fix: local model switching, mid-run settings, deterministic chess endings; release 0.0.12

[siarheidudko]

Fixes the demo problems reported after the last release.

Problem: switching local models

Report: after loading one local model, switching to another had no effect, and there was no way to load it.

Cause: useWebLLMModel kept the loaded model across modelId changes. So ready stayed true, the card said "loaded", the Load button never appeared, and the agent kept using the old model.

Fix:

• Everything the hook reports is now about the current modelId: model, ready, loading, progress and error.
• The previous model stays in memory (new loadedModelId), so switching back is instant.
• Loading the new model first frees the old model's GPU memory, through the core's unloadWebLLMModel. There is also a new unload().
• A load that another load supersedes frees its engine. Calling load() twice for the same id shares one download.

Problem: settings not applying

Report: settings didn't seem to apply.

Cause:

• With a cloud model they do apply: a probe showed the thinking level and skill toggles reach the next request.
• With a local Qwen they did not. WebLLM ignores the AI SDK's portable reasoning. That is fixed in the core (agent-web 0.0.21, merged as feat: local models with their real windows; useLocalModel; refreshed demo models (Kimi, Groq, Mistral, OpenRouter, Gemini Nano, transformers.js); fast answers; release 0.0.13 #19) by mapping thinking onto extra_body.enable_thinking.
• Changing a setting while a run was in flight rebuilt the agent and flipped the panel from "running" to "ready" (the stop button disappeared).

Fix:

• useAgent keeps the run's status through a rebuild. The run keeps its agent, and the next run gets the new one.
• New "No thinking" level (none) in the composer chip, the labels and the demo settings.

Chess

• Endings come from the rules (chess.js), not the model.
  • A mate, stalemate or draw ends the game on the spot, and a mating move by White never starts the agent.
  • The result covers the board ("Checkmate — you win (1-0)" + New game) and is added to the chat.
  • When the agent's own move ends the game, make_move says the game is over.
• A failed agent turn is reported as failed.
  • If the agent's turn ends with Black still to move (API error or spent quota, a stop, a limit), the board shows the reason with Retry and Engine move, plus a notice in the chat. The board stays locked until Black moves.
  • Before, a 429 on the executor was swallowed: the turn "succeeded" with "I will wait."

Demo model card

• Reopens when the newly selected model needs a key or a load.
• Names the local model still in GPU memory ("… is in GPU memory — loading this one frees it first").
• Has an Unload button.

Version

main already published 0.0.11 (the MCP SDK autoupdate, #17), so this ships as 0.0.12. I merged main in:

• the devDependencies take both sides: @dudko.dev/agent-web ^0.0.21 and @modelcontextprotocol/sdk ^1.32.0;
• the lockfile is regenerated with npm;
• CHANGELOG has a 0.0.12 entry.

The peer stays >=0.0.20.

Checks

These are after the merge.

• typecheck, format:check, build, npm test 43/43, demo typecheck and build.
• npm run test:e2e 15/15, with 3 new tests:
  • Scholar's mate: no agent run after the mate, the result is shown, New game resets the board.
  • A 429 on the executor: the board stays locked with the reason, and Retry plays the move.
  • A setting changed mid-run: the run keeps going, and the next run uses the new thinking level.
• The hook, driven in Chromium with a fake WebLLM, passes 13 checks: switch, switch back, free before load, unload, a superseded load, double load. On main's hook the same harness shows the bug: after switching to B, ready: true with model A.

🤖 Generated with Claude Code

https://claude.ai/code/session_01A7jHa5Pim4ufrdgxg561G5