@mindstudio-ai/remy 0.1.177 → 0.1.178

package/dist/headless.d.ts

@@ -131,6 +131,12 @@ declare class HeadlessSession {
      */
     private kickDrain;
     private handleClear;
+    /** Archive the current session and seed a fresh one with the given
+     * per-agent model overrides. Models are immutable for the life of a
+     * session — this is the only way to change them. Omitting `models`
+     * (or sending an empty object) resets to "use server defaults for
+     * every agent". */
+    private handleNewSession;
     /** Cancel the running turn and drain the queue. Returns the drained items. */
     private handleCancel;
     private handleStdinLine;

package/dist/headless.js

@@ -400,16 +400,17 @@ ${loadPlanStatus()}
 // src/api.ts
 var log2 = createLogger("api");
 async function* streamChat(params) {
-  const { baseUrl: baseUrl2, apiKey, signal, requestId, ...body } = params;
+  const { baseUrl: baseUrl2, apiKey, signal, requestId, model, ...rest } = params;
   const url = `${baseUrl2}/_internal/v2/agent/remy/chat`;
   const startTime = Date.now();
-  const subAgentId = body.subAgentId;
+  const subAgentId = rest.subAgentId;
+  const requestBody = { ...rest, ...model && { modelId: model } };
   log2.info("API request", {
     requestId,
     ...subAgentId && { subAgentId },
-    model: body.model,
-    messageCount: body.messages.length,
-    toolCount: body.tools.length
+    model,
+    messageCount: rest.messages.length,
+    toolCount: rest.tools.length
   });
   let res;
   try {
@@ -419,7 +420,7 @@ async function* streamChat(params) {
         "Content-Type": "application/json",
         Authorization: `Bearer ${apiKey}`
       },
-      body: JSON.stringify(body),
+      body: JSON.stringify(requestBody),
       signal
     });
   } catch (err) {
@@ -447,13 +448,21 @@ async function* streamChat(params) {
   });
   if (!res.ok) {
     let errorMessage = `HTTP ${res.status}`;
+    let errorCode;
+    let badModelId;
     try {
-      const body2 = await res.json();
-      if (body2.error) {
-        errorMessage = body2.error;
+      const body = await res.json();
+      if (body.error) {
+        errorMessage = body.error;
+      }
+      if (body.errorMessage) {
+        errorMessage = body.errorMessage;
+      }
+      if (typeof body.code === "string") {
+        errorCode = body.code;
       }
-      if (body2.errorMessage) {
-        errorMessage = body2.errorMessage;
+      if (typeof body.modelId === "string") {
+        badModelId = body.modelId;
       }
     } catch {
     }
@@ -461,9 +470,16 @@ async function* streamChat(params) {
       requestId,
       ...subAgentId && { subAgentId },
       status: res.status,
-      error: errorMessage
+      error: errorMessage,
+      ...errorCode && { code: errorCode },
+      ...badModelId && { badModelId }
     });
-    yield { type: "error", error: errorMessage };
+    yield {
+      type: "error",
+      error: errorMessage,
+      ...errorCode && { code: errorCode },
+      ...badModelId && { badModelId }
+    };
     return;
   }
   const STALL_TIMEOUT_MS = 3e5;
@@ -654,7 +670,7 @@ var log3 = createLogger("compaction");
 var CONVERSATION_SUMMARY_PROMPT = readAsset("compaction", "conversation.md");
 var SUBAGENT_SUMMARY_PROMPT = readAsset("compaction", "subagent.md");
 var SUMMARIZABLE_SUBAGENTS = ["visualDesignExpert", "productVision"];
-async function compactConversation(messages, apiConfig, system, tools2) {
+async function compactConversation(messages, apiConfig, system, tools2, model) {
   const endIndex = findSafeInsertionPoint(messages);
   const summaries = [];
   const tasks = [];
@@ -670,7 +686,8 @@ async function compactConversation(messages, apiConfig, system, tools2) {
         CONVERSATION_SUMMARY_PROMPT,
         conversationMessages,
         system,
-        tools2
+        tools2,
+        model
       ).then((text) => {
         if (text) {
           summaries.push({ name: "conversation", text });
@@ -692,7 +709,8 @@ async function compactConversation(messages, apiConfig, system, tools2) {
           SUBAGENT_SUMMARY_PROMPT,
           subagentMessages,
           system,
-          tools2
+          tools2,
+          model
         ).then((text) => {
           if (text) {
             summaries.push({ name, text });
@@ -814,7 +832,7 @@ function serializeForSummary(messages) {
   }).join("\n\n");
 }
 var CHUNK_CHAR_LIMIT = 24e5;
-async function generateSummary(apiConfig, name, compactionPrompt, messagesToSummarize, mainSystem, mainTools) {
+async function generateSummary(apiConfig, name, compactionPrompt, messagesToSummarize, mainSystem, mainTools, model) {
   const serialized = serializeForSummary(messagesToSummarize);
   if (!serialized.trim()) {
     return null;
@@ -833,7 +851,8 @@ async function generateSummary(apiConfig, name, compactionPrompt, messagesToSumm
         compactionPrompt,
         messagesToSummarize.slice(0, mid),
         mainSystem,
-        mainTools
+        mainTools,
+        model
       ),
       generateSummary(
         apiConfig,
@@ -841,7 +860,8 @@ async function generateSummary(apiConfig, name, compactionPrompt, messagesToSumm
         compactionPrompt,
         messagesToSummarize.slice(mid),
         mainSystem,
-        mainTools
+        mainTools,
+        model
       )
     ]);
     const parts = [first, second].filter((p) => !!p);
@@ -866,6 +886,7 @@ ${serialized}` : serialized;
   const iterStart = Date.now();
   for await (const event of streamChat({
     ...apiConfig,
+    model,
     subAgentId: "conversationSummarizer",
     system,
     messages: [{ role: "user", content: userContent }],
@@ -3110,6 +3131,9 @@ ${content}` : attachmentHeader;
     if (thinking.length > 0) {
       cleaned2.thinking = thinking;
     }
+    if (msg.providerMetadata) {
+      cleaned2.providerMetadata = msg.providerMetadata;
+    }
     if (msg.hidden) {
       cleaned2.hidden = true;
     }
@@ -3196,6 +3220,7 @@ ${partial}` : "[INTERRUPTED] Agent was interrupted before producing output.",
       let stopReason = "end_turn";
       let currentToolNames = "";
       let lastUsage;
+      let lastProviderMetadata;
       const statusWatcher = startStatusWatcher({
         apiConfig,
         getContext: () => {
@@ -3308,6 +3333,7 @@ ${partial}` : "[INTERRUPTED] Agent was interrupted before producing output.",
                 cacheReadTokens: event.usage.cacheReadTokens,
                 llmCalls: 1
               };
+              lastProviderMetadata = event.providerMetadata;
               recordUsage({
                 ts: Date.now(),
                 requestId,
@@ -3345,7 +3371,8 @@ ${partial}` : "[INTERRUPTED] Agent was interrupted before producing output.",
       messages.push({
         role: "assistant",
         content: contentBlocks,
-        ...lastUsage ? { usage: lastUsage } : {}
+        ...lastUsage ? { usage: lastUsage } : {},
+        ...lastProviderMetadata ? { providerMetadata: lastProviderMetadata } : {}
       });
       const toolCalls = contentBlocks.filter(
         (b) => b.type === "tool"
@@ -3746,7 +3773,7 @@ var browserAutomationTool = {
           return `Error: unknown local tool "${name}"`;
         },
         apiConfig: context.apiConfig,
-        model: context.model,
+        model: context.models?.browserAutomation ?? context.model,
         subAgentId: "browserAutomation",
         signal: context.signal,
         parentToolId: context.toolCallId,
@@ -4962,7 +4989,7 @@ var designExpertTool = {
         );
       },
       apiConfig: context.apiConfig,
-      model: context.model,
+      model: context.models?.visualDesignExpert ?? context.model,
       subAgentId: "visualDesignExpert",
       signal: context.signal,
       parentToolId: context.toolCallId,
@@ -5180,7 +5207,7 @@ var productVisionTool = {
         return executeVisionTool(name, input2, childCtx);
       },
       apiConfig: context.apiConfig,
-      model: context.model,
+      model: context.models?.productVision ?? context.model,
       subAgentId: "productVision",
       signal: context.signal,
       parentToolId: context.toolCallId,
@@ -5288,7 +5315,7 @@ var codeSanityCheckTool = {
       externalTools: /* @__PURE__ */ new Set(),
       executeTool: (name, toolInput) => executeTool(name, toolInput, context),
       apiConfig: context.apiConfig,
-      model: context.model,
+      model: context.models?.codeSanityCheck ?? context.model,
       subAgentId: "codeSanityCheck",
       signal: context.signal,
       parentToolId: context.toolCallId,
@@ -5419,7 +5446,7 @@ function triggerCompaction(state, apiConfig, opts = {}) {
   if (inflightCompaction) {
     return inflightCompaction;
   }
-  const { blocking = false, requestId } = opts;
+  const { blocking = false, requestId, model } = opts;
   listener?.({ type: "started", blocking, requestId });
   const system = buildSystemPrompt("onboardingFinished");
   const tools2 = getToolDefinitions("onboardingFinished");
@@ -5427,7 +5454,8 @@ function triggerCompaction(state, apiConfig, opts = {}) {
     state.messages,
     apiConfig,
     system,
-    tools2
+    tools2,
+    state.models?.conversationSummarizer ?? model
   ).then((summaries) => {
     pendingSummaries.push(...summaries);
     listener?.({ type: "complete", requestId });
@@ -5451,7 +5479,7 @@ var log8 = createLogger("brandExtraction");
 var EXTRACT_PROMPT = readAsset("brandExtraction", "extract.md");
 var BRAND_FILE = ".remy-brand.json";
 var CACHE_FILE = ".remy-brand.cache.json";
-async function runExtraction(apiConfig) {
+async function runExtraction(apiConfig, model) {
   const inputHash = computeInputHash();
   const cached2 = readCache();
   if (cached2 && cached2.inputHash === inputHash) {
@@ -5459,7 +5487,7 @@ async function runExtraction(apiConfig) {
     return null;
   }
   log8.info("Extracting brand", { inputHash });
-  const brand = await extractBrand(apiConfig);
+  const brand = await extractBrand(apiConfig, model);
   if (!brand) {
     log8.warn("Brand extraction failed \u2014 leaving cache untouched");
     return null;
@@ -5528,7 +5556,7 @@ function parseFrontmatter3(filePath) {
     return { type: "" };
   }
 }
-async function extractBrand(apiConfig) {
+async function extractBrand(apiConfig, model) {
   const corpus = buildCorpus();
   if (!corpus.trim()) {
     log8.debug("No spec corpus \u2014 emitting empty brand");
@@ -5539,6 +5567,7 @@ async function extractBrand(apiConfig) {
   try {
     for await (const event of streamChat({
       ...apiConfig,
+      model,
       subAgentId: "brandExtractor",
       system: EXTRACT_PROMPT,
       messages: [{ role: "user", content: corpus }],
@@ -5716,19 +5745,19 @@ function readCache() {
 var log9 = createLogger("brandExtraction:trigger");
 var inflight = false;
 var dirty = false;
-function triggerBrandExtraction(apiConfig) {
+function triggerBrandExtraction(apiConfig, model) {
   if (inflight) {
     dirty = true;
     return;
   }
   inflight = true;
-  void runExtraction(apiConfig).catch((err) => {
+  void runExtraction(apiConfig, model).catch((err) => {
     log9.error("Brand extraction failed", { error: err?.message });
   }).finally(() => {
     inflight = false;
     if (dirty) {
       dirty = false;
-      triggerBrandExtraction(apiConfig);
+      triggerBrandExtraction(apiConfig, model);
     }
   });
 }
@@ -5743,9 +5772,15 @@ function loadSession(state) {
   try {
     const raw = fs21.readFileSync(SESSION_FILE, "utf-8");
     const data = JSON.parse(raw);
+    if (data.models && typeof data.models === "object") {
+      state.models = data.models;
+    }
     if (Array.isArray(data.messages) && data.messages.length > 0) {
       state.messages = sanitizeMessages(data.messages);
-      log10.info("Session loaded", { messageCount: state.messages.length });
+      log10.info("Session loaded", {
+        messageCount: state.messages.length,
+        ...state.models && { models: state.models }
+      });
       return true;
     }
   } catch {
@@ -5790,11 +5825,11 @@ function sanitizeMessages(messages) {
 }
 function saveSession(state) {
   try {
-    fs21.writeFileSync(
-      SESSION_FILE,
-      JSON.stringify({ messages: state.messages }, null, 2),
-      "utf-8"
-    );
+    const payload = { messages: state.messages };
+    if (state.models && Object.keys(state.models).length > 0) {
+      payload.models = state.models;
+    }
+    fs21.writeFileSync(SESSION_FILE, JSON.stringify(payload, null, 2), "utf-8");
     log10.info("Session saved", { messageCount: state.messages.length });
   } catch (err) {
     log10.warn("Session save failed", { error: err.message });
@@ -6135,6 +6170,7 @@ async function runTurn(params) {
     let textBlockOpen = false;
     const toolInputAccumulators = /* @__PURE__ */ new Map();
     let stopReason = "end_turn";
+    let turnProviderMetadata;
     let subAgentText = "";
     let currentToolNames = "";
     const statusWatcher = isFirstMessage ? { stop() {
@@ -6216,11 +6252,12 @@ async function runTurn(params) {
         onEvent({ type: "tool_input_delta", id, name, result: content });
       }
     }
+    const parentModel = state.models?.parent ?? model;
     try {
       for await (const event of streamChatWithRetry(
         {
           ...apiConfig,
-          model,
+          model: parentModel,
           requestId,
           system,
           messages: cleanMessagesForApi(state.messages),
@@ -6332,6 +6369,7 @@ async function runTurn(params) {
           }
           case "done":
             stopReason = event.stopReason;
+            turnProviderMetadata = event.providerMetadata;
             turnLlmCalls++;
             lastCallInputTokens = event.usage.inputTokens;
             lastCallCacheCreation = event.usage.cacheCreationTokens ?? 0;
@@ -6385,6 +6423,9 @@ async function runTurn(params) {
             cacheCreationTokens: turnCacheCreation || void 0,
             cacheReadTokens: turnCacheRead || void 0,
             llmCalls: turnLlmCalls
+          },
+          ...turnProviderMetadata && {
+            providerMetadata: turnProviderMetadata
           }
         });
       }
@@ -6402,7 +6443,8 @@ async function runTurn(params) {
           cacheCreationTokens: turnCacheCreation || void 0,
           cacheReadTokens: turnCacheRead || void 0,
           llmCalls: turnLlmCalls
-        }
+        },
+        ...turnProviderMetadata && { providerMetadata: turnProviderMetadata }
       });
     }
     const toolCalls = getToolCalls(contentBlocks);
@@ -6491,6 +6533,7 @@ async function runTurn(params) {
               result = await executeTool(tc.name, input, {
                 apiConfig,
                 model,
+                models: state.models,
                 signal: toolAbort.signal,
                 onEvent: wrappedOnEvent,
                 resolveExternalTool,
@@ -6995,10 +7038,14 @@ var HeadlessSession = class {
     if (resumed) {
       this.emit("session_restored", {
         messageCount: this.state.messages.length,
+        ...this.state.models && { models: this.state.models },
         ...this.queueFields()
       });
     }
-    triggerBrandExtraction(this.config);
+    triggerBrandExtraction(
+      this.config,
+      this.state.models?.brandExtractor ?? this.opts.model
+    );
     this.toolRegistry.onEvent = this.onEvent;
     setCompactionListener((event) => {
       if (event.type === "started") {
@@ -7151,7 +7198,8 @@ var HeadlessSession = class {
     try {
       await triggerCompaction(this.state, this.config, {
         blocking: true,
-        requestId
+        requestId,
+        model: this.opts.model
       });
       this.applyPendingSummaries();
     } catch {
@@ -7589,6 +7637,17 @@ var HeadlessSession = class {
     clearSession(this.state);
     return {};
   }
+  /** Archive the current session and seed a fresh one with the given
+   * per-agent model overrides. Models are immutable for the life of a
+   * session — this is the only way to change them. Omitting `models`
+   * (or sending an empty object) resets to "use server defaults for
+   * every agent". */
+  handleNewSession(models) {
+    clearSession(this.state);
+    this.state.models = models && Object.keys(models).length > 0 ? models : void 0;
+    saveSession(this.state);
+    return {};
+  }
   /** Cancel the running turn and drain the queue. Returns the drained items. */
   handleCancel() {
     if (this.currentAbort) {
@@ -7662,6 +7721,7 @@ var HeadlessSession = class {
         totalMessageCount: total,
         running: this.running,
         ...this.running && this.currentRequestId ? { currentRequestId: this.currentRequestId } : {},
+        ...this.state.models && { models: this.state.models },
         ...this.queueFields()
       }));
       return;
@@ -7674,6 +7734,15 @@ var HeadlessSession = class {
       );
       return;
     }
+    if (action === "newSession") {
+      const models = parsed.models;
+      this.dispatchSimple(
+        requestId,
+        "session_cleared",
+        () => this.handleNewSession(models)
+      );
+      return;
+    }
     if (action === "cancel") {
       const cancelled = this.handleCancel();
       this.emit(
@@ -7720,7 +7789,8 @@ var HeadlessSession = class {
       try {
         await triggerCompaction(this.state, this.config, {
           blocking: false,
-          requestId
+          requestId,
+          model: this.opts.model
         });
         if (!this.running) {
           this.applyPendingSummaries();

package/dist/index.js

@@ -86,16 +86,17 @@ var init_logger = __esm({
 
 // src/api.ts
 async function* streamChat(params) {
-  const { baseUrl: baseUrl2, apiKey, signal, requestId, ...body } = params;
+  const { baseUrl: baseUrl2, apiKey, signal, requestId, model, ...rest } = params;
   const url = `${baseUrl2}/_internal/v2/agent/remy/chat`;
   const startTime = Date.now();
-  const subAgentId = body.subAgentId;
+  const subAgentId = rest.subAgentId;
+  const requestBody = { ...rest, ...model && { modelId: model } };
   log.info("API request", {
     requestId,
     ...subAgentId && { subAgentId },
-    model: body.model,
-    messageCount: body.messages.length,
-    toolCount: body.tools.length
+    model,
+    messageCount: rest.messages.length,
+    toolCount: rest.tools.length
   });
   let res;
   try {
@@ -105,7 +106,7 @@ async function* streamChat(params) {
         "Content-Type": "application/json",
         Authorization: `Bearer ${apiKey}`
       },
-      body: JSON.stringify(body),
+      body: JSON.stringify(requestBody),
       signal
     });
   } catch (err) {
@@ -133,13 +134,21 @@ async function* streamChat(params) {
   });
   if (!res.ok) {
     let errorMessage = `HTTP ${res.status}`;
+    let errorCode;
+    let badModelId;
     try {
-      const body2 = await res.json();
-      if (body2.error) {
-        errorMessage = body2.error;
+      const body = await res.json();
+      if (body.error) {
+        errorMessage = body.error;
+      }
+      if (body.errorMessage) {
+        errorMessage = body.errorMessage;
       }
-      if (body2.errorMessage) {
-        errorMessage = body2.errorMessage;
+      if (typeof body.code === "string") {
+        errorCode = body.code;
+      }
+      if (typeof body.modelId === "string") {
+        badModelId = body.modelId;
       }
     } catch {
     }
@@ -147,9 +156,16 @@ async function* streamChat(params) {
       requestId,
       ...subAgentId && { subAgentId },
       status: res.status,
-      error: errorMessage
+      error: errorMessage,
+      ...errorCode && { code: errorCode },
+      ...badModelId && { badModelId }
     });
-    yield { type: "error", error: errorMessage };
+    yield {
+      type: "error",
+      error: errorMessage,
+      ...errorCode && { code: errorCode },
+      ...badModelId && { badModelId }
+    };
     return;
   }
   const STALL_TIMEOUT_MS = 3e5;
@@ -1471,7 +1487,7 @@ var init_assets = __esm({
 });
 
 // src/compaction/index.ts
-async function compactConversation(messages, apiConfig, system, tools2) {
+async function compactConversation(messages, apiConfig, system, tools2, model) {
   const endIndex = findSafeInsertionPoint(messages);
   const summaries = [];
   const tasks = [];
@@ -1487,7 +1503,8 @@ async function compactConversation(messages, apiConfig, system, tools2) {
         CONVERSATION_SUMMARY_PROMPT,
         conversationMessages,
         system,
-        tools2
+        tools2,
+        model
       ).then((text) => {
         if (text) {
           summaries.push({ name: "conversation", text });
@@ -1509,7 +1526,8 @@ async function compactConversation(messages, apiConfig, system, tools2) {
           SUBAGENT_SUMMARY_PROMPT,
           subagentMessages,
           system,
-          tools2
+          tools2,
+          model
         ).then((text) => {
           if (text) {
             summaries.push({ name, text });
@@ -1630,7 +1648,7 @@ function serializeForSummary(messages) {
     return `[${msg.role}]: ${parts.join("\n")}`;
   }).join("\n\n");
 }
-async function generateSummary(apiConfig, name, compactionPrompt, messagesToSummarize, mainSystem, mainTools) {
+async function generateSummary(apiConfig, name, compactionPrompt, messagesToSummarize, mainSystem, mainTools, model) {
   const serialized = serializeForSummary(messagesToSummarize);
   if (!serialized.trim()) {
     return null;
@@ -1649,7 +1667,8 @@ async function generateSummary(apiConfig, name, compactionPrompt, messagesToSumm
         compactionPrompt,
         messagesToSummarize.slice(0, mid),
         mainSystem,
-        mainTools
+        mainTools,
+        model
       ),
       generateSummary(
         apiConfig,
@@ -1657,7 +1676,8 @@ async function generateSummary(apiConfig, name, compactionPrompt, messagesToSumm
         compactionPrompt,
         messagesToSummarize.slice(mid),
         mainSystem,
-        mainTools
+        mainTools,
+        model
       )
     ]);
     const parts = [first, second].filter((p) => !!p);
@@ -1682,6 +1702,7 @@ ${serialized}` : serialized;
   const iterStart = Date.now();
   for await (const event of streamChat({
     ...apiConfig,
+    model,
     subAgentId: "conversationSummarizer",
     system,
     messages: [{ role: "user", content: userContent }],
@@ -2014,7 +2035,7 @@ function triggerCompaction(state, apiConfig, opts = {}) {
   if (inflightCompaction) {
     return inflightCompaction;
   }
-  const { blocking = false, requestId } = opts;
+  const { blocking = false, requestId, model } = opts;
   listener?.({ type: "started", blocking, requestId });
   const system = buildSystemPrompt("onboardingFinished");
   const tools2 = getToolDefinitions("onboardingFinished");
@@ -2022,7 +2043,8 @@ function triggerCompaction(state, apiConfig, opts = {}) {
     state.messages,
     apiConfig,
     system,
-    tools2
+    tools2,
+    state.models?.conversationSummarizer ?? model
   ).then((summaries) => {
     pendingSummaries.push(...summaries);
     listener?.({ type: "complete", requestId });
@@ -3437,6 +3459,9 @@ ${content}` : attachmentHeader;
     if (thinking.length > 0) {
       cleaned2.thinking = thinking;
     }
+    if (msg.providerMetadata) {
+      cleaned2.providerMetadata = msg.providerMetadata;
+    }
     if (msg.hidden) {
       cleaned2.hidden = true;
     }
@@ -3528,6 +3553,7 @@ ${partial}` : "[INTERRUPTED] Agent was interrupted before producing output.",
       let stopReason = "end_turn";
       let currentToolNames = "";
       let lastUsage;
+      let lastProviderMetadata;
       const statusWatcher = startStatusWatcher({
         apiConfig,
         getContext: () => {
@@ -3640,6 +3666,7 @@ ${partial}` : "[INTERRUPTED] Agent was interrupted before producing output.",
                 cacheReadTokens: event.usage.cacheReadTokens,
                 llmCalls: 1
               };
+              lastProviderMetadata = event.providerMetadata;
               recordUsage({
                 ts: Date.now(),
                 requestId,
@@ -3677,7 +3704,8 @@ ${partial}` : "[INTERRUPTED] Agent was interrupted before producing output.",
       messages.push({
         role: "assistant",
         content: contentBlocks,
-        ...lastUsage ? { usage: lastUsage } : {}
+        ...lastUsage ? { usage: lastUsage } : {},
+        ...lastProviderMetadata ? { providerMetadata: lastProviderMetadata } : {}
       });
       const toolCalls = contentBlocks.filter(
         (b) => b.type === "tool"
@@ -4116,7 +4144,7 @@ var init_browserAutomation = __esm({
               return `Error: unknown local tool "${name}"`;
             },
             apiConfig: context.apiConfig,
-            model: context.model,
+            model: context.models?.browserAutomation ?? context.model,
             subAgentId: "browserAutomation",
             signal: context.signal,
             parentToolId: context.toolCallId,
@@ -5494,7 +5522,7 @@ Visual design expert. Describe the situation and what you need \u2014 the agent
             );
           },
           apiConfig: context.apiConfig,
-          model: context.model,
+          model: context.models?.visualDesignExpert ?? context.model,
           subAgentId: "visualDesignExpert",
           signal: context.signal,
           parentToolId: context.toolCallId,
@@ -5751,7 +5779,7 @@ var init_productVision = __esm({
             return executeVisionTool(name, input2, childCtx);
           },
           apiConfig: context.apiConfig,
-          model: context.model,
+          model: context.models?.productVision ?? context.model,
           subAgentId: "productVision",
           signal: context.signal,
           parentToolId: context.toolCallId,
@@ -5877,7 +5905,7 @@ var init_codeSanityCheck = __esm({
           externalTools: /* @__PURE__ */ new Set(),
           executeTool: (name, toolInput) => executeTool(name, toolInput, context),
           apiConfig: context.apiConfig,
-          model: context.model,
+          model: context.models?.codeSanityCheck ?? context.model,
           subAgentId: "codeSanityCheck",
           signal: context.signal,
           parentToolId: context.toolCallId,
@@ -6048,9 +6076,15 @@ function loadSession(state) {
   try {
     const raw = fs19.readFileSync(SESSION_FILE, "utf-8");
     const data = JSON.parse(raw);
+    if (data.models && typeof data.models === "object") {
+      state.models = data.models;
+    }
     if (Array.isArray(data.messages) && data.messages.length > 0) {
       state.messages = sanitizeMessages(data.messages);
-      log7.info("Session loaded", { messageCount: state.messages.length });
+      log7.info("Session loaded", {
+        messageCount: state.messages.length,
+        ...state.models && { models: state.models }
+      });
       return true;
     }
   } catch {
@@ -6095,11 +6129,11 @@ function sanitizeMessages(messages) {
 }
 function saveSession(state) {
   try {
-    fs19.writeFileSync(
-      SESSION_FILE,
-      JSON.stringify({ messages: state.messages }, null, 2),
-      "utf-8"
-    );
+    const payload = { messages: state.messages };
+    if (state.models && Object.keys(state.models).length > 0) {
+      payload.models = state.models;
+    }
+    fs19.writeFileSync(SESSION_FILE, JSON.stringify(payload, null, 2), "utf-8");
     log7.info("Session saved", { messageCount: state.messages.length });
   } catch (err) {
     log7.warn("Session save failed", { error: err.message });
@@ -6345,7 +6379,7 @@ var init_errors = __esm({
 import fs20 from "fs";
 import path10 from "path";
 import { createHash } from "crypto";
-async function runExtraction(apiConfig) {
+async function runExtraction(apiConfig, model) {
   const inputHash = computeInputHash();
   const cached2 = readCache();
   if (cached2 && cached2.inputHash === inputHash) {
@@ -6353,7 +6387,7 @@ async function runExtraction(apiConfig) {
     return null;
   }
   log8.info("Extracting brand", { inputHash });
-  const brand = await extractBrand(apiConfig);
+  const brand = await extractBrand(apiConfig, model);
   if (!brand) {
     log8.warn("Brand extraction failed \u2014 leaving cache untouched");
     return null;
@@ -6422,7 +6456,7 @@ function parseFrontmatter3(filePath) {
     return { type: "" };
   }
 }
-async function extractBrand(apiConfig) {
+async function extractBrand(apiConfig, model) {
   const corpus = buildCorpus();
   if (!corpus.trim()) {
     log8.debug("No spec corpus \u2014 emitting empty brand");
@@ -6433,6 +6467,7 @@ async function extractBrand(apiConfig) {
   try {
     for await (const event of streamChat({
       ...apiConfig,
+      model,
       subAgentId: "brandExtractor",
       system: EXTRACT_PROMPT,
       messages: [{ role: "user", content: corpus }],
@@ -6621,19 +6656,19 @@ var init_brandExtraction = __esm({
 });
 
 // src/brandExtraction/trigger.ts
-function triggerBrandExtraction(apiConfig) {
+function triggerBrandExtraction(apiConfig, model) {
   if (inflight) {
     dirty = true;
     return;
   }
   inflight = true;
-  void runExtraction(apiConfig).catch((err) => {
+  void runExtraction(apiConfig, model).catch((err) => {
     log9.error("Brand extraction failed", { error: err?.message });
   }).finally(() => {
     inflight = false;
     if (dirty) {
       dirty = false;
-      triggerBrandExtraction(apiConfig);
+      triggerBrandExtraction(apiConfig, model);
     }
   });
 }
@@ -6752,6 +6787,7 @@ async function runTurn(params) {
     let textBlockOpen = false;
     const toolInputAccumulators = /* @__PURE__ */ new Map();
     let stopReason = "end_turn";
+    let turnProviderMetadata;
     let subAgentText = "";
     let currentToolNames = "";
     const statusWatcher = isFirstMessage ? { stop() {
@@ -6833,11 +6869,12 @@ async function runTurn(params) {
         onEvent({ type: "tool_input_delta", id, name, result: content });
       }
     }
+    const parentModel = state.models?.parent ?? model;
     try {
       for await (const event of streamChatWithRetry(
         {
           ...apiConfig,
-          model,
+          model: parentModel,
           requestId,
           system,
           messages: cleanMessagesForApi(state.messages),
@@ -6949,6 +6986,7 @@ async function runTurn(params) {
           }
           case "done":
             stopReason = event.stopReason;
+            turnProviderMetadata = event.providerMetadata;
             turnLlmCalls++;
             lastCallInputTokens = event.usage.inputTokens;
             lastCallCacheCreation = event.usage.cacheCreationTokens ?? 0;
@@ -7002,6 +7040,9 @@ async function runTurn(params) {
             cacheCreationTokens: turnCacheCreation || void 0,
             cacheReadTokens: turnCacheRead || void 0,
             llmCalls: turnLlmCalls
+          },
+          ...turnProviderMetadata && {
+            providerMetadata: turnProviderMetadata
           }
         });
       }
@@ -7019,7 +7060,8 @@ async function runTurn(params) {
           cacheCreationTokens: turnCacheCreation || void 0,
           cacheReadTokens: turnCacheRead || void 0,
           llmCalls: turnLlmCalls
-        }
+        },
+        ...turnProviderMetadata && { providerMetadata: turnProviderMetadata }
       });
     }
     const toolCalls = getToolCalls(contentBlocks);
@@ -7108,6 +7150,7 @@ async function runTurn(params) {
               result = await executeTool(tc.name, input, {
                 apiConfig,
                 model,
+                models: state.models,
                 signal: toolAbort.signal,
                 onEvent: wrappedOnEvent,
                 resolveExternalTool,
@@ -7767,10 +7810,14 @@ var init_headless = __esm({
         if (resumed) {
           this.emit("session_restored", {
             messageCount: this.state.messages.length,
+            ...this.state.models && { models: this.state.models },
             ...this.queueFields()
           });
         }
-        triggerBrandExtraction(this.config);
+        triggerBrandExtraction(
+          this.config,
+          this.state.models?.brandExtractor ?? this.opts.model
+        );
         this.toolRegistry.onEvent = this.onEvent;
         setCompactionListener((event) => {
           if (event.type === "started") {
@@ -7923,7 +7970,8 @@ var init_headless = __esm({
         try {
           await triggerCompaction(this.state, this.config, {
             blocking: true,
-            requestId
+            requestId,
+            model: this.opts.model
           });
           this.applyPendingSummaries();
         } catch {
@@ -8361,6 +8409,17 @@ var init_headless = __esm({
         clearSession(this.state);
         return {};
       }
+      /** Archive the current session and seed a fresh one with the given
+       * per-agent model overrides. Models are immutable for the life of a
+       * session — this is the only way to change them. Omitting `models`
+       * (or sending an empty object) resets to "use server defaults for
+       * every agent". */
+      handleNewSession(models) {
+        clearSession(this.state);
+        this.state.models = models && Object.keys(models).length > 0 ? models : void 0;
+        saveSession(this.state);
+        return {};
+      }
       /** Cancel the running turn and drain the queue. Returns the drained items. */
       handleCancel() {
         if (this.currentAbort) {
@@ -8434,6 +8493,7 @@ var init_headless = __esm({
             totalMessageCount: total,
             running: this.running,
             ...this.running && this.currentRequestId ? { currentRequestId: this.currentRequestId } : {},
+            ...this.state.models && { models: this.state.models },
             ...this.queueFields()
           }));
           return;
@@ -8446,6 +8506,15 @@ var init_headless = __esm({
           );
           return;
         }
+        if (action === "newSession") {
+          const models = parsed.models;
+          this.dispatchSimple(
+            requestId,
+            "session_cleared",
+            () => this.handleNewSession(models)
+          );
+          return;
+        }
         if (action === "cancel") {
           const cancelled = this.handleCancel();
           this.emit(
@@ -8492,7 +8561,8 @@ var init_headless = __esm({
           try {
             await triggerCompaction(this.state, this.config, {
               blocking: false,
-              requestId
+              requestId,
+              model: this.opts.model
             });
             if (!this.running) {
               this.applyPendingSummaries();

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@mindstudio-ai/remy",
-  "version": "0.1.177",
+  "version": "0.1.178",
   "description": "MindStudio coding agent",
   "repository": {
     "type": "git",