✅ Done! Here's the current state: What's in the UI today • Model selection is fully surfaced: Settings → Agent lets you pick the ACP preset, credentials, and model (Agent default / provider's models / custom model ID), and the in-chat picker lets you switch models mid-conversation. • Reasoning effort: no dedicated control — it always runs at the provider's default. The only workaround is for Codex, where you can encode effort in the model ID itself (e.g.

gpt-5.5/high

, with

low/medium/high/xhigh

) via a custom model entry — the SDK splits that into the

reasoning_effort

config option. For Claude Code, its CLI exposes an

effort

option, but it's not dispatched today, so

opus[1m]/high

won't work yet. Open issues / PRs to make it easier • 📌 OpenHands#15578 — "ACP: switch reasoning effort (settings/agent + mid-conversation)" — the main tracking issue. Proposes a generic config-option mechanism so effort gets a persisted default in Settings → Agent and live mid-conversation switching, with values read from the server rather than hardcoded. • SDK-side, three open PRs are in flight: ◦ software-agent-sdk#4384 — surface session effort state + model/effort switching via config options (pairs with a frontend change) ◦ #6107 — supports effort levels for Claude ACP models (#6106 is the request) ◦ #4989 — simpler alternative: extend the Codex-style

model/effort

ID splitting to Claude Code • Context: OpenHands#16350 (frontend for dynamic model lists + effort selection) was closed unmerged because it needed those paired SDK changes first. So: the tracking issue is open but hasn't landed, and the SDK-side PRs are the current blockers — no merged support yet beyond the Codex model-ID suffix trick.

cc @strong-ice-84145 since you opened the original 15578