Riding Out a Transient API Outage

This first cut carried the classification surface without changing the run's behavior. A fresh agent could hold it because the brief kept the boundary at parsing: add the outcome class, feed captured stderr into every harness parser, and collapse the new result back to the old failure class before any caller could observe it. That made the risky part the text classifier, not the control flow above it. The sprint stayed green, including an extra gate pass after review, so the next unit inherited a proven parser contract rather than a speculative one.

With classification already landed, this sprint could be a contained swap inside the single agent runner. Its brief asked for the former collapse shim to become a retry loop with per-attempt timeout, abort-aware sleeping, injectable waits for tests, and a config schedule, while keeping the transient class invisible above the runner. That is a narrow control-flow job for a clean-context agent because the callers were explicitly out of scope. The sprint did not need a fix or retry pass, so the stress the brief named was absorbed by the runner tests and config parsing.

The last sprint was intentionally thinner: the behavior already existed, so the unit only had to make waiting observable. The brief used the event channel as the boundary, adding one retry event and wiring the build context to emit it while leaving utility one-shots quiet. Because the front-ends already consumed the same run event stream, this was a propagation task across history, headless output, TUI state, and docs rather than another change to retry semantics. It also stayed green, matching that lower-risk shape.