bug(tui): persisted per-model variant state in model.json shadows agent.variant config for primary agents

Description

Changing agent.<name>.variant in opencode.json has no effect on primary agents driven from the TUI, even after a full restart. The config is persisted and served correctly (GET /config shows the new variant), but requests keep running with the old one.

Cause — a precedence inversion between the TUI's persisted per-model variant memory and the agent config (file/line refs against the v1.18.0 tag):

1. The TUI keeps a per-model variant map in ~/.local/state/opencode/model.json (packages/tui/src/context/local.tsx:155-189).
2. The map is auto-seeded: opening or switching to a session copies the last user message's variant into it and saves (local.model.variant.set(msg.model.variant), packages/tui/src/component/prompt/index.tsx:328). No native variant pick is needed — using a session once is enough.
3. Every prompt sends the map value as the explicit variant (prompt/index.tsx:988). Unlike model resolution (local.tsx:236-245, which falls back to the agent's configured a.model), variant resolution (local.tsx:362-374) never consults the agent's configured variant.
4. Server-side the client-sent variant always wins: input.variant ?? (ag.variant && ...) (packages/opencode/src/session/prompt.ts:654).

So once the map has an entry for a model — automatic after the first message — later agent.<name>.variant changes are silently shadowed forever for TUI-driven agents. Restarting doesn't help (model.json is persisted). Picking via /variants "works" because it writes the map that actually wins, which makes the config path look broken by comparison.

A per-model memory also can't represent this config at all — several agents legitimately share a model with different variants:

"agent": {
  "orchestrator": { "model": "openai/gpt-5.6-sol", "variant": "low" },
  "sdd-design":   { "model": "openai/gpt-5.6-sol", "variant": "medium" },
  "sdd-spec":     { "model": "openai/gpt-5.6-sol", "variant": "high" }
}

One "openai/gpt-5.6-sol": "medium" entry in model.json overrides all three.

Expected behavior — mirror the model precedence: treat a native pick as an ephemeral per-agent override (like the in-memory /model override), fall back to the agent's configured variant otherwise, and don't auto-seed the persisted map from session sync (or key it per agent+model and invalidate it when the configured variant changes).

Related: #13718 reported this same shadowing on v1.2.4 (closed not_planned — this adds the root-cause trace and confirms it on v1.18.0). #28803 is the adjacent session-lock case (SessionTable.model consulted first, session/prompt.ts:615-627). #36250 covers the subagent vector (handleSubtask resolves taskModel before taskAgent, prompt.ts:267 vs :313, and stamps the parent's variant at :275 — also still present at v1.18.0).

Plugins

None — reproduces on a clean install with no plugins

OpenCode version

1.18.0

Steps to reproduce

With isolated XDG_CONFIG_HOME / XDG_STATE_HOME / XDG_DATA_HOME:

1. Config: agent.build.model = "openai/gpt-5.5", agent.build.variant = "high". Start the TUI, send one message — it correctly runs with high (the state map is still empty, so config applies). The stored message carries model.variant: "high".
2. Restart the TUI and resume/switch to that session. model.json now contains "variant": { "openai/gpt-5.5": "high" }, seeded without any user action.
3. Change the config to "variant": "medium". Restart the TUI.
4. Send a message. Actual: still high (persisted map beats ag.variant). Expected: medium.
5. Pick medium via /variants — now it applies, confirming the inversion.

Also observed on a real setup: config orchestrator → openai/gpt-5.6-sol variant=low, model.json holding "openai/gpt-5.6-sol": "medium", orchestrator observably running at medium across restarts.

Screenshot and/or share link

No response

Operating System

macOS 26 (Darwin 25.4.0)

Terminal

Ghostty