Repository navigation
Live model: room to finish the brief, and a guaranteed fallback - #6
Merged
Merged
Conversation
- MapperFace reads reduced motion through useSyncExternalStore and derives the blink instead of resetting state inside an effect. - WorkflowChat sends the intent per request, opens the brief from onFinish, and tracks streamed chunks during render instead of in an effect. - factsAndAssumptions de-duplicates evidence lines, so React keys stay unique and the brief no longer repeats a line. - Rendered-HTML tests now assert the current greeting and the data: URL download path that replaced blob URLs. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DbmD6cAA88NEhhUnPW3ew9
The route parsed the raw body with req.json() and forwarded whatever arrived: any role (system included), any part type, remote file URLs and unbounded sizes. Malformed JSON surfaced as a 500. - lib/optimize-request.ts checks content type, body size (4 MB), message count, text length, file count, roles (user/assistant), part types, and inline data: attachments against one allowlist shared with the composer. Assistant history keeps only text parts. Rejections are 400/413/415 with a message the chat shows to the user. - The composer normalizes attachment media types, refuses unsupported images and caps PDFs and images at 2.5 MB before upload. - tests/optimize-request.test.mjs covers the UI's real payloads, injection of system turns, SSRF-style remote URLs, mislabelled binaries and size limits. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DbmD6cAA88NEhhUnPW3ew9
- GitHub Actions runs lint, then tsc, next build and node --test, with no model key. - README states what is live (demo mode), what is built but off (live model), what the tests cover and what they do not. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DbmD6cAA88NEhhUnPW3ew9
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DbmD6cAA88NEhhUnPW3ew9
The earlier capture came from next dev and showed development issue badges. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DbmD6cAA88NEhhUnPW3ew9
gpt-5-mini spends reasoning tokens against maxOutputTokens, so the 1,400 cap could end the reply before the fenced JSON brief closed. temperature was also set on a reasoning model that ignores it. - lib/agent-stream.ts streams every reply: 6,000 output tokens, low reasoning effort, one retry. If a model fails before its first token, the next model is tried; if all fail, the demo brief streams with a one-line notice. Output is buffered until the first token, so a failed attempt leaves nothing behind. - tests/agent-stream.test.mjs covers success, model fallback, demo fallback, no-key demo and a mid-stream failure, using the SDK's mock model. Verified against the production build with an unreachable provider: both models fail and the demo brief arrives with the notice. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DbmD6cAA88NEhhUnPW3ew9
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 8b9e6081e3
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
Each file passed the 2.5 MB cap, but two 1.55 MB PDFs encode to about 4.13 MB together and the API guard rejected them with 413. The composer had already cleared the files, so the advice to remove one could not be followed. Earlier turns' attachments are re-sent on every turn and counted too. The composer now sums inline attachment data across the conversation and the new files before clearing anything, and keeps the files in place with a message when they don't fit. A test pins the budget below the API body limit. Reported by Codex review on #6. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DbmD6cAA88NEhhUnPW3ew9
This branch was successfully deployed
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
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.
Makes the live-model path safe to switch on. Merge after #5.
Why:
gpt-5-minicounts reasoning tokens againstmaxOutputTokens. At the old 1,400 cap the reply could stop before the fenced JSON brief closed, so the brief would never render. Any provider error also left an empty reply.What changed
lib/agent-stream.tsstreams every reply, demo or live. The live path gets 6,000 output tokens, low reasoning effort and one retry.gpt-5-mini, thengpt-4.1-mini, then the deterministic demo brief with the notice "The live model is unavailable right now, so this is the rule-based demo brief." Chunks are buffered until the first real token, so a failed attempt leaves nothing half-written.Verification
npm run lint: clean.npm test: 51/51.tests/agent-stream.test.mjsuses the SDK mock model to cover success, model fallback, demo fallback, no-key demo and a mid-stream failure.Not checked: a real OpenAI response. There is no key in this environment. After adding
OPENAI_API_KEYin Vercel, run one starter and confirm the brief renders.Before turning the key on
Set a hard monthly budget on the OpenAI project. The site has no rate limit yet.
Generated by Claude Code