[BUG] v2.0.16: agent-level variant is resolved by debug agents but never applied to requests #51326

In v2.0.16 the agent-level variant config key has no effect on outgoing requests, although the config schema documents it and opencode debug agents resolves it correctly.

Schema (https://opencode.ai/config.json, $defs/AgentConfig/properties/variant):

Default model variant for this agent (applies only when using the agent's configured model).

Config used

{
  "model": "opencode-go/deepseek-v4.1-flash",
  "agent": {
    "build":   { "model": "opencode-go/deepseek-v4.1-flash", "variant": "max" },
    "plan":    { "model": "opencode-go/deepseek-v4.1-flash", "variant": "max" },
    "general": { "model": "opencode-go/deepseek-v4.1-flash", "variant": "max" }
  }
}

opencode debug agents resolves it fine:

build -> {"id": "deepseek-v4.1-flash", "providerID": "opencode-go", "variant": "max"}

Reproduction### Reproduction

I pointed provider.opencode-go.options.baseURL at a local recording proxy (project-level opencode.json) and captured the raw POST /chat/completions bodies.

Agent config in place for these runs:

"agent": { "build": { "model": "opencode-go/deepseek-v4.1-flash", "variant": "max" } }

opencode debug agents resolves that to {"id":"deepseek-v4.1-flash","providerID":"opencode-go","variant":"max"}.

A field-by-field diff of the two bodies (agent config vs explicit #max, same model, same prompt shape) shows reasoning_effort is the only top-level field present in the explicit case and absent in the agent-config case. Scanning the entire agent-config body for max outside messages finds only unrelated tool-parameter descriptions ("Very short label (max 30 chars)" etc.), so the value is not being sent under some other key.

Two extra probes, to rule out "invalid value is simply dropped":

• The same capture was also run while the agent variant was an invalid value ("bogus") — identical result: no error, no effort field.
• With a valid value ("max") — again no error and no effort field, while the explicit -m ...#max path in the same session does send reasoning_effort: "max".

So config-provided variants are neither validated nor applied, for valid and invalid values alike.

Expected

A session started with an agent that configures both model and variant should send the corresponding effort (e.g. reasoning_effort: "max"), as the schema description states.

Related

• core: legacy Markdown agent silently drops "model: provider/model#variant" — subagent inherits parent model #50849 / fix(core): keep embedded variant in legacy markdown agent model #51311 cover the legacy Markdown agent surface, where an embedded model: provider/model#variant is dropped. This report is the JSON-config surface (agent.<id>.variant), where the variant is resolved by debug agents but still never reaches the request.
• Putting #variant inside the model string in config ("model": "opencode-go/deepseek-v4.1-flash#max", in both the agent blocks and the top-level model) makes debug agents report model: null for every agent — the embedded form appears to be CLI-only (related: fix(core): parse embedded "#variant" in model selections #50853).
• Qwen3.8 Flash: variant=xhigh silently disables reasoning by sending output_config.effort without thinking #49079 looks like a sibling report for a different model (variant accepted but not reflected in the request).

Environment

• opencode v2.0.16, Windows x64, installed via npm i -g @opencode/cli
• Provider/model: opencode-go/deepseek-v4.1-flash (the registry accepts #max and rejects #xhigh with Variant unavailable, so variants are declared for this model — it is not a "model has no variants" case)