The scenario
The current retry loop sometimes calls the operation four times. Existing retryable-error classification must remain unchanged.
Required context: a selected repository matching this scenario, relevant task evidence, and access to its instructions and checks. Use redacted or synthetic data where appropriate.
Adapt the complete prompt
Edits stay in this browser tab. Saving bookmarks the original example; export your edited version to keep it elsewhere.
Why this is a strong example
Defining total attempts avoids the common ambiguity between retries and attempts.
A concrete outcome
Enforce the documented maximum of three total attempts, including the first attempt.
The result can be assessed against an observable task rather than prompt length or confidence.
Boundaries and a method
Reproduce with a counted fake operation and fix the boundary narrowly. Preserve cancellation and non-retryable-error behavior; do not change timeout policy.
This leaves implementation judgment while constraining the changes that would exceed the task.
Evidence that can disagree
Assert exact call counts for immediate success, repeated retryable failure, non-retryable failure and cancellation.
These checks describe what would make acceptance justified; asking for them does not mean they ran.
A safe way to encounter uncertainty
If the repository contradicts these assumptions, required access is unavailable, or a consequential product decision is missing, explain the conflict and pause the dependent work. Preserve unrelated changes.
Consequential uncertainty stays visible rather than being converted into an invented requirement.