replicas-engine 0.1.321 → 0.1.322

package/dist/src/index.js

@@ -287,7 +287,7 @@ var WORKSPACE_SIZES = ["small", "large"];
 var INVALID_WORKSPACE_SIZE_ERROR = `Invalid size: must be one of ${WORKSPACE_SIZES.join(", ")}`;
 
 // ../shared/src/e2b.ts
-var E2B_TEMPLATE_NAME = "replicas-sandbox-2026-06-15-v5";
+var E2B_TEMPLATE_NAME = "replicas-sandbox-2026-06-16-v1";
 
 // ../shared/src/runtime-env.ts
 function parsePosixEnvFile(content) {
@@ -6457,7 +6457,8 @@ var ClaudeManager = class _ClaudeManager extends CodingAgentManager {
       images,
       permissionMode,
       thinkingLevel,
-      enableInteractiveTools
+      enableInteractiveTools,
+      fastMode
     } = request;
     if (!message || !message.trim()) {
       throw new Error("Message cannot be empty");
@@ -6504,7 +6505,8 @@ var ClaudeManager = class _ClaudeManager extends CodingAgentManager {
       // enableInteractiveTools only matters when permissionMode is 'read'; the
       // session signature tracks the effective flag so a permissionMode change
       // can hot-swap via setPermissionMode without restarting.
-      enableInteractiveTools: Boolean(enableInteractiveTools)
+      enableInteractiveTools: Boolean(enableInteractiveTools),
+      fastMode: Boolean(fastMode)
     };
     await this.ensureSession({
       signature,
@@ -6570,7 +6572,7 @@ var ClaudeManager = class _ClaudeManager extends CodingAgentManager {
     });
   }
   sessionSignaturesMatch(a, b) {
-    return a.combinedInstructions === b.combinedInstructions && a.thinkingLevel === b.thinkingLevel && a.enableInteractiveTools === b.enableInteractiveTools;
+    return a.combinedInstructions === b.combinedInstructions && a.thinkingLevel === b.thinkingLevel && a.enableInteractiveTools === b.enableInteractiveTools && a.fastMode === b.fastMode;
   }
   async startSession(args) {
     const {
@@ -6627,6 +6629,7 @@ var ClaudeManager = class _ClaudeManager extends CodingAgentManager {
         ...enableRegistrySkills ? { skills: "all" } : {},
         env: queryEnv,
         model: resolvedModel,
+        settings: { fastMode: signature.fastMode },
         includePartialMessages: true,
         ...supportsClaudeThinkingDisplay(resolvedModel) ? { thinking: { type: "adaptive", display: "summarized" } } : {},
         ...thinkingLevel ? { effort: thinkingLevel } : {},
@@ -7146,7 +7149,7 @@ var AspClient = class {
 // src/managers/codex-asp/app-server-process.ts
 var DEFAULT_CODEX_BINARY = "codex";
 var DEFAULT_CODEX_ARGS = ["app-server", "--listen", "stdio://"];
-var ENGINE_PACKAGE_VERSION = "0.1.321";
+var ENGINE_PACKAGE_VERSION = "0.1.322";
 var INITIALIZE_METHOD = "initialize";
 var INITIALIZED_NOTIFICATION = "initialized";
 var ACCOUNT_LOGIN_START_METHOD = "account/login/start";
@@ -7396,9 +7399,11 @@ var THREAD_READ_METHOD = "thread/read";
 var THREAD_GOAL_SET_METHOD = "thread/goal/set";
 var THREAD_GOAL_GET_METHOD = "thread/goal/get";
 var THREAD_GOAL_CLEAR_METHOD = "thread/goal/clear";
+var THREAD_SETTINGS_UPDATE_METHOD = "thread/settings/update";
 var TURN_START_METHOD = "turn/start";
 var TURN_INTERRUPT_METHOD = "turn/interrupt";
 var ACCOUNT_RATE_LIMITS_READ_METHOD = "account/rateLimits/read";
+var MODEL_LIST_METHOD = "model/list";
 var MAX_CODEX_ASP_TRANSCRIPT_OUTPUT_CHARS = DEFAULT_HOOK_OUTPUT_PREVIEW_CHARS;
 function codexApprovalPolicyOverrides() {
   if (!ENGINE_ENV.REPLICAS_DISABLE_GH_PR_MERGE) {
@@ -7797,10 +7802,11 @@ function mergeCodexAspTranscripts(primary, supplemental) {
     turns: turns.sort((a, b) => Date.parse(a.startedAt) - Date.parse(b.startedAt))
   });
 }
-async function buildThreadStartParams(workingDirectory, request, developerInstructions) {
+async function buildThreadStartParams(workingDirectory, request, developerInstructions, serviceTier) {
   const additionalDirectories = await getAgentAdditionalDirectories();
   return {
     model: request.model ?? DEFAULT_MODEL,
+    ...serviceTier !== void 0 ? { serviceTier } : {},
     cwd: workingDirectory,
     runtimeWorkspaceRoots: additionalDirectories,
     sandbox: "danger-full-access",
@@ -7811,11 +7817,12 @@ async function buildThreadStartParams(workingDirectory, request, developerInstru
     persistExtendedHistory: false
   };
 }
-async function buildThreadResumeParams(workingDirectory, threadId, request, developerInstructions) {
+async function buildThreadResumeParams(workingDirectory, threadId, request, developerInstructions, serviceTier) {
   const additionalDirectories = await getAgentAdditionalDirectories();
   return {
     threadId,
     model: request.model ?? DEFAULT_MODEL,
+    ...serviceTier !== void 0 ? { serviceTier } : {},
     cwd: workingDirectory,
     runtimeWorkspaceRoots: additionalDirectories,
     sandbox: "danger-full-access",
@@ -7843,7 +7850,7 @@ async function buildTurnInput(request) {
   })));
   return { input, tempImagePaths };
 }
-async function buildTurnStartParams(threadId, request, developerInstructions) {
+async function buildTurnStartParams(threadId, request, developerInstructions, serviceTier) {
   const effort = toReasoningEffort(request.thinkingLevel);
   const model = request.model ?? DEFAULT_MODEL;
   const { input, tempImagePaths } = await buildTurnInput(request);
@@ -7852,6 +7859,7 @@ async function buildTurnStartParams(threadId, request, developerInstructions) {
       threadId,
       input,
       model,
+      ...serviceTier !== void 0 ? { serviceTier } : {},
       ...codexApprovalPolicyOverrides(),
       ...effort ? { effort } : {},
       ...developerInstructions ? {
@@ -7994,6 +8002,8 @@ var CodexAspManager = class extends CodingAgentManager {
     this.flushTranscriptUpdated(threadId);
   });
   skillRegistriesApplied = false;
+  modelServiceTierCache = null;
+  activeServiceTier;
   constructor(options) {
     super(options);
     this.historyFile = options.historyFilePath ? new CodexHistoryFile(options.historyFilePath) : null;
@@ -8159,7 +8169,9 @@ var CodexAspManager = class extends CodingAgentManager {
     const developerInstructions = this.buildCombinedInstructions(request.customInstructions);
     recordUserMessage(options.userMessagePayload);
     const threadId = await this.ensureThread(host, request, developerInstructions);
-    const runTurn = options.runTurn ?? ((aspHost, aspThreadId, aspInstructions) => this.runTurn(aspHost, aspThreadId, request, aspInstructions));
+    const serviceTier = await this.resolveRequestedServiceTier(host, request);
+    await this.applyThreadServiceTier(host, threadId, serviceTier);
+    const runTurn = options.runTurn ?? ((aspHost, aspThreadId, aspInstructions) => this.runTurn(aspHost, aspThreadId, request, aspInstructions, serviceTier));
     let completedTurn;
     try {
       completedTurn = await runTurn(host, threadId, developerInstructions);
@@ -8183,27 +8195,68 @@ var CodexAspManager = class extends CodingAgentManager {
     const existingThreadId = this.currentThreadId;
     if (existingThreadId) {
       if (!this.threadAttached) {
+        const serviceTier2 = await this.resolveRequestedServiceTier(host, request);
         const response = await host.client.request(
           THREAD_RESUME_METHOD,
-          await buildThreadResumeParams(this.workingDirectory, existingThreadId, request, developerInstructions)
+          await buildThreadResumeParams(this.workingDirectory, existingThreadId, request, developerInstructions, serviceTier2)
         );
         this.currentThreadId = response.thread.id;
+        this.activeServiceTier = response.serviceTier;
         this.threadAttached = true;
         this.seedHistoryFromThread(response.thread);
         await this.onSaveSessionId(this.currentThreadId);
       }
       return this.currentThreadId ?? existingThreadId;
     }
+    const serviceTier = await this.resolveRequestedServiceTier(host, request);
     const threadStartResponse = await host.client.request(
       THREAD_START_METHOD,
-      await buildThreadStartParams(this.workingDirectory, request, developerInstructions)
+      await buildThreadStartParams(this.workingDirectory, request, developerInstructions, serviceTier)
     );
     const threadId = threadStartResponse.thread.id;
     this.currentThreadId = threadId;
+    this.activeServiceTier = threadStartResponse.serviceTier;
     this.threadAttached = true;
     await this.onSaveSessionId(this.currentThreadId);
     return threadId;
   }
+  async resolveRequestedServiceTier(host, request) {
+    if (request.fastMode === void 0) return void 0;
+    if (!request.fastMode) return null;
+    const model = request.model ?? DEFAULT_MODEL;
+    const models = await this.listModels(host);
+    const catalogModel = models.find((candidate) => candidate.model === model || candidate.id === model);
+    if (!catalogModel) return void 0;
+    const fastTier = catalogModel.serviceTiers.find((tier) => /fast/i.test(tier.id) || /fast/i.test(tier.name));
+    if (fastTier) return fastTier.id;
+    if (catalogModel.defaultServiceTier && catalogModel.serviceTiers.length === 1) {
+      return catalogModel.defaultServiceTier;
+    }
+    return void 0;
+  }
+  async listModels(host) {
+    if (this.modelServiceTierCache) return this.modelServiceTierCache;
+    const models = [];
+    let cursor = null;
+    do {
+      const response = await host.client.request(
+        MODEL_LIST_METHOD,
+        { cursor, includeHidden: true }
+      );
+      models.push(...response.data);
+      cursor = response.nextCursor;
+    } while (cursor);
+    this.modelServiceTierCache = models;
+    return models;
+  }
+  async applyThreadServiceTier(host, threadId, serviceTier) {
+    if (serviceTier === void 0 || serviceTier === this.activeServiceTier) return;
+    await host.client.request(
+      THREAD_SETTINGS_UPDATE_METHOD,
+      { threadId, serviceTier }
+    );
+    this.activeServiceTier = serviceTier;
+  }
   async applySkillRegistries(host) {
     if (this.skillRegistriesApplied) return;
     try {
@@ -8216,8 +8269,8 @@ var CodexAspManager = class extends CodingAgentManager {
       console.warn("[CodexAspManager] Failed to apply skill registries:", error);
     }
   }
-  async runTurn(host, threadId, request, developerInstructions) {
-    const { params, tempImagePaths } = await buildTurnStartParams(threadId, request, developerInstructions);
+  async runTurn(host, threadId, request, developerInstructions, serviceTier) {
+    const { params, tempImagePaths } = await buildTurnStartParams(threadId, request, developerInstructions, serviceTier);
     return this.observeTurn(host, threadId, request, async () => {
       const turnStartResponse = await host.client.request(
         TURN_START_METHOD,
@@ -11141,6 +11194,7 @@ var sendMessageSchema = z2.object({
   })).optional(),
   thinkingLevel: z2.enum(["low", "medium", "high", "max"]).optional(),
   goalMode: z2.boolean().optional(),
+  fastMode: z2.boolean().optional(),
   enableInteractiveTools: z2.boolean().optional(),
   type: z2.string().min(1).optional(),
   merge: z2.boolean().optional(),

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "replicas-engine",
-  "version": "0.1.321",
+  "version": "0.1.322",
   "description": "Lightweight API server for Replicas workspaces",
   "type": "module",
   "main": "dist/src/index.js",