Model Tasks and Reviewed Default Routes - Tools for Enlightenment developer docs

Every request to cl-router carries a task. This page explains the task vocabulary that @claritylabs/cl-router-policy exports and how a specific call resolves to a reviewed default route.

Providers

MODEL_PROVIDERS lists every provider identifier in the vocabulary: openai, anthropic, google, xai, mistral, cohere, fireworks, moonshot, and deepseek. DIRECT_MODEL_PROVIDERS is the subset with an executable direct-provider adapter in cl-router; it excludes moonshot. A route pairs a provider with a model identifier:

type ModelRoute = {
  provider: ModelProvider;
  model: string;
};

Tasks and task kinds

MODEL_TASKS defines the coarse route vocabulary: chat, chat_vision, voice_transcription, email_draft, email_reply, extraction, extraction_preview, extraction_coverage_recovery, classification, requirement_extraction, org_memory_extraction, analysis, summary, triage, email_extraction, document_extraction, security, mailbox_coordinator, and embeddings. KNOWN_MODEL_TASK_KINDS names the more specific kinds emitted by the current CL SDK and Spot integration, such as extraction_classify, query_reason, and pce_impact_analysis. The ModelTaskKind type keeps an open string tail, so the vocabulary stays forward-compatible with newer SDK kinds.

Resolving a call

Use modelTaskForCall to map a base task plus an optional task kind to the route vocabulary:

import { modelTaskForCall, taskFamilyForCall } from "@claritylabs/cl-router-policy";

modelTaskForCall("extraction", "extraction_classify"); // "classification"
modelTaskForCall("extraction", "extraction_focused");  // "extraction"
modelTaskForCall("chat", "query_reason");              // "chat"
modelTaskForCall("chat", "pce_impact_analysis");       // "analysis"

The resolution rules are:

extraction_classify and query_classify resolve to classification.
extraction_coverage_recovery keeps its own task.
Other extraction_* kinds resolve to extraction.
Other query_* kinds resolve to chat, or chat_vision when the base task is chat_vision.
pce_* kinds resolve to analysis.
Anything else keeps the base task.

taskFamilyForCall returns the most specific stable identifier for quality tracking: the task kind when supplied, otherwise the resolved task. Routing safety helpers take this family value.

Default routes

MODEL_TASK_ROUTES maps every task to its reviewed default route. For example, chat routes to Fireworks DeepSeek V4 Flash, chat_vision routes to OpenAI gpt-5.6-terra, and embeddings routes to OpenAI text-embedding-3-small. FIREWORKS_MODEL_IDS holds the reviewed serverless Fireworks model identifiers, including Muse Glimmer 30B and Inkling. SPECIAL_MODEL_ROUTES adds routes outside the per-task map:

Route ID	Purpose
`fallback`	The shared deterministic recovery route (`FALLBACK_MODEL`).
`extraction_quality`	The quality-primary route for select extraction kinds (`EXTRACTION_QUALITY_MODEL`).
`extraction_coverage_cleanup`	The coverage-cleanup route, also exported as `QUERY_REASON_FALLBACK_MODEL`.

Use defaultModelRouteForId to resolve either kind of route ID:

import { defaultModelRouteForId, MODEL_TASK_ROUTES } from "@claritylabs/cl-router-policy";

const chatRoute = MODEL_TASK_ROUTES.chat;
const fallbackRoute = defaultModelRouteForId("fallback");

Route helpers

directProviderModelForRoute(route) returns the provider-specific model identifier used for direct execution, or null when the provider has no direct adapter for that model. For example, Anthropic dotted versions rewrite to dashed identifiers.
sameModelRoute(left, right) compares two routes by provider and model.
isModelTask(value) narrows a string to ModelTask.

​Providers

​Tasks and task kinds

​Resolving a call

​Default routes

​Route helpers

Providers

Tasks and task kinds

Resolving a call

Default routes

Route helpers