@reasoning-router/opencode

English | 简体中文 | 日本語 | 한국어

An OpenCode V2 plugin that asks a classifier how much reasoning each step needs, then applies that effort to the outgoing model request without breaking the prompt cache. Every classifier in @reasoning-router/classifiers is available.

Alpha. OpenCode installs the plugin from npm by package name; list it under plugins as shown below.

Ported from opencode-jev-router.

Requirements

• OpenCode V2 2.0.4 or newer (smoke-tested with 2.0.18).
• Whatever the chosen classifier needs (a key, or a server you run); see classifiers.
• A Responses API endpoint serving GPT-6 Astra, Luna, or Sol, or an Anthropic Messages endpoint with the mid-conversation output-config beta.

Usage

{
  "$schema": "https://opencode.ai/config.json",
  "plugins": [{ "package": "@reasoning-router/opencode", "options": {
    "classifier": { "provider": "jev", "apiKey": "{env:REASONING_ROUTER_CLASSIFIER_API_KEY}" },
    "wrap": { "openai": ["openai/gpt-6-astra"] },
    "decisionsLogPath": "/tmp/reasoning-decisions.jsonl"
  }}],
  "model": "reasoning-router/gpt-6-astra"
}

The plugin registers reasoning-router/<profile> aliases only for the source models listed in wrap (openai and/or anthropic arrays of provider/model refs). Source models are left untouched. Each primary request to an alias gets one classification; if the classifier fails, the request continues at the fallback effort (high by default). With decisionsLogPath, each routed request appends a metadata-only ReasoningDecision event that records the deciding classifier.

For OpenAI Decisions (public beta), replace the classifier block with:

"classifier": { "provider": "openai-decisions", "apiKey": "{env:REASONING_ROUTER_CLASSIFIER_API_KEY}" }

See examples/opencode.jsonc for a full example.

Options

| Option | Default | Purpose | | --- | --- | --- | | classifier.provider | REASONING_ROUTER_CLASSIFIER env, then jev | Classifier provider name, such as jev (see classifiers). | | classifier.apiKey | REASONING_ROUTER_CLASSIFIER_API_KEY env | Classifier key, if the provider needs one (not needed when fixedEffort is set). | | classifier.baseUrl | REASONING_ROUTER_CLASSIFIER_BASE_URL env, then the provider default | Classifier endpoint; defaults to the provider's. | | classifier.accountId | REASONING_ROUTER_CLASSIFIER_ACCOUNT_ID env | Account ID, for providers that take one. | | classifier.model | REASONING_ROUTER_CLASSIFIER_MODEL env | Classifier model, for providers that take one. | | classifier.timeoutMs | 4000 | Total classification budget, including retries. | | wrap | none | Required nonempty object of openai/anthropic source refs. | | decisionsLogPath | off | Absolute path for ReasoningDecision JSONL. | | baseEffort | profile default | Request-level effort reported by responses. | | fixedEffort | none | Skip the classifier and always use this effort. | | maxRetries | 1 | Extra attempts after transient classifier errors. | | fallbackMode | fixed | fixed, previous, or error. | | fallbackEffort | high | Effort used when classification fails. | | maxRequestBytes | 33554432 | Largest request body. | | maxInFlight | 32 | Concurrent requests. | | upstreamHeaderTimeoutMs | 10000 | Wait for endpoint response headers. | | upstreamIdleTimeoutMs | 60000 | Longest gap between streamed chunks. |

Wire behavior, logging, and limits are documented in docs/behavior.md and docs/classification-policy.md.

License

MIT