Let a webhook request choose the scheduled task's model (opt-in, from an allowed list)

[X2Borgo]

Before submitting

• I searched existing issues, PRs and discussions and did not find a duplicate.
• I am describing a concrete problem or use case, not just a vague idea.

Area

apps/server (scheduledTasks: webhook route, ScheduledTaskService), packages/contracts, scheduled-task editor on web and mobile.

Problem or use case

A webhook task is set to a single model when you create it. The incoming request can shape the prompt through {{…}} placeholders, but it can't change which provider/model, reasoning level or other options the run uses. Often the sender knows better than the task what a delivery needs:

• A CI or GitHub workflow wants a cheap, fast model for a lint failure and a strong one for a broken integration test.
• A single "review this PR" hook is called by different repos or labels that want different models (e.g. review:deep → Opus, default → Sonnet).
• A script or another agent sends T3 a job over HTTP and wants to choose the model the way delegate_task lets an agent choose one.

Today you need one webhook task per model, each with its own URL, secret and prompt, and the sender has to pick the right URL. This gets messy quickly.

Proposed solution

An opt-in "Let the request choose the model" toggle on webhook tasks. It's off by default, and when it's off nothing changes.

When it's on:

• The task has a list of allowed models (provider instance + model, plus their saved options). The task's own model is the default and is always in the list.
• The request picks one by a short key, with the same lookup syntax as the prompt template, so no new syntax is needed. For example, a model field on the task set to {{body.model}}, {{query.model}} or {{headers.x-t3-model}}.
• A key that isn't in the allowed list, or is missing, falls back to the task's default model. The delivery log records which model ran and why (requested / default / rejected: not allowed). A bad value never fails the delivery, so senders like GitHub don't keep retrying.

The allowed list matters because anyone with the URL could otherwise run any model the environment has credentials for. That could mean expensive models, or providers the user never meant to expose to automation. Restricting choices to a list on the task keeps the "token in the URL" security model as it is.

Smallest useful scope

Provider + model only, picked from the allowed list with {{query.model}} / {{body.model}} / a header. Reasoning, fast mode and permissions come from that allowed entry's saved options, and the request can never set them. Permission mode especially should stay tied to the task. Web editor first; mobile can show the toggle as read-only like it does for signatures.

Current workaround

Duplicate the task for each model and have the sender pick the URL.

Related

• feat(web): choose reasoning, speed, and permissions for scheduled tasks #16592: choose reasoning, speed and permissions in the web scheduled-task editor. The allowed-list entries would reuse that UI.

I'm happy to implement this if the direction and scope are approved.

[Bombatomica64]

would love this to start a thread like openclaw or hermes from telegram and such