the problem
You're using Claude Code
for everything.
One harness, one model, one prompt style — for architectural decisions and JSON formatting alike. It works. But a single Claude Code session has no oversight, no debate, no memory across runs. Important decisions get made without scrutiny. Context resets at every session boundary.
the proto.type approach
| Architectural reasoning | Hotaru | → | Opus 4.6 |
| Adversarial review | Seori + Kael | → | GPT-5.4 / o3 |
| Output review | Iseul + Brynja | → | GPT-5.4 |
| Context compression | Sepehr | → | Sonnet 4.6 |
| Everything | you, probably | → | Claude Code on everything |
Each agent uses the right model for their role. Architectural reasoning gets Opus. Context compression gets Sonnet. The morning report shows what the oversight cost you — usually less than the mistake it prevented.
model routing
Intelligent model assignment
proto.type automatically assigns each agent to the best model for their role. Hotaru gets Opus. Sepehr gets Sonnet. Any provider, any setup.
providers
Every provider. Intelligently routed.
Works with Anthropic, OpenAI, Google, and dozens more. Automatically assigns each agent to the right model for their task.
context management
Context management across sessions
Claude Code has a context window. Long runs exhaust it. proto.type checkpoints, Sepehr compresses the state, and the next session starts with the Mission Object plus compressed context. Claude Code never sees the seam.
soul layer
Agents accumulate memory across runs
The soul layer. Each agent stores structured lessons from every run — not event logs, what was learned. Injected into agent context on every run start. The team gets better the more you build.
transparency
Every token accounted for
Every agent call is logged with tokens used and cost. Full breakdown by agent in every morning report. No surprises.
skill library
Powered by gstack
Agents use gstack — a library of engineering best practices. They cite these practices in debates with the same constitutional weight as Mission Object citations.
single binary
Single binary. Runs everywhere.
Built on Zig and Shoji. No runtime, no GC, no cloud dependency. Your keys and your code stay local. Ships on macOS, Windows, and Linux from one binary.
“gstack was instrumental in conceiving this vision. The idea that engineering knowledge could be encoded as a skill library — something agents cite like law — came directly from seeing what Garry built.”