Jev-assisted model and effort selection with Suggest and Auto-select modes

[vidhatrashukla]

Choosing a model and reasoning effort for every task takes attention, especially when moving between small edits and difficult investigations. I would like an optional Jev-assisted selector that recommends a configured coding model and effort, with an explicit option to make that selection automatically for future tasks.

Jev would make the routing decision; the selected coding agent would still do the work. A typo fix might use a lighter model/effort combination, while a difficult debugging task might use a stronger combination. These are intended outcomes to evaluate, not claims that Jev already makes the right choice.

Proposed behavior

Three modes, remembered per project:

• Off: use the existing manual model and effort selection. This remains the default until enabled.
• Suggest: show Jev's recommendation before submission, with accept and override controls.
• Auto-select: automatically apply its recommendation to future eligible tasks. Offer an explicit “Automatically select for future tasks” option from Suggest mode, and make it easy to switch back or disable.

A manual choice for a specific task always takes precedence without changing the project's mode. In Auto-select, show the selected model and effort with the submitted task so the decision remains visible.

Small initial scope

• Limit selection to a user-approved shortlist of currently configured, available coding models and supported effort values. Respect user-defined effort limits and preserve permission settings.
• Start with the task prompt and available choices as routing input. Explain what is sent to TypeSafe and require an explicit opt-in and configured credentials. Do not silently include repository files or full thread history.
• Use a configured fallback model/effort if Jev times out, errors, returns an unavailable choice, or does not meet a configured confidence threshold. Make fallback visible. If the fallback is unavailable too, return to manual selection.
• Initially apply routing only to the first submission of a new thread, before execution starts. Do not switch the model of a running thread or overtake queued work. Later scope could consider explicit routing of new tasks in existing threads after lifecycle behavior is agreed.
• Cancellation or navigation must not dispatch the abandoned task, change another thread's selection, or leave its composer locked. Failed routing/submission must preserve a recoverable draft and avoid phantom sent messages.

Why this shape

Suggest mode lets users evaluate recommendations before trusting automatic selection. Auto-select then removes repeated manual decisions while retaining explicit overrides and predictable failure behavior. Any latency or cost benefit should be measured on representative tasks; typed outputs and confidence alone do not establish routing quality.

Related work

• PR #15663 attempted Jev model and reasoning-effort orchestration and was closed without merging. Its review raised cancellation, queue-ordering, navigation-state, and send-recovery concerns; the initial scope should account for these.
• Discussion #14501 proposes a Jev archive checker, a separate use case.
• TypeSafe's Jev documentation describes typed decisions and probabilities rather than text generation.

Would this focused, opt-in selector fit T3's direction, and would maintainers prefer it as a native capability or through an extension? This is a proposal for direction and scope before implementation.

Prepared with GPT-6 Astra through the Codex harness in T3 Code.