Most tools stop at suggestions. Dinopsys is aimed at systems that take a goal, inspect the real environment, figure out the steps, do the work, check whether it actually worked, recover when it didn’t, and get sharper the next time around.
The point is a system that owns the whole job — not one that hands you a plan and walks away. It should finish the computer-based work, not just describe how a human could do it.
Not locked to a single language, project type, or toolset. The same loop should stretch across different kinds of work.
Built as cooperating parts, not one giant black box. Each subsystem owns a clear job inside the same loop: discover, decide, act, check, recover.
Orchestrator — gathers from ATLAS & ORACLE, drives agents and the loop.
File atlas — system-wide browser with permissions for grounded discovery.
RAG layer — retrieves stored knowledge when the current task needs it.
Aider chat — runs the programming AI that implements the changes.
Compiles & tests Aider output, returns evidence to KORA.
Make a change, then look at what actually happened. That result is the evidence for the next move — not a one-shot guess.
Not reckless self-rewriting. A measured loop: watch how it performed, name the weak spots, try a concrete change in a safe place, check whether it actually helped, and keep only what holds up.
Built in stages on purpose. Each one unlocks the next instead of trying to jump straight to the end state.
Eventually this has to leave the original machine. Clean installs, portable pieces, versioned releases, and support for the places where the work actually happens.
Build one that figures out what the job actually is, finds a way to do it, runs the work, checks the outcome, recovers when it misses, and shows up sharper the next time.