@outputai/llm 0.7.1-next.2a4105c.0 → 0.7.1-next.5485680.0

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@outputai/llm",
-  "version": "0.7.1-next.2a4105c.0",
+  "version": "0.7.1-next.5485680.0",
   "description": "Framework abstraction to interact with LLM models",
   "type": "module",
   "main": "src/index.js",
@@ -13,7 +13,7 @@
     "gray-matter": "4.0.3",
     "liquidjs": "10.25.7",
     "undici": "8.1.0",
-    "@outputai/core": "0.7.1-next.2a4105c.0"
+    "@outputai/core": "0.7.1-next.5485680.0"
   },
   "devDependencies": {
     "ai": "6.0.168",

package/src/agent.js

@@ -5,7 +5,7 @@ import { loadAiSdkTextOptions } from './ai_sdk_options.js';
 import { prepareTextPrompt } from './prompt/prepare_text.js';
 import { startTrace, endTraceWithError } from './utils/trace.js';
 import { wrapTextResponse, wrapStreamOnFinishResponse } from './utils/response_wrappers.js';
-import { ROLE, isRole, getContent } from './utils/message.js';
+import { ROLE, isRole } from './utils/message.js';
 export { skill } from './prompt/skill.js';
 
 export const createMemoryConversationStore = () => {
@@ -43,16 +43,13 @@ export class Agent extends AIToolLoopAgent {
 
     const { loadedPrompt, tools } = prepareTextPrompt( { prompt, variables, promptDir: resolvedPromptDir, skills, tools: toolsArg } );
 
-    const { messages: allMessages, ...constructorOptions } = loadAiSdkTextOptions( loadedPrompt );
-
-    // Extract system messages as `instructions` for the ToolLoopAgent constructor
-    // and keep user messages for generate() calls — avoids provider errors
-    // with multiple system messages during multi-step tool loops
-    const systemContent = allMessages.filter( isRole( ROLE.SYSTEM ) ).map( getContent ).join( '\n\n' );
+    const { system, messages, ...constructorOptions } = loadAiSdkTextOptions( loadedPrompt );
 
+    // loadAiSdkTextOptions routes system blocks to the `system` slot (preserving
+    // per-message providerOptions); pass them as the agent's `instructions`.
     super( {
       ...constructorOptions,
-      ...( systemContent ? { instructions: systemContent } : {} ),
+      ...( system.length > 0 ? { instructions: system } : {} ),
       ...( tools ? { tools } : {} ),
       stopWhen: stopWhen ?? stepCountIs( maxSteps ),
       ...rest
@@ -60,7 +57,9 @@ export class Agent extends AIToolLoopAgent {
 
     this.#prompt = prompt;
     this.#modelId = loadedPrompt.config.model;
-    this.#initialMessages = allMessages.filter( isRole( ROLE.USER ) );
+    // `messages` is system-free but may still hold authored <assistant>/<tool>
+    // blocks; seed only <user> turns into each generate()/stream() call.
+    this.#initialMessages = messages.filter( isRole( ROLE.USER ) );
     this.#store = conversationStore ?? null;
   }
 
@@ -79,7 +78,7 @@ export class Agent extends AIToolLoopAgent {
     const traceId = startTrace( { name: 'Agent.generate', prompt: this.#prompt } );
     try {
       const messages = await this.#fetchMessages( userMessages );
-      const response = await super.generate( { messages, ...callOptions } );
+      const response = await super.generate( { messages, allowSystemInMessages: true, ...callOptions } );
       const wrapped = await wrapTextResponse( { traceId, response, modelId: this.#modelId } );
       await this.#storeMessages( userMessages, wrapped );
       return wrapped;
@@ -95,6 +94,7 @@ export class Agent extends AIToolLoopAgent {
       const messages = await this.#fetchMessages( userMessages );
       return super.stream( {
         messages,
+        allowSystemInMessages: true,
         ...callOptions,
         ...wrapStreamOnFinishResponse( { traceId, modelId: this.#modelId, onFinish } ),
         onError( event ) {

package/src/agent.spec.js

@@ -86,7 +86,8 @@ vi.mock( './utils/trace.js', () => ( {
 
 vi.mock( './utils/response_wrappers.js', () => ( {
   wrapTextResponse: ( ...args ) => wrapMocks.wrapTextResponse( ...args ),
-  wrapStreamOnFinishResponse: ( ...args ) => wrapMocks.wrapStreamOnFinishResponse( ...args )
+  wrapStreamOnFinishResponse: ( ...args ) =>
+    wrapMocks.wrapStreamOnFinishResponse( ...args )
 } ) );
 
 vi.mock( './prompt/skill.js', () => ( {
@@ -112,7 +113,8 @@ const model = { id: 'MODEL' };
 
 const textOptions = {
   model,
-  messages: loadedPrompt.messages,
+  system: [ { role: 'system', content: 'You are concise.' } ],
+  messages: [ { role: 'user', content: 'Initial user message' } ],
   providerOptions: { test: true },
   temperature: 0.3
 };
@@ -131,7 +133,9 @@ describe( 'Agent', () => {
     aiMocks.superConstructor.mockReset();
     aiMocks.superGenerate.mockReset().mockResolvedValue( aiResponse );
     aiMocks.superStream.mockReset().mockReturnValue( { textStream: 'stream' } );
-    aiMocks.stepCountIs.mockReset().mockImplementation( count => ( { type: 'step-count', count } ) );
+    aiMocks.stepCountIs
+      .mockReset()
+      .mockImplementation( count => ( { type: 'step-count', count } ) );
 
     promptMocks.prepareTextPrompt.mockReset().mockReturnValue( {
       loadedPrompt,
@@ -143,7 +147,9 @@ describe( 'Agent', () => {
     traceMocks.startTrace.mockReset().mockReturnValue( 'trace-id' );
     traceMocks.endTraceWithError.mockReset();
 
-    wrapMocks.wrapTextResponse.mockReset().mockImplementation( async ( { response } ) => response );
+    wrapMocks.wrapTextResponse
+      .mockReset()
+      .mockImplementation( async ( { response } ) => response );
     wrapMocks.wrapStreamOnFinishResponse.mockReset().mockReturnValue( {
       onFinish: vi.fn()
     } );
@@ -175,7 +181,9 @@ describe( 'Agent', () => {
 
   it( 'prepares the prompt using the resolved invocation dir', async () => {
     const { Agent } = await importSut();
-    const skills = [ { name: 'style', description: 'Style', instructions: '# Style' } ];
+    const skills = [
+      { name: 'style', description: 'Style', instructions: '# Style' }
+    ];
     const tools = { search: { description: 'Search' } };
 
     new Agent( {
@@ -199,9 +207,11 @@ describe( 'Agent', () => {
 
     new Agent( { prompt: 'test@v1', promptDir: '/explicit/prompts' } );
 
-    expect( promptMocks.prepareTextPrompt ).toHaveBeenCalledWith( expect.objectContaining( {
-      promptDir: '/explicit/prompts'
-    } ) );
+    expect( promptMocks.prepareTextPrompt ).toHaveBeenCalledWith(
+      expect.objectContaining( {
+        promptDir: '/explicit/prompts'
+      } )
+    );
   } );
 
   it( 'constructs ToolLoopAgent with text options, instructions, tools, and default stopWhen', async () => {
@@ -215,12 +225,34 @@ describe( 'Agent', () => {
       model,
       providerOptions: { test: true },
       temperature: 0.3,
-      instructions: 'You are concise.',
+      instructions: [ { role: 'system', content: 'You are concise.' } ],
       tools: preparedTools,
       stopWhen: { type: 'step-count', count: 10 }
     } );
   } );
 
+  it( 'preserves per-message providerOptions on system messages passed as instructions', async () => {
+    const { Agent } = await importSut();
+    const systemMessage = {
+      role: 'system',
+      content: 'You are concise.',
+      providerOptions: { anthropic: { cacheControl: { type: 'ephemeral' } } }
+    };
+    optionMocks.loadAiSdkTextOptions.mockReturnValueOnce( {
+      model,
+      system: [ systemMessage ],
+      messages: [ { role: 'user', content: 'Hello' } ]
+    } );
+
+    new Agent( { prompt: 'test@v1' } );
+
+    expect( aiMocks.superConstructor ).toHaveBeenCalledWith(
+      expect.objectContaining( {
+        instructions: [ systemMessage ]
+      } )
+    );
+  } );
+
   it( 'omits tools when prompt preparation returns null tools', async () => {
     const { Agent } = await importSut();
     promptMocks.prepareTextPrompt.mockReturnValueOnce( {
@@ -234,7 +266,7 @@ describe( 'Agent', () => {
       model,
       providerOptions: { test: true },
       temperature: 0.3,
-      instructions: 'You are concise.',
+      instructions: [ { role: 'system', content: 'You are concise.' } ],
       stopWhen: { type: 'step-count', count: 10 }
     } );
   } );
@@ -246,9 +278,11 @@ describe( 'Agent', () => {
     new Agent( { prompt: 'test@v1', stopWhen } );
 
     expect( aiMocks.stepCountIs ).not.toHaveBeenCalled();
-    expect( aiMocks.superConstructor ).toHaveBeenCalledWith( expect.objectContaining( {
-      stopWhen
-    } ) );
+    expect( aiMocks.superConstructor ).toHaveBeenCalledWith(
+      expect.objectContaining( {
+        stopWhen
+      } )
+    );
   } );
 
   it( 'passes custom constructor options through', async () => {
@@ -256,10 +290,12 @@ describe( 'Agent', () => {
 
     new Agent( { prompt: 'test@v1', temperature: 0.8, seed: 42 } );
 
-    expect( aiMocks.superConstructor ).toHaveBeenCalledWith( expect.objectContaining( {
-      temperature: 0.8,
-      seed: 42
-    } ) );
+    expect( aiMocks.superConstructor ).toHaveBeenCalledWith(
+      expect.objectContaining( {
+        temperature: 0.8,
+        seed: 42
+      } )
+    );
   } );
 
   it( 'keeps only user prompt messages as initial generate messages', async () => {
@@ -269,13 +305,36 @@ describe( 'Agent', () => {
     await agent.generate();
 
     expect( aiMocks.superGenerate ).toHaveBeenCalledWith( {
-      messages: [ { role: 'user', content: 'Initial user message' } ]
+      messages: [ { role: 'user', content: 'Initial user message' } ],
+      allowSystemInMessages: true
+    } );
+  } );
+
+  it( 'excludes authored assistant/tool blocks from the initial generate messages', async () => {
+    const { Agent } = await importSut();
+    optionMocks.loadAiSdkTextOptions.mockReturnValueOnce( {
+      model,
+      system: [ { role: 'system', content: 'You are concise.' } ],
+      messages: [
+        { role: 'user', content: 'Initial user message' },
+        { role: 'assistant', content: 'Authored assistant block' }
+      ]
+    } );
+    const agent = new Agent( { prompt: 'test@v1' } );
+
+    await agent.generate();
+
+    expect( aiMocks.superGenerate ).toHaveBeenCalledWith( {
+      messages: [ { role: 'user', content: 'Initial user message' } ],
+      allowSystemInMessages: true
     } );
   } );
 
   it( 'combines initial, stored, and caller messages for generate', async () => {
     const store = {
-      getMessages: vi.fn( () => [ { role: 'assistant', content: 'Stored reply' } ] ),
+      getMessages: vi.fn( () => [
+        { role: 'assistant', content: 'Stored reply' }
+      ] ),
       addMessages: vi.fn()
     };
     const callerMessage = { role: 'user', content: 'New question' };
@@ -290,6 +349,7 @@ describe( 'Agent', () => {
         { role: 'assistant', content: 'Stored reply' },
         callerMessage
       ],
+      allowSystemInMessages: true,
       maxRetries: 1
     } );
   } );
@@ -336,7 +396,9 @@ describe( 'Agent', () => {
 
   it( 'streams with initial, stored, and caller messages', async () => {
     const store = {
-      getMessages: vi.fn( () => [ { role: 'assistant', content: 'Stored reply' } ] ),
+      getMessages: vi.fn( () => [
+        { role: 'assistant', content: 'Stored reply' }
+      ] ),
       addMessages: vi.fn()
     };
     const onFinish = vi.fn();
@@ -345,7 +407,12 @@ describe( 'Agent', () => {
     const { Agent } = await importSut();
     const agent = new Agent( { prompt: 'test@v1', conversationStore: store } );
 
-    const result = await agent.stream( { messages: [ callerMessage ], onFinish, onError, maxRetries: 1 } );
+    const result = await agent.stream( {
+      messages: [ callerMessage ],
+      onFinish,
+      onError,
+      maxRetries: 1
+    } );
 
     expect( traceMocks.startTrace ).toHaveBeenCalledWith( {
       name: 'Agent.stream',
@@ -362,6 +429,7 @@ describe( 'Agent', () => {
         { role: 'assistant', content: 'Stored reply' },
         callerMessage
       ],
+      allowSystemInMessages: true,
       maxRetries: 1,
       onFinish: expect.any( Function ),
       onError: expect.any( Function )

package/src/ai_sdk.js

@@ -22,6 +22,7 @@ export async function generateText( { prompt, variables, promptDir, skills = [],
   try {
     const response = await AI.generateText( {
       ...loadAiSdkTextOptions( loadedPrompt ),
+      allowSystemInMessages: true,
       maxRetries: 0,
       ...aiSdkArgs,
       ...( tools && { tools } ),
@@ -50,6 +51,7 @@ export function streamText( { prompt, variables, promptDir, skills = [], maxStep
   try {
     return AI.streamText( {
       ...loadAiSdkTextOptions( loadedPrompt ),
+      allowSystemInMessages: true,
       maxRetries: 0,
       ...aiSdkArgs,
       ...( tools && { tools } ),

package/src/ai_sdk.spec.js

@@ -190,6 +190,7 @@ describe( 'ai_sdk', () => {
       expect( aiFns.stepCountIs ).toHaveBeenCalledWith( 4 );
       expect( aiFns.generateText ).toHaveBeenCalledWith( {
         ...textOptions,
+        allowSystemInMessages: true,
         maxRetries: 0,
         tools,
         temperature: 0.2,
@@ -225,6 +226,7 @@ describe( 'ai_sdk', () => {
       expect( aiFns.stepCountIs ).not.toHaveBeenCalled();
       expect( aiFns.generateText ).toHaveBeenCalledWith( {
         ...textOptions,
+        allowSystemInMessages: true,
         maxRetries: 0
       } );
     } );
@@ -326,6 +328,7 @@ describe( 'ai_sdk', () => {
       expect( aiFns.stepCountIs ).toHaveBeenCalledWith( 4 );
       expect( aiFns.streamText ).toHaveBeenCalledWith( {
         ...textOptions,
+        allowSystemInMessages: true,
         maxRetries: 0,
         tools,
         temperature: 0.2,
@@ -373,6 +376,7 @@ describe( 'ai_sdk', () => {
       expect( aiFns.stepCountIs ).not.toHaveBeenCalled();
       expect( aiFns.streamText ).toHaveBeenCalledWith( {
         ...textOptions,
+        allowSystemInMessages: true,
         maxRetries: 0,
         onFinish: expect.any( Function ),
         onError: expect.any( Function )

package/src/ai_sdk_options.js

@@ -1,10 +1,16 @@
 import { loadImageModel, loadTextModel, loadTools } from './ai_model.js';
 import { resolveMessageProviderOptions } from './prompt/block_options.js';
+import { ROLE, isRole } from './utils/message.js';
 import { FatalError } from '@outputai/core';
 
 /**
  * Convert a loaded prompt into AI SDK text generation options.
  *
+ * System blocks are routed to the `system` option (as `SystemModelMessage[]`, so
+ * per-message providerOptions like `cacheControl` are preserved) rather than left
+ * in `messages` — the AI SDK flags system roles inside `messages` as a prompt
+ * injection risk, and `system` is the provider-recommended slot.
+ *
  * @param {object} prompt - Loaded prompt object
  * @returns {object} Options for AI SDK text calls
  */
@@ -12,9 +18,13 @@ export const loadAiSdkTextOptions = prompt => {
   if ( prompt.messages.length === 0 ) {
     throw new FatalError( `Prompt "${prompt.name}" has no chat-style messages. Add role-tagged blocks like <system> or <user>.` );
   }
+  const isSystem = isRole( ROLE.SYSTEM );
+  const resolvedMessages = resolveMessageProviderOptions( prompt );
+
   const options = {
     model: loadTextModel( prompt ),
-    messages: resolveMessageProviderOptions( prompt ),
+    system: resolvedMessages.filter( isSystem ),
+    messages: resolvedMessages.filter( message => !isSystem( message ) ),
     providerOptions: prompt.config.providerOptions
   };
 

package/src/ai_sdk_options.spec.js

@@ -60,7 +60,8 @@ describe( 'ai_sdk_options', () => {
     expect( loadToolsImpl ).toHaveBeenCalledWith( prompt );
     expect( result ).toEqual( {
       model: 'MODEL',
-      messages: prompt.messages,
+      system: [ { role: 'system', content: 'You are concise.' } ],
+      messages: [ { role: 'user', content: 'Hello' } ],
       providerOptions: prompt.config.providerOptions,
       temperature: 0.3,
       maxOutputTokens: 1000
@@ -180,13 +181,50 @@ describe( 'ai_sdk_options', () => {
     const { loadAiSdkTextOptions } = await importSut();
     const result = loadAiSdkTextOptions( prompt );
 
-    expect( result.messages ).toEqual( [
+    expect( result.system ).toEqual( [
       {
         role: 'system',
         content: 'Static',
         providerOptions: { anthropic: { cacheControl: { type: 'ephemeral', ttl: '1h' } } }
-      },
-      { role: 'user', content: 'Hello' }
+      }
+    ] );
+    expect( result.messages ).toEqual( [ { role: 'user', content: 'Hello' } ] );
+  } );
+
+  it( 'returns an empty system array when the prompt has no system block', async () => {
+    const prompt = {
+      name: 'no-system@v1',
+      config: { provider: 'anthropic', model: 'claude-haiku-4-5' },
+      messages: [ { role: 'user', content: 'Hello' } ],
+      instructions: null
+    };
+
+    const { loadAiSdkTextOptions } = await importSut();
+    const result = loadAiSdkTextOptions( prompt );
+
+    expect( result.system ).toEqual( [] );
+    expect( result.messages ).toEqual( [ { role: 'user', content: 'Hello' } ] );
+  } );
+
+  it( 'groups multiple system blocks into the system option and keeps them out of messages', async () => {
+    const prompt = {
+      name: 'multi-system@v1',
+      config: { provider: 'anthropic', model: 'claude-haiku-4-5' },
+      messages: [
+        { role: 'system', content: 'First' },
+        { role: 'system', content: 'Second' },
+        { role: 'user', content: 'Hello' }
+      ],
+      instructions: null
+    };
+
+    const { loadAiSdkTextOptions } = await importSut();
+    const result = loadAiSdkTextOptions( prompt );
+
+    expect( result.system ).toEqual( [
+      { role: 'system', content: 'First' },
+      { role: 'system', content: 'Second' }
     ] );
+    expect( result.messages ).toEqual( [ { role: 'user', content: 'Hello' } ] );
   } );
 } );

package/src/index.d.ts

@@ -62,7 +62,7 @@ export type PromptMessage = {
    * frontmatter `messageOptions` set names — which is resolved into per-message `providerOptions`
    * at call time and stripped before the request is sent. Authored as `<system options="set_a set_b">`.
    */
-  attributes?: Record<string, string | boolean>;
+  attributes?: Record<string, string | true>;
 };
 
 /**