16 — Goals

The /goal long-horizon autonomous mode. Provider-agnostic. The agent runs until a completion contract is satisfied — minutes, hours, or days — planning, acting, testing, and self-correcting without manual user intervention. OpenAI's Codex and Anthropic's Claude Code both ship /goal for their own models. Picasso brings it to every provider the abstraction supports — Anthropic, OpenAI, Google, xAI, OpenRouter, Ollama, vLLM — through one harness, one contract format, one observability surface.

The /goal mode's visual identity is locked to --cyan. Scene 04 in design/Picasso TUI.html shows the full monitor: a contract band at the top with the objective, completion criteria, and time/cost ceilings; four budget tiles (wall time, cost, steps, subagents) each with a colored <Bar> showing consumption; an iteration strip (one rectangle per iteration, color-coded by outcome — green for success, amber for retry, danger for failure, klein for in-flight); and a live tool + memory feed below. The /goal badge in the status line uses cyan; the live iteration uses a cyan <Pulse>. See design/scenes-a.jsx.

What ships

• /goal <objective> starts a goal; /goal pause|resume|edit <objective>|continue [n]|rounds <n>|complete|clear|monitor controls it. From the shell: picasso goal create|run|status|events|logs|pause|resume|stop|templates.
• Each round shows as a "Goal round N of M" card, live and after /resume, and goal sessions are titled by their objective.
• /loop <objective> works an objective in self-paced rounds without a written contract.
• Contracts and templates: goal quickstart and goal contract reference.

Runtime note

Goal state is owned by the engine (durable objective state in the session log); Picasso's goal surfaces drive the engine's goal rounds through the adapter. See architecture.