@reasoning-router/classifiers

English | 简体中文 | 日本語 | 한국어

The reasoning-effort classifiers for reasoning-router. Each one asks a decision model one choice question, limited to the target model's supported efforts, over plain fetch. OpenAI Decisions uses its own request format.

Select one with provider in a classifier config block.

Jev (provider: "jev")

Jev, through TypeSafe or Vercel AI Gateway.

| Field | Default | Purpose | | --- | --- | --- | | apiKey | none | TypeSafe key, or a Vercel AI Gateway key. Required. | | baseUrl | https://api.typesafe.ai | Set to https://ai-gateway.vercel.sh/typesafe for a Vercel key. No other values are accepted. | | model | the endpoint's model | Optional. Must be jev-latest for TypeSafe or typesafe-ai/jev for Vercel. | | timeoutMs | 4000 | Total classification budget, including retries. |

The key is not inspected and the endpoint is not detected automatically, so baseUrl must match the credential.

Clef (provider: "clef")

Cloudflare's Clef, on Workers AI.

| Field | Default | Purpose | | --- | --- | --- | | apiKey | none | Cloudflare API token with Workers AI permissions. Required. | | accountId | none | Cloudflare account ID. Required. | | model | none | clef, clef-omni, or clef-flash. Required. | | timeoutMs | 4000 | Total classification budget, including retries. |

The key is used with the account ID to call Workers AI at https://api.cloudflare.com/client/v4/accounts/{accountId}/ai/run/@cf/cloudflare/{model}. baseUrl is ignored.

Laya (provider: "laya")

Laya, an open-source Jev-compatible model served by a Laya server that you run, such as laya-serve. This package only calls it over HTTP.

| Field | Default | Purpose | | --- | --- | --- | | baseUrl | http://127.0.0.1:8000 | Laya server. HTTPS, or plain HTTP to a loopback host only. | | apiKey | none | Optional; set only when the server sets LAYA_API_KEY. | | model | server's choice | Optional checkpoint: english, multilingual, or typed-decisions. When unset, the server chooses by the language of the state. | | timeoutMs | 4000 | Total classification budget, including retries. |

Calls {baseUrl}/v1/systemone. The key is sent as a bearer token; a blank key counts as unset.

Kev (provider: "kev")

Kev, an open-source Jev-compatible family of decision models served by a kev.serve server that you run. This package only calls it over HTTP.

| Field | Default | Purpose | | --- | --- | --- | | baseUrl | http://127.0.0.1:8008 | Kev server. HTTPS, or plain HTTP to a loopback host only. | | apiKey | none | Optional; set only when the server sets KEV_API_KEY. | | model | none | Optional. Only echoed back; the server picks the checkpoint at startup (--run). | | timeoutMs | 4000 | Total classification budget, including retries. |

Calls {baseUrl}/v1/systemone, with the same key rules as Laya (a bearer token; a blank key counts as unset). See the Kev proposal.

OpenAI Decisions (provider: "openai-decisions")

Calls OpenAI's Decisions API. Users bring an OpenAI API key with Decisions access.

| Field | Default | Purpose | | --- | --- | --- | | apiKey | none | OpenAI API key. Required. | | baseUrl | https://api.openai.com/v1 | Also accepts https://us.api.openai.com/v1 or https://eu.api.openai.com/v1. No other values are accepted. | | model | gpt-6-luna | Only accepted value: gpt-6-luna. Independent of the target generation model. | | timeoutMs | 4000 | Total classification budget, including retries. |

The provider calls POST {baseUrl}/decisions. The bounded classifier state is JSON text in input; a choice question named effort offers only the target model's supported efforts, with the existing effort descriptions. Answers are matched by name. Refusals, missing or duplicate answers, non-choice answers, and unsupported efforts fall back as classifier_invalid_output. Authentication errors are not retried; 429 and 5xx responses retry within the deadline, honoring Retry-After. Client cancellation aborts without fallback.

OpenAI receives the same bounded summary as other classifiers: recent user text, assistant progress, up to 8 tool results, a failure summary, and the target model ID. Zero Data Retention and residency depend on project eligibility; see OpenAI's data guide. The us. and eu. endpoints need a project or organization that meets OpenAI's regional requirements (for eu., Modified Abuse Monitoring or Zero Data Retention); otherwise requests fail and the router falls back. The API is public beta; gpt-6-luna is the only supported model. Pricing, checked 2026-10-07, is $0.10 per 1M input tokens with no output or cache charges; regional premiums and long-context multipliers can apply. Recheck OpenAI pricing.

Verification used mocked HTTP contract tests only; no live or authenticated calls were made, so live compatibility is unverified. OpenAI's claim of about 10x lower latency than the Responses API is its own and was not measured under the router's deadline. Effort-selection quality has not been evaluated against labeled coding-agent decisions, so Jev remains the default.

The Decisions guide was checked 2026-10-07. Request and response types came from the openai npm package 7.30.0 (src/resources/decisions.ts, generated from OpenAI's OpenAPI spec); no /v1/decisions API reference page was published then. Regional URLs come from the data controls guide, checked the same day. See issue #33.

CLM (provider: "clm")

CLM, served by a clm-serve server that you run. clm-serve needs a separate pooling backend (Qwen3-8B embeddings, for example vLLM) that you also run. This package only calls clm-serve over HTTP. See docs/classifiers/clm.md for the pinned version and limits. Only API compatibility is verified, not effort-decision quality.

| Field | Default | Purpose | | --- | --- | --- | | baseUrl | http://127.0.0.1:8700 | CLM server. HTTPS, or plain HTTP to a loopback host only. | | apiKey | none | Optional; set only when the server sets CLM_API_KEY. | | model | server's clm-latest | Optional CLM head: clm-latest, clm-raw, or a head loaded with --model NAME=PATH. | | timeoutMs | 4000 | Total classification budget, including retries. |

Calls {baseUrl}/v1/systemone, with the same key rules as Laya (a bearer token; a blank key counts as unset).

Microsoft-Decision-1 (provider: "microsoft-decision")

Microsoft's Microsoft-Decision-1, deployed in your own Microsoft Foundry resource. Deploy the model, then copy the resource endpoint, deployment name, and API key from the deployment details page.

| Field | Default | Purpose | | --- | --- | --- | | baseUrl | none | Foundry resource endpoint, e.g. https://<resource>.services.ai.azure.com. Required. HTTPS, without credentials, query, or fragment. | | apiKey | none | The resource's API key, sent in an api-key header. Required. | | model | none | Your deployment name, not the model name. Required. | | timeoutMs | 4000 | Total classification budget, including retries. |

The provider calls POST {baseUrl}/providers/microsoft/v1/systemone with the System One request: the bounded state and one choice question named effort over the target model's supported efforts. It reads answers.effort.choice; an unsupported or missing choice falls back as classifier_invalid_output. Error bodies are discarded. Authentication errors are not retried; 429 and 5xx responses retry within the deadline, honoring Retry-After. Microsoft Entra ID bearer tokens are not supported, because the router does not acquire or refresh them. Pricing was not rechecked; see Foundry's model catalog before use.

The contract was checked against Microsoft's how-to guide on 2026-10-11 (see issue #64). Verification used mocked HTTP contract tests only; no live or authenticated calls were made. Microsoft's accuracy and latency claims are its own, and effort-selection quality has not been evaluated, so Jev remains the default.

Liquid d1 (provider: "liquid")

Liquid d1, Liquid AI's hosted decision model, through Liquid's API or Vercel AI Gateway.

| Field | Default | Purpose | | --- | --- | --- | | apiKey | none | Liquid key (liquid_...), or a Vercel AI Gateway key. Required. | | baseUrl | https://api.liquid.ai | Set to https://ai-gateway.vercel.sh/typesafe for a Vercel key. No other values are accepted. | | model | the endpoint's model | Optional. d1 (default) or the text-only d1:free for Liquid; liquid/d1 for Vercel. | | timeoutMs | 4000 | Total classification budget, including retries. |

The provider calls POST https://api.liquid.ai/decisions/v1/systemone or POST https://ai-gateway.vercel.sh/typesafe/v1/systemone with the System One request and a bearer key, and reads answers.effort.choice. As with Jev, the key is not inspected, so baseUrl must match the credential. An unsupported or missing choice falls back as classifier_invalid_output. Error bodies are discarded. Authentication errors are not retried; 429 and 5xx responses retry within the deadline, honoring Retry-After. d1's probabilities and confidence are ignored.

The contract was checked against Liquid's d1 guide and Vercel's TypeSafe API page and d1 changelog on 2026-10-11 (see issue #68). Verification used mocked HTTP contract tests only; no live or authenticated calls were made. Liquid's quality claims are its own, and effort-selection quality has not been evaluated, so Jev remains the default.

SGLang (provider: "sglang")

An SGLang server that you run, serving a chat model (v0.5.21 and later) or a decision checkpoint such as PPLX-Decider (nightly builds). This package only calls it over HTTP. See docs/classifiers/sglang.md for the pinned version and limits. Only API compatibility is verified, not effort-decision quality.

| Field | Default | Purpose | | --- | --- | --- | | baseUrl | http://127.0.0.1:30000 | SGLang server. HTTPS, or plain HTTP to a loopback host only. | | apiKey | none | Optional; set only when the server runs with --api-key. | | model | default | Sent because SGLang requires it, but SGLang ignores the value. Must not contain a colon, which SGLang reads as a LoRA adapter. | | timeoutMs | 4000 | Total classification budget, including retries. |

Calls {baseUrl}/v1/systemone, with the same key rules as Laya (a bearer token; a blank key counts as unset). SGLang refuses some servers and chat templates with a 400, which falls back without a retry.

Exports

• classifierProviders: every classifier, for createConfiguredSelector in @reasoning-router/core.
• jevClassifierProvider, clefClassifierProvider, layaClassifierProvider, kevClassifierProvider: each provider.
• openAIDecisionsClassifierProvider, createOpenAIDecisionsTransport, resolveOpenAIDecisionsConnection, and the OpenAIDecisionsConnection type.
• clmClassifierProvider, createClmTransport, resolveClmConnection, and the ClmConnection type.
• microsoftDecisionClassifierProvider, createMicrosoftDecisionTransport, resolveMicrosoftDecisionConnection, and the MicrosoftDecisionConnection type.
• liquidClassifierProvider, createLiquidTransport, resolveLiquidConnection, and the LiquidConnection type.
• sglangClassifierProvider, createSglangTransport, resolveSglangConnection, and the SglangConnection type.
• createJevClassifier, createJevTransport, createClefTransport, createLayaTransport, createKevTransport, resolveJevConnection, resolveClefConnection, resolveLayaConnection, resolveKevConnection: lower-level helpers.
• ClassifierRequestError and the Fetch type, used by the transports.

License

MIT