Copilot

RoRo Copilot lives in the last-but-one rail icon: an agent that sees your project and the platform around it and proposes changes — honest about what it did, what it cost, and what it may not do alone.

#What it can see

Every read is automatic and shows as an activity line.

• Your project — list_files, read_file (the live buffers, unsaved edits included), search_files (string or regex), read_output (the dock's log).
• The machines — list_machines: codename (the id the SDK, the API and the Run rail use), qubits, list price per shot, status, queue.
• A price, from the server — quote_run has the platform price a .qasm for a machine and shot count and relays it verbatim. A quote holds nothing and submits nothing, and the copilot never multiplies numbers itself.
• This project's runs — list_runs / read_run: status, cost, the credit line the Output pane prints (a refund is certain only when the record says approved), the counts.
• The library — list_library / read_library_item pull an example's README and program in, so "add a QAOA example" starts from verified code.

#What it can do

• Write — write_file pauses the loop for an approval card: Edit or New file, the reason, a compact +/− diff. Apply writes through the ordinary save path; Decline is an answer the model must live with.
• A change set — several files proposed in one round become one card headed N files: a row each with +/−, its own diff and a tick, then Apply all / Apply selected / Decline all. Applying is sequential and the copilot is told file by file what landed — one question for one edit, never a half-applied refactor.
• Run Python — run_python asks first, then runs the file on the browser kernel: local and free. The output goes to the model, not the dock.
• Explain — the Output pane's Explain this result / error is the same metered turn.

#From the editor

Select code and a small bar appears — or press ⌘., which takes the current line when nothing is selected: Explain · Fix this · Add comments · Write a test · Ask about this…. Nothing is sent until you click one. It starts an ordinary turn: the file and line range are quoted in the transcript, it costs credits, Stop ends it, and any edit still arrives as a card.

#What it cannot do

No way to submit a run or spend run credits — a quote is a price check; hardware stays behind the Run rail's confirm. Nothing written or run without a card. The model is anonymous: the platform picks it.

#Cost, visibly

Every turn is metered as AI credits. The footer shows N cr this session · balance · metered per turn, read when the panel opens. There is no per-message ceiling: the cost trail, Stop and your balance are the controls. Out of credits, the server refuses the next request before it is charged and the feed shows a Top up in the console card. After 40 tool rounds it pauses; Continue resumes.

#Conversations and asks

Sessions are saved per project (Conversations…, +, the bin); switching rails never cuts off a streaming answer. Three starter chips, or a persona (hackathon, integration, research). Good asks: "What would 4000 shots cost on the cheapest real machine?", "Why did my last run fail?" — then read the diff before Apply.

Related: Python in the browser · Machines & credits · Run a circuit · Library