 Matthew Meszaros
|
069ab15194
|
feat: address the second review pass on the Edit with AI fix by reading a markdown destination with balanced parentheses so a link the model normalises out of the angle-bracket form it was given comes back whole instead of truncated at the first paren with a stray one left in the copy, restoring only spaces and tabs around the model's answer so a selection that ran to the start of the next paragraph joins the two the way a paste would rather than gaining a blank paragraph nobody typed, and capping the completion at a flat 4096 tokens, above the 3072 it was and below the smallest completion limit in common use, because a request over a backend's own cap earns a 400 naming max_tokens that openAIProvider.adaptParams latches for the life of the process and degrades every later call
|
2026-09-11 21:21:40 -07:00 |
|
|
Matthew Meszaros
|
1bdd1da800
|
feat: address the CodeRabbit review on the Edit with AI fix by leaving the caret after text written at a collapsed position instead of in front of it (an insertion maps to itself unless the position associates rightwards, so continuing to type went before the insert), splitting the model's blocks one separator at a time so a blank paragraph the author used as spacing survives a rewrite instead of being swallowed by a greedy newline run, carrying a link destination that holds a paren or a space through in markdown's angle-bracket form rather than dropping the link, putting the author's boundary whitespace back on the model's trimmed answer in both hosts so a selection ending on a space does not glue the rewrite to the next word and "did anything change?" compares exactly what was written, and sizing the completion cap from the passage's own rune count so an 8000-rune CJK body is not truncated by a cap chosen for English
|
2026-09-11 20:23:53 -07:00 |
|
|
Matthew Meszaros
|
6501cb599c
|
feat: fix the "Edit with AI" rewrite in the campaign body and the unibox composer for issue #432 by running /generation/edit on a new generation.BuildEditRules system prompt through AIProvider.Complete instead of the cold-outreach writer prompt that redefined the model's role, capped it at 80 words and imposed a five-part email skeleton on every instruction, raising the completion cap so a full-body rewrite is no longer truncated at 1024 tokens and counting the request limits in runes rather than bytes, carrying merge variables, AI blocks, conditionals, form links and link destinations through the round trip in web/src/components/app/ai/richTextPassage.ts instead of deleting every atom node via doc.textBetween, replacing the passage with paste semantics so a phrase rewritten inside a sentence stops splitting its paragraph into three, saying "No change" when the model hands the passage back untouched, and clamping the floating AI card to the surface it is editing so it no longer draws outside the step drawer over the flow canvas
|
2026-09-11 20:04:32 -07:00 |
|