Skip to content

Models and effort ​

The catalog comes from a live model/list call during plugin setup — never a hardcoded list. Each row becomes an OpenCode model built from the Muse profile plus model ID (for example muse/muse-profile-tbh-model-muse-spark-1.3), carrying the display label, context/output limits, release date, and known effort tiers.

Effort variants ​

Reasoning effort tiers (minimal … max, per-model subsets) appear as OpenCode model variants. The selected tier rides on that turn only (turn/start.reasoningEffort); the standing session default is untouched unless explicitly changed. If you pick nothing, the default variant resolves to the catalog row's default effort. Unknown variant names fail closed.

Logged out? No models, no trap ​

If the CLI reports no usable catalog (logged out, no subscription), the provider stays visible but disabled with a "Muse setup required" placeholder. OpenCode hides providers with zero models, so without the placeholder the Muse entry would vanish silently — the placeholder keeps the fix obvious.

Titles and summaries are local ​

Thread titles and compaction summaries never reach Muse: the bridge answers them deterministically from the request itself. A Muse turn costs real inference, and Muse already owns the authoritative conversation — spending a turn on bookkeeping OpenCode does for its own transcript would pollute Muse's history for nothing.

MIT licensed. Auth and inference stay inside the official Muse Code CLI.