Skip to main content
A normal turn ends when the model stops. A goal does not: /goal attaches an objective to the session, and the agent keeps working toward it across turns — planning, acting, verifying — until the work is evidence-complete, until you pause or clear it, or until the budget runs out. The goal survives the end of a turn, context compaction, and closing and resuming the TUI. This is Cortex session state, not a second coding provider. Agent mode still talks to the Cortex API; /goal is harness state — persist, continue, wrap up.
/goal landed on the CLI main branch on 8 September 2026 (CortexLM/cli #54) and ships with the next tagged release after 0.1.10. If /goal is not in your palette yet, run cortex upgrade once that release is out.

​
Commands

A reserved word is only special when it is the entire argument. /goal pause the deploy sets an objective; it does not pause.

​
What you see

The composer carries a chip with the goal’s progress, painted in Cortex green on the border next to the mode:
Cortex CLI composer with the Goal · 2/8 chip on its border

An active goal: the composer border reads Agent — Goal · 2/8 after the kickoff turn “ship the rate limiter and prove it with tests”.

The chip reads Goal · 2/8 while active, and paused, done, budget, or blocked in the other states.
Cortex CLI composer with the Goal · done chip on its border

Evidence-complete: the chip reads Goal · done and auto-continuation stops.

​
States

​
Evidence, not vibes

Completion is evidence-based. To mark a goal done the model must call UpdateGoal with a reason and at least one of: a file path, a command, or a test. A confident paragraph is not enough.

​
Budget

The default budget is 8 turns. Near the limit — one turn left, or 85% of a token cap — the next continuation asks the agent to wrap up rather than open new scope.

​
Where it is stored

goal.json lives next to the session files:
Resuming the session reloads it. See Sessions and Configuration → Data locations.

​
When to use a goal

​
Related