v0.1.10 — Never Leave a Doc Behind

Author

🗓️ July 2026

Free and overloaded models sometimes return empty responses without signaling an error. Previously, Aether trusted the ✓ and wrote 0-byte files — the genesis said "complete" but half the docs were hollow. This release makes generation resilient: every doc the planner decides on must be generated, or the run fails visibly.

🐛 Fixes

Empty responses no longer produce 0-byte docs. When a model returns empty content, Aether now retries up to 5 times with escalating backoff (10s, 20s, 30s, 40s) before declaring the step failed. A doc file is never written unless it has content — no more phantom docs that look generated but are empty
Failed steps no longer kill the rest of the run. The generation pool now continues generating other docs even if one step fails. Only user cancellation (ESC) stops everything. At the end, failed docs are clearly marked with ✗ so you know exactly what to retry
Cancel (ESC) is now instant during backoff waits. Previously, pressing ESC while Aether waited between retries would block until the sleep finished. The sleep now responds to the abort signal immediately

🏗️ Under the hood

Resilient worker pool — runPooled no longer uses fail-fast (Promise.all + rethrow on first error). Workers now catch errors per-step, mark them ✗, and keep processing. Abort errors still halt all workers immediately — only generation failures are tolerated
abortableSleep — a new utility that sleeps for N ms but resolves instantly when the abort signal fires, so ESC always responds within one event-loop tick even during a 40s backoff wait
Empty-content retry loop — wraps each doc generation call: checks response.content.trim(), and if empty, waits and retries with a visible spinner update (empty response — retry 2/4). After exhausting retries, throws a clear error naming the doc that failed