@opencow-ai/opencow-agent-sdk 0.4.6 → 0.4.8

package/dist/providers/openai/shim.d.ts

@@ -20,7 +20,8 @@
  *   GITHUB_TOKEN or GH_TOKEN         — PAT with models access (mapped to Bearer auth)
  *   OPENAI_MODEL                     — optional; use github:copilot or openai/gpt-4.1 style IDs
  */
-import { type AnthropicStreamEvent, type ShimCreateParams } from '../../providers/codex/shim.js';
+import type { ProviderOverride } from '../shared/routing.js';
+import { type AnthropicStreamEvent, type AnthropicUsage, type ShimCreateParams } from '../../providers/codex/shim.js';
 interface OpenAIMessage {
     role: 'system' | 'user' | 'assistant' | 'tool';
     content?: string | null | Array<{
@@ -67,6 +68,21 @@ export declare function convertTools(tools: Array<{
     input_schema?: Record<string, unknown>;
 }>, model?: string): OpenAITool[];
 export declare function makeMessageId(): string;
+/**
+ * 把 OpenAI(chat_completions)usage 折算成 Anthropic 语义的 usage。
+ *
+ * 关键：OpenAI 的 `prompt_tokens` 已包含缓存命中（`cached_tokens` 是它的子集）；而
+ * Anthropic 语义里 input_tokens / cache_read / cache_creation 互不相交、三者之和才是
+ * 完整输入上下文。所以 input_tokens 必须取「非缓存」部分 = prompt_tokens − cached，否则
+ * 上层（messageMapper 占用环、成本核算）把 input + cache_read 求和会把缓存算两遍。
+ */
+export declare function openaiUsageToAnthropicUsage(usage: {
+    prompt_tokens?: number;
+    completion_tokens?: number;
+    prompt_tokens_details?: {
+        cached_tokens?: number;
+    };
+} | undefined): AnthropicUsage;
 /**
  * Async generator that transforms an OpenAI SSE stream into
  * Anthropic-format BetaRawMessageStreamEvent objects.
@@ -171,12 +187,7 @@ export declare function convertOpenAIResponseToAnthropic(data: {
     model: string;
     stop_reason: string;
     stop_sequence: any;
-    usage: {
-        input_tokens: number;
-        output_tokens: number;
-        cache_creation_input_tokens: number;
-        cache_read_input_tokens: number;
-    };
+    usage: AnthropicUsage;
 };
 /**
  * @internal γ.6c
@@ -199,10 +210,6 @@ export declare function createOpenAIShimClient(options: {
     maxRetries?: number;
     timeout?: number;
     reasoningEffort?: 'low' | 'medium' | 'high' | 'xhigh';
-    providerOverride?: {
-        model: string;
-        baseURL: string;
-        apiKey: string;
-    };
+    providerOverride?: ProviderOverride;
 }): unknown;
 export {};

package/dist/providers/provider.d.ts

@@ -1,4 +1,5 @@
 import type Anthropic from '@anthropic-ai/sdk';
+import type { ProviderOverride } from './shared/routing.js';
 import type { ClientOptions } from '@anthropic-ai/sdk';
 import type { ModelOutputTokenCap, NormalizedRequest, NormalizedResponse, ProviderError, ProviderFeature, ProviderId, ProviderProtocol, StreamEvent } from './types.js';
 /**
@@ -19,11 +20,7 @@ export interface CreateClientOptions {
      * without forcing hosts to rebuild a whole Provider object when
      * the swap is temporary.
      */
-    readonly providerOverride?: {
-        readonly model: string;
-        readonly baseURL: string;
-        readonly apiKey: string;
-    };
+    readonly providerOverride?: ProviderOverride;
 }
 export interface Provider {
     /** Stable id for logs / telemetry / per-session pinning. Not used for dispatch. */

package/dist/providers/shared/clientFactory.d.ts

@@ -1,4 +1,5 @@
 import Anthropic, { type ClientOptions } from '@anthropic-ai/sdk';
+import type { ProviderOverride } from './routing.js';
 /**
  * @internal γ.6c
  *
@@ -22,10 +23,6 @@ export declare function getNormalizedClient({ apiKey, maxRetries, model, fetchOv
     model?: string;
     fetchOverride?: ClientOptions['fetch'];
     source?: string;
-    providerOverride?: {
-        model: string;
-        baseURL: string;
-        apiKey: string;
-    };
+    providerOverride?: ProviderOverride;
 }): Promise<Anthropic>;
 export declare const CLIENT_REQUEST_ID_HEADER = "x-client-request-id";

package/dist/providers/shared/logging.d.ts

@@ -5,9 +5,9 @@ import type { EffortLevel } from 'src/lib/effort.js';
 import type { PermissionMode } from 'src/permissions/PermissionMode.js';
 import { type Span } from 'src/audit/telemetry/sessionTracing.js';
 import type { NonNullableUsage } from '../../entrypoints/sdk/sdkUtilityTypes.js';
-import { EMPTY_USAGE } from './usage.js';
+import { EMPTY_USAGE, contextWindowTokens } from './usage.js';
 export type { NonNullableUsage };
-export { EMPTY_USAGE };
+export { EMPTY_USAGE, contextWindowTokens };
 export type GlobalCacheStrategy = 'tool_based' | 'system_prompt' | 'none';
 export declare function logAPIQuery({ model, messagesLength, temperature, betas, permissionMode, querySource, queryTracking, thinkingType, effortValue, fastMode, previousRequestId, }: {
     model: string;

package/dist/providers/shared/model/providers.d.ts

@@ -19,5 +19,10 @@ export declare function getAPIProviderForStatsig(): AnalyticsMetadata_I_VERIFIED
  * Check if ANTHROPIC_BASE_URL is a first-party Anthropic API URL.
  * Returns true if not set (default API) or points to api.anthropic.com
  * (or api-staging.anthropic.com for ant users).
+ *
+ * Reads the session-aware env (Options.env → process.env fallback): in-process
+ * hosts put the proxy base URL ONLY in createSession({env}), so a raw
+ * process.env read would misclassify their sessions as first-party and leak
+ * beta wire shapes (tool_reference / defer_loading) to the proxy.
  */
 export declare function isFirstPartyAnthropicBaseUrl(): boolean;

package/dist/providers/shared/routing.d.ts

@@ -1,16 +1,47 @@
 import type { SettingsJson } from '../../session/settings/types.js';
+import type { ProviderTransport } from './config.js';
 /**
- * Provider override resolved from agent routing config.
- * When present, the API client should use these instead of global env vars.
+ * Provider override resolved for a specific agent/model.
+ * When present, the API client uses these instead of the session-global
+ * env-derived provider config (baseURL/key/transport).
  */
 export interface ProviderOverride {
     /** Model name to send to the API (e.g. "deepseek-chat", "gpt-4o") */
     model: string;
-    /** OpenAI-compatible base URL */
-    baseURL: string;
-    /** API key for this provider */
-    apiKey: string;
+    /** Base URL. Optional — falls back to the session's OPENAI_BASE_URL env. */
+    baseURL?: string;
+    /** API key. Optional — falls back to the session's env credential chain. */
+    apiKey?: string;
+    /**
+     * Wire transport for this override. Without it the session-level
+     * transport/sniff applies — which is exactly the cross-protocol subagent
+     * bug: a tier model living on a different wire (e.g. openai_responses)
+     * than the main session (chat_completions) hits the wrong endpoint and
+     * 404s. 'anthropic' routes to the native Anthropic client instead of the
+     * OpenAI shim.
+     */
+    transport?: ProviderTransport | 'anthropic';
+    /** Per-wire extras (e.g. openai-responses reasoning summary config). */
+    providerSpecific?: {
+        openaiResponses?: Record<string, unknown>;
+    };
 }
+/**
+ * Host-supplied model → provider-route catalog (`Options.modelProviders`).
+ * Key is the model id as it appears after subagent model resolution (tier
+ * env values / Agent tool `model` param). The host owns the model catalog
+ * (which channel/protocol each model lives on); the SDK consults this AFTER
+ * resolving a subagent's model so any resolution path — tier alias, explicit
+ * model param, inherit — lands on the right wire.
+ */
+export type ModelProviderConfig = Omit<ProviderOverride, 'model'>;
+export type ModelProviders = Record<string, ModelProviderConfig>;
+/**
+ * Resolve a model's provider route from the host catalog. Returns null when
+ * the catalog has no entry — the model is served by the session's default
+ * provider config (the common case: same wire as the main session).
+ */
+export declare function resolveModelProvider(model: string, modelProviders: ModelProviders | undefined): ProviderOverride | null;
 /**
  * Look up agent.routing by name or subagent_type, then resolve via agent.models.
  *

package/dist/providers/shared/usage.d.ts

@@ -5,3 +5,14 @@ import type { NonNullableUsage } from '../../entrypoints/sdk/sdkUtilityTypes.js'
  * api/errors.ts → utils/messages.ts → BashTool.tsx → the world.
  */
 export declare const EMPTY_USAGE: Readonly<NonNullableUsage>;
+/**
+ * Tokens currently occupying the context window for a single API response.
+ *
+ * Context occupancy is a point-in-time snapshot = the input side of the most
+ * recent message: input_tokens + cache_read_input_tokens +
+ * cache_creation_input_tokens. output_tokens are excluded — they only enter
+ * the window on the *next* request. Callers must pass per-message usage, not a
+ * turn-cumulative total (summing every message's input over-counts the window
+ * because each tool round re-sends the growing context).
+ */
+export declare function contextWindowTokens(usage: NonNullableUsage): number;

package/dist/query.d.ts

@@ -24,6 +24,9 @@ export type QueryParams = {
     taskBudget?: {
         total: number;
     };
+    compactRequest?: {
+        instructions?: string;
+    };
     deps?: QueryDeps;
 };
 export declare function query(params: QueryParams): AsyncGenerator<StreamEvent | RequestStartEvent | Message | TombstoneMessage | ToolUseSummaryMessage, Terminal>;