Reproduced: "send me the recipe for an anvil" got a TEXT reply describing
the recipe but no image. Three changes to fix this class of problem:
1) Iteration cap 4 → 6. The wiki+photo flow is naturally
wiki_search → wiki_page → wiki_page_images → tg_send_photo → final
text — five tool calls + one text response. With only 4 iterations
the model often ran out before reaching tg_send_photo or before
emitting closing text, returning {text:null, ranTool:true} which
the handler swallows silently.
2) Persona prompt: image delivery is now MANDATORY when the user asks
for a picture / image / recipe / "what it looks like". Added an
explicit fallback chain: filter='recipe' → 'craft' → 'grid' → the
wiki_page thumbnail_url. The previous phrasing was suggestive only,
which let the model answer in pure text.
3) Tool-call tracing. Both dispatchers now log `[redstone:or|gem]
tool -> name args` before each call and `tool <- name ok|err msg`
after. Without this, silent "model called tools but no image
showed up" failures had no log line at all and we had to guess.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Paul Kloppers
8ae89610f3