Hey HN. I built this to solve context loss across Claude Code sessions. Every session starts cold, so I needed a persistence layer that could carry forward priorities, decisions, blockers, and institutional memory without depending on anything outside the terminal.
The core design decision was bounded responsibility per agent rather than one monolithic prompt. 38 agents in 5 categories (governance, knowledge, technical, business, evolution), each with a defined role and explicit collaboration interfaces. Workflows are pipelines with ordered handoffs, not free-form agent swarms. A technical task goes Architect → Security-Expert → Engineer → Security-Expert (second pass) → QA → Reviewer → Documenter → Historian. The routing is deterministic.
State is plain markdown in an `active/` directory that loads on session start. Git handles persistence and backup. Two-remote model separates the public framework from private user data so you can receive updates without exposing your files.
The part I find most interesting architecturally: the evolution loop. An evaluator agent measures outcomes, an improver proposes changes through a formal queue, and nothing gets implemented without user approval. The system's structure is mutable by design but gated on human review.
Tradeoffs worth noting: it's opinionated. The agent definitions and workflow ordering reflect how I work as a solo founder. If your workflow is fundamentally different you'll want to modify the agent prompts and pipeline ordering. The `/setup` wizard handles initial customization but it's not infinitely flexible out of the box.
MIT licensed. Happy to discuss architecture decisions.
slythefox•1h ago
The core design decision was bounded responsibility per agent rather than one monolithic prompt. 38 agents in 5 categories (governance, knowledge, technical, business, evolution), each with a defined role and explicit collaboration interfaces. Workflows are pipelines with ordered handoffs, not free-form agent swarms. A technical task goes Architect → Security-Expert → Engineer → Security-Expert (second pass) → QA → Reviewer → Documenter → Historian. The routing is deterministic.
State is plain markdown in an `active/` directory that loads on session start. Git handles persistence and backup. Two-remote model separates the public framework from private user data so you can receive updates without exposing your files.
The part I find most interesting architecturally: the evolution loop. An evaluator agent measures outcomes, an improver proposes changes through a formal queue, and nothing gets implemented without user approval. The system's structure is mutable by design but gated on human review.
Tradeoffs worth noting: it's opinionated. The agent definitions and workflow ordering reflect how I work as a solo founder. If your workflow is fundamentally different you'll want to modify the agent prompts and pipeline ordering. The `/setup` wizard handles initial customization but it's not infinitely flexible out of the box.
MIT licensed. Happy to discuss architecture decisions.