Preview and debug

Use the Agent workspace to test draft configuration before publishing.

Preview

Enter a representative task in the left pane. Test:

• a normal request and expected output;
• a request that uses each attached Skill or MCP server;
• file input and generated artifacts;
• missing information and clarification behavior;
• tool, credential, or Environment failure paths.

Draft changes affect Preview immediately. A green Ready state means the editor is ready to accept a task; it is not proof that every external dependency works.

Logs

Open Logs to inspect session and execution history. Use it to correlate runtime startup, tool use, and failures. Secrets should not appear in logs; if a raw credential is present in user-authored content, remove it before sharing diagnostics.

Cost

Open Cost for Agent-level spend estimates, model mix, and recent usage events. Filter production and debug use separately. Cost values are estimates from recorded model calls, not Provider invoices.

Terminal and reset

Assistant Agents expose a Terminal and working-state controls because their environment can persist across sessions. Use reset when you intentionally want a clean state; it can remove local workspace state, caches, or sign-ins. Task Agents start clean for every Run and do not expose the same persistent controls.

Diagnose common failures

1. Runtime needs a key: configure the matching Provider in this App.
2. Model is unavailable: verify the Provider's model list and endpoint compatibility.
3. MCP tool fails: confirm the connection is enabled and authorized, then retry a direct tool task.
4. Environment startup fails: check exact package versions, setup script, and required variables.
5. Old session behaves differently: start a new session to use the current published version.