Feature request: model and reasoning-effort controls for Hosted OpenBot

Start by tracing the Hosted/server settings UI and API, built-in coworker configuration, and run setup; the issue does not name specific files or tests. Then inspect how model and reasoning-effort settings reach the TypeScript and Python harnesses, especially agent-langgraph-agui on the ChatGPT-plan path. Done means the requested defaults, overrides, validation, audit visibility, and backward compatibility work across supported Hosted paths.

Summary

Hosted/server OpenBot deployments need administrator-controlled model and reasoning-effort settings in the web UI, with optional overrides per built-in coworker and per task.

This request is specifically for the browser-accessed Hosted/server product, rather than the Desktop setup flow.

Current behaviour

A Hosted OpenBot deployment can select a shared model through deployment configuration such as BOT_MODEL. Built-in coworkers inherit that shared runtime model.

The Hosted UI currently does not let an administrator or user:

• view or change the deployment's default chat model and reasoning effort;
• assign a different model or effort to an individual built-in coworker;
• select an allowed model or effort when starting a task or conversation; or
• see which model and effort actually handled a run.

Existing work covers parts of this problem but not the Hosted UI or all harnesses:

• #212 and #268 added BOT_REASONING_EFFORT for the TypeScript agent-langgraph Responses API Bot.
• #451 added model-provider selection to Desktop setup.
• #649 centralised defaults for the separate Bot/harness implementations; it explicitly did not add per-coworker model settings.
• #736 requests supported ChatGPT-plan harness lifecycle for self-hosted/server deployments.

The Python agent-langgraph-agui harness used by the ChatGPT-plan path does not currently expose the same reasoning-effort control added to the TypeScript Responses API Bot.

Requested behaviour

Add Hosted/server model controls with this resolution order:

1. task or conversation override;
2. built-in coworker override;
3. Hosted deployment default;
4. provider default.

Hosted administration

Provide an Admin setting for the default provider, model and reasoning effort. The page should show the effective values and whether each value comes from Hosted configuration or the provider default.

Administrators should be able to define an allow-list of models and effort levels available to users. Credentials must remain server-side and must never be returned to the browser.

Per-coworker settings

Add optional model and reasoning-effort fields to built-in coworker configuration. Leaving either field unset should inherit the Hosted deployment default.

Remote AG-UI and Mastra coworkers should continue to report that their endpoint owns model selection unless OpenBot has a truthful capability contract for changing it.

Per-task settings

When starting a task or conversation, let the user choose from the administrator-approved models and effort levels. Existing conversations should show their effective selection, and changing it should apply only to subsequent runs.

An optional administrator-defined routing policy could select a model or effort from task characteristics, but the chosen effective values must remain visible and auditable.

Harness support

Pass model and effort settings through every supported Hosted harness that can honour them, including agent-langgraph-agui when it is using a ChatGPT plan. Unsupported provider/model/effort combinations should be rejected before a run with a clear explanation rather than silently ignored.

Acceptance criteria

• A Hosted administrator can set and change the deployment-wide default model and reasoning effort from the web UI.
• A built-in coworker can inherit those defaults or define an approved override.
• A user can select an approved model and effort for a task or conversation.
• The effective provider, model and effort are displayed for each run and recorded in the audit trail.
• The same settings are available through the server API for managed deployments.
• agent-langgraph-agui forwards supported effort settings on the ChatGPT-plan path.
• Invalid or unsupported combinations fail before inference with an actionable error.
• Existing deployments retain their current behaviour when no new settings are configured.

Why this matters

Different work has different latency, capability and cost requirements. Routine classification or drafting may suit a faster model and lower effort, while research, planning or infrastructure work may require a stronger model and higher effort. Hosted OpenBot already lets organisations define specialised coworkers and govern their tools; model and effort policy should be governed at the same levels rather than requiring environment-file edits and container recreation for the whole deployment.