Aider keeps looping? Bound the task and verify one outcome.

This guide covers repeated Aider actions that are not driven by the same lint or test error. Capture the repeated action, stop condition, and last useful repository change before starting again.

Do this first: Pause after one repeated action that produces no new evidence. Save the current diff, note the last completed step, and define one observable result for the next run.

What the failure looks like.

Aider continues exploring, proposing, or applying similar changes without reaching a stable result or satisfying a stated completion condition.

Why it happens.

A broad goal without a stop condition can leave progress difficult to judge. This page is not the lint-specific case; repeated lint/test feedback has its own focused diagnostic guide.

Fix it without losing the task.

Record the goal, last completed action, and observable output from at least two repeated attempts.

Inspect git status and the diff to find the last meaningful change. Separate useful edits from repeated or reverted work.

Narrow the next request to one file or decision and state the check that will end the task.

If the same lint/test diagnostic is causing the repetition, use the lint-specific guide. Otherwise, verify the one bounded action before expanding scope.

What the next AI should receive.

The original goal, the repeated action and its output, current diff, decisions that still hold, approaches not to repeat, exact completion criterion, and one smallest next check.

First safe action

Stop after the repeated action yields no new result. Preserve the last useful diff and write down what the next attempt must prove.

Evidence boundary

Issue #3614 is one Aider report and is also referenced by the lint-loop page for context. It is not an independent second incident or evidence of prevalence.

Diagnostic checkpoint

If the same linter output is driving each edit, switch to the lint-specific guide; this page addresses broader non-converging actions.

How ShardStitch helps.

ShardStitch can turn the current goal, observed changes, failed approaches, and next check into a bounded handoff. It cannot cap Aiders runtime or decide whether an open-ended plan is complete.

Source trail.

Aider issue #3614 and Aider usage documentation.

FAQ.

Does a long Aider run prove it is stuck?

No. Compare observable progress, changed files, and the stated completion condition. A run is not necessarily stuck simply because it takes longer than expected.

When should I use the Aider lint-loop guide instead?

Use that guide when the same lint or test result is repeatedly returned after edits. This page is for loops without a single recurring check result driving the cycle.

Scope note: This is a task-control workflow for an Aider run that is not converging. It does not claim every long run is a product defect; the evidence cited is a single report.