Orca

About

Single-model coding pipelines blur responsibility: the same context plans, codes, reviews, and talks to the user, so quality control becomes an afterthought. Orca assigns each function a named role in a fixed pipeline — Brain decomposes the request, Miranda runs a PLAN-ANSWER-CRITIQUE-REWRITE compliance loop, Pappy issues PASS/WARN/FAIL quality verdicts, and Benson handles intent parsing — with a Maestro router managing a nine-role subagent pool that includes strong and cheap model tiers, debugger, reviewer, and vision roles. Runs persist in SQLite, MCP connects external tools, and a distillation loop exports interactions as training data for fine-tuning. The runtime is a TypeScript pnpm monorepo with an Electron desktop app and a CLI runner keyed to OpenRouter or local Ollama, under a PolyForm Noncommercial license and aimed primarily at Windows. It is an early personal project (39 stars, 442 commits) exploring role-decomposed agent architectures.