@mupt-ai/dari-router 0.1.0

package/dist/auto_router.d.ts

@@ -0,0 +1,10 @@
+import { type DariRoutingPolicyOptions } from "./dari_policy.js";
+import type { RoutingPolicy } from "./framework_types.js";
+export declare const DEFAULT_AUTO_ROUTER_ENDPOINT = "https://routing.dari.dev/v1/auto-router";
+export declare const DEFAULT_AUTO_ROUTER_MODEL = "dari/auto-router";
+export declare const DEFAULT_AUTO_ROUTER_CONTEXT_WINDOW_CHARS = 32000;
+export type CreateAutoRouterOptions<Metadata = unknown> = {
+    apiKey: string;
+    endpoint?: string;
+} & Omit<DariRoutingPolicyOptions<Metadata>, "selector" | "runtime" | "selectorModel" | "selectorContextWindowChars" | "strategy" | "customConfig">;
+export declare function createAutoRouter<Metadata = unknown>(options: CreateAutoRouterOptions<Metadata>): RoutingPolicy<Metadata>;

package/dist/auto_router.js

@@ -0,0 +1,56 @@
+// Dari Auto Router — a hosted routing policy for self-hosted createRouter
+// users. Dari bills the calling org per selection call.
+import { RouterFrameworkError } from "./framework_error.js";
+import { createDariRoutingPolicy } from "./dari_policy.js";
+export const DEFAULT_AUTO_ROUTER_ENDPOINT = "https://routing.dari.dev/v1/auto-router";
+export const DEFAULT_AUTO_ROUTER_MODEL = "dari/auto-router";
+export const DEFAULT_AUTO_ROUTER_CONTEXT_WINDOW_CHARS = 32_000;
+export function createAutoRouter(options) {
+    if (!options.apiKey?.trim()) {
+        throw new RouterFrameworkError("configuration", "createAutoRouter requires an apiKey.", "auto_router_api_key_missing");
+    }
+    // The hosted Auto Router serves only the default routing policy and drops
+    // custom rules server-side, so accepting them here would silently ignore
+    // them. Reject loudly instead of routing with rules the user thinks apply.
+    const unsupported = options;
+    if (unsupported.strategy !== undefined || unsupported.customConfig !== undefined) {
+        throw new RouterFrameworkError("configuration", "createAutoRouter does not support custom routing rules (strategy/customConfig): " +
+            "the hosted Dari Auto Router serves only the default policy. Use " +
+            "createDariRoutingPolicy with a self-hosted selector or the Dari managed " +
+            "platform for custom routing.", "auto_router_custom_rules_unsupported");
+    }
+    const endpoint = (options.endpoint ?? DEFAULT_AUTO_ROUTER_ENDPOINT)
+        .replace(/\/+$/, "")
+        .replace(/\/select$/, "");
+    const { apiKey, endpoint: _endpoint, ...policyOptions } = options;
+    return createDariRoutingPolicy({
+        ...policyOptions,
+        selectorModel: DEFAULT_AUTO_ROUTER_MODEL,
+        selectorContextWindowChars: DEFAULT_AUTO_ROUTER_CONTEXT_WINDOW_CHARS,
+        selector: async (request, signal) => {
+            let response;
+            try {
+                response = await fetch(`${endpoint}/select`, {
+                    method: "POST",
+                    headers: {
+                        Authorization: `Bearer ${apiKey}`,
+                        "Content-Type": "application/json",
+                    },
+                    body: JSON.stringify(request),
+                    signal,
+                });
+            }
+            catch (error) {
+                if (error instanceof Error && error.name === "AbortError")
+                    throw error;
+                throw new RouterFrameworkError("policy", `Dari Auto Router request failed: ${error instanceof Error ? error.message : String(error)}`, "auto_router_request_failed", undefined, { cause: error });
+            }
+            if (!response.ok) {
+                const body = await response.text().catch(() => "");
+                throw new RouterFrameworkError("policy", `Dari Auto Router returned HTTP ${response.status}: ${body}`, "auto_router_http_error");
+            }
+            return await response.text();
+        },
+    });
+}
+//# sourceMappingURL=auto_router.js.map

package/dist/auto_router.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"auto_router.js","sourceRoot":"","sources":["../src/auto_router.ts"],"names":[],"mappings":"AAAA,0EAA0E;AAC1E,wDAAwD;AAExD,OAAO,EAAE,oBAAoB,EAAE,MAAM,sBAAsB,CAAC;AAC5D,OAAO,EAAE,uBAAuB,EAAiC,MAAM,kBAAkB,CAAC;AAG1F,MAAM,CAAC,MAAM,4BAA4B,GAAG,yCAAyC,CAAC;AACtF,MAAM,CAAC,MAAM,yBAAyB,GAAG,kBAAkB,CAAC;AAC5D,MAAM,CAAC,MAAM,wCAAwC,GAAG,MAAM,CAAC;AAe/D,MAAM,UAAU,gBAAgB,CAC9B,OAA0C;IAE1C,IAAI,CAAC,OAAO,CAAC,MAAM,EAAE,IAAI,EAAE,EAAE,CAAC;QAC5B,MAAM,IAAI,oBAAoB,CAC5B,eAAe,EACf,sCAAsC,EACtC,6BAA6B,CAC9B,CAAC;IACJ,CAAC;IACD,0EAA0E;IAC1E,yEAAyE;IACzE,2EAA2E;IAC3E,MAAM,WAAW,GAAG,OAAyD,CAAC;IAC9E,IAAI,WAAW,CAAC,QAAQ,KAAK,SAAS,IAAI,WAAW,CAAC,YAAY,KAAK,SAAS,EAAE,CAAC;QACjF,MAAM,IAAI,oBAAoB,CAC5B,eAAe,EACf,kFAAkF;YAChF,kEAAkE;YAClE,0EAA0E;YAC1E,8BAA8B,EAChC,sCAAsC,CACvC,CAAC;IACJ,CAAC;IACD,MAAM,QAAQ,GAAG,CAAC,OAAO,CAAC,QAAQ,IAAI,4BAA4B,CAAC;SAChE,OAAO,CAAC,MAAM,EAAE,EAAE,CAAC;SACnB,OAAO,CAAC,WAAW,EAAE,EAAE,CAAC,CAAC;IAC5B,MAAM,EACJ,MAAM,EACN,QAAQ,EAAE,SAAS,EACnB,GAAG,aAAa,EACjB,GAAG,OAAO,CAAC;IAEZ,OAAO,uBAAuB,CAAW;QACvC,GAAG,aAAa;QAChB,aAAa,EAAE,yBAAyB;QACxC,0BAA0B,EAAE,wCAAwC;QACpE,QAAQ,EAAE,KAAK,EAAE,OAAO,EAAE,MAAM,EAAE,EAAE;YAClC,IAAI,QAAkB,CAAC;YACvB,IAAI,CAAC;gBACH,QAAQ,GAAG,MAAM,KAAK,CAAC,GAAG,QAAQ,SAAS,EAAE;oBAC3C,MAAM,EAAE,MAAM;oBACd,OAAO,EAAE;wBACP,aAAa,EAAE,UAAU,MAAM,EAAE;wBACjC,cAAc,EAAE,kBAAkB;qBACnC;oBACD,IAAI,EAAE,IAAI,CAAC,SAAS,CAAC,OAAO,CAAC;oBAC7B,MAAM;iBACP,CAAC,CAAC;YACL,CAAC;YAAC,OAAO,KAAK,EAAE,CAAC;gBACf,IAAI,KAAK,YAAY,KAAK,IAAI,KAAK,CAAC,IAAI,KAAK,YAAY;oBAAE,MAAM,KAAK,CAAC;gBACvE,MAAM,IAAI,oBAAoB,CAC5B,QAAQ,EACR,oCAAoC,KAAK,YAAY,KAAK,CAAC,CAAC,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,MAAM,CAAC,KAAK,CAAC,EAAE,EAC5F,4BAA4B,EAC5B,SAAS,EACT,EAAE,KAAK,EAAE,KAAK,EAAE,CACjB,CAAC;YACJ,CAAC;YACD,IAAI,CAAC,QAAQ,CAAC,EAAE,EAAE,CAAC;gBACjB,MAAM,IAAI,GAAG,MAAM,QAAQ,CAAC,IAAI,EAAE,CAAC,KAAK,CAAC,GAAG,EAAE,CAAC,EAAE,CAAC,CAAC;gBACnD,MAAM,IAAI,oBAAoB,CAC5B,QAAQ,EACR,kCAAkC,QAAQ,CAAC,MAAM,KAAK,IAAI,EAAE,EAC5D,wBAAwB,CACzB,CAAC;YACJ,CAAC;YACD,OAAO,MAAM,QAAQ,CAAC,IAAI,EAAE,CAAC;QAC/B,CAAC;KACF,CAAC,CAAC;AACL,CAAC"}

package/dist/cache_behavior.d.ts

@@ -0,0 +1,14 @@
+import type { RouterPrefixHit } from "./types.js";
+export declare const PREFIX_WARMTH_WINDOW_MS: number;
+export declare const ANTHROPIC_LOOKBACK_BLOCKS = 19;
+export declare const CROSS_TOKENIZER_RATIO = 1.1;
+export type PromptCacheProvider = "openai" | "anthropic" | "fireworks" | "meta";
+export declare function isAnthropicFamily(model: string): boolean;
+export declare function isOpenAiFamily(model: string): boolean;
+export declare function promptCacheProviderForModel(model: string): PromptCacheProvider | null;
+export declare function openAiCachedPrefixTokens(promptTokens: number): number;
+export declare function anthropicMinCacheTokens(model: string): number;
+export declare function tokenizerFamily(model: string): string;
+export declare function providerMinCacheTokens(model: string): number;
+export declare function warmPrefixHits(hits: readonly RouterPrefixHit[] | undefined, nowMs: number): RouterPrefixHit[];
+export declare function providerCacheableTokens(model: string, provider: PromptCacheProvider, prefixTokens: number): number;

package/dist/cache_behavior.js

@@ -0,0 +1,112 @@
+import { nativeModelId, providerForModel } from "./model_ids.js";
+// Provider prompt caches stop serving a prefix roughly five minutes after it
+// was last written. Stored prefix entries outlive that: a host's lookback
+// keeps them long enough to still name the conversation and the decision it
+// continued from. Only entries inside this window describe warmth.
+export const PREFIX_WARMTH_WINDOW_MS = 5 * 60 * 1000;
+// Anthropic checks 20 positions per breakpoint, counting the breakpoint
+// itself as the first, so an old cached breakpoint hits only when at most
+// 19 content blocks were appended after it. These are routing estimates, never
+// billing.
+export const ANTHROPIC_LOOKBACK_BLOCKS = 19;
+export const CROSS_TOKENIZER_RATIO = 1.1;
+const OPENAI_MIN_CACHE_TOKENS = 1024;
+const OPENAI_CACHE_INCREMENT_TOKENS = 128;
+// providerForModel() lowercases the provider prefix, so family checks must
+// compare case-insensitively too. Unlike providerForModel this never throws;
+// malformed ids classify as no family.
+function providerPrefix(model) {
+    const slash = model.indexOf("/");
+    return slash <= 0 ? "" : model.slice(0, slash).toLowerCase();
+}
+export function isAnthropicFamily(model) {
+    return providerPrefix(model) === "anthropic";
+}
+export function isOpenAiFamily(model) {
+    return providerPrefix(model) === "openai";
+}
+function isFireworksFamily(model) {
+    return providerPrefix(model) === "fireworks";
+}
+function isMetaFamily(model) {
+    return providerPrefix(model) === "meta";
+}
+export function promptCacheProviderForModel(model) {
+    if (isOpenAiFamily(model))
+        return "openai";
+    if (isAnthropicFamily(model))
+        return "anthropic";
+    if (isFireworksFamily(model))
+        return "fireworks";
+    if (isMetaFamily(model))
+        return "meta";
+    return null;
+}
+export function openAiCachedPrefixTokens(promptTokens) {
+    if (promptTokens < OPENAI_MIN_CACHE_TOKENS)
+        return 0;
+    const extra = promptTokens - OPENAI_MIN_CACHE_TOKENS;
+    return (OPENAI_MIN_CACHE_TOKENS +
+        Math.floor(extra / OPENAI_CACHE_INCREMENT_TOKENS) * OPENAI_CACHE_INCREMENT_TOKENS);
+}
+// Anthropic cache minimums vary by model family. Keep this routing-side
+// estimate conservative for known higher-minimum models; unknown future
+// direct Anthropic models fall back to the common 1024-token threshold.
+export function anthropicMinCacheTokens(model) {
+    if (!isAnthropicFamily(model))
+        return OPENAI_MIN_CACHE_TOKENS;
+    const native = nativeModelId(model)
+        .toLowerCase()
+        .replace(/(?<=\d)\.(?=\d)/g, "-");
+    if (native.includes("opus-4-6") ||
+        native.includes("opus-4-5") ||
+        native.includes("haiku-4-5")) {
+        return 4096;
+    }
+    if (native.includes("fable-5")) {
+        return 512;
+    }
+    return OPENAI_MIN_CACHE_TOKENS;
+}
+export function tokenizerFamily(model) {
+    const provider = providerForModel(model);
+    if (provider === "fireworks") {
+        const native = nativeModelId(model).toLowerCase();
+        const accountPrefix = "accounts/fireworks/models/";
+        const modelFamily = native.startsWith(accountPrefix)
+            ? native.slice(accountPrefix.length)
+            : native;
+        const slash = modelFamily.indexOf("/");
+        return `${provider}/${slash === -1 ? modelFamily : modelFamily.slice(0, slash)}`;
+    }
+    return provider;
+}
+export function providerMinCacheTokens(model) {
+    if (isAnthropicFamily(model))
+        return anthropicMinCacheTokens(model);
+    if (isFireworksFamily(model) || isMetaFamily(model))
+        return 0;
+    return OPENAI_MIN_CACHE_TOKENS;
+}
+// The stored entries a provider cache can still be re-reading at nowMs. This
+// is the definition of "warm", and every cost estimate rests on it, so an
+// unparseable timestamp reads as cold: warmth is proven, never assumed.
+// nowMs is supplied by the caller because this package reads no clock.
+export function warmPrefixHits(hits, nowMs) {
+    const cutoffMs = nowMs - PREFIX_WARMTH_WINDOW_MS;
+    return (hits ?? []).filter((hit) => {
+        const updatedAtMs = Date.parse(hit.updated_at);
+        return Number.isFinite(updatedAtMs) && updatedAtMs >= cutoffMs;
+    });
+}
+export function providerCacheableTokens(model, provider, prefixTokens) {
+    if (prefixTokens <= 0)
+        return 0;
+    if (provider === "openai")
+        return openAiCachedPrefixTokens(prefixTokens);
+    if (provider === "anthropic") {
+        return prefixTokens >= anthropicMinCacheTokens(model) ? Math.floor(prefixTokens) : 0;
+    }
+    return Math.floor(prefixTokens);
+}
+//# sourceMappingURL=cache_behavior.js.map

package/dist/cache_behavior.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"cache_behavior.js","sourceRoot":"","sources":["../src/cache_behavior.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,aAAa,EAAE,gBAAgB,EAAE,MAAM,gBAAgB,CAAC;AAGjE,6EAA6E;AAC7E,0EAA0E;AAC1E,4EAA4E;AAC5E,mEAAmE;AACnE,MAAM,CAAC,MAAM,uBAAuB,GAAG,CAAC,GAAG,EAAE,GAAG,IAAI,CAAC;AAErD,wEAAwE;AACxE,0EAA0E;AAC1E,+EAA+E;AAC/E,WAAW;AACX,MAAM,CAAC,MAAM,yBAAyB,GAAG,EAAE,CAAC;AAC5C,MAAM,CAAC,MAAM,qBAAqB,GAAG,GAAG,CAAC;AACzC,MAAM,uBAAuB,GAAG,IAAI,CAAC;AACrC,MAAM,6BAA6B,GAAG,GAAG,CAAC;AAI1C,2EAA2E;AAC3E,6EAA6E;AAC7E,uCAAuC;AACvC,SAAS,cAAc,CAAC,KAAa;IACnC,MAAM,KAAK,GAAG,KAAK,CAAC,OAAO,CAAC,GAAG,CAAC,CAAC;IACjC,OAAO,KAAK,IAAI,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,KAAK,CAAC,KAAK,CAAC,CAAC,EAAE,KAAK,CAAC,CAAC,WAAW,EAAE,CAAC;AAC/D,CAAC;AAED,MAAM,UAAU,iBAAiB,CAAC,KAAa;IAC7C,OAAO,cAAc,CAAC,KAAK,CAAC,KAAK,WAAW,CAAC;AAC/C,CAAC;AAED,MAAM,UAAU,cAAc,CAAC,KAAa;IAC1C,OAAO,cAAc,CAAC,KAAK,CAAC,KAAK,QAAQ,CAAC;AAC5C,CAAC;AAED,SAAS,iBAAiB,CAAC,KAAa;IACtC,OAAO,cAAc,CAAC,KAAK,CAAC,KAAK,WAAW,CAAC;AAC/C,CAAC;AAED,SAAS,YAAY,CAAC,KAAa;IACjC,OAAO,cAAc,CAAC,KAAK,CAAC,KAAK,MAAM,CAAC;AAC1C,CAAC;AAED,MAAM,UAAU,2BAA2B,CAAC,KAAa;IACvD,IAAI,cAAc,CAAC,KAAK,CAAC;QAAE,OAAO,QAAQ,CAAC;IAC3C,IAAI,iBAAiB,CAAC,KAAK,CAAC;QAAE,OAAO,WAAW,CAAC;IACjD,IAAI,iBAAiB,CAAC,KAAK,CAAC;QAAE,OAAO,WAAW,CAAC;IACjD,IAAI,YAAY,CAAC,KAAK,CAAC;QAAE,OAAO,MAAM,CAAC;IACvC,OAAO,IAAI,CAAC;AACd,CAAC;AAED,MAAM,UAAU,wBAAwB,CAAC,YAAoB;IAC3D,IAAI,YAAY,GAAG,uBAAuB;QAAE,OAAO,CAAC,CAAC;IACrD,MAAM,KAAK,GAAG,YAAY,GAAG,uBAAuB,CAAC;IACrD,OAAO,CACL,uBAAuB;QACvB,IAAI,CAAC,KAAK,CAAC,KAAK,GAAG,6BAA6B,CAAC,GAAG,6BAA6B,CAClF,CAAC;AACJ,CAAC;AAED,wEAAwE;AACxE,wEAAwE;AACxE,wEAAwE;AACxE,MAAM,UAAU,uBAAuB,CAAC,KAAa;IACnD,IAAI,CAAC,iBAAiB,CAAC,KAAK,CAAC;QAAE,OAAO,uBAAuB,CAAC;IAC9D,MAAM,MAAM,GAAG,aAAa,CAAC,KAAK,CAAC;SAChC,WAAW,EAAE;SACb,OAAO,CAAC,kBAAkB,EAAE,GAAG,CAAC,CAAC;IAEpC,IACE,MAAM,CAAC,QAAQ,CAAC,UAAU,CAAC;QAC3B,MAAM,CAAC,QAAQ,CAAC,UAAU,CAAC;QAC3B,MAAM,CAAC,QAAQ,CAAC,WAAW,CAAC,EAC5B,CAAC;QACD,OAAO,IAAI,CAAC;IACd,CAAC;IACD,IAAI,MAAM,CAAC,QAAQ,CAAC,SAAS,CAAC,EAAE,CAAC;QAC/B,OAAO,GAAG,CAAC;IACb,CAAC;IACD,OAAO,uBAAuB,CAAC;AACjC,CAAC;AAED,MAAM,UAAU,eAAe,CAAC,KAAa;IAC3C,MAAM,QAAQ,GAAG,gBAAgB,CAAC,KAAK,CAAC,CAAC;IACzC,IAAI,QAAQ,KAAK,WAAW,EAAE,CAAC;QAC7B,MAAM,MAAM,GAAG,aAAa,CAAC,KAAK,CAAC,CAAC,WAAW,EAAE,CAAC;QAClD,MAAM,aAAa,GAAG,4BAA4B,CAAC;QACnD,MAAM,WAAW,GAAG,MAAM,CAAC,UAAU,CAAC,aAAa,CAAC;YAClD,CAAC,CAAC,MAAM,CAAC,KAAK,CAAC,aAAa,CAAC,MAAM,CAAC;YACpC,CAAC,CAAC,MAAM,CAAC;QACX,MAAM,KAAK,GAAG,WAAW,CAAC,OAAO,CAAC,GAAG,CAAC,CAAC;QACvC,OAAO,GAAG,QAAQ,IAAI,KAAK,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC,WAAW,CAAC,CAAC,CAAC,WAAW,CAAC,KAAK,CAAC,CAAC,EAAE,KAAK,CAAC,EAAE,CAAC;IACnF,CAAC;IACD,OAAO,QAAQ,CAAC;AAClB,CAAC;AAED,MAAM,UAAU,sBAAsB,CAAC,KAAa;IAClD,IAAI,iBAAiB,CAAC,KAAK,CAAC;QAAE,OAAO,uBAAuB,CAAC,KAAK,CAAC,CAAC;IACpE,IAAI,iBAAiB,CAAC,KAAK,CAAC,IAAI,YAAY,CAAC,KAAK,CAAC;QAAE,OAAO,CAAC,CAAC;IAC9D,OAAO,uBAAuB,CAAC;AACjC,CAAC;AAED,6EAA6E;AAC7E,0EAA0E;AAC1E,wEAAwE;AACxE,uEAAuE;AACvE,MAAM,UAAU,cAAc,CAC5B,IAA4C,EAC5C,KAAa;IAEb,MAAM,QAAQ,GAAG,KAAK,GAAG,uBAAuB,CAAC;IACjD,OAAO,CAAC,IAAI,IAAI,EAAE,CAAC,CAAC,MAAM,CAAC,CAAC,GAAG,EAAE,EAAE;QACjC,MAAM,WAAW,GAAG,IAAI,CAAC,KAAK,CAAC,GAAG,CAAC,UAAU,CAAC,CAAC;QAC/C,OAAO,MAAM,CAAC,QAAQ,CAAC,WAAW,CAAC,IAAI,WAAW,IAAI,QAAQ,CAAC;IACjE,CAAC,CAAC,CAAC;AACL,CAAC;AAED,MAAM,UAAU,uBAAuB,CACrC,KAAa,EACb,QAA6B,EAC7B,YAAoB;IAEpB,IAAI,YAAY,IAAI,CAAC;QAAE,OAAO,CAAC,CAAC;IAChC,IAAI,QAAQ,KAAK,QAAQ;QAAE,OAAO,wBAAwB,CAAC,YAAY,CAAC,CAAC;IACzE,IAAI,QAAQ,KAAK,WAAW,EAAE,CAAC;QAC7B,OAAO,YAAY,IAAI,uBAAuB,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,KAAK,CAAC,YAAY,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;IACvF,CAAC;IACD,OAAO,IAAI,CAAC,KAAK,CAAC,YAAY,CAAC,CAAC;AAClC,CAAC"}

package/dist/cache_scope.d.ts

@@ -0,0 +1,5 @@
+export type ReasoningCacheScope = "effort_keyed" | "shared";
+export declare const DEFAULT_REASONING_CACHE_SCOPES: Readonly<Record<string, ReasoningCacheScope>>;
+export type ReasoningCacheScopeLookup = (modelId: string) => ReasoningCacheScope;
+export declare function reasoningCacheScope(modelId: string): ReasoningCacheScope;
+export declare function cachePartitionKey(modelId: string, reasoningBucket: string | null | undefined, cacheScope?: ReasoningCacheScopeLookup): string;

package/dist/cache_scope.js

@@ -0,0 +1,49 @@
+import { canonicalModelId } from "./model_ids.js";
+// How a model's provider prompt cache behaves when the reasoning payload
+// changes between otherwise identical requests, from live-provider and
+// router-path validation:
+//   effort_keyed - each effective reasoning effort warms its own cache
+//                  partition; switching effort misses/rewrites, switching back
+//                  reuses that effort's still-warm partition.
+//   shared       - one cache partition regardless of reasoning effort.
+// OpenAI reasoning models are effort_keyed (a first-ever effort missed a
+// steady-warm prefix on every model tested). Anthropic models are effort_keyed
+// because pi-ai places cache_control on message blocks, which Anthropic
+// invalidates on thinking changes; budget-thinking Haiku would share under a
+// system-block breakpoint, so revalidate its entry if pi-ai's Anthropic cache
+// placement changes. Fireworks caching ignored reasoning entirely.
+// Unlisted models default to effort_keyed: the safe direction is to
+// under-claim warmth, never to price a cold prefix as a cache read.
+export const DEFAULT_REASONING_CACHE_SCOPES = {
+    "openai/gpt-5.6-sol": "effort_keyed",
+    "openai/gpt-5.6-terra": "effort_keyed",
+    "openai/gpt-5.6-luna": "effort_keyed",
+    "openai/gpt-5.5": "effort_keyed",
+    "openai/gpt-4.1-mini": "shared", // single reasoning mode; nothing to key on
+    "anthropic/claude-opus-4-8": "effort_keyed",
+    "anthropic/claude-sonnet-4-6": "effort_keyed",
+    "anthropic/claude-haiku-4-5": "effort_keyed", // via pi-ai message-block cache_control; see note above
+    "fireworks/deepseek-ai/DeepSeek-V4-Pro": "shared",
+    "fireworks/deepseek-ai/DeepSeek-V4-Flash": "shared",
+    "fireworks/deepseek-ai/DeepSeek-V4-Flash-0731": "shared",
+    "fireworks/zai-org/GLM-5.2": "shared",
+};
+export function reasoningCacheScope(modelId) {
+    return DEFAULT_REASONING_CACHE_SCOPES[canonicalModelId(modelId)] ?? "effort_keyed";
+}
+// The provider cache partition a prefix entry belongs to. An effort-keyed
+// provider warms one partition per effective reasoning effort, so the pair
+// identifies it; a shared-scope provider holds a single partition per model,
+// and keying that by effort would remember several entries describing one
+// cache. Hosts that store warmth per partition derive their key here so a
+// reclassified provider moves the estimator and their storage together.
+// modelId is used verbatim, matching how ChainHits keys warm prefixes: ids that
+// differ only in spelling stay separate rather than cross-contaminating. The
+// separator cannot appear in a model id or a bucket, so the pair is injective
+// and can never collide with a shared-scope key.
+export function cachePartitionKey(modelId, reasoningBucket, cacheScope = reasoningCacheScope) {
+    if (cacheScope(modelId) === "shared")
+        return modelId;
+    return `${modelId}\u0000${reasoningBucket ?? ""}`;
+}
+//# sourceMappingURL=cache_scope.js.map

package/dist/cache_scope.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"cache_scope.js","sourceRoot":"","sources":["../src/cache_scope.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,gBAAgB,EAAE,MAAM,gBAAgB,CAAC;AAIlD,yEAAyE;AACzE,uEAAuE;AACvE,0BAA0B;AAC1B,uEAAuE;AACvE,+EAA+E;AAC/E,8DAA8D;AAC9D,uEAAuE;AACvE,yEAAyE;AACzE,+EAA+E;AAC/E,wEAAwE;AACxE,6EAA6E;AAC7E,8EAA8E;AAC9E,mEAAmE;AACnE,oEAAoE;AACpE,oEAAoE;AACpE,MAAM,CAAC,MAAM,8BAA8B,GAAkD;IAC3F,oBAAoB,EAAE,cAAc;IACpC,sBAAsB,EAAE,cAAc;IACtC,qBAAqB,EAAE,cAAc;IACrC,gBAAgB,EAAE,cAAc;IAChC,qBAAqB,EAAE,QAAQ,EAAE,2CAA2C;IAC5E,2BAA2B,EAAE,cAAc;IAC3C,6BAA6B,EAAE,cAAc;IAC7C,4BAA4B,EAAE,cAAc,EAAE,wDAAwD;IACtG,uCAAuC,EAAE,QAAQ;IACjD,yCAAyC,EAAE,QAAQ;IACnD,8CAA8C,EAAE,QAAQ;IACxD,2BAA2B,EAAE,QAAQ;CACtC,CAAC;AAIF,MAAM,UAAU,mBAAmB,CAAC,OAAe;IACjD,OAAO,8BAA8B,CAAC,gBAAgB,CAAC,OAAO,CAAC,CAAC,IAAI,cAAc,CAAC;AACrF,CAAC;AAED,0EAA0E;AAC1E,2EAA2E;AAC3E,6EAA6E;AAC7E,0EAA0E;AAC1E,0EAA0E;AAC1E,wEAAwE;AACxE,gFAAgF;AAChF,6EAA6E;AAC7E,8EAA8E;AAC9E,iDAAiD;AACjD,MAAM,UAAU,iBAAiB,CAC/B,OAAe,EACf,eAA0C,EAC1C,aAAwC,mBAAmB;IAE3D,IAAI,UAAU,CAAC,OAAO,CAAC,KAAK,QAAQ;QAAE,OAAO,OAAO,CAAC;IACrD,OAAO,GAAG,OAAO,SAAS,eAAe,IAAI,EAAE,EAAE,CAAC;AACpD,CAAC"}

package/dist/cache_switch_decision.d.ts

@@ -0,0 +1,24 @@
+export type WarmIncumbentCost = {
+    fixedTurnCostUsd: number;
+    outputCostPerMtok: number;
+};
+export type SwitchCandidateCost = {
+    fixedTurnCostUsd: number | null;
+    outputCostPerMtok: number | null;
+};
+export type CacheSwitchDecision = {
+    action: "keep";
+    reason: "unknown_switch_cost" | "unknown_switch_output_price" | "not_cheaper_output_tier" | "saves_more_than_threshold";
+    fixedTurnCostUsd?: number;
+    savingsRatio?: number;
+} | {
+    action: "prune";
+    reason: "insufficient_savings";
+    fixedTurnCostUsd: number;
+    savingsRatio: number;
+};
+export declare function decideCacheSwitch(args: {
+    warmIncumbent: WarmIncumbentCost;
+    switchCandidate: SwitchCandidateCost;
+    minSwitchSavingsRatio: number;
+}): CacheSwitchDecision;

package/dist/cache_switch_decision.js

@@ -0,0 +1,21 @@
+export function decideCacheSwitch(args) {
+    const fixedTurnCostUsd = args.switchCandidate.fixedTurnCostUsd;
+    if (fixedTurnCostUsd === null) {
+        return { action: "keep", reason: "unknown_switch_cost" };
+    }
+    const outputCostPerMtok = args.switchCandidate.outputCostPerMtok;
+    if (outputCostPerMtok === null) {
+        return { action: "keep", reason: "unknown_switch_output_price" };
+    }
+    // Higher output price is our capability-upgrade guard. The pruning rule is
+    // only for cheaper-tier switches that look cheap before cache effects.
+    if (outputCostPerMtok >= args.warmIncumbent.outputCostPerMtok) {
+        return { action: "keep", reason: "not_cheaper_output_tier", fixedTurnCostUsd };
+    }
+    const savingsRatio = (args.warmIncumbent.fixedTurnCostUsd - fixedTurnCostUsd) / args.warmIncumbent.fixedTurnCostUsd;
+    if (savingsRatio > args.minSwitchSavingsRatio) {
+        return { action: "keep", reason: "saves_more_than_threshold", fixedTurnCostUsd, savingsRatio };
+    }
+    return { action: "prune", reason: "insufficient_savings", fixedTurnCostUsd, savingsRatio };
+}
+//# sourceMappingURL=cache_switch_decision.js.map

package/dist/cache_switch_decision.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"cache_switch_decision.js","sourceRoot":"","sources":["../src/cache_switch_decision.ts"],"names":[],"mappings":"AAwBA,MAAM,UAAU,iBAAiB,CAAC,IAIjC;IACC,MAAM,gBAAgB,GAAG,IAAI,CAAC,eAAe,CAAC,gBAAgB,CAAC;IAC/D,IAAI,gBAAgB,KAAK,IAAI,EAAE,CAAC;QAC9B,OAAO,EAAE,MAAM,EAAE,MAAM,EAAE,MAAM,EAAE,qBAAqB,EAAE,CAAC;IAC3D,CAAC;IAED,MAAM,iBAAiB,GAAG,IAAI,CAAC,eAAe,CAAC,iBAAiB,CAAC;IACjE,IAAI,iBAAiB,KAAK,IAAI,EAAE,CAAC;QAC/B,OAAO,EAAE,MAAM,EAAE,MAAM,EAAE,MAAM,EAAE,6BAA6B,EAAE,CAAC;IACnE,CAAC;IAED,2EAA2E;IAC3E,uEAAuE;IACvE,IAAI,iBAAiB,IAAI,IAAI,CAAC,aAAa,CAAC,iBAAiB,EAAE,CAAC;QAC9D,OAAO,EAAE,MAAM,EAAE,MAAM,EAAE,MAAM,EAAE,yBAAyB,EAAE,gBAAgB,EAAE,CAAC;IACjF,CAAC;IAED,MAAM,YAAY,GAChB,CAAC,IAAI,CAAC,aAAa,CAAC,gBAAgB,GAAG,gBAAgB,CAAC,GAAG,IAAI,CAAC,aAAa,CAAC,gBAAgB,CAAC;IACjG,IAAI,YAAY,GAAG,IAAI,CAAC,qBAAqB,EAAE,CAAC;QAC9C,OAAO,EAAE,MAAM,EAAE,MAAM,EAAE,MAAM,EAAE,2BAA2B,EAAE,gBAAgB,EAAE,YAAY,EAAE,CAAC;IACjG,CAAC;IAED,OAAO,EAAE,MAAM,EAAE,OAAO,EAAE,MAAM,EAAE,sBAAsB,EAAE,gBAAgB,EAAE,YAAY,EAAE,CAAC;AAC7F,CAAC"}

package/dist/candidate_pruning.d.ts

@@ -0,0 +1,45 @@
+import { type CacheSwitchDecision } from "./cache_switch_decision.js";
+import { type CandidateCostEstimate, type PreviousDecision, type ReasoningEffort, type RoutingCandidate } from "./types.js";
+export declare const MIN_SWITCH_SAVINGS_RATIO = 0.1;
+type PrunedCandidate = {
+    model: string;
+    reasoning_effort: ReasoningEffort;
+    fixed_turn_cost_usd: number;
+    savings_ratio: number;
+};
+type CandidateSwitchKeepDecision = {
+    action: "keep";
+    reason: Extract<CacheSwitchDecision, {
+        action: "keep";
+    }>["reason"] | "same_model_upgrade";
+    fixedTurnCostUsd?: number;
+    savingsRatio?: number;
+};
+type KeptReason = "incumbent" | CandidateSwitchKeepDecision["reason"];
+type KeptCandidate = {
+    model: string;
+    reasoning_effort: ReasoningEffort;
+    reason: KeptReason;
+    savings_ratio?: number;
+};
+export type CandidatePruningAudit = {
+    min_switch_savings_ratio: number;
+    incumbent_model: string | null;
+    incumbent_reasoning_effort: ReasoningEffort | null;
+    incumbent_fixed_turn_cost_usd: number | null;
+    skip_reason: "no_previous_decision" | "incumbent_not_warm" | "incumbent_unpriced" | null;
+    pruned_candidates: PrunedCandidate[];
+    kept_candidates: KeptCandidate[];
+};
+export type CandidatePruning = {
+    candidates: RoutingCandidate[];
+    costEstimates: CandidateCostEstimate[];
+    audit: CandidatePruningAudit;
+};
+export declare function pruneSwitchCandidates(args: {
+    candidates: RoutingCandidate[];
+    costEstimates: CandidateCostEstimate[];
+    previousDecision?: PreviousDecision;
+    minSwitchSavingsRatio: number;
+}): CandidatePruning;
+export {};

package/dist/candidate_pruning.js

@@ -0,0 +1,154 @@
+// Cache-aware pruning for selector-based routing strategies. When the previous
+// model/thinking-level pair is still warm, a nominally cheaper pair can be a
+// fake-cheap switch: the cold start erases most of the sticker-price saving.
+import { decideCacheSwitch, } from "./cache_switch_decision.js";
+import { turnCostUsdAt } from "./cost.js";
+import { FIXED_TURN_COST_COMPARISON_TURNS } from "./fixed_turn_cost_config.js";
+import { REASONING_EFFORTS, routingCandidateKey, } from "./types.js";
+export const MIN_SWITCH_SAVINGS_RATIO = 0.1;
+export function pruneSwitchCandidates(args) {
+    const incumbentCandidate = args.previousDecision
+        ? {
+            model: args.previousDecision.model,
+            reasoningEffort: args.previousDecision.reasoningEffort,
+        }
+        : null;
+    const candidateKeys = new Set(args.candidates.map(routingCandidateKey));
+    const incumbentKey = incumbentCandidate
+        ? routingCandidateKey(incumbentCandidate)
+        : null;
+    const estimateByKey = new Map(args.costEstimates.map((estimate) => [
+        routingCandidateKey({
+            model: estimate.model,
+            reasoningEffort: estimate.reasoning_effort,
+        }),
+        estimate,
+    ]));
+    const incumbent = incumbentKey && candidateKeys.has(incumbentKey)
+        ? estimateByKey.get(incumbentKey)
+        : undefined;
+    if (!incumbent || !incumbentCandidate) {
+        return keepAll(args, incumbentCandidate, "no_previous_decision");
+    }
+    if (incumbent.warm_tokens <= 0) {
+        return keepAll(args, incumbentCandidate, "incumbent_not_warm");
+    }
+    const incumbentCost = turnCostUsdAt(incumbent.fixed_turn_cost_estimate, FIXED_TURN_COST_COMPARISON_TURNS);
+    const incumbentOutputPrice = incumbent.output_cost_per_mtok;
+    if (incumbentCost === null ||
+        incumbentCost <= 0 ||
+        incumbentOutputPrice === null) {
+        return keepAll(args, incumbentCandidate, "incumbent_unpriced");
+    }
+    const prunedCandidates = [];
+    const keptCandidates = [];
+    const survivors = args.candidates.filter((candidate) => {
+        const key = routingCandidateKey(candidate);
+        if (key === incumbentKey) {
+            keptCandidates.push({
+                model: candidate.model,
+                reasoning_effort: candidate.reasoningEffort,
+                reason: "incumbent",
+            });
+            return true;
+        }
+        const decision = candidateSwitchDecision({
+            candidate,
+            estimate: estimateByKey.get(key),
+            incumbent: incumbentCandidate,
+            incumbentCost,
+            incumbentOutputPrice,
+            minSwitchSavingsRatio: args.minSwitchSavingsRatio,
+        });
+        if (decision.action === "keep") {
+            keptCandidates.push({
+                model: candidate.model,
+                reasoning_effort: candidate.reasoningEffort,
+                reason: decision.reason,
+                ...(decision.savingsRatio === undefined
+                    ? {}
+                    : { savings_ratio: decision.savingsRatio }),
+            });
+            return true;
+        }
+        prunedCandidates.push({
+            model: candidate.model,
+            reasoning_effort: candidate.reasoningEffort,
+            fixed_turn_cost_usd: decision.fixedTurnCostUsd,
+            savings_ratio: decision.savingsRatio,
+        });
+        return false;
+    });
+    const survivorKeys = new Set(survivors.map(routingCandidateKey));
+    return {
+        candidates: survivors,
+        costEstimates: args.costEstimates.filter((estimate) => survivorKeys.has(routingCandidateKey({
+            model: estimate.model,
+            reasoningEffort: estimate.reasoning_effort,
+        }))),
+        audit: {
+            min_switch_savings_ratio: args.minSwitchSavingsRatio,
+            incumbent_model: incumbentCandidate.model,
+            incumbent_reasoning_effort: incumbentCandidate.reasoningEffort,
+            incumbent_fixed_turn_cost_usd: incumbentCost,
+            skip_reason: null,
+            pruned_candidates: prunedCandidates,
+            kept_candidates: keptCandidates,
+        },
+    };
+}
+function candidateSwitchDecision(args) {
+    const candidateCost = turnCostUsdAt(args.estimate?.fixed_turn_cost_estimate, FIXED_TURN_COST_COMPARISON_TURNS);
+    if (args.candidate.model !== args.incumbent.model) {
+        return decideCacheSwitch({
+            warmIncumbent: {
+                fixedTurnCostUsd: args.incumbentCost,
+                outputCostPerMtok: args.incumbentOutputPrice,
+            },
+            switchCandidate: {
+                fixedTurnCostUsd: candidateCost,
+                outputCostPerMtok: args.estimate?.output_cost_per_mtok ?? null,
+            },
+            minSwitchSavingsRatio: args.minSwitchSavingsRatio,
+        });
+    }
+    // Output price cannot distinguish thinking levels on one model, so preserve
+    // capability upgrades and judge lower-effort switches by projected cost.
+    if (REASONING_EFFORTS.indexOf(args.candidate.reasoningEffort) >=
+        REASONING_EFFORTS.indexOf(args.incumbent.reasoningEffort)) {
+        return { action: "keep", reason: "same_model_upgrade" };
+    }
+    if (candidateCost === null) {
+        return { action: "keep", reason: "unknown_switch_cost" };
+    }
+    const savingsRatio = (args.incumbentCost - candidateCost) / args.incumbentCost;
+    return savingsRatio > args.minSwitchSavingsRatio
+        ? {
+            action: "keep",
+            reason: "saves_more_than_threshold",
+            fixedTurnCostUsd: candidateCost,
+            savingsRatio,
+        }
+        : {
+            action: "prune",
+            reason: "insufficient_savings",
+            fixedTurnCostUsd: candidateCost,
+            savingsRatio,
+        };
+}
+function keepAll(args, incumbent, skipReason) {
+    return {
+        candidates: args.candidates,
+        costEstimates: args.costEstimates,
+        audit: {
+            min_switch_savings_ratio: args.minSwitchSavingsRatio,
+            incumbent_model: incumbent?.model ?? null,
+            incumbent_reasoning_effort: incumbent?.reasoningEffort ?? null,
+            incumbent_fixed_turn_cost_usd: null,
+            skip_reason: skipReason,
+            pruned_candidates: [],
+            kept_candidates: [],
+        },
+    };
+}
+//# sourceMappingURL=candidate_pruning.js.map

package/dist/candidate_pruning.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"candidate_pruning.js","sourceRoot":"","sources":["../src/candidate_pruning.ts"],"names":[],"mappings":"AAAA,+EAA+E;AAC/E,6EAA6E;AAC7E,6EAA6E;AAE7E,OAAO,EACL,iBAAiB,GAElB,MAAM,4BAA4B,CAAC;AACpC,OAAO,EAAE,aAAa,EAAE,MAAM,WAAW,CAAC;AAC1C,OAAO,EAAE,gCAAgC,EAAE,MAAM,6BAA6B,CAAC;AAC/E,OAAO,EACL,iBAAiB,EACjB,mBAAmB,GAKpB,MAAM,YAAY,CAAC;AAEpB,MAAM,CAAC,MAAM,wBAAwB,GAAG,GAAG,CAAC;AAmD5C,MAAM,UAAU,qBAAqB,CAAC,IAKrC;IACC,MAAM,kBAAkB,GAAG,IAAI,CAAC,gBAAgB;QAC9C,CAAC,CAAC;YACE,KAAK,EAAE,IAAI,CAAC,gBAAgB,CAAC,KAAK;YAClC,eAAe,EAAE,IAAI,CAAC,gBAAgB,CAAC,eAAe;SACvD;QACH,CAAC,CAAC,IAAI,CAAC;IACT,MAAM,aAAa,GAAG,IAAI,GAAG,CAAC,IAAI,CAAC,UAAU,CAAC,GAAG,CAAC,mBAAmB,CAAC,CAAC,CAAC;IACxE,MAAM,YAAY,GAAG,kBAAkB;QACrC,CAAC,CAAC,mBAAmB,CAAC,kBAAkB,CAAC;QACzC,CAAC,CAAC,IAAI,CAAC;IACT,MAAM,aAAa,GAAG,IAAI,GAAG,CAC3B,IAAI,CAAC,aAAa,CAAC,GAAG,CAAC,CAAC,QAAQ,EAAE,EAAE,CAAC;QACnC,mBAAmB,CAAC;YAClB,KAAK,EAAE,QAAQ,CAAC,KAAK;YACrB,eAAe,EAAE,QAAQ,CAAC,gBAAgB;SAC3C,CAAC;QACF,QAAQ;KACT,CAAC,CACH,CAAC;IACF,MAAM,SAAS,GAAG,YAAY,IAAI,aAAa,CAAC,GAAG,CAAC,YAAY,CAAC;QAC/D,CAAC,CAAC,aAAa,CAAC,GAAG,CAAC,YAAY,CAAC;QACjC,CAAC,CAAC,SAAS,CAAC;IACd,IAAI,CAAC,SAAS,IAAI,CAAC,kBAAkB,EAAE,CAAC;QACtC,OAAO,OAAO,CAAC,IAAI,EAAE,kBAAkB,EAAE,sBAAsB,CAAC,CAAC;IACnE,CAAC;IACD,IAAI,SAAS,CAAC,WAAW,IAAI,CAAC,EAAE,CAAC;QAC/B,OAAO,OAAO,CAAC,IAAI,EAAE,kBAAkB,EAAE,oBAAoB,CAAC,CAAC;IACjE,CAAC;IACD,MAAM,aAAa,GAAG,aAAa,CACjC,SAAS,CAAC,wBAAwB,EAClC,gCAAgC,CACjC,CAAC;IACF,MAAM,oBAAoB,GAAG,SAAS,CAAC,oBAAoB,CAAC;IAC5D,IACE,aAAa,KAAK,IAAI;QACtB,aAAa,IAAI,CAAC;QAClB,oBAAoB,KAAK,IAAI,EAC7B,CAAC;QACD,OAAO,OAAO,CAAC,IAAI,EAAE,kBAAkB,EAAE,oBAAoB,CAAC,CAAC;IACjE,CAAC;IAED,MAAM,gBAAgB,GAAsB,EAAE,CAAC;IAC/C,MAAM,cAAc,GAAoB,EAAE,CAAC;IAC3C,MAAM,SAAS,GAAG,IAAI,CAAC,UAAU,CAAC,MAAM,CAAC,CAAC,SAAS,EAAE,EAAE;QACrD,MAAM,GAAG,GAAG,mBAAmB,CAAC,SAAS,CAAC,CAAC;QAC3C,IAAI,GAAG,KAAK,YAAY,EAAE,CAAC;YACzB,cAAc,CAAC,IAAI,CAAC;gBAClB,KAAK,EAAE,SAAS,CAAC,KAAK;gBACtB,gBAAgB,EAAE,SAAS,CAAC,eAAe;gBAC3C,MAAM,EAAE,WAAW;aACpB,CAAC,CAAC;YACH,OAAO,IAAI,CAAC;QACd,CAAC;QACD,MAAM,QAAQ,GAAG,uBAAuB,CAAC;YACvC,SAAS;YACT,QAAQ,EAAE,aAAa,CAAC,GAAG,CAAC,GAAG,CAAC;YAChC,SAAS,EAAE,kBAAkB;YAC7B,aAAa;YACb,oBAAoB;YACpB,qBAAqB,EAAE,IAAI,CAAC,qBAAqB;SAClD,CAAC,CAAC;QACH,IAAI,QAAQ,CAAC,MAAM,KAAK,MAAM,EAAE,CAAC;YAC/B,cAAc,CAAC,IAAI,CAAC;gBAClB,KAAK,EAAE,SAAS,CAAC,KAAK;gBACtB,gBAAgB,EAAE,SAAS,CAAC,eAAe;gBAC3C,MAAM,EAAE,QAAQ,CAAC,MAAM;gBACvB,GAAG,CAAC,QAAQ,CAAC,YAAY,KAAK,SAAS;oBACrC,CAAC,CAAC,EAAE;oBACJ,CAAC,CAAC,EAAE,aAAa,EAAE,QAAQ,CAAC,YAAY,EAAE,CAAC;aAC9C,CAAC,CAAC;YACH,OAAO,IAAI,CAAC;QACd,CAAC;QACD,gBAAgB,CAAC,IAAI,CAAC;YACpB,KAAK,EAAE,SAAS,CAAC,KAAK;YACtB,gBAAgB,EAAE,SAAS,CAAC,eAAe;YAC3C,mBAAmB,EAAE,QAAQ,CAAC,gBAAgB;YAC9C,aAAa,EAAE,QAAQ,CAAC,YAAY;SACrC,CAAC,CAAC;QACH,OAAO,KAAK,CAAC;IACf,CAAC,CAAC,CAAC;IACH,MAAM,YAAY,GAAG,IAAI,GAAG,CAAC,SAAS,CAAC,GAAG,CAAC,mBAAmB,CAAC,CAAC,CAAC;IAEjE,OAAO;QACL,UAAU,EAAE,SAAS;QACrB,aAAa,EAAE,IAAI,CAAC,aAAa,CAAC,MAAM,CAAC,CAAC,QAAQ,EAAE,EAAE,CACpD,YAAY,CAAC,GAAG,CACd,mBAAmB,CAAC;YAClB,KAAK,EAAE,QAAQ,CAAC,KAAK;YACrB,eAAe,EAAE,QAAQ,CAAC,gBAAgB;SAC3C,CAAC,CACH,CACF;QACD,KAAK,EAAE;YACL,wBAAwB,EAAE,IAAI,CAAC,qBAAqB;YACpD,eAAe,EAAE,kBAAkB,CAAC,KAAK;YACzC,0BAA0B,EAAE,kBAAkB,CAAC,eAAe;YAC9D,6BAA6B,EAAE,aAAa;YAC5C,WAAW,EAAE,IAAI;YACjB,iBAAiB,EAAE,gBAAgB;YACnC,eAAe,EAAE,cAAc;SAChC;KACF,CAAC;AACJ,CAAC;AAED,SAAS,uBAAuB,CAAC,IAOhC;IACC,MAAM,aAAa,GAAG,aAAa,CACjC,IAAI,CAAC,QAAQ,EAAE,wBAAwB,EACvC,gCAAgC,CACjC,CAAC;IACF,IAAI,IAAI,CAAC,SAAS,CAAC,KAAK,KAAK,IAAI,CAAC,SAAS,CAAC,KAAK,EAAE,CAAC;QAClD,OAAO,iBAAiB,CAAC;YACvB,aAAa,EAAE;gBACb,gBAAgB,EAAE,IAAI,CAAC,aAAa;gBACpC,iBAAiB,EAAE,IAAI,CAAC,oBAAoB;aAC7C;YACD,eAAe,EAAE;gBACf,gBAAgB,EAAE,aAAa;gBAC/B,iBAAiB,EAAE,IAAI,CAAC,QAAQ,EAAE,oBAAoB,IAAI,IAAI;aAC/D;YACD,qBAAqB,EAAE,IAAI,CAAC,qBAAqB;SAClD,CAAC,CAAC;IACL,CAAC;IAED,4EAA4E;IAC5E,yEAAyE;IACzE,IACE,iBAAiB,CAAC,OAAO,CAAC,IAAI,CAAC,SAAS,CAAC,eAAe,CAAC;QACzD,iBAAiB,CAAC,OAAO,CAAC,IAAI,CAAC,SAAS,CAAC,eAAe,CAAC,EACzD,CAAC;QACD,OAAO,EAAE,MAAM,EAAE,MAAM,EAAE,MAAM,EAAE,oBAAoB,EAAE,CAAC;IAC1D,CAAC;IACD,IAAI,aAAa,KAAK,IAAI,EAAE,CAAC;QAC3B,OAAO,EAAE,MAAM,EAAE,MAAM,EAAE,MAAM,EAAE,qBAAqB,EAAE,CAAC;IAC3D,CAAC;IACD,MAAM,YAAY,GAChB,CAAC,IAAI,CAAC,aAAa,GAAG,aAAa,CAAC,GAAG,IAAI,CAAC,aAAa,CAAC;IAC5D,OAAO,YAAY,GAAG,IAAI,CAAC,qBAAqB;QAC9C,CAAC,CAAC;YACE,MAAM,EAAE,MAAM;YACd,MAAM,EAAE,2BAA2B;YACnC,gBAAgB,EAAE,aAAa;YAC/B,YAAY;SACb;QACH,CAAC,CAAC;YACE,MAAM,EAAE,OAAO;YACf,MAAM,EAAE,sBAAsB;YAC9B,gBAAgB,EAAE,aAAa;YAC/B,YAAY;SACb,CAAC;AACR,CAAC;AAED,SAAS,OAAO,CACd,IAIC,EACD,SAAkC,EAClC,UAA6D;IAE7D,OAAO;QACL,UAAU,EAAE,IAAI,CAAC,UAAU;QAC3B,aAAa,EAAE,IAAI,CAAC,aAAa;QACjC,KAAK,EAAE;YACL,wBAAwB,EAAE,IAAI,CAAC,qBAAqB;YACpD,eAAe,EAAE,SAAS,EAAE,KAAK,IAAI,IAAI;YACzC,0BAA0B,EAAE,SAAS,EAAE,eAAe,IAAI,IAAI;YAC9D,6BAA6B,EAAE,IAAI;YACnC,WAAW,EAAE,UAAU;YACvB,iBAAiB,EAAE,EAAE;YACrB,eAAe,EAAE,EAAE;SACpB;KACF,CAAC;AACJ,CAAC"}

package/dist/candidate_resolution.d.ts

@@ -0,0 +1,41 @@
+import { type CandidatePruningAudit } from "./candidate_pruning.js";
+import { type CandidateCostEstimate, type CustomRouterConfig, type CustomRouterRule, type PreviousDecision, type ReasoningEffort, type RouterModelPrice, type RoutingCandidate, type RoutingStrategy } from "./types.js";
+export type CustomDefaultTarget = {
+    model: string;
+    thinkingLevel: ReasoningEffort | null;
+};
+type CandidateSet = {
+    candidates: RoutingCandidate[];
+    costEstimates: CandidateCostEstimate[];
+};
+type CustomDetails = {
+    rules: CustomRouterRule[];
+    defaultTarget: CustomDefaultTarget | null;
+    audit: {
+        dropped_rules: CustomRouterRule[];
+        fallback_all_candidates: boolean;
+        no_routable_rule_models: boolean;
+    };
+};
+export type StrategyCandidateResolution = CandidateSet & ({
+    strategy: "slm";
+    pruning: CandidatePruningAudit;
+} | {
+    strategy: "custom";
+    pruning: null;
+    custom: null;
+} | {
+    strategy: "custom";
+    pruning: CandidatePruningAudit;
+    custom: CustomDetails;
+});
+type CandidateResolutionArgs = {
+    strategy: RoutingStrategy;
+    candidates: RoutingCandidate[];
+    costEstimates: CandidateCostEstimate[];
+    previousDecision?: PreviousDecision;
+    customConfig?: CustomRouterConfig | null;
+    modelPrices: Record<string, RouterModelPrice>;
+};
+export declare function resolveStrategyCandidates(args: CandidateResolutionArgs): StrategyCandidateResolution;
+export {};