Previous wiki_recipe only returned the page thumbnail, so "send me the
recipe image" delivered a picture of the item, not the recipe itself.
minecraft.wiki renders recipes as live HTML grids, not standalone PNGs,
so we now generate the image ourselves.
mcp/lib/minecraft-wiki-tools.ts:
- parseCraftingGrid(html): pulls the 3×3 input grid + result item URL
from <span class="mcui mcui-Crafting_Table">. Each <span class="invslot">
is either empty (chunk starts with </span>) or has an <img src="/images/
Invicon_X.png?h">. Reading-order chunks → grid; first non-null after
the 9th input slot → result.
- composeRecipePng(grid, result): fetches each unique sprite, resizes
to 64×64 via sharp (nearest-neighbour to keep pixel art crisp),
composites onto a #8b8b8b inventory-style background with #6f6f6f
slot squares + dark borders, places an SVG-rendered "→" arrow,
then the result cell next to it. PNG buffer returned.
- New tool wiki_recipe_image(title, chat, caption?) that runs the
pipeline end-to-end and uploads via tg.sendPhotoBytes — one tool
call, image already in the chat when it returns.
mcp/lib/types.ts:
- TgClient gains sendPhotoBytes({ chat_id, bytes, filename?, caption?,
parse_mode?, reply_to_message_id? }). Multipart/form-data POST to
sendPhoto, since Telegram's URL-mode can only fetch public URLs and
the composited PNG only lives in memory. 30s timeout (vs the 20s on
the JSON call) because uploads are bigger.
mcp/package.json + bot/package.json: add sharp ^0.33.5. mcp/ pulls it
in too so /mc/mcp/lib resolves it at the bot's runtime (the bot's
node_modules is at /app while the lib files live at /mc/mcp/lib).
Persona prompt: wiki_recipe_image is now the FIRST call for any
"recipe" / "send me the recipe image" request. Single tool turn ends
with the photo already in the chat. wiki_recipe stays as a fallback
when the page has no 3×3 crafting widget (smelting / brewing / smithing
use different mcui classes).
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Adds a new MCP tool that parses minecraft.wiki's <table data-description=
"Crafting recipes"> directly, pulling the ingredients column out of each
row (e.g. "Block of Iron + Iron Ingot" for Anvil), then chases up the
page's main item thumbnail via prop=pageimages so the bot can deliver
text + a real image in one tool round.
Why a dedicated tool: minecraft.wiki does not host standalone "recipe.png"
files — recipes are rendered live as HTML grids composed from individual
BlockSprite icons. Filtering wiki_page_images for 'recipe' / 'craft' /
'grid' returned empty for every craftable item, leaving the model with
nothing to call tg_send_photo with. Confirmed by inspecting the Anvil
page's image inventory (Anvil1-6.png + GUI + sound files, no recipe).
Tool returns {
title, page_url,
recipes: [{ ingredients: "…" }, …], // strips <br>/<a>/ cleanly
thumbnail_url, // for tg_send_photo
note?, // present when no recipe rows
}
Persona prompt tightened:
- wiki_recipe is now the FIRST call for "how do I craft X" / "recipe for X"
- wiki_page_images is downgraded to "specific non-thumbnail images only"
- IMAGE DELIVERY MANDATORY clause: must end the turn with tg_send_photo
using a real URL or say plainly that no thumbnail was found — no more
describing-the-image-in-prose-and-stopping failures.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Redstone can now query the official Minecraft wiki and deliver
crafting recipe / item images straight into the chat.
New: mcp/lib/minecraft-wiki-tools.ts exports three tools, all hitting
https://minecraft.wiki/api.php (MediaWiki API, formatversion=2):
- wiki_search(query, limit?) — opensearch top results, namespace 0,
returns { title, description, url } for each hit.
- wiki_page(title, thumb_size?) — extracts|pageimages → intro extract
(plain text, ~600 chars) + thumbnail_url + page_url.
- wiki_page_images(title, filter?, limit?) — generator=images +
prop=imageinfo (iiprop=url|mime|size) returns every image embedded
on the page along with its direct https URL. The `filter`
substring is case-insensitive against the filename — pass 'recipe'
/ 'craft' / 'grid' to surface crafting-recipe diagrams.
Bot wiring (bot/bot.ts):
- import minecraftWikiTools, merge into allTools so it flows through
the existing OpenRouter + Gemini function/tool schemas automatically.
- Persona prompt teaches the model the canonical flow:
wiki_search → wiki_page → wiki_page_images(filter='recipe') →
tg_send_photo(url, caption). Composes with the already-existing
tg_send_photo tool — no new telegram tool needed.
User-Agent on every wiki request, 15s timeout, JSON-only.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Two related changes to how Redstone behaves in chat.
Streaming: replies now animate word-by-word into Telegram using the new
sendMessageDraft endpoint (Bot API 9.5, March 2026) via the @grammyjs/stream
plugin, with @grammyjs/auto-retry on the API layer to swallow 429s
transparently. The previous editMessageText-based approach is gone —
sendMessageDraft is designed for this and animates natively on the client
without hitting the 1/sec-per-chat edit limit. Pace lives in
STREAM_WORD_DELAY_MS=50; tunable in one spot.
server_down footgun: the MCP tool was running `compose down` with no
service arg, which tore down the whole project — including the bot
container running the tool call, which then got SIGTERM mid-conversation
and didn't come back (unless-stopped doesn't restart on a clean compose
down). Behaviour now matches the existing /stop slash command:
compose stop minecraft autossh, leaving the bot up. Pass an explicit
service to stop just that one.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Redstone previously only saw the 23 tg_* tools — it had no idea the
Minecraft stack existed, so questions like "is the server up?" went
unanswered. This change extracts the 15 Minecraft tools (lifecycle,
rcon, backup, players, seen) into the same shared catalog the Telegram
tools already used, so both Gemini and external MCP clients see them.
mcp/lib/types.ts (new) holds the shared shape: Tool<P>, ToolCtx,
createTgClient, createMcRuntime (a Bun.spawn-based wrapper for docker
compose / docker exec), and toolToGeminiFunction (zod → Gemini schema,
now also stripping exclusiveMinimum/Maximum since Gemini rejects them).
mcp/lib/minecraft-tools.ts (new) is the catalog itself. Eight handlers
are flagged requiresAdmin: rcon, backup, and all destructive player_*
writes plus seen_list. mcp/server.ts trusts the caller (Claude / Paul
on the host) and ignores the flag; bot/bot.ts honours it at dispatch
time, returning {error: "...requires admin role..."} to Gemini so it
can explain to the user instead of attempting the call.
mcp/server.ts shrinks from 423 lines to 70 — a single loop over both
catalogs replaces the hand-rolled registrations.
bot/bot.ts wires both catalogs into the function declarations and adds
the admin gate. It also gains a defensive re-check on every incoming
group/DM text: the Redstone handler now does its own lookup of
ctx.from.id against the users table and refuses to reply unless
status='active'. This is belt-and-braces — the auth middleware already
short-circuits unauthorized callers earlier in the chain, but with
privacy-mode-off groups now feeding every message through, a future
refactor that reorders middleware shouldn't be able to make Redstone
respond to strangers.
Drops thinkingConfig from the Gemini call (gemini-2.5-flash-lite
rejects it outright with HTTP 400).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Drops two new capabilities on top of the existing Telegram bot + Minecraft
stack, both built around a shared toolkit so there's one source of truth
for the actions either side can take.
mcp/ — a Bun-native MCP server exposing 38 tools to MCP clients (Claude
Code, etc.): server lifecycle, rcon, backup, player/db CRUD, plus 23
Telegram Bot API methods (messaging, reactions, polls, dice, photos,
stickers, pins, forum topics, chat config). Runs over stdio.
mcp/lib/telegram-tools.ts — the Telegram tool catalog as Zod-typed
handlers. Imported by both mcp/server.ts (registers each as an MCP tool)
and bot/bot.ts (exposes each as a Gemini function declaration), so adding
a tool in one place lights it up everywhere.
bot/bot.ts — replaces the silent-on-unknown-text behaviour with Redstone,
an in-bot persona driven by gemini-2.5-flash-lite with native function
calling. In DMs it always responds; in groups only when @-mentioned or
replied to. The tool-use loop (max 4 rounds) lets it decide to send a
poll, react with an emoji, roll dice, etc. via the shared handlers rather
than just text. Thinking budget zeroed and system prompt locked down so
the model doesn't leak its reasoning into replies.
docker-compose.yml — adds google_key as a docker secret and passes
GEMINI_API_KEY_FILE + GEMINI_MODEL to the bot.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
