect bypasses end_session() — needs to either route through the runtime or gain its own feedback hook.

Planned​

Messaging Channels​

Adapters for Telegram, WhatsApp, and Discord. The Channel enum already has slots (Channel.TELEGRAM, Channel.WHATSAPP); the agent graph is channel-agnostic. Each adapter maps platform message formats to AgentInput / AgentOutput. Crisis responses would need channel-specific formatting (inline buttons, embeds).

Acoustic Crisis Detection​

Voice mode currently uses transcript-only crisis detection. Real gaps: voice cracking, sobbing, pressured speech, prosodic flatness. A user saying "I'm fine" through tears scores level 0.

Requires either a curated distressed-voice dataset (ethically fraught) or a validated off-the-shelf acoustic classifier (not a solved problem). Calendar-gated on dataset and model maturity.

Graph Memory​

Graphiti + Neo4j for entity/relationship extraction from semantic facts. Enables relational reasoning: "you mentioned your sister and your work stress — they tend to co-occur." The wire frame exists (agent/memory/graph_store.py with NullGraphMemoryStore); the graphiti-core dependency is in pyproject.toml but the integration is intentionally disabled pending design.

Background Consolidation​

Automatic fact merging, dormant marking, and a consolidation_runs log. Schema is defined (ConsolidationProposal, ConsolidationRunRecord in agent/memory/models.py); the implementation node is sketched but not wired into the graph. Adds /memory restore as an undo for destructive operations.

Session Intent, Stage, and Response Guidance​

Three state fields (progress.intent, progress.stage, response.guidance) are defined in the schema but not yet populated by any node. When implemented, they enable session-level steering: the agent knows whether to deepen, stabilize, or close based on conversation arc rather than just the current message. The eval runner already supports assertions for all three — just re-add the dataset expectations.

Crisis Gate Production Telemetry​

Model ID, prompt version, raw/normalized levels, confidence values, deterministic shadow results, disagreement rates, timeout/parse failure counters, and degraded-mode alerts. The shadow monitoring infrastructure is in place; the production telemetry layer is not.

Clinical Review​

A trained clinician reviews the knowledge/response_modes/*.md files, the prompt builders in agent/therapeutic/prompts.py, and agent responses across dogfood sessions. This is the gate before "a trusted friend could try it" becomes a defensible claim. Calendar dependency, not engineering.