mTm-Ch

Context

Working across multiple Claude interfaces (web, code, API) and multiple AI providers means constantly re-establishing context. Every new session starts from zero. Months of shared understanding — architecture decisions, project conventions, debugging history — evaporates at session boundaries. The token is cheap; the context is expensive.

Approach

Built a multi-agent Claude infrastructure with specialized instances:

- Mountaintop Monk: Conversational instance for thinking, planning, and strategic work
- Cave Hermit: Code-focused instance for development and implementation

Both share a "charter document" system — fast-loading context documents that instantly establish shared understanding regardless of session state. The charter acts as a cold-start protocol: any new session inherits the full project context without manual briefing.

Currently in deep testing: pushing the limits of shared multi-agent context to determine how much persistent state can survive handoffs between specialized agents. UI redesign underway based on testing findings.

Outcome

Basic functions live. Deep testing underway with active UI redesign. Architecture validated through daily personal use — dogfooding the product before any external release. Testing is specifically targeting the boundaries of multi-agent context sharing to document what works, what breaks, and where the hard limits are.