Use the Convergence Gate
The Convergence Gate takes a Sweep reviewer verdict and runs a bounded fix-and-recheck loop, stopping when the surface is clean or a human needs to decide. You never have to manually re-queue “fix the Sweep finding” prompts.
Start a Sweep review
Section titled “Start a Sweep review”- Open the Sweep drawer from the left sidebar.
- Orbit starts the Sweep sidecar and waits for the viewer to respond.
- Once the viewer returns verdicts, you’ll see grounded visual-review findings for your target surface.
Choose a convergence mode
Section titled “Choose a convergence mode”Convergence Gate supports two modes.
Mode A — human-in-the-loop per fix: Each Sweep finding becomes one sandboxed fix prompt staged for your approval. You review and approve each fix before it runs. Use Mode A when the surface is sensitive or the fixes are non-trivial.
Mode B — auto-iterate to clean: Orbit auto-approves the fix prompt, waits for the agent to finish, enqueues a rebuild, reruns the Sweep reviewer, and checks the result. It keeps looping until one of these stopping conditions is met:
- The surface is clean — no more findings
- No progress — the last fix didn’t reduce the finding count
- Oscillation — findings are bouncing between two states
- Max iterations reached
When the loop stops, Orbit persists the result and surfaces a pending human-merge card instead of merging automatically. You decide whether to merge.
Start the gate from the Pipeline
Section titled “Start the gate from the Pipeline”- In the Pipeline panel, find the prompt or Sweep result you want to converge.
- Click Convergence Gate and select Mode A or Mode B.
- For Mode B, set your max-iterations limit if you want a tighter bound than the default.
- Orbit runs the loop. Each iteration appears as a new pipeline entry so you can follow progress.
Reviewing and merging the result
Section titled “Reviewing and merging the result”When the gate stops, a pending human-merge card appears in the Pipeline. Open it to:
- See the final Sweep verdict (clean or stopped-with-findings)
- Review the cumulative diff across all fix iterations
- Approve the merge or discard the run