nexu-io/open-design

What problem are you trying to solve? On Home, the compact model switcher (the top-bar chip) opens a popover that lists the current CLI agent's models as a short scrollable radio list (six rows visible). With providers that use provider/model ids and many same-family variants, every row ellipsises to the same shared prefix — with opencode-go, rows all read opencode-go/deepseek-… / opencode-go/glm-… — so you can't tell rows apart, or see which model is active, without hovering one by one. The chip itself truncates the same way, so the selected model is unreadable at a glance. On this machine the OpenCode agent exposes 77 models, which makes the missing filter worse. Reproduction (Open Design 0.24.1, macOS Apple Silicon; also reproduced on a current main dev run): Use a CLI/provider whose catalog has many same-family ids (e.g. opencode-go with several opencode-go/deepseek-* models). Click the model chip in the Home top bar. The popover shows Default (CLI config) + rows that all

Issue 2 — A run that produced an artifact is reported as failed, and the failure card hides the artifact Related to Issue 1, but a UX/contract problem on its own: state.json for the run: artifactCount: 1, artifactPaths: ["image-…png"], failureCategory: "empty_output". The conversation shows only a red "未能生成内容 — AI 未能生成内容，请重新发起任务，或更换模型后再试。" card, with buttons 联系我们 / 导出日志 / 切换到 OpenDesign Cloud. Nothing in the UI mentions that a usable artifact was produced and is sitting in the project. Requested fix If artifactCount > 0, show the artifact result card alongside the failure state (or downgrade to a warning), and state which artifacts were produced. Consider not classifying "artifact produced, no prose" as a run failure.

Environment OpenDesign docker image ghcr.io/nexu-io/od:0.24.1, self-hosted (daemon + web in Docker via Dokploy) Media provider: custom-image → Alibaba Cloud Model Studio (Bailian) OpenAI-compatible endpoint (https://<workspace>.cn-beijing.maas.aliyuncs.com/compatible-mode/v1), model qwen-image-3.0-pro Dispatched by the agent via od media generate --surface image --model custom-image --image <project-relative ref> ... Steps to reproduce Settings → Media providers → Custom Image API: Aliyun Model Studio compatible-mode base URL, API key, model qwen-image-3.0-pro. Text-to-image works fine: od media generate --surface image --model custom-image --aspect 16:9 --prompt "…" → 200, file written. ✅ Pass a reference image: … --image ref.jpeg --aspect 16:9 --prompt "…" → task fails immediately. Actual result task failed: custom image 404: {"taskId":"…","status":"failed","error":{"message":"custom image 404: ","status":400,"nextStep":"switch-model"}} Note the empty messag

Summary On Windows, when an ACP-based agent runtime resolves to an npm-installed CLI shim (kimi.CMD), detectAcpModels crashes with spawn EINVAL, so /api/agents reports modelsSource: "fallback" and the UI model picker shows only the runtime def's hard-coded fallback IDs — none of which exist in the current CLI, leaving users with only "Default (CLI config)". Environment OpenDesign 0.24.1 (Windows x64 desktop app, release-stable-win) Agent: kimi runtime, Kimi Code 2.0.2 installed via npm global (C:\Users\<user>\AppData\Roaming\npm\kimi.CMD) Node 24.16.0 (also reproduces on Node ≥ 20.12) Reproduction On Windows, install Kimi Code via npm i -g (so kimi resolves to a .cmd shim). Open the app, go to agent/model selection for the kimi runtime. GET /api/agents shows: { "id": "kimi", "available": true, "version": "2.0.2", "modelsSource": "fallback", "models": ["default", "kimi-k2-turbo-preview", "moonshot-v1-8k", "moonshot-v1-32k"] } These fallback IDs are from the legacy kimi-cli era;

What happens An agent connection test for the Reasonix CLI runs reasonix acp inside a throwaway od-conn-test-* temp dir (apps/daemon/src/connectionTest.ts), with cwd pointed at the temp dir. The probe sends a real smoke prompt, so the Reasonix CLI starts an actual session — and persists it into the user's real state home (%APPDATA%\reasonix\sessions\*.jsonl), because the test isolates cwd but not the child's state root. Reasonix's session import then picks that file up; since the probe's cwd belongs to no project, the conversation gets bound to the default workspace. Net effect: every "Test connection" click deposits a ghost conversation into the user's Reasonix app — one the user never created and can't attribute to OpenDesign. Proposed fix One line in connectionTest.ts: set REASONIX_STATE_HOME=<tempDir> in the env handed to the reasonix probe. Per Reasonix's documented path model (its docs/CONFIG_PATHS.md), REASONIX_STATE_HOME relocates state (sessions/archive/memory) only; con