Fixed2026-05-29

Long pages no longer time out

You can now analyze long pages with dense concept coverage and get back a complete report. Previously these analyses failed before producing writing guidance, with no recovery path.

Problem

Pages with twenty or more surfaced concepts, explainers, listicles, comparison guides, failed at the writing-guidance step with "Analysis failed: The analysis output exceeded the response limit." The analysis never reached the report, never appeared in history, and consumed a credit that had to be refunded. Even on pages that did finish, the brief sometimes included tasks for concepts that were already well-covered, working against its positioning as the gap-closing agenda.

Context

The three Phase 2 stages, the framework, the writing brief, and the architecture map, each had no upper bound on what they could emit. On dense pages the model wrote past the response limit and the entire run was discarded. The brief was also given every framework concept without distinguishing the gaps from what the page already handles well, so guidance went to concepts that did not need any.

What changed

Every Phase 2 stage now has a hard ceiling on what it emits and trims deterministically when the model overshoots: at most 25 concepts in the framework (anchor first, then by priority), 40 relationships, 5 items in each writing-guidance bucket. The brief is now built around the gap: well-integrated concepts are shown to the model for context but excluded from the actionable buckets. A short line at the top of the writing guidance calls out how many concepts are already well-integrated, so the brief stays focused on what needs work. Analyses that previously failed now complete on the same input.