How to do looping

Any edge pointing back at an earlier node is a loop. The work is in stopping it. This graph reviews an answer, sends rejections to a repair turn, and gives up after five unsuccessful rounds instead of arguing with itself forever.

Prerequisites

An authenticated harness and a local Git project folder.

review passes to done or rejects into repair; repair returns to review. Five unsuccessful repair rounds lead to exhausted.

How the bound works

Three edges leave review and all three name the same repair_budget="answer". The retry edge starts a round. The exhausted edge carries the same condition and a lower weight, and runs once after five unsuccessful rounds. The reset edge on the passing verdict clears the budget so a later episode starts from five again.

Every retry edge needs an exhausted partner with the same explicit condition, the same budget and a different weight. Put both after the review that decides, never on a fan-out or inside a parallel branch. Graph and node visit limits stay independent of the budget and remain the outer guard.

Failure routing is separate. A failed stage takes a matching conditioned edge and never the ordinary fallback, so route outcome == 'failed' to a repair prompt that reads {{ context.failure_reason }} when the graph knows how to recover. Read the failure reason first: an authentication failure needs credentials, not another attempt.

Save, validate and run

Download all files. Create a local Git practice repository. It needs no remote:

mkdir -p ~/orbital-bounded-retries/.orbital
git init ~/orbital-bounded-retries
cd ~/orbital-bounded-retries

Extract the archive into ~/orbital-bounded-retries/.orbital/. Keep every relative path. From ~/orbital-bounded-retries run:

orbital validate .orbital/workflow.dot

Add the project folder in Orbital. On New run select the project and workflow graph, choose your authenticated harness and supply the goal. Use a separate practice project for each example so the workflow name is unambiguous.

Expected behaviour

The first passing verdict ends successfully. Rejection starts up to five repairs. A sixth rejecting review ends at exhausted with a failed outcome.

Complete source

Every DOT file below is a complete runnable graph. The archive contains the same files. Prompt files are companions, not separate workflows.

workflow.dot

workflow.dot

digraph retries {
  graph [version="1", entry="review", max_visits="30"]
  review [prompt="Review the goal and any previous repair. Choose pass or reject.", outputs="verdict:choice"]
  repair [prompt="Revise the proposed answer to the goal. Do not change files."]
  done [shape="Msquare"]
  exhausted [shape="Msquare", outcome="failed"]
  review -> repair [condition="verdict == 'reject'", weight="3", repair_budget="answer", repair_round="retry"]
  review -> exhausted [condition="verdict == 'reject'", weight="2", repair_budget="answer", repair_round="exhausted"]
  review -> done [condition="verdict == 'pass'", weight="1", repair_budget="answer", repair_round="reset"]
  repair -> review
}

Download workflow.dot

For a loop that watches something outside the run, add an insulator wait so polling spends no agent turns. See loops and error handling.