Add vision: pipe attached Slack images into the LLM

[scarolan]

Both Ollama and Gemini take images natively. Slack messages with file attachments arrive on the event payload. ~30 lines of glue unlocks a whole feature class.

Use cases

• "Data, what's in this screenshot?"
• "Data, what's broken about this stack trace?"
• "Identify this Star Trek species."
• "Translate the text in this photo."

Plan

1. In the message handler, check message.files for image MIME types (image/png, image/jpeg, image/webp).
2. Fetch each via app.client.files.info + the URL with the bot token, base64 encode.
3. Extend the canonical message shape to allow images: [base64] on user turns.
4. Adapter translation:
   • Ollama: native — { role: 'user', content, images: [...] }
   • Gemini: parts: [{ text }, { inlineData: { mimeType, data } }]
5. Persist message text only in convoStore (not the image bytes) — assistant's reply implicitly captures what was discussed.
6. Document in BOT_PERSONALITY that Data can see images.

Out of scope (for now)

Image generation in mid-conversation (covered by tool calling once #1 lands).

[scarolan]

Shipped on main as part of the work merged via PR #30.

Slack image attachments are extracted in app.js::extractMessageImages, base64-encoded, and attached to the user turn in the canonical chat shape as images: [{ mimeType, data }]. Adapters translate to Ollama's message.images and Gemini's inlineData parts. Image bytes are not persisted to convoStore.

Working live against gemma4:31b — verified with a meme: correctly identified hammer/sickle symbology and the joke premise unprompted. See CLAUDE.md "Model selection notes" for lessons learned (Slack subtype: 'file_share' gotcha, observability logging added).