Conversation
Server-side resizing used fit: 'cover', which crops to the requested box and fights the rendered element's CSS object-fit (e.g. a single ?width= param returned a distorted sliver). Switch to fit: 'inside' so images scale proportionally, and extract an isResizableBitmap helper to gate the sharp path.
A failed agent run (e.g. provider out of credits) only set a transient store-level error, so the turn was saved as an empty assistant message and reloaded transcripts gave no clue why the AI never replied. Record the humanized error on the turn itself so it persists to ai_chats and renders in the bubble after reloads.
- Add Grok 4.5 and Grok 4.3 to the model picker with cost estimates - Reuse the OpenAI Chat Completions provider against the x.ai base URL - Validate xAI keys via /v1/api-key: per-endpoint ACLs 403 the models list even for keys that work for chat, and blocked teams (usually missing credits) get an actionable message - Show key verification errors in an alert with a Save anyway button instead of a cramped inline row
Replayed as history, the review instruction ("reply with at most one
short sentence", "do not make changes for the sake of it") reads like a
standing order — weaker models then answer later requests with "Looks
good — no changes needed." instead of doing the work. Review turns are
self-contained, and their fixes reach the next turn via the fresh page
snapshot injected per request.
Vercel kills the chat function at maxDuration without running catch/finally, so long builds were silently truncated mid-stream with no error and no change snapshots. Raise the route limit to 800s (Pro GA maximum) and stop the tool loop at a 740s wall-clock budget so runs end cleanly: snapshots and usage are emitted and the user gets a resumable "ran out of time" error. Also fix the vercel.json functions glob, which was missing the (builder) segment and matched no routes.
- Add Fable 5 ($10/$50) and Opus 5 ($5/$25) with cost estimates - Make Opus 5 the default model - Mark Opus 4.8 legacy: hidden from settings and pickers unless the project's stored allowlist already includes it
Current Claude models run adaptive thinking at high effort by default, spending most of a turn reasoning silently while thinking bills as output tokens — one Fable 5 build burned the full 8192-token cap in a single truncated 91s turn. Request medium effort via output_config (visual building doesn't need deep reasoning, matching the OpenAI provider's low reasoning_effort) and raise DEFAULT_MAX_TOKENS to 16384 so thinking plus large batch payloads fit.
Replace the static "Working…" with rotating phrases and a ticking elapsed-time counter, so long silent thinking stretches still show visible progress between tool rows.
- Scale tall page captures to fit vision-model size limits (Anthropic rejects sides over 8000px), which failed the self-review on long pages - Explain 413 responses as attached images being too large instead of surfacing a bare status code
ysamcode
requested review from
liamwalder,
lunenas and
tristan-mouchet
as code owners
July 28, 2026 06:05
Typing a fraction like 1/2 in the width setting previously produced an invalid arbitrary class (w-[1/2]). Fractions now map to native Tailwind utilities (w-1/2, min-w-5/12), and whole fractions like 12/12 normalize to the full keyword (w-full) since Tailwind no longer ships n/n classes.
MCP instructions and the design reference did not mention fraction widths, so agents had no way to discover them. Also add common width fractions to the class autocomplete suggestions.
…lete" This reverts commit f4580d1.
MCP instructions and the design reference did not mention fraction widths, so agents had no way to discover them. Also add common width fractions to the class autocomplete suggestions.
The post-build screenshot critique nearly always returns "looks good" now that main models produce solid first passes — and the reviewer is deliberately a cheaper, weaker tier that mostly rubber-stamps stronger builders' work. Each pass still cost a full extra agent turn (system prompt, tools, full-page screenshot) and 10-20s of "Reviewing the result" after the build already finished. Old chats keep working: the review flag, compact rendering, and model-history exclusion stay, and providerOfModel still resolves the legacy review-only model ids found in persisted turns.
feat: support fraction widths like 1/2 and 12/12
bgGradientVars was the only var channel wired to the backgroundImage design property, so image URLs passed the same way never displayed. Handle bgImageVars alongside it and wrap bare URLs in url(...) so the values are always valid CSS.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Add xAI Grok as a bring-your-own-key AI provider and refresh the Claude model lineup with Opus 5 (new default) and Fable 5. Also includes agent reliability and UX improvements: graceful run timeouts, faster responses, removal of the low-yield post-build self-review pass, persisted turn errors, and a live status indicator in the chat — plus fraction width support in sizing inputs.
Changes
AI providers & models
Agent reliability & performance
UI
Builder
1/2,2/3) in sizing inputs, surfaced to AI agents and autocompleteOther fixes
bgImageVars), instead of only setting the CSS variableTest plan
1/2in the sizing panel and confirm it applies