System Maturation and UI Capabilities

Intent Fulfillment: Lists and Scheduling

This update brings significant progress to Chalie’s ability to manage state over time. We’ve introduced a robust List and Scheduler API, allowing for the creation, tracking, and automated firing of intents. These aren’t just simple CRUD endpoints; they are integrated into the cognitive triage system, allowing Chalie to proactively manage tasks and reminders. The Brain UI (admin dashboard) now includes dedicated interfaces for these features, and the primary chat interface now supports specialized cards for rendering list items and scheduled events.

To support these new features in the UI, I’ve enriched the tool parsing logic. The backend now maps raw manifest schemas to UI-friendly fields—adding labels, hints, and placeholders—ensuring that when Chalie asks for tool parameters, the interface remains intuitive.

Refined Cognitive Routing

We spent a lot of time debugging the ACT loop and the orchestrator. One persistent issue was SSE (Server-Sent Events) timing out before long-running tool jobs could finish. I’ve extended the timeout windows and implemented a ‘close’ signal for the background threads to ensure the frontend knows exactly when a processing cycle is complete.

I also refined the deterministic mode router. By decoupling the decision of how to engage from the LLM’s response generation, we’ve reduced latency significantly. The documentation now better reflects this ‘Deterministic Routing’ philosophy, explaining how we use observable signals to select engagement modes in milliseconds rather than waiting for an LLM to decide its own next steps.

Backend Stability and Infrastructure

As the project grows, so does the number of background processes. With 18+ independent processes running, we were quickly exhausting the PostgreSQL connection limit. I’ve tuned the database service to use a more lightweight pooling strategy—dropping the per-process pool size to 3—ensuring the system remains stable under load without hitting the 100-connection ceiling.

We also addressed several provider-specific friction points. LLM provider strings were removed from individual agent configurations to favor a more centralized, swappable provider architecture. Additionally, I fixed a critical issue with VAPID key deserialization for push notifications, ensuring that cognitive drift and proactive alerts actually reach the user’s device.

Documentation and Cleanup

The project docs received a major overhaul. We’ve replaced the internal plan.md and E2E test logs with formalized architecture guides. These new documents provide a clear map for developers, covering the 15-step request pipeline, the layered memory system, and the sandboxed tool framework. The README now emphasizes our ‘Privacy First’ approach, making it clear that all memory and learned traits stay on the local machine.