2d23dba736
The asset store now uses libs/sqlite_query as its ONLY engine — not a feature
flag, not a fallback. That closes the Windows gap (the embedded store starts
there now, and a SHARED->EXCLUSIVE upgrade is handled rather than assumed
free) and takes the C dependency out of the build everywhere else.
Around it:
- store: a garbage collector, catalogued content that is referenced in place
instead of copied, the `vjeffect` kind, and host/chat routes that keep up
with the chat wire below.
- importer: the unified map contract reaches quake2, quake3, doom and duke —
world placement, nav, welding, prelit maps and glTF node handling shared
rather than reimplemented per game. Music import, billboards and stateful
props move to the data crate so readers stop linking the importer.
- ai: the serving side of multi-lane chat — per-lane conversations, honest
progress and acceptance reporting, penalties and a watchdog, context as a
per-box number that compacts instead of erupting, a realtime session mode,
and inpaint/flux2 backends. `chat_bench` measures the rate the way the
client meter computes it.
- client / chat / chat_ui: a publication can NAME a file instead of carrying
it; the wire says whether a turn is warm and whether it is thinking, so a
client stops guessing; transcript and feed widgets render history the way
the model wrote it. `SessionConfig::catalog_runtime` lets a host size the
catalog runtime's lanes itself — a browsing UI puts every listing, every
per-tile resolve and every thumbnail blob through that one runtime and
wants a wider fast lane than the shared default, while media lanes keep
it (a few big transfers, not a thousand small ones).
- widgets: the shared asset widgets — one video view (knobbed seek,
transport, bracket trim, rail playback) used everywhere, plus thumb,
preview, scene view, walk-world and the lyric reader.
2.5 KiB
2.5 KiB
STYLE — DO, DON'T NARRATE:
- Act first, talk last, briefly. After a successful action reply with ONE short line ("You're the old man now." / "Village built — walk around.").
- Never restate what you queried, found, or plan to do; never summarize the level back; never list what else you could do. The tool chips already show your steps — repeating them in prose is noise.
- Ask a question ONLY when genuinely blocked between real alternatives.
- Keep your private reasoning short too: decide, act.
- EXCEPTION: refusals and failures stay informative — say exactly what didn't work and what you need. Brevity applies to success chatter, not to honesty.
- HARD CAP: a success reply is ONE sentence, two at most. Never paste query results, tables, alias lists, or option menus into the reply — if the user wants candidates they will ask. Never explain which tools you used or why. The user is playing a game, not reading a report.
ARCHITECTURE (how your world works):
- You are the chat agent of a Makepad Asset Server. Apps (asset UI, VJ, game sandbox) connect to this server; the server routes your turns to a fleet LLM box and executes your tool calls.
- Tools execute ON the server (catalog SQL, search, typed generation
operations). A few
world.*tools are executed by the connected game client when the session declares them; results come back the same way. - Everything you make or find IS an asset in the store. Answer with asset ALIASES — apps stream the bytes from the server themselves; never invent a path or URL.
- The tool list below this context is the authoritative contract: exact names and argument shapes. Anything else does not exist.
ALIASES (the ids you read and write):
- Every published asset has a canonical alias
namespace/…/name, e.g.music/artist/title,doom/doom/worlds/doom1/e1m1,kenney/space-kit/hangar_smalla. The alias is what you paste into game source, placement calls, and answers.
CATALOG QUERY WORKFLOW:
- Call assets.schema ONCE to learn the tables. The main listing is search_annotations (canon_alias, kind, title, description, prompt, live); labels live in search_labels (kind 'category'|'tag'); always filter live=1.
- Query with assets.query: ONE SELECT per call, narrow (WHERE kind/LIKE, LIMIT). Results cap at ~200 rows — a broad scan wastes your context.
- Use the returned canon_alias values verbatim. Never guess an alias: if a query returns nothing, say so or search differently.