Agents

Run the shared checks once per agent, then that agent’s own section. After bumping an agent’s version, also compare its event stream with what its *Events class expects (see AGENTS.md).

Every agent

Start a new project on the agent and model under test, with a small prompt.

A turn runs and finishes

• The chat shows the agent working within a few seconds: thinking, tool calls and file edits stream in as they happen.
• The preview leaves the placeholder and shows the app once the agent starts a dev server.
• The turn ends cleanly: no error line, the send button comes back.
• A follow-up message (“make the heading blue”) continues the same conversation; the agent knows what it built.
• Tools → Git shows a checkpoint commit for the turn.

Stop and queue

1. Send a message that takes a while (“add a settings page with five fields”).
2. While it runs, type and send a second message.
3. Click stop.

• The second message waits in the queue instead of interrupting.
• Stop ends the turn within a few seconds and the chat says it was stopped.
• The queued message runs next, or can be removed before it does.

Reload during a turn

1. Send a message, then reload the page while the agent works.

• The chat comes back with everything so far and keeps streaming.
• Closing the tab and reopening the project mid-turn does the same.

Errors reach the agent

1. Ask the agent to add a page that throws an error, then open it in the preview.

• A bar over the preview shows the error, and Fix it sends it to the chat.
• With Autofix on, a new error after a turn is sent to the agent once by itself; with it off, nothing is sent.

A bad connection says so

1. In Settings → AI, reconnect the provider with an invalid key (or revoke it at the provider).
2. Send a message.

• The chat says the key was rejected, in plain words, not a stack trace.
• Reconnecting a good key and sending again works.

Model and reasoning

• Changing the model in the picker applies to the next message (the chat or the agent’s output says which model ran).
• Auto picks a model per message and shows which one it picked.
• A reasoning level other than Default is accepted and the turn runs.

OpenCode

• Runs on an OpenRouter key, an Anthropic or OpenAI API key, a Gemini key, an Ollama key, and a ChatGPT sign-in (pick an OpenAI model).
• With a local Ollama server (http://localhost:11434 on a local install), its pulled models show in the picker and a turn runs on one.
• Without any Claude API key, OpenCode isn’t offered a Claude subscription.

Claude Code

Claude subscription sign-in

Needs a Claude Pro or Max account. Disconnect Claude first if DEV_CLAUDE_CREDENTIAL seeded one.

1. Connect Claude with Use my Claude subscription and open a project on Claude Code.
2. Click Sign in to Claude above the chat box.

• A Shell tab opens running claude auth login.
• After pasting Claude’s code, the Shell says Login successful, the workspace switches back to Preview, and the line above the chat says who is signed in.
• A message runs on the subscription.
• On Docker sandboxes, a second project is already signed in.
• A message sent before signing in picks itself back up after sign-in, even after a page reload.
• Disconnecting Claude in Settings → AI signs Claude Code out; the next message asks to sign in again.

API key

• With only an Anthropic API key, Claude Code runs and the picker lists only Claude models.

Codex

ChatGPT sign-in

1. In Settings → AI, click Sign in with ChatGPT and approve the code at OpenAI.

• OneDrop connects as soon as you approve, without a reload.
• A project on Codex runs a turn, and the picker lists only the models OpenAI includes with Codex.
• Get a new code gives a fresh code when the first expires.

API key

• With only an OpenAI API key, Codex runs a turn.

Switching agents

1. In a project with a few turns, switch from one agent to another in the picker.

• The chat notes the switch and the new agent starts a fresh conversation.
• The app’s files are unchanged and the preview keeps working.
• Switching is refused while the agent is working.
• Repeat for each pair you changed: OpenCode ↔ Claude Code ↔ Codex.

AI credits

Only on installs with AI credits on.

• A new organization with no connection can build on AI credits, and the balance beside the model picker goes down after a turn.
• The Usage page shows the same balance.
• Once credits run out, the chat says when the next $1 arrives and points to Settings → AI.