A fourth repair attempt needs a new decision, even when the assistant says the fix is obvious. Without that boundary, a small form change can become an open-ended negotiation between an implementer and a reviewer. Each response sounds productive. The person responsible for accepting the work gets another revision and a less reliable account of what remains broken.
I would give this workflow at most three repair-and-retest rounds after the initial review. Three is a chosen operating limit, not a measured optimum or a claim about model capability. Its purpose is to put a visible end on automatic work while keeping the acceptance standard intact. A failed third round produces a handoff with evidence, not permission to ship.
Consider a hypothetical event form with a start date, an end date and a required event name. The agreed task is narrow: reject an end date earlier than the start date, preserve keyboard operation and identify the required field clearly. The assistant may edit the form component and its focused tests. Changing storage, replacing the component library or deploying the result falls outside this attempt.
The initial review examines build A and records that the form accepts an end date before its start date. The fixture uses November 10 as the start and November 2 as the end; the expected result is rejection with an understandable error. These are invented inputs and outcomes for a paper walkthrough. No application was run to produce them, and no result below establishes that an autonomous repair system works.
Before the first repair, create the ledger that will outlive the conversation. Give the attempt an identifier, name its owner and freeze the acceptance conditions. Each round needs an input build, a finding, the permitted repair, an output build, a retest result and a disposition. The initial review is round zero. It consumes none of the three repair rounds, but its evidence is the starting point for every later comparison.
Round one takes build A and the reversed-date finding. The proposed repair adds the missing comparison, producing build B. In this hypothetical ledger, the retest rejects the reversed dates and still accepts an ordinary valid interval. The keyboard walkthrough then exposes a separate problem: after a validation error, focus moves through the controls in an order that makes correcting the form confusing. Record that finding against B rather than describing the whole form as fixed.
The keyboard finding needs a concrete expectation. W3C’s explanation of Focus Order says sequential navigation must preserve meaning and operability where the sequence affects them. It does not require every form to follow one universal visual order. The review packet should therefore record the actual sequence and how it obstructs this task, rather than treating a preferred arrangement as a standards violation.
Round two addresses that finding and produces build C. Its hypothetical retest preserves the date check and restores a usable keyboard sequence. Inspection also finds that the required event-name field has visible text beside it but lacks the agreed programmatic label association. That becomes the next blocker. The ledger retains the passed checks and their build identifiers, while making clear that C has not passed the complete acceptance set.
This distinction matters when an assistant offers a cosmetic shortcut. Making nearby text bolder does not establish that the control has a usable label association. W3C’s form-labeling tutorial describes explicit association through matching label and control identifiers, alongside other labeling approaches. The chosen implementation still needs inspection. A reference to the tutorial is support for the technique, not evidence that the current form uses it correctly.
Round three attempts the label repair and produces build D. In the worked ledger, the identifiers still do not match, so the required-label defect persists. The date and keyboard checks remain recorded as passing in this hypothetical retest, but the overall disposition is blocked. The automatic attempt ends. There is no fourth row labeled “quick cleanup,” no new conversation that resets the counter, and no promotion of the unresolved label to optional polish.
The counter must describe work that happened, including work that ended badly. Record the start of a round before authorizing its patch, then attach the output and retest when they exist. If execution stops after an edit but before verification, the round remains consumed with an incomplete result. Reopening the session should recover that state. Otherwise, every interruption becomes an accidental extension of the repair budget.
Three rounds are a ceiling, not a quota to spend. Stop sooner when the same blocker survives an attempted repair, when the next change exceeds the permitted files or effects, or when the reviewer cannot identify which build was tested. An accepted result can also end the loop early. In the example, the label problem first appears at C and survives its attempted repair at D; both persistence and the round limit support the stop.
A reviewer can discover more work without authorizing it. Suppose build B also reveals that the submit button’s wording could be shorter. Record the suggestion separately if the existing wording still satisfies the task. That preference does not reopen a passed check or buy another repair round. Correctness and accessibility failures retain their blocking status because the attempt ran out of room, even if the remaining change looks small.
The owner now receives build D, the preserved diff, the label reproduction, the completed checks and the exhausted round count. They can defer the change, narrow the task through an explicit new decision, or authorize another bounded attempt with a reason. A new attempt references the old ledger and carries forward its unresolved findings. It does not erase the failed history or imply that a different prompt will solve the defect.
This approach has a cost: some easy fixes will wait for a person who could have let the assistant continue. That is a deliberate tradeoff when review effort and scope are part of the task’s constraints. The ledger makes the cost inspectable without inventing a savings estimate. The useful output of the third round is a specific remaining defect on a specific build, delivered to someone who can decide what happens next.
Use the W3C form-labeling tutorial to define one concrete check for the next form you review.
If this was useful, forward it to one engineer who needs less noise in their feed.


