feat(grok): add image command for grok.com image generation

[flizzywine]

Summary

Adds opencli grok image <prompt>, a new command that submits a prompt via the existing grok.com browser session and returns the generated image URLs from the latest assistant message bubble. Optionally downloads the images to a local directory.

This is a natural complement to the existing grok ask command: the current ask only extracts innerText from the latest bubble, so when Grok's response is an image the command silently returns an empty / truncated string. grok image scrapes <img> elements from the same bubble instead.

Design notes

• Submission path reuses the same composer interaction as ask.ts — it prefers the ProseMirror composer (.ProseMirror[contenteditable="true"]) when available and falls back to the legacy <textarea> flow otherwise.
• Stability detection: polls readLastBubbleImages(page) every ~3s and returns once the set of image URLs has been stable across two consecutive reads. This mirrors the ask stabilization logic.
• Downloading: assets.grok.com/users/.../image.jpg is gated by Cloudflare; plain HTTP clients (curl, node-fetch) receive HTTP 403. When --out <dir> is provided, the command performs the fetch inside the page via fetch(url, { credentials: 'include', referrer: 'https://grok.com/' }), converts the blob to base64, and writes the decoded bytes to disk from Node. This is the minimal-viable way to bypass the auth without juggling cookie jars.
• Avatar/UI filter: images with naturalWidth < 128 are dropped so UI chrome inside the bubble doesn't pollute the output.

Flags

Flag	Default	Description
prompt (positional)	—	Image generation prompt
--new	false	Start a fresh chat before sending
--timeout	240	Max seconds to wait
--count	1	Minimum images to wait for before returning
--out	""	Directory to save downloaded images (triggers in-page fetch)

Columns: url, width, height, path.

Test plan

• npx tsc --noEmit — clean
• npx vitest run --project unit — 579 passed, 1 skipped (no regressions)
• npx vitest run clis/grok/image.test.ts — 9 passed (new tests for isOnGrok, normalizeBooleanFlag, dedupeBySrc, imagesSignature, extFromContentType, buildFilename)
• Manual end-to-end smoke test against a live grok.com session: opencli grok image "a cyberpunk mechanical owl, neon purple and blue" --new true --out /tmp/grok-img --timeout 300 --format json — returned a path pointing at a valid 784×1168 JPEG on disk.

No new runtime dependencies.

[hiSandog]

I think sendPrompt() here is a bit less robust than the existing Grok web flow in clis/grok/ask.ts. ask.ts's sendPromptViaExplicitWeb() waits/retries for the ProseMirror composer and for a visible enabled Submit button; this new code does a single DOM query and immediately falls back to textarea.

On a cold session right after goto(GROK_URL) / tryStartFreshChat(), the composer often is not mounted yet but appears a second later, so grok image can return [BLOCKED] send failed: no composer even though the session is healthy. I would reuse the same readiness loop / visible-button check here so grok image behaves more like grok ask --web true.

[hiSandog]

I think clis/grok/image.ts needs the same baseline/new-bubble guard that the explicit web flow in clis/grok/ask.ts already has (getBubbleTexts() + pickLatestAssistantCandidate()).

Right now the image loop polls readLastBubbleImages() against whatever the current last bubble is. In an existing chat that already has a previous image reply, if sendPrompt() does not stick immediately or Grok takes a moment to append the new user/assistant bubbles, the recorder can stabilize on the previous assistant image set and return stale URLs after ~6s.

I would capture a baseline bubble count/signature before sendPrompt(), then only accept images from bubbles that appeared after that baseline. That keeps grok image --new false aligned with the stale-response protection that grok ask --web true already uses.