The orchestration layer
Maestro
Build the systems that build the systems.
Six products run out of one company on one workstation. Most of the code that ships is written, reviewed and landed by models. Maestro is the part that decides which model, in which harness, against which plan, and whether the result is allowed through.
A model is only as good as its cradle
A cradle is the harness a model's makers built around it. Its own tools, its own sense of what a file is, its own memory of what it did an hour ago. The difference between a model in its cradle and the same weights behind a bare API is not small, and it does not show up in a benchmark, because a benchmark is a single turn and real work is four hundred of them.
So most orchestration has the trade backwards. It standardises on the API, throws the cradle away, and meters you per token for the privilege. Maestro keeps the cradle and gives up the uniformity. Every harness it drives is different, and handling that difference is most of what the code does.
Three things it does
- 01
Every runner is a real CLI
Every runner is a vendor CLI, started as a real process with its own workspace. The model arrives holding the tools, the file discipline and the memory its makers gave it. Nothing is flattened into a chat completion first, and no request ever leaves on a metered key.
- 02
It routes on what is actually left
Before a task is handed out, every runner reports the headroom on its own subscription, read from that vendor in its own words. Both windows count, the short one and the weekly one. Work goes where there is room, and a reserve is held back so a plan is never drained to zero by a background job.
- 03
A different model checks the work
Acceptance is a second pass by a runner that did not do the job, on a different harness where one is free. A model reviewing its own output is not review. If no independent reviewer is available the task waits, and waiting is a state it can sit in safely.
Where it breaks
A lease that expires means unknown, not stopped. If a runner stops reporting mid task, the honest reading is that a process somewhere may still be writing to a branch, so the task is frozen rather than reassigned. Recovering it is a manual job, and that is the correct trade, because the alternative is two workers editing one tree.
Quota reads are the other soft spot. A vendor that will not say what is left gets no work, which is safe but wasteful, and a vendor that reports optimistically will be believed. The layer measures rather than guesses, and the cost of that honesty is idle capacity on any plan that keeps its numbers to itself.
It is not a product
Maestro runs one company's work on one operator's machine. There is no tenancy, no billing and no support, and nothing here is for sale. It is on this site because the interesting part of running six products with a small team is what that takes and where it gives way, and this is the load bearing piece.
The build log covers the rest, including the decisions that turned out to be wrong.