neurolink: Skill for Claude Code

The decide inference type

Moved out of the root CLAUDE.md. The transcribe inference type is in src/lib/voice/CLAUDE.md.

generate, stream, decide and transcribe are peers. The first two assume the model emits text; decide does not — a decision model takes one state plus a map of named typed questions and returns one typed, calibrated answer each in a single parallel pass, with no text anywhere. transcribe starts from audio instead of text (see src/lib/voice/CLAUDE.md). InferenceKind includes "transcribe", but no ProviderDescriptor declares it: STT providers are chosen from their own list, STT_PROVIDER_DESCRIPTORS.

The discriminator is one field on ProviderDescriptor:

inferenceKinds?: readonly InferenceKind[] // "generate" | "stream" | "decide" | "transcribe"

Omitted means ["generate","stream"], so every text provider keeps its meaning. For the text and decision providers this is the only declarative statement of provider modality in the codebase (STT has its own list, above) — before it, modality was implied by toolSupport, healthCheck and the auto-select ranks, which is why embedding-only providers (voyage, jina) have to implement getAISDKModel() as a throw. Anything that builds a generation fallback chain, runs the health sweep or offers model choices must filter on inferenceKinds rather than special-casing a provider name.

AIProvider.decide?() is optional and BaseProvider supplies a throwing default, exactly as embed() does. Public surface is neurolink.decide() and neurolink.tryDecide() — the public fail-open variant, which returns null instead of throwing (a question map over the provider's cap is split into batches; a failed batch costs only its own answers, and null comes back only when every batch fails); the internal consumers get the same contract through the private siteDecide (see below) — plus the CLI's neurolink decide [state], a thin wrapper over decide() (src/cli/commands/decide.ts).

Decision providers; descriptor order is precedence. TypeSafe (Jev, hosted), Laya (open weights, at a configured base URL — there is no default), XOR (open weights that also reads an image and audio and extracts typed values, likewise at a configured base URL), SubQ (subq, described below), Perplexity (perplexity-decider, the hosted Decisions API; it also reads images, no video, at the public https://api.perplexity.ai, so a key alone configures it), Cloudflare Clef (cloudflare-clef, on Workers AI; it also reads images, no video, at Cloudflare's own API, so a token and an account id configure it) and OpenAI Decisions (openai-decisions, OpenAI's hosted Decisions API in public beta; it also reads images, no video, at the public https://api.openai.com, so a key alone configures it) and ScaleDown (scaledown, hosted Classify models, text only, at the public https://api.scaledown.xyz, so a key alone configures it; described below) all extend SystemOneDecisionProvider in src/lib/providers/systemOneDecision.ts, which owns the request loop, retries, the auth circuit breaker and answer parsing; each provider supplies only its endpoint, headers, body and error parsing. OpenAI Decisions is the one whose wire is not System One's (named questions in an array, an array of answers), so its catalog entry names decide.dialect: "openai-decisions", translated by src/lib/providers/decide/dialects/openaiDecisions.ts; the Laya, XOR and SubQ entries use the system-one dialect, as does ScaleDown. resolveDefaultDecisionProvider() returns the first DECISION_PROVIDERS entry that is fully configured (a key, plus a base URL for Laya and XOR, which have no built-in endpoint), in decideOrder order (the list is sorted by each descriptor's decideOrder, not by its position), so TypeSafe's 1 before Laya's 2, Laya's before XOR's, XOR's before SubQ's, SubQ's before Perplexity's, Perplexity's before Clef's, Clef's before OpenAI Decisions', and OpenAI Decisions' before ScaleDown's, is what makes TypeSafe win whenever either of its keys (TYPESAFE_API_KEY, AI_GATEWAY_API_KEY) is set alongside LAYA_API_KEY, XOR_API_KEY, SUBQ_API_KEY, PERPLEXITY_API_KEY, CLOUDFLARE_API_KEY, OPENAI_API_KEY or SCALEDOWN_API_KEY. Reordering them changes which model every built-in consumer uses. The three providers whose keys are shared with a text provider (Perplexity, Clef and OpenAI Decisions) stay behind every provider a host sets up on purpose, except ScaleDown (below): a provider added after them would be displaced by an ambient key. SubQ sits ahead of them only because SUBQ_API_KEY belongs to no other provider. ScaleDown's key belongs to no other provider either, yet ScaleDown sits last, after those three, by the owner's decision of 2026-10-15 (decide.order 8): an ambient OPENAI_API_KEY, PERPLEXITY_API_KEY or Cloudflare pair outranks a SCALEDOWN_API_KEY a host set on purpose, and ScaleDown is reached by naming it. Test 18.4 pins scaledown as the last entry.