opencode-optimal-model-temps

OpenCode plugin for conservative model sampling profiles.

The plugin hooks the request parameters (chat.params on OpenCode V1, the session context hook on OpenCode V2) and manages only:

• temperature
• top_p (topP in the request options)
• top_k (topK in the request options)

It does not change reasoning effort, thinking mode, output-token limits, tools, prompts, context compaction, or model routing.

Thinking and non-thinking models can have separate presets. The plugin reads the mode selected by OpenCode; it never enables or disables thinking itself. When no mode field is present, the request is treated as non-thinking; an explicitly present undefined value is treated as thinking. If the selected mode has no preset, the fields remain unchanged. OPENCODE_SAMPLING_MODE=thinking or OPENCODE_SAMPLING_MODE=nonThinking can be used as an explicit override.

Installation

Add the plugin to ~/.config/opencode/opencode.json (OpenCode V2):

{
  "plugins": [
    "opencode-optimal-model-temps"
  ]
}

Local checkouts use a path entry (a directory containing a root index.js):

{
  "plugins": [
    "/path/to/opencode-optimal-model-temps"
  ]
}

OpenCode V1 (1.18.29 and newer) still loads the same package through the legacy entry point:

{
  "plugin": [
    "opencode-optimal-model-temps"
  ]
}

Profiles

The active profiles are family-first: GLM-5.x, Qwen3.x, DeepSeek V4.x, Kimi K2.x/K3.x, MiniMax M2.x/M3.x, MiMo V2.x, Grok 4.x, grok-build, Tencent Hunyuan Hy3, Tencent Hunyuan Hy4, and Meituan LongCat 2.x. Route-specific subfamilies are retained only when a provider constraint or probe requires them. The original Gemini 3 Pro rule remains available for compatibility.

Model-specific entries are retained only as safety overrides when a live OpenCode Go probe rejects the family value or when a model has a documented thinking-only constraint. The registry deliberately registers no profile for hosted Claude, StepFun, or Meta Llama models, because Claude rejects non-default sampling while StepFun's hosted API and Meta Llama publish no usable recommendation.

Hosted gateways can namespace model ids by vendor (for example commandcode/z-ai/glm-5.3-flash). A profile matches both the full id and the vendor-stripped basename, so the same family rule applies across providers.

See docs/sampling-profiles.md for the complete table, source URLs, evidence levels, route-specific limitations, and the policy for adding new values.

Precedence

Profile selection is two-tier: a model-specific rule (scope: "model") always wins, and family rules are a fallback used only when no model-specific rule matches — independent of entry order in the registry. The plugin then preserves explicit sampling values and only fills fields that are unset. A defined value is never replaced, even when it equals a provider baseline. The plugin does not add fields without an applicable profile value.

On OpenCode V2 the context hook receives request options that are treated as explicit configuration: any field already set to a non-null value in event.options (including values written by an earlier hook) is never overwritten. null is treated as unset, matching V1 behavior. The hook has been observed to arrive with fields prepopulated (for example maxTokens), so the plugin does not assume the object is empty.

When configuration layers disagree, the more specific layer wins, highest first:

1. fields already set to a non-null value in event.options (including earlier hooks) — never overwritten (null counts as unset)
2. the agent's request settings
3. the selected variant's settings
4. model settings

Profile values fill only the sampling fields left unset by all of those layers.

When a model reference matches a profile but its model record cannot be resolved, the plugin fails closed: it writes no profile values, because it cannot distinguish an unset field from a user-configured one. The skip is logged once per model reference; the plugin retries after its 30-second negative-cache entry expires, and the cache is cleared immediately when the model list or agent list changes (model.updated / agent.updated). The same applies when the agent record cannot be resolved; an agent-read failure also retries after that short TTL.

V1 note: a V1 model record exposes both modelID and api.id, and either id may match a profile. A record whose two ids disagree therefore still resolves its profile; earlier versions matched only the first non-empty id. This is deliberate: failing to match would silently keep the wrong sampling values.

OpenCode V2 model capabilities expose only tools, input, and output; there is no temperature capability signal to inspect. The capabilities.temperature === false guard therefore only applies to V1-shaped models.

For the legacy Gemini profile, this environment variable remains supported:

• OPENCODE_GEMINI3_TEMPERATURE: target temperature, default 0.35.

Development

npm install
npm test
node --check index.js
node --check src/index.js
node --check src/profiles.js
npm pack --dry-run

The package ships two entrypoints from one default export:

• Plugin.define({ id, setup }) for OpenCode V2, registering the session context hook.
• server() for OpenCode V1 (1.18.29 and newer), returning the chat.params hook.