@mupt-ai/dari-router 0.1.0 → 0.2.0

package/src/selector_truncation.ts

@@ -1,48 +0,0 @@
-// Selector prompt sizing. The selector can see a front-trimmed view of the
-// conversation; the eventual provider call still receives the full request.
-
-import { isRecord, type ChatMessage } from "./types.js";
-
-const SELECTOR_TRUNCATION_MARKER = "<earlier messages truncated for routing>";
-const MARKER_MESSAGE: ChatMessage = { role: "user", content: SELECTOR_TRUNCATION_MARKER };
-
-// Drop oldest messages first — in multi-turn conversations the latest
-// messages carry the routing signal.
-export function trimMessagesFromFront(messages: ChatMessage[], dropChars: number): ChatMessage[] {
-  const targetLength = Math.max(0, JSON.stringify(messages).length - Math.max(0, dropChars));
-  const kept = [...messages];
-  while (kept.length > 1 && JSON.stringify([MARKER_MESSAGE, ...kept]).length > targetLength) {
-    kept.shift();
-  }
-
-  const marked = [MARKER_MESSAGE, ...kept];
-  if (JSON.stringify(marked).length <= targetLength) return marked;
-
-  if (kept.length === 1) {
-    const first = kept[0]!;
-    const trimmableContent = trimSourceContent(first);
-    if (trimmableContent === null) return [MARKER_MESSAGE];
-    const prefix = `${SELECTOR_TRUNCATION_MARKER}\n`;
-    let content = `${prefix}${trimmableContent}`;
-    while (content.length > prefix.length) {
-      const excess = JSON.stringify([{ ...first, content }]).length - targetLength;
-      if (excess <= 0) break;
-      content = `${prefix}${content.slice(prefix.length + excess)}`;
-    }
-    return [{ ...first, content }];
-  }
-
-  return [MARKER_MESSAGE];
-}
-
-function trimSourceContent(message: ChatMessage): string | null {
-  if (typeof message.content === "string") return message.content;
-  if (!Array.isArray(message.content)) return null;
-  return message.content.map(formatContentPartForTrim).join("\n");
-}
-
-function formatContentPartForTrim(part: Record<string, unknown>): string {
-  if (isRecord(part) && part.type === "text" && typeof part.text === "string") return part.text;
-  if (isRecord(part) && part.type === "image_url") return "<image omitted>";
-  return JSON.stringify(part);
-}

package/src/selector_wire.ts

@@ -1,167 +0,0 @@
-// Inverse codec for the selector-request wire format buildSelectorRequest
-// produces: the selector input travels as a JSON-stringified user message
-// inside a chat-completion request.
-
-import {
-  anonymizeSelectorInput,
-  assignAnonymousActions,
-  type Rng,
-} from "./anonymous_actions.js";
-import { RouterFrameworkError } from "./framework_error.js";
-import type { JsonObject } from "./json.js";
-import type { SelectorInput } from "./selector_input.js";
-import {
-  isReasoningEffort,
-  isRecord,
-  routingCandidateKey,
-  type PreviousDecision,
-  type RoutingCandidate,
-} from "./types.js";
-
-export type DecodedSelectorRequest = {
-  candidates: RoutingCandidate[];
-  previousDecision?: PreviousDecision;
-  selectorInput: SelectorInput;
-};
-
-// Decodes an untrusted POSTed selector request body; throws kind
-// "invalid_request" on malformed input.
-export function decodeUntrustedSelectorRequest(payload: unknown): DecodedSelectorRequest {
-  const selectorInput = parseSelectorInput(payload);
-  const candidates: RoutingCandidate[] = [];
-  const seenPairs = new Set<string>();
-  for (const [index, pair] of (selectorInput.candidate_pairs as unknown[]).entries()) {
-    const param = `candidate_pairs[${index}]`;
-    if (!isRecord(pair)) {
-      throw invalidSelectorRequest("Selector candidate must be an object.", param);
-    }
-    const candidate = parseCandidatePair(pair, "candidate", param);
-    const key = routingCandidateKey(candidate);
-    if (seenPairs.has(key)) {
-      throw invalidSelectorRequest(
-        `Selector input contains duplicate candidate ${candidate.model}/${candidate.reasoningEffort}.`,
-        param,
-      );
-    }
-    seenPairs.add(key);
-    candidates.push(candidate);
-  }
-  const previousDecision = parsePreviousDecision(selectorInput.previous_decision);
-  if (candidates.length === 0) {
-    throw invalidSelectorRequest("Selector request contains no candidates.");
-  }
-  return { candidates, ...(previousDecision !== undefined ? { previousDecision } : {}), selectorInput };
-}
-
-// anonymizeSelectorInput's validators throw plain Errors; dry-run it here so
-// malformed input surfaces as invalid_request instead of an internal error.
-export function assertAnonymizableSelectorInput(
-  selectorInput: SelectorInput,
-  candidates: RoutingCandidate[],
-  rng: Rng = Math.random,
-): void {
-  if (candidates.length < 2) return;
-  try {
-    anonymizeSelectorInput(
-      selectorInput as unknown as JsonObject,
-      assignAnonymousActions(candidates, rng),
-    );
-  } catch (error) {
-    throw invalidSelectorRequest(
-      `Selector input is invalid: ${error instanceof Error ? error.message : String(error)}`,
-    );
-  }
-}
-
-function parsePreviousDecision(value: unknown): PreviousDecision | undefined {
-  if (value === null || value === undefined) return undefined;
-  if (!isRecord(value)) {
-    throw invalidSelectorRequest(
-      "Selector previous_decision must be an object or null.",
-      "previous_decision",
-    );
-  }
-  const candidate = parseCandidatePair(value, "previous_decision", "previous_decision");
-  if (typeof value.reason !== "string") {
-    throw invalidSelectorRequest(
-      "Selector previous_decision reason must be a string.",
-      "previous_decision.reason",
-    );
-  }
-  return { ...candidate, reason: value.reason };
-}
-
-function parseCandidatePair(
-  value: Record<string, unknown>,
-  subject: "candidate" | "previous_decision",
-  param: string,
-): RoutingCandidate {
-  if (typeof value.model !== "string" || value.model === "") {
-    throw invalidSelectorRequest(`Selector ${subject} is missing a model name.`, `${param}.model`);
-  }
-  if (!isReasoningEffort(value.thinking_level)) {
-    throw invalidSelectorRequest(
-      `Selector ${subject} has invalid thinking_level: ${String(value.thinking_level)}`,
-      `${param}.thinking_level`,
-    );
-  }
-  return { model: value.model, reasoningEffort: value.thinking_level };
-}
-
-function parseSelectorInput(payload: unknown): SelectorInput {
-  if (!isRecord(payload)) {
-    throw invalidSelectorRequest("Selector request must be a JSON object.");
-  }
-  const messages: unknown = payload.messages ?? [];
-  if (!Array.isArray(messages)) {
-    throw invalidSelectorRequest("Selector request messages must be an array.", "messages");
-  }
-  const userMessage = messages.find(
-    (message): message is Record<string, unknown> => isRecord(message) && message.role === "user",
-  );
-  if (userMessage === undefined) {
-    throw invalidSelectorRequest("Selector request is missing the user message.");
-  }
-  const content = userMessage.content;
-  if (typeof content !== "string") {
-    throw invalidSelectorRequest("Selector request user message content must be a string.");
-  }
-  let parsed: unknown;
-  try {
-    parsed = JSON.parse(content);
-  } catch {
-    throw invalidSelectorRequest("Selector request user message is not valid JSON.");
-  }
-  if (!isRecord(parsed)) {
-    throw invalidSelectorRequest("Selector request user message is not a JSON object.");
-  }
-  if (!Array.isArray(parsed.candidate_pairs)) {
-    throw invalidSelectorRequest("Selector input is missing candidate_pairs.");
-  }
-  // Anonymization would silently drop custom rules rather than apply them.
-  if ("custom_rules" in parsed || "default_target" in parsed) {
-    throw new RouterFrameworkError(
-      "invalid_request",
-      "The Dari Auto Router serves only the default routing policy and cannot apply " +
-        "custom routing rules. Self-host the selector or use the Dari managed platform " +
-        "for custom routing.",
-      "custom_rules_not_supported",
-    );
-  }
-  if (!Array.isArray(parsed.messages)) {
-    throw invalidSelectorRequest("Selector input messages must be an array.", "messages");
-  }
-  for (const [index, message] of (parsed.messages as unknown[]).entries()) {
-    if (!isRecord(message) || typeof message.role !== "string") {
-      throw invalidSelectorRequest(
-        "Selector input messages must be objects with a string role.",
-        `messages[${index}]`,
-      );
-    }
-  }
-  return parsed as unknown as SelectorInput;
-}
-
-function invalidSelectorRequest(message: string, param?: string): RouterFrameworkError {
-  return new RouterFrameworkError("invalid_request", message, "invalid_request_error", param);
-}

package/src/types.ts

@@ -1,231 +0,0 @@
-export function isRecord(value: unknown): value is Record<string, unknown> {
-  return typeof value === "object" && value !== null && !Array.isArray(value);
-}
-
-// Provider ids are the lowercase prefix of a `provider/native-model-id` model
-// string, e.g. "openai" in "openai/gpt-5.6-luna".
-export type Provider = string;
-
-export type ChatMessage = {
-  role: string;
-  content?: string | Array<Record<string, unknown>> | null;
-  tool_call_id?: string;
-  tool_calls?: Array<Record<string, unknown>>;
-  reasoning_content?: string;
-  reasoning?: string;
-  reasoning_text?: string;
-  reasoning_details?: Array<Record<string, unknown>>;
-};
-
-export const REASONING_EFFORTS = ["off", "minimal", "low", "medium", "high", "xhigh", "max"] as const;
-export type ReasoningEffort = (typeof REASONING_EFFORTS)[number];
-
-export function isReasoningEffort(value: unknown): value is ReasoningEffort {
-  return typeof value === "string" && REASONING_EFFORTS.some((effort) => effort === value);
-}
-
-export type ResponseFormat =
-  | { type: "text" }
-  | { type: "json_object" }
-  | {
-      type: "json_schema";
-      json_schema: {
-        name: string;
-        schema: Record<string, unknown>;
-        strict?: boolean;
-        description?: string;
-      };
-    };
-
-export type ChatCompletionRequest = {
-  model?: string;
-  messages?: ChatMessage[];
-  stream?: boolean;
-  source_protocol?: "anthropic_messages" | "openai_chat" | "openai_responses";
-  prompt_cache_key?: string;
-  temperature?: number;
-  top_p?: number;
-  max_tokens?: number;
-  max_completion_tokens?: number;
-  stop?: string | string[];
-  tools?: Array<Record<string, unknown>>;
-  tool_choice?: string | Record<string, unknown>;
-  parallel_tool_calls?: boolean;
-  response_format?: ResponseFormat;
-  reasoning_effort?: ReasoningEffort;
-  metadata?: Record<string, unknown>;
-  user?: string;
-};
-
-export type RoutingCandidate = {
-  model: string;
-  reasoningEffort: ReasoningEffort;
-};
-
-export function routingCandidateKey(candidate: RoutingCandidate): string {
-  return `${candidate.model}\u0000${candidate.reasoningEffort}`;
-}
-
-export function sameRoutingCandidate(
-  left: RoutingCandidate,
-  right: RoutingCandidate,
-): boolean {
-  return left.model === right.model && left.reasoningEffort === right.reasoningEffort;
-}
-
-export type RoutingDecision = {
-  selectedModel: string;
-  reasoningEffort: ReasoningEffort;
-  reason: string;
-  // Turns still committed to this selection after the current one, when a
-  // lease-aware policy made (or is holding) a commitment. The serving layer
-  // persists the countdown on the conversation's prefix chain.
-  leaseTurnsRemaining?: number;
-};
-
-export type RouterEvalScore = {
-  model_id: string;
-  score: number;
-  notes?: string | null;
-  thinking_level?: ReasoningEffort | null;
-};
-
-export type RouterEval = {
-  id: string;
-  name: string;
-  description?: string | null;
-  min_score: number;
-  max_score: number;
-  scores: RouterEvalScore[];
-};
-
-// One stored prefix entry for a conversation chain, as resolved by the
-// host's prefix-hit storage. The core never persists these; it only consumes
-// them as explicit input to cost estimation and conversation-identity
-// recovery. Hosts may resolve entries past the cache-warmth window; `updated_at`
-// is what separates the warm ones from the identity-only ones.
-export type RouterPrefixHit = {
-  hash: string;
-  // Client-durable hash tier (no reasoning payloads, item ids, or
-  // provider-specific canonicalization). Matches the host's loose chain for
-  // conversation identity even when the client mutated its echo. Null on
-  // entries stored before the loose tier existed.
-  loose_hash?: string | null;
-  conversation_id: string;
-  model: string;
-  // Selector recommendation to use for the next turn. The serving model
-  // remains separate because it owns this prefix's provider-cache metadata.
-  next_model?: string | null;
-  next_reasoning_effort?: ReasoningEffort | null;
-  message_depth: number;
-  // Legacy Chat-projection block depth, retained for compatibility with
-  // existing hosts and rolling deployments.
-  block_depth: number;
-  // Candidate-provider-visible block depth. Null/absent on entries written
-  // before provider-aware prompt accounting was introduced.
-  provider_block_depth?: number | null;
-  // Depth of the provider input represented by prompt_tokens. The entry hash
-  // may extend through provider output for conversation identity.
-  prompt_anchor_depth?: number | null;
-  prompt_tokens: number;
-  // Running maximum prompt depth of the hypothetical sticky benchmark.
-  // Hosts persist this with prefix state so future turns can price the
-  // benchmark incrementally without replaying request history.
-  benchmark_replay_depth?: number | null;
-  input_tokens: number;
-  cache_read_tokens: number;
-  cache_write_tokens: number;
-  tool_choice_fp: string;
-  response_format_fp: string;
-  // Effective reasoning bucket the serving request ran with ("off" groups
-  // omitted/off payloads). Null on entries written before buckets existed.
-  reasoning_bucket?: string | null;
-  // Turns still committed to next_model/next_reasoning_effort by a lease-aware
-  // policy. Positive while a lease holds; absent or null when routing decides
-  // per turn.
-  lease_turns_remaining?: number | null;
-  // Next lease decided ahead of expiry (the serving-side analogue of
-  // training's n-lag prefetch). Adopted when the active lease runs out, so the
-  // expiry turn pays no decision latency.
-  pending_lease_model?: string | null;
-  pending_lease_reasoning_effort?: ReasoningEffort | null;
-  pending_lease_turns?: number | null;
-  pending_lease_reason?: string | null;
-  pending_lease_output_text?: string | null;
-  reason: string | null;
-  updated_at: string;
-};
-
-export type RoutingStrategy = "slm" | "custom";
-
-export type CustomRouterRule = {
-  // Natural-language condition describing when this rule applies, e.g.
-  // "planning and architecture" or "implementation".
-  when: string;
-  // The enabled model serving matching requests. Null means the router
-  // chooses among that model's enabled thinking levels automatically.
-  use: string;
-  thinking_level?: ReasoningEffort | null;
-};
-
-export type CustomRouterConfig = {
-  rules: CustomRouterRule[];
-  default?: string | null;
-  default_thinking_level?: ReasoningEffort | null;
-};
-
-export type RouterModelPrice = {
-  input: number;
-  output: number;
-  cached_input?: number | null;
-  cache_write?: number | null;
-};
-
-export type PrefixHit = {
-  entry: RouterPrefixHit;
-  depth: number;
-};
-
-export type ChainHits = {
-  deepest?: PrefixHit;
-  perModel: Map<string, PrefixHit>;
-  // Deepest hit per model per reasoning bucket. Effort-keyed provider caches
-  // keep one warm partition per effective effort (partitions coexist:
-  // A -> B -> A reuses A's still-warm entry), so warm-prefix selection must
-  // not let a deeper cross-bucket entry shadow a valid same-bucket one.
-  // Entries written before buckets existed live under null.
-  perModelBuckets: Map<string, Map<string | null, PrefixHit>>;
-};
-
-export type TurnCostProjection = {
-  projected_turns: number;
-  total_cost_usd: number;
-};
-
-export type FixedTurnCostEstimate = {
-  output_tokens_per_turn: number;
-  assumed_reasoning_effort: ReasoningEffort;
-  // Cumulative loop cost at each horizon in FIXED_TURN_COST_PROJECTED_TURNS,
-  // ascending by turn count.
-  projections: TurnCostProjection[];
-};
-
-export type CandidateCostEstimate = {
-  model: string;
-  reasoning_effort: ReasoningEffort;
-  warm_tokens: number;
-  est_prompt_tokens: number;
-  est_input_cost_usd: number | null;
-  output_cost_per_mtok: number | null;
-  pricing_known: boolean;
-  fixed_turn_cost_estimate: FixedTurnCostEstimate | null;
-};
-
-export type PreviousDecision = {
-  model: string;
-  reasoningEffort: ReasoningEffort;
-  reason: string;
-};
-
-// Request capabilities that hard-constrain candidate eligibility.
-export type HardCapability = "image_input" | "structured_output" | "openai_hosted_web_search";