@opencode/ai 0.0.0-dev-20867 → 0.0.0-dev-20869

package/dist/cache-policy.js

@@ -143,17 +143,7 @@ const countHints = (request) => countToolHints(request.tools) +
     request.messages.reduce((count, message) => count +
         message.content.reduce((contentCount, part) => contentCount + ("cache" in part && part.cache !== undefined ? 1 : 0), 0), 0);
 export const applyCachePolicy = (request) => {
-    const route = request.model.route.id;
-    if (!RESPECTS_INLINE_HINTS.has(route))
-        return request;
-    const policy = (route === "openrouter" || route === "vercel-ai-gateway-messages") &&
-        (request.cache === undefined || request.cache === "auto")
-        ? gatewayPolicy(request.model.id)
-        : route === "alibaba-chat" && (request.cache === undefined || request.cache === "auto")
-            ? request.model.id.toLowerCase().startsWith("qwen")
-                ? QWEN
-                : NONE
-            : resolve(request.cache);
+    const policy = selectPolicy(request);
     if (!policy.tools && !policy.system && !policy.messages)
         return request;
     const hint = makeHint(policy.ttlSeconds);
@@ -165,3 +155,22 @@ export const applyCachePolicy = (request) => {
         return request;
     return LLMRequest.update(request, { tools, system, messages });
 };
+// A declared cache format overrides the route allowlist and its model-ID heuristics.
+const selectPolicy = (request) => {
+    const format = request.model.compatibility?.cacheFormat;
+    const route = request.model.route.id;
+    const auto = request.cache === undefined || request.cache === "auto";
+    if (format === "alibaba")
+        return auto ? QWEN : { ...resolve(request.cache), tools: false };
+    if (format === "anthropic")
+        return resolve(request.cache);
+    if (!RESPECTS_INLINE_HINTS.has(route))
+        return NONE;
+    if (!auto)
+        return resolve(request.cache);
+    if (route === "openrouter" || route === "vercel-ai-gateway-messages")
+        return gatewayPolicy(request.model.id);
+    if (route === "alibaba-chat")
+        return request.model.id.toLowerCase().startsWith("qwen") ? QWEN : NONE;
+    return AUTO;
+};

package/dist/protocols/openai-chat.js

@@ -9,6 +9,7 @@ import { Protocol } from "../route/protocol.js";
 import { AIError, AIErrorReason, InvalidProviderOutputError, LLMEvent, ProviderInternalError, UnknownProviderError, Usage, } from "../schema/index.js";
 import { classifyProviderFailure } from "../provider-error.js";
 import { isRecord, JsonObject, optionalArray, optionalNull, ProviderShared } from "./shared.js";
+import { cacheControl } from "./utils/cache.js";
 import { OpenAIOptions } from "./utils/openai-options.js";
 import { Lifecycle } from "./utils/lifecycle.js";
 import { ToolStream } from "./utils/tool-stream.js";
@@ -646,14 +647,15 @@ export const fromRequest = Effect.fn("OpenAIChat.fromRequest")(function* (reques
     const zaiToolStream = request.model.compatibility?.zaiToolStream ?? detectZaiToolStream(provider, baseURL, request.model.id);
     const hasHistory = hasToolHistory(request.messages);
     const hasActiveTools = flattened.tools.length > 0;
+    const lowering = { ...options, cacheControl: options.cacheControl ?? declaredCacheControl(request) };
     return {
         model: request.model.id,
-        messages: yield* lowerMessages(flattened.request, options),
+        messages: yield* lowerMessages(flattened.request, lowering),
         tools: flattened.tools.length === 0
             ? hasHistory
                 ? []
                 : undefined
-            : flattened.tools.map((tool) => lowerTool(tool, options, supportsStrictMode)),
+            : flattened.tools.map((tool) => lowerTool(tool, lowering, supportsStrictMode)),
         tool_choice: hasActiveTools && request.toolChoice ? yield* lowerToolChoice(request.toolChoice) : undefined,
         stream: true,
         ...(supportsUsageInStreaming ? { stream_options: { include_usage: true } } : {}),
@@ -670,6 +672,8 @@ export const fromRequest = Effect.fn("OpenAIChat.fromRequest")(function* (reques
         ...lowerOptions(request, supportsStore),
     };
 });
+// Provider wrappers that pass their own marker writer keep it; otherwise a declared cache format adds one.
+const declaredCacheControl = (request) => request.model.compatibility?.cacheFormat === undefined ? undefined : cacheControl();
 // =============================================================================
 // Stream Parsing
 // =============================================================================

package/dist/schema/options.d.ts

@@ -94,6 +94,8 @@ export declare const LanguageModelSanitizerCompatibility: Schema.Literals<readon
 export type LanguageModelSanitizerCompatibility = Schema.Schema.Type<typeof LanguageModelSanitizerCompatibility>;
 export declare const LanguageModelMaxTokensFieldCompatibility: Schema.Literals<readonly ["max_completion_tokens", "max_tokens"]>;
 export type LanguageModelMaxTokensFieldCompatibility = Schema.Schema.Type<typeof LanguageModelMaxTokensFieldCompatibility>;
+export declare const LanguageModelCacheFormatCompatibility: Schema.Literals<readonly ["anthropic", "alibaba"]>;
+export type LanguageModelCacheFormatCompatibility = Schema.Schema.Type<typeof LanguageModelCacheFormatCompatibility>;
 declare const LanguageModelCompatibility_base: Schema.Class<LanguageModelCompatibility, Schema.Struct<{
     readonly sanitizer: Schema.optional<Schema.Literals<readonly ["gemini", "moonshot", "none"]>>;
     readonly reasoningField: Schema.optional<Schema.String>;
@@ -106,6 +108,8 @@ declare const LanguageModelCompatibility_base: Schema.Class<LanguageModelCompati
     readonly supportsUsageInStreaming: Schema.optional<Schema.Boolean>;
     readonly supportsStrictMode: Schema.optional<Schema.Boolean>;
     readonly supportsPromptCacheKey: Schema.optional<Schema.Boolean>;
+    /** Prompt-cache marker syntax the endpoint accepts. Overrides the route default; both send `cache_control`, and `alibaba` skips tool definitions. */
+    readonly cacheFormat: Schema.optional<Schema.Literals<readonly ["anthropic", "alibaba"]>>;
     readonly zaiToolStream: Schema.optional<Schema.Boolean>;
     readonly requireSignature: Schema.optional<Schema.Boolean>;
     /** Supports Anthropic's thinking-prefix mismatch controls. Overrides model-ID detection. */

package/dist/schema/options.js

@@ -130,6 +130,7 @@ export const ReasoningEffort = Schema.declare((value) => typeof value === "strin
 /** Tool schema sanitizer for a model family. `none` opts out of the protocol and model-name defaults. */
 export const LanguageModelSanitizerCompatibility = Schema.Literals(["gemini", "moonshot", "none"]);
 export const LanguageModelMaxTokensFieldCompatibility = Schema.Literals(["max_completion_tokens", "max_tokens"]);
+export const LanguageModelCacheFormatCompatibility = Schema.Literals(["anthropic", "alibaba"]);
 export class LanguageModelCompatibility extends Schema.Class("LLM.LanguageModelCompatibility")({
     sanitizer: Schema.optional(LanguageModelSanitizerCompatibility),
     reasoningField: Schema.optional(Schema.String),
@@ -145,6 +146,8 @@ export class LanguageModelCompatibility extends Schema.Class("LLM.LanguageModelC
     // this is true; other protocols omit it only when this is false.
     // Session-affinity headers still flow regardless.
     supportsPromptCacheKey: Schema.optional(Schema.Boolean),
+    /** Prompt-cache marker syntax the endpoint accepts. Overrides the route default; both send `cache_control`, and `alibaba` skips tool definitions. */
+    cacheFormat: Schema.optional(LanguageModelCacheFormatCompatibility),
     zaiToolStream: Schema.optional(Schema.Boolean),
     requireSignature: Schema.optional(Schema.Boolean),
     /** Supports Anthropic's thinking-prefix mismatch controls. Overrides model-ID detection. */

package/package.json

@@ -1,6 +1,6 @@
 {
   "$schema": "https://json.schemastore.org/package.json",
-  "version": "0.0.0-dev-20867",
+  "version": "0.0.0-dev-20869",
   "name": "@opencode/ai",
   "type": "module",
   "license": "MIT",
@@ -34,7 +34,7 @@
   "devDependencies": {
     "@clack/prompts": "1.0.0-alpha.1",
     "@effect/platform-node": "4.0.1",
-    "@opencode/http-recorder": "0.0.0-dev-20867",
+    "@opencode/http-recorder": "0.0.0-dev-20869",
     "@tsconfig/bun": "1.0.9",
     "@types/bun": "1.4.0",
     "@typescript/native-preview": "7.0.0-dev.20251207.1",
@@ -44,7 +44,7 @@
     "@aws-sdk/credential-providers": "3.1057.0",
     "@smithy/eventstream-codec": "4.2.14",
     "@smithy/util-utf8": "4.2.2",
-    "@opencode/schema": "0.0.0-dev-20867",
+    "@opencode/schema": "0.0.0-dev-20869",
     "aws4fetch": "1.0.20",
     "effect": "4.0.1",
     "google-auth-library": "10.5.0"