init

The draft becomes the deployment's next version, with its starter tests. The interview ends with its closing words: what the draft sets Bash up to do, the tools your code must answer, the spending cap per task, how many tests it wrote, and the next steps. The CLI prints them. You can also read them with draft.closing() in an SDK, or in the version's interview field over HTTP and MCP.

GET /v1/deployments/{name}?format=yaml returns the draft as vestra.yaml, ready to keep in your repo and publish with deploy. After you edit the file, vestra test saves your edit as a new version and tests it, without promoting it. vestra deploy then promotes it, which makes it the version your agent runs, without running the tests again.

When the job mentions one of your own tools, the file declares it: its name, what it does, its input and what it returns. Vestra never writes the tool's code. Your code runs the tool, with vestra dev --tools ./tools.mjs or with the SDK's tools option.

The interview asks what each tool returns. Answer with one example result that uses your code's own field names and value formats, with made-up values. Never paste a real record. For example: {customer: 'Maya Chen', due_back: '2026-10-04 18:00'}. The tool's description keeps this example. Each starter test that calls the tool gets tool_results copied from it, with no field, unit or time zone the example lacks. That way test can play the tool before your code exists, and answer the way your code will.

Some tools need your user's approval before they run. Such a tool also takes, as inputs, each value the user agrees to, such as charge_usd and new_due_back on extend_rental. The instructions say where each value comes from, such as daily_price_usd from get_rental. The approval shows every input, so the user always sees the price before saying yes. If none of your lookups returns such a value (because your code works out the price only as it acts), the interview asks for a tool that does, such as a quote, and marks it read_only. Bash never guesses a value for an approval. The closing words remind you that your code must check these values against its own before it acts.

With --answers, vestra init reads your answers as text instead of asking at the terminal. Use it from a coding agent or CI: vestra init "<the job>" --answers answers.txt, or --answers - to read them from stdin. If the interview asks a question after your answers run out, it is told to use the defaults. The draft holds only the rules your answers give, so include every rule the job needs. Init adds one rule without asking, which you can edit: how Bash checks who is asking before it shares a user's record.

Say in the job description whether this agent works for your users, inside your product, or for your own company. The interview runs in a conversation of its own, kept apart from your company's own Bash and its memory.