studio: agent window, a chat with a model that edits the scene

✦ Agent in the Studio's top bar opens a floating chat window. The model
(any OpenAI-compatible endpoint with tool calling; default B4L's inference
server, https://inference.p0.contact/v1) gets the 15 MCP tools plus two page
tools (select_object, set_view) and the Studio's state (selection, level,
rotation) in its system prompt. Answers and reasoning stream in; tool calls
show as rows with arguments and results; object ids link to the object;
Stop ends a turn; Undo takes the agent's edits back. The window moves by
its title bar and resizes at its top-left corner.

- studio/agent.js: shared by both front ends (waits for studio.js's
  isokit:studio event); MCP over /mcp or isokit.mcp in the web Studio
- isokit studio: /api/agent/{config,models,chat} relay the calls on their
  own threads (ureq, chunked and flushed per read), so the key stays on the
  machine and the endpoint needs no CORS; settings from ISOKIT_LLM_URL /
  _KEY / _MODEL or agent.json in the user's config folder
- web Studio: calls the endpoint from the page with a key kept in the
  browser; build_web.py --agent-url names a same-site relay as default; the
  deploy builds with /llm/v1, which deploy/Caddyfile forwards to the
  inference server (its nginx refuses the CORS preflight)
- tools/mock_llm.py + tools/agent_checks.py: the agent window checks in both
  check_studio.py and check_web.py

README.md

@@ -9,6 +9,8 @@ click for details. One Rust binary does everything:
 - **Web Studio**: the same Studio running entirely in the browser, with the
   engine compiled to WebAssembly. Nothing to install: projects live in the
   browser, or in a folder on disk (Chrome / Edge) that agents edit alongside.
+- **Agent window**: chat with a language model right in the Studio (by default
+  B4L's inference server); it edits the scene with the MCP tools while you watch.
 - **MCP server**: AI agents (Hermes and others) edit the same scene with the
   same operations; their changes show up live in an open Studio.
 - **Build**: a self-contained HTML page with every view prerendered, or
@@ -80,6 +82,33 @@ MCP tools also run inside the page: `await isokit.mcp({jsonrpc: "2.0", id: 1,
 method: "tools/list"})`, and browsers with WebMCP (`navigator.modelContext`)
 get them as page tools.
 
+### Agent window
+
+**✦ Agent** in the Studio's top bar opens a chat with a language model that
+edits the scene for you: "plant a row of trees along the road", "make the
+workshop one storey taller", "what is on this site?". It works with the 15 MCP
+tools below plus two page tools (point at an object, turn the view), streams
+its answer and its reasoning, and shows each tool call (click one for the
+arguments and the result, click an object id to select it). Every edit lands
+on the map at once, and Undo takes it back. Stop ends a turn; the window moves
+by its title bar and resizes at its top-left corner.
+
+Any OpenAI-compatible endpoint with tool calling works (vLLM, llama.cpp,
+Ollama, LM Studio, hosted APIs). The default is B4L's inference server,
+`https://inference.p0.contact/v1`; ⚙ sets endpoint, key, model and whether
+the model thinks first.
+
+- `isokit studio` passes the calls through its own server (`/api/agent/*`), so
+  the key never reaches the page and the endpoint needs no CORS. Settings:
+  `ISOKIT_LLM_URL`, `ISOKIT_LLM_KEY`, `ISOKIT_LLM_MODEL`, or the window's ⚙,
+  which saves them in `%APPDATA%\isokit\agent.json` (`~/.config/isokit` elsewhere).
+- The web Studio calls the endpoint from the page; the key stays in that
+  browser. The endpoint must allow the page (CORS, preflight included), or the
+  site relays it: https://isokit.buildfor.life/ is built with
+  `--agent-url /llm/v1`, a same-site path that Caddy forwards to B4L's
+  inference server (`deploy/Caddyfile`). The relay adds no key: each browser
+  enters its own once.
+
 ### Agents (MCP)
 
 While the Studio runs, it also serves MCP over HTTP at `http://127.0.0.1:7878/mcp`.
@@ -127,7 +156,7 @@ storey.
 | Command | |
 |---|---|
 | `isokit new <dir>` | starter project |
-| `isokit studio [dir] [--port N] [--no-open]` | GUI editor + `/mcp` + `/view` live preview |
+| `isokit studio [dir] [--port N] [--no-open]` | GUI editor + `/mcp` + `/view` live preview + the agent window's link to the model |
 | `isokit mcp [dir]` | MCP over stdio |
 | `isokit build [dir] [-o file]` | self-contained HTML (all levels × 4 rotations) |
 | `isokit render [dir] -o x.svg [--level gf] [--rot 2] [--light]` | one view as SVG |
@@ -144,7 +173,7 @@ crates/isokit-project  project folder, document ops, export, check, Workspace (u
 crates/isokit-cli      the `isokit` binary: Studio HTTP server, MCP stdio, build/render/check
 crates/isokit-wasm     the engine for the browser (wasm-bindgen): a project in memory behind the Studio API
 viewer/                isokit-view.js/.css: pan, zoom, levels, rotation, tooltips (any page can embed it)
-studio/                the Studio GUI (served by `isokit studio`, and by the web Studio)
+studio/                the Studio GUI (served by `isokit studio`, and by the web Studio); agent.js = agent window
 web/                   web Studio shell: WASM backend for studio.js, projects, folders, zip, code editor
 examples/demo          demo project
 deploy/Caddyfile       the web Studio's vhost on the buildfor.life web server (s1)
@@ -165,12 +194,16 @@ Plans: [docs/ROADMAP.md](docs/ROADMAP.md).
 ```bash
 cargo test                                   # engine, runtime, ops, demo, MCP
 cargo clippy --all-targets -- -D warnings
-isokit studio examples/demo --port 7890 --no-open &
-uv run --with playwright python tools/check_studio.py <shot-dir> http://127.0.0.1:7890   # GUI + MCP end to end (Edge)
+ISOKIT_LLM_URL=http://127.0.0.1:7932/v1 ISOKIT_LLM_KEY=test-key ISOKIT_CONFIG_DIR=<scratch> \
+  isokit studio <copy of examples/demo> --port 7890 --no-open &     # agent window on tools/mock_llm.py
+uv run --with playwright python tools/check_studio.py <shot-dir> http://127.0.0.1:7890   # GUI + MCP + agent window end to end (Edge)
 cargo clippy -p isokit-wasm --target wasm32-unknown-unknown -- -D warnings
 python tools/build_web.py && uv run --with playwright python tools/check_web.py <shot-dir>  # web Studio end to end
 ```
 
+The agent window checks run on `tools/mock_llm.py`, a scripted stand-in for an
+OpenAI-compatible server that the check scripts start on port 7932.
+
 `python tools/build_web.py --skip-cargo` re-copies the front end (studio/,
 viewer/, web/) without rebuilding the WASM. Code that runs in the browser must
 not use `std::time` (it traps on wasm32-unknown-unknown): use `web_time`.