@ixo/oracle-runtime-workers 0.13.0 → 0.16.0

package/src/core/llm.ts

@@ -82,7 +82,7 @@ export const TIER_DISPLAY: Record<
 export const MODEL_CATALOG: readonly ModelCatalogEntry[] = [
   // ── $ Everyday — fast and cheap ─────────────────────────────────────────
   {
-    id: DEFAULT_MODEL_ID,
+    id: 'openai/gpt-5.4-nano',
     label: 'GPT-5.4 Nano',
     family: 'openai',
     tier: 'everyday',
@@ -120,7 +120,7 @@ export const MODEL_CATALOG: readonly ModelCatalogEntry[] = [
 
   // ── $$ Balanced — smarter, still affordable ─────────────────────────────
   {
-    id: 'openai/gpt-5.6-luna',
+    id: DEFAULT_MODEL_ID,
     label: 'GPT-5.6 Luna',
     family: 'openai',
     tier: 'balanced',
@@ -239,7 +239,10 @@ export function getDefaultModelId(
   return override.length > 0 ? override : DEFAULT_MODEL_ID;
 }
 
-const MODEL_INPUT_CAPS: Record<string, ModelInputCapabilities> = {
+/** Native input capabilities per catalog id (every key is a `MODEL_CATALOG` id). */
+export const MODEL_INPUT_CAPS: Readonly<
+  Record<string, ModelInputCapabilities>
+> = {
   'openai/gpt-5.4-nano': {
     image: true,
     file: true,
@@ -396,12 +399,19 @@ export async function listModels(
         OPEN_ROUTER_API_KEY?: unknown;
       }
     | undefined,
+  options: {
+    /** Keeps a background price refresh alive past the response (`ctx.waitUntil`). */
+    waitUntil?: (work: Promise<unknown>) => void;
+  } = {},
 ): Promise<ModelListing> {
   const apiKey =
     typeof env?.OPEN_ROUTER_API_KEY === 'string' && env.OPEN_ROUTER_API_KEY
       ? env.OPEN_ROUTER_API_KEY
       : undefined;
-  const livePrices = await fetchOpenRouterPrices({ apiKey });
+  const livePrices = await fetchOpenRouterPrices({
+    apiKey,
+    ...(options.waitUntil ? { waitUntil: options.waitUntil } : {}),
+  });
   return listModelCatalog(env, livePrices);
 }
 
@@ -446,6 +456,20 @@ export const OPENROUTER_MAIN_FALLBACKS: readonly string[] = [
 
 export const OPENROUTER_BASE_URL = 'https://openrouter.ai/api/v1';
 
+/**
+ * Attribution headers OpenRouter asks callers to send. Shared by the chat
+ * models and the `openrouter-jev` Decision adapter so every request the
+ * oracle makes is credited to the same app.
+ */
+export function openRouterAttributionHeaders(
+  oracleName?: string,
+): Record<string, string> {
+  return {
+    'HTTP-Referer': 'oracle-app.com',
+    'X-Title': oracleName ?? 'Oracle App',
+  };
+}
+
 // ── Nebius (self-hosted deployments) ────────────────────────────────────────
 
 export const NEBIUS_BASE_URL = 'https://api.tokenfactory.nebius.com/v1/';
@@ -561,10 +585,7 @@ export function createLlmAdapter(
   if (env.LLM_PROVIDER === 'nebius') return createNebiusAdapter(env, logger);
   const defaultModelId = getDefaultModelId(env);
   const mainEffort = env.MAIN_REASONING_EFFORT ?? 'medium';
-  const headers: Record<string, string> = {
-    'HTTP-Referer': 'oracle-app.com',
-    'X-Title': env.ORACLE_NAME ?? 'Oracle App',
-  };
+  const headers = openRouterAttributionHeaders(env.ORACLE_NAME);
 
   const modelForRole = (role: ModelRole): string => {
     if (role === 'main') return defaultModelId;
@@ -758,7 +779,7 @@ export interface ResolveLangsmithTracingArgs {
   /** The turn's user DID — from the authenticated payload, never client-set metadata. */
   userDid: string;
   /** Ingress surface, so latency can be sliced per client in LangSmith. */
-  client: 'portal' | 'matrix' | 'slack';
+  client: 'portal' | 'matrix' | 'slack' | 'channel';
   env: LangsmithTracingEnv;
 }
 

package/src/core/main-agent-types.ts

@@ -1,8 +1,10 @@
+import type { TaskExecutionProfile } from '../core/execution-profile';
 import type { ContextBudget } from './context-budget';
 import type { ContextGuardEvent } from './middlewares/context-guard';
 import type { ResultCapConfig } from './middlewares/result-cap';
 import type { ReadResultOutcome } from '../do/result-store';
 import type { AgentMiddleware } from 'langchain';
+import type { TurnBudget } from './turn-budget';
 import type { BaseChatModel } from '@langchain/core/language_models/chat_models';
 import type { BaseCheckpointSaver } from '@langchain/langgraph';
 import type { ReactAgent } from 'langchain';
@@ -12,7 +14,9 @@ import type {
   MergedConfig,
   ModelRole,
   OracleIdentity,
+  PluginTool,
 } from '../plugin-api/types';
+import type { DeliveryProfile } from '../delivery/types';
 import type { Registries } from './registries';
 import type { AmbientServices, RunConfigContext } from './runtime-context';
 import type { TMainAgentGraphState } from './state';
@@ -30,7 +34,7 @@ export const mainAgentRequestContextSchema = z.object({
   }),
   session: z.object({
     id: z.string(),
-    client: z.enum(['portal', 'matrix', 'slack']),
+    client: z.enum(['portal', 'matrix', 'slack', 'channel']),
     requestId: z.string(),
     wsId: z.string().optional(),
     roomId: z.string().optional(),
@@ -85,6 +89,14 @@ export interface MainAgentHooks {
    * write-ahead tool marks of durable runs and the result cap live here.
    */
   toolMiddlewares?: AgentMiddleware[];
+  /**
+   * The tool-execution middleware (budget, scheduling, write claims). It
+   * must see a tool's own thrown error, so on the main agent it is placed
+   * innermost — inside the retry middlewares, which turn a thrown error into
+   * an error ToolMessage — and each retry attempt passes through it. Every
+   * sub-agent gets it after `toolMiddlewares`.
+   */
+  toolExecution?: AgentMiddleware;
   /**
    * The turn's result cap (result-cap.ts), applied inside every wrapped
    * tool — plugin, meta and sub-agent inner tools — so a large result is
@@ -110,9 +122,16 @@ export interface MainAgentHooks {
     offset: number,
     length: number,
   ) => Promise<ReadResultOutcome>;
+  /**
+   * Tools the host binds for this turn only, outside every plugin and the
+   * capability gate (chat delivery's `create_artifact`). `returnDirect` ends
+   * the run once the tool has answered, without another model call.
+   */
+  turnTools?: Array<{ tool: PluginTool; returnDirect?: boolean }>;
 }
 
 export interface MainAgentArgs {
+  executionProfile?: TaskExecutionProfile;
   registries: MainAgentRegistries;
   identity: OracleIdentity;
   config: MergedConfig;
@@ -120,6 +139,12 @@ export interface MainAgentArgs {
   ambient: AmbientServices;
   state: Partial<TMainAgentGraphState>;
   availablePlugins: ReadonlySet<string>;
+  /**
+   * Plugins the capability router chose for this turn. The gate admits their
+   * tools and tool handlers see them as loaded, but they are never written
+   * to the graph's `loadedPlugins` channel: the preload lasts one turn.
+   */
+  preloadedPlugins?: ReadonlySet<string>;
   /**
    * Checkpointer for the user's thread state. On Workers the Durable Object
    * owns it (SQLite over DO storage) and passes it in per turn; omit for a
@@ -141,6 +166,19 @@ export interface MainAgentArgs {
    * size. Omitted → the legacy fixed thresholds (tests, stateless builds).
    */
   contextBudget?: ContextBudget;
+  /**
+   * The turn's resource budget (turn-budget.ts). Model calls are charged by
+   * the metered LLM adapter and tool calls by the host's tool-execution
+   * middleware; the build itself only uses it to skip a summary after the
+   * deadline. Omitted → unbudgeted (tests, stateless builds).
+   */
+  turnBudget?: TurnBudget;
+  /**
+   * Where the reply lands (chat delivery). A `chat` profile adds the surface
+   * section to the prompt; tools see it as `ctx.session.surface`. Omitted →
+   * a Portal turn.
+   */
+  delivery?: DeliveryProfile;
   hooks?: MainAgentHooks;
 }
 
@@ -160,6 +198,8 @@ export interface MainAgentBuildResult {
    * turn, `write` never does). Declared by the plugin, else by name.
    */
   toolEffects: Map<string, 'read' | 'write'>;
+  /** Names of the sub-agent dispatch tools (`call_*_agent`): scheduled in their own lane. */
+  subAgentToolNames: ReadonlySet<string>;
   /**
    * The request's `{ user, session }` channel. Tools already fall back to it,
    * but pass it as `context` in the `invoke` / `stream` config so LangChain