How Celeris gets faster
This comes with the Celeris desktop app, with nothing to switch on. The Mac App Store version does not include it.
The first time you ask Celeris for something it works out what you meant, which tool fits, and where "my repo" points. The second time it should not have to. So it keeps a record of what happened, and reads a small slice of that record back into the next task.
The model does not change while you work. A new model arrives on the schedule training runs allow: gather the data, train, evaluate, ship. Celeris records the task you just finished in time to use it on the next one.
What it keeps
One row per task: what you asked, which tools ran, with what arguments, how it turned out, and whether you took the answer, changed it or threw it away. Celeris reads four things back out of that record.
A shortcut. You have done this exact thing before, so Celeris replays what it worked out last time instead of reasoning it out again. The second "raise a PR" here already knows the path, the branch and the commands.
Your words. What your names point at. "My repo", "the project folder", the channel you call by its short name. Checked when it is used, so a name that has gone stale gets resolved again rather than trusted.
Which tool fits. Which tools have worked on work like this, from the ones you reached for and the ones that finished. A weighting that moves with evidence, not a list.
When to ask. Whether to just do it, offer you two options, or ask one question. Learned from how your ambiguous asks actually resolved. These fade and get re-checked; they never harden into a rule.
Rejecting is the useful signal
If Celeris gets something wrong, reject it rather than quietly fixing it by hand. A rejection is the clearest thing you can tell it, and a hand-edit outside Celeris is invisible to the record.
What is shared, and what is not
Your record stays on your machine. Celeris reads a slice of it into the turn you asked for, the same way your prompt reaches the model, and that slice is covered by the same settings your prompts are.
Separately, and only if you leave it on, an install can contribute the shape of an outcome, counts and labels, never your text, paths, commands or content, to a shared pool. What comes back is instincts: what already works for work like yours, on the tasks you have not done yet, which is most of them. Privacy lists exactly what leaves the device and how to turn it off.