feat(image): switch image generation to Meta Muse Image

[NickB03]

Summary

Switches inline image generation from google/gemini-2.5-flash-image to Meta Muse Image (gateway:meta/muse-image-1.0) and reworks the tool to fit it.

Stacked on #269 (both PRs edit the same README and DECISIONS lines). GitHub retargets this to main once #269 merges.

	$/image	Generation Elo (rank)	Editing Elo (rank)
Gemini 2.5 Flash Image (before)	0.039	985 (#50)	987 (#45)
Muse Image 1.0 (after)	0.010	1112 (#7)	1116 (#6)

Source: Artificial Analysis image arenas, 2026-09-18.

What changed

• API: Muse is a Gateway image-type model, so the tool calls the AI SDK's generateImage instead of generateText + result.files. Added getImageModel to lib/utils/registry.ts.
• Aspect ratios: Muse ignores aspectRatio (the Gateway warns "use size instead"). The tool maps each ratio to a size hint. Muse snaps output to about 2.4 MP at that ratio.
• Editing: the source image is passed as an http(s) string, which the SDK forwards as a URL file for the Gateway to fetch. The ownership check and signed-URL flow are unchanged, and this server still never downloads model-supplied URLs.
• Telemetry: generateImage has no telemetry option, so the tool annotates the enclosing tool span (image.model, image.is_edit, image.size), the same pattern the search tool uses.
• No-image case: NoImageGeneratedError maps to the existing "No image was generated" tool error.
• Output format: WebP. The storage bucket already allows image/webp; files are 60–330 KB against a 5 MB limit.

Verification

• bun lint, bun typecheck, bun format, bun run test (193 files / 1650 tests), and graph freshness all pass.
• Live against the Gateway with the same calls the tool makes:

Requested	size sent	Output
1:1	1024x1024	1600x1600
16:9	1792x1024	2016x1152
9:16	1024x1792	1152x2016
4:3	1536x1152	1792x1344
3:4	768x1024	1344x1792

Edit from bytes and edit from a public URL both succeeded; a visual check showed the edit recolored the subject and kept the scene. Latency was 9–19 s per image.

Not verified locally (check on the preview)

Local Supabase was not running, so the upload and signed-URL steps ran only under unit tests:

• Generate an image in a signed-in chat and confirm it renders and downloads as .webp
• Ask for an edit of that image (exercises Gateway fetch of a Supabase signed URL)
• Generate as a guest

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
polymorph	Ready	Preview	Sep 18, 2026 4:08pm UTC

[coderabbitai]

Important

• 🔍 Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Advanced

Run ID: d2cde6dc-3a5a-4c96-8632-6d5a2ca8b6bd

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-09-18T16:08:28.640601Z	d7f06e2	PR opened

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" or "@codex security review".

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]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: d7f06e2b24

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".

[chatgpt-codex-connector]

Refresh shifted graph source locations

The graph refresh adds the new symbols but leaves existing lib/tools/generate-image/server.ts nodes at their pre-change positions: ImageToolContext, resolveSourceImageUrl, createGenerateImageTool, and serverTool still point to L19/L36/L61/L149 even though this commit moves them to L38/L55/L80/L168. Consequently, Graphify queries return incorrect source locations for this module; regenerate or repair all affected node metadata, not only the newly introduced nodes.

AGENTS.md reference: AGENTS.md:L122-L122

Useful? React with 👍 / 👎.