Hand off a goal. Check in when it suits you.

Describe the outcome and get on with your day. The agent works toward it in checkpoints on a draft pull request, takes your corrections while it runs, and finishes with a PR you review like any other. Goals run on Claude Code, Codex and Antigravity CLI.

Describe

Run isolated

Review in GitHub

You start with: An outcome that takes several steps
You use: New Goal · create_goal · send_goal_input · wait_goal
Takes: Minutes to hours of agent time; #2670 finished in about 12 minutes
Real run: propr #2670

State the outcome

Open Goals → New Goal, pick the repository and write the objective. Attach specs or screenshots. Choose the agent and model; ProPR offers only agents whose configured runtime runs goals.

Start a goal on northwind/storefront-web: the objective, an outcome that takes many steps

Let it work in checkpoints

When the agent implements directly, ProPR opens a draft PR on the goal branch before the agent starts. When a coherent set of changes is ready, the agent requests a checkpoint and ProPR commits exactly that scope and pushes it. Checkpoints target a 15-minute cadence.

Prefer a breakdown? Let the agent orchestrate: it files issues and runs them through ProPR.

Two checkpoint commits at a 15-minute target cadence, with the latest declaration: three files included, one held back

Check in

The goal console shows current activity, checkpoints, logs and visual previews. A paused goal, or a Codex goal with a question, lands under Needs you on the dashboard. Nothing else interrupts you.

A goal waiting on its operator: the agent asks whether offline photos upload over cellular, with suggested answers and Answer, Pause and Cancel

Steer it

Send a correction or a question. Claude Code and Codex receive it live; Antigravity CLI takes it at the next finished step. The timeline records your words verbatim.

Pause, resume or cancel, and switch the model between steps. Committed checkpoints keep their place on the branch.

The goal timeline records the correction verbatim, followed by the agent acting on it

Or run it from chat

Over MCP, your assistant drives the same goal: create_goal starts it, send_goal_input steers it, and wait_goal returns at a confirmed state or a new checkpoint, with a deadline, so a check-in costs one call. The propr goal CLI covers the same controls. MCP ships disabled; an administrator turns it on.

Review the draft PR

ProPR finishes only after it validates an open draft PR on the goal branch against the expected base; the agent's say-so never skips that check. Run /review, comment for changes, mark it ready and merge it yourself. Goals never auto-merge.

A real goal, start to merge.

On public propr #2670, GPT-5.6 Sol took a two-sentence objective: add the newest GPT models to the Codex agent list and update the pinned Codex version. ProPR opened the draft with an initializing commit, pushed the agent's single checkpoint 626c366 twelve minutes later, and rewrote the description with the result. After a /review and a /fix pass, the maintainer marked it ready and merged it.