Refactor canned triggers into native tool calls

[scarolan]

The regex matchers in lib/responses.js (isImageRequest, isDanceParty, isPodBayDoor, isLoveYou, isRickroll, isTikTok, dad-joke trigger) are 2018-vintage and increasingly out of place now that Data runs on native Ollama/Gemini SDKs with proper tool support.

Idea

Define a small set of tools and let the model decide when to call them:

• generate_image(prompt) — calls Gemini Nano Banana
• tell_dad_joke() — fetches from icanhazdadjoke.com
• start_dance_party() — emits the emoji rain
• show_help() — returns the help text
• state_asimovs_laws() — Asimov rule recitation
• play_rickroll() / play_tiktok() — the button blocks

Both Ollama (tools: [...] → message.tool_calls) and Gemini (config.tools with function declarations) have native tool calling. Translation happens inside lib/chat-backends.js so handleMessage only deals with the canonical shape.

Biggest immediate win

The IMAGE_REQUEST_GUIDANCE shim where Data has to say "please use the /image command" instead of just generating the image. That whole interaction goes away — user says "Data, picture of a sehlat", Data calls the tool, image appears in the channel.

Plan

1. New lib/tools.js — tool definitions (JSON schemas + implementations).
2. Extend the chat-backend adapter interface: chat({ messages, tools }) → { text, toolCalls? }. Adapters translate to/from each provider's native shape.
3. handleMessage becomes a small loop: call the model → if toolCalls, execute them and feed the results back as a tool-role message → repeat until plain text. Cap iterations to prevent runaways.
4. Delete isImageRequest/IMAGE_REQUEST_GUIDANCE, the love-you/pod-bay/danceparty/rickroll/tiktok matchers and their call sites in app.js. Keep the rules, help, dad joke matchers as fallbacks for now (or convert them too).
5. Tests: adapter translation of tool calls for both Ollama + Gemini, and a fake-tool-calling chat that exercises the multi-turn dispatch loop in handleMessage.

What this unlocks downstream

Future tools (web search, weather, Slack user lookups, etc.) just slot into lib/tools.js — no new dispatch code, no new regex.

[scarolan]

Built, shipped, then reverted. Won't pursue further with the current model lineup.

What we learned:

1. gemma4 (Data's preferred backend for its strong vision) cannot reliably emit Ollama's structured tool_calls field. It writes calls as inline text like <call:generate_image{prompt:"..."}><tool_call|> instead. Same behavior on gemma4-openhands:latest (the agent fine-tune of the same family), so it's a family-wide wire-format / template issue, not a prompting issue.

2. 6 of the 7 tools added in PR Tool calling, vision, thinking traces, reaction UX, threaded replies #30 duplicated regex matchers that already worked. tell_dad_joke, state_asimovs_laws, show_help, start_dance_party, play_rickroll, play_tiktok — all of these were already exact-phrase triggers that were faster, more reliable, and model-agnostic. Adding tools for them was solving a non-problem.

3. The one tool that genuinely benefited from LLM intent extraction — generate_image — was the unreliable one (gemma flubs the structured call). And /image slash command already handles image gen cleanly with no LLM round-trip.

4. The reliable alternative (qwen3.6:27b for tools) isn't vision-capable. Picking "tools work but no vision" was a worse trade than "vision works, no tools."

Outcome: tool calling code (lib/tools.js, dispatch loop, adapter translation) removed. IMAGE_REQUEST_GUIDANCE restored — Data politely directs users to /image when they ask for an image in chat. ~800 lines of code retired for net-positive reliability.

If someone ever wants to revisit: pull a model that does both vision and tool calling well (e.g. qwen2.5-vl:32b or mistral-small3.1) on kepler.local first, verify it emits structured tool_calls against a 2-tool test registry, then re-wire. Don't ship tools without that pre-check. Notes captured in CLAUDE.md "Model selection notes".