@runtypelabs/cli 2.31.1 → 2.31.3

package/dist/index.js

@@ -15260,7 +15260,7 @@ var edgeContextPayloadSchema = external_exports.object({
 // ../shared/dist/chunk-SCJRL6LP.mjs
 var MAX_DURABLE_POLL_ATTEMPTS = 1e3;
 
-// ../shared/dist/chunk-K5ARRYN3.mjs
+// ../shared/dist/chunk-R7RM64AL.mjs
 var DEFAULT_MODELS_FOR_NEW_ACCOUNTS = [
   // Mixlayer models
   {
@@ -15325,7 +15325,7 @@ var DEFAULT_MODELS_FOR_NEW_ACCOUNTS = [
   },
   {
     provider: "runtype",
-    modelId: "gemini-3.5-flash",
+    modelId: "gemini-3.6-flash",
     isDefault: false,
     strengths: "Fast, low-cost, and natively multimodal with good reasoning, tool-use, and a 1M context \u2014 a strong choice for vision and high-throughput steps.",
     bestFor: ["default", "vision", "fast-cheap", "agentic-tools", "long-context"]
@@ -15786,7 +15786,7 @@ function resolveFallback(parsed, context, tracking, options = {}) {
 }
 var templateEngine = new SimpleTemplateEngine();
 
-// ../shared/dist/chunk-PL7A5DM4.mjs
+// ../shared/dist/chunk-TJITC2J6.mjs
 var tokensSchema = external_exports.object({ input: external_exports.number(), output: external_exports.number() });
 var wireStopReasonSchema = external_exports.enum([
   "end_turn",
@@ -15807,6 +15807,10 @@ var unifiedBase = {
   seq: external_exports.number()
 };
 var executionKindSchema = external_exports.enum(["agent", "flow"]);
+var unifiedExternalAgentContextSchema = external_exports.object({
+  contextId: external_exports.string().optional(),
+  taskId: external_exports.string().optional()
+});
 var fallbackAttemptSchema = external_exports.object({
   attempt: external_exports.number(),
   type: external_exports.enum(["retry", "model", "message", "flow"]),
@@ -15861,7 +15865,10 @@ var unifiedExecutionCompleteEventSchema = external_exports.object({
   failedSteps: external_exports.number().optional(),
   // Inline Claude Managed create dispatches return the reusable Anthropic
   // agent id on the terminal frame so callers can connect on a later turn.
-  claudeManagedAgentId: external_exports.string().optional()
+  claudeManagedAgentId: external_exports.string().optional(),
+  // External A2A agents return the conversation handle callers must persist in
+  // order to resume the same upstream task on the next execution.
+  externalAgent: unifiedExternalAgentContextSchema.optional()
 });
 var unifiedExecutionErrorEventSchema = external_exports.object({
   ...unifiedBase,
@@ -16135,7 +16142,10 @@ var unifiedApprovalStartEventSchema = external_exports.object({
   iteration: external_exports.number().optional(),
   // Present when the approval originated inside a subagent's execution.
   // Attribution only — never a control signal.
-  subagent: unifiedSubagentAttributionSchema.optional()
+  subagent: unifiedSubagentAttributionSchema.optional(),
+  // External A2A input/auth pauses must expose the handle used to resume the
+  // paused upstream conversation after approval.
+  externalAgent: unifiedExternalAgentContextSchema.optional()
 });
 var unifiedApprovalCompleteEventSchema = external_exports.object({
   ...unifiedBase,
@@ -16428,6 +16438,7 @@ var agentToolCompleteEventSchema = external_exports.object({
   toolName: external_exports.string(),
   success: external_exports.boolean(),
   result: external_exports.unknown().optional(),
+  error: external_exports.string().optional(),
   executionTime: external_exports.number().optional()
 });
 var agentMediaEventSchema = external_exports.object({
@@ -35069,6 +35080,36 @@ var CORE_BUILTIN_TOOLS_REGISTRY = [
     // Requires OpenAI key for embeddings
     executionHint: "platform"
   },
+  // Runtype documentation search — semantic RAG over the published docs corpus.
+  // Executes server-side against Fern's hosted docs MCP server (see
+  // apps/api/src/services/docs-search-service.ts). Free (no per-call cost, no
+  // provider key); gated by the `enable-docs-search` Flagship flag at execution.
+  {
+    id: "search_runtype_docs",
+    name: "Search Runtype Docs",
+    description: "Search Runtype's own product documentation and get back a synthesized answer with citations. Use this to answer how-to and conceptual questions about building on Runtype \u2014 flows, agents, surfaces, tools, records, schedules, and platform capabilities. Prefer this over generic web search for anything about how Runtype itself works. Returns { success, content } where content carries the answer text and its source citations.",
+    category: BuiltInToolCategory.KNOWLEDGE_RETRIEVAL,
+    providers: [BuiltInToolProvider.MULTI],
+    parametersSchema: {
+      type: "object",
+      properties: {
+        query: {
+          type: "string",
+          description: 'Natural-language documentation question about how to build or configure something on Runtype. Example: "How do I add an approval step to an agent?"',
+          minLength: 1
+        }
+      },
+      required: ["query"]
+    },
+    executionHint: "platform",
+    requiresApiKey: false,
+    platformKeySupport: false,
+    // Hidden from production tool discovery until `enable-docs-search` rolls out
+    // (the flag fails closed in production, so a discoverable tool would just
+    // error "not enabled" for every prod account). Execution path is
+    // unaffected — the Generator Agent and non-prod accounts still use it.
+    experimental: true
+  },
   // ElevenLabs Text-to-Speech
   {
     id: "elevenlabs-tts",
@@ -38740,10 +38781,18 @@ var caseCheckpointSchema = external_exports.object({
    */
   forkMessageIndex: external_exports.number().int().nonnegative()
 }).strict();
+var requiredJsonValueSchema = external_exports.union([
+  external_exports.null(),
+  external_exports.boolean(),
+  external_exports.number(),
+  external_exports.string(),
+  external_exports.array(external_exports.any()),
+  external_exports.record(external_exports.string(), external_exports.any())
+]);
 var toolMockSchema = external_exports.object({
   toolName: external_exports.string().min(1),
   input: external_exports.unknown().optional(),
-  output: external_exports.unknown(),
+  output: requiredJsonValueSchema,
   /**
    * True when the recorded call FAILED — `output` is then the error payload,
    * and a replay must serve it as a failure, not a success. Captured now so
@@ -38782,7 +38831,7 @@ var recordedToolUnmatchedPolicySchema = external_exports.enum(RECORDED_TOOL_UNMA
 var EVAL_JUDGE_SYSTEM_PROMPT = [
   "You are Runtype's evaluation expert: a rigorous, impartial judge of AI agent and flow behavior. Your job is to decide whether ONE run of an AI system met ONE plain-language criterion written by the system's builder.",
   "",
-  'You are given the criterion (what this flow or agent is supposed to do \u2014 the only standard you grade against), the case name, an optional expected answer for reference, and the run transcript: an indexed record of the steps that ran and every tool call with its inputs and outputs (labeled "[message N]" and "[tool_call N]"), followed by the final output.',
+  `You are given the criterion (what this flow or agent is supposed to do \u2014 the only standard you grade against), the case name, an optional expected answer for reference, and the run transcript: an indexed record of the steps that ran and every tool call with its inputs and outputs (labeled "[message N]" and "[tool_call N]"), followed by the final output. For conversation-seeded cases, the transcript opens with the seeded conversation \u2014 the dialogue the run was replying to. It is context the system was given, not something the run produced: use it to understand what the final output is responding to, and never grade the seeded turns themselves against the criterion. The transcript may also include a "Deterministic check results" section: the pass/fail outcomes of the suite's automated checks, already decided before you were called. Treat those results as settled ground truth.`,
   "",
   "How to judge:",
   '1. Read the criterion first and decide what evidence it demands. A process criterion ("confirms X before Y", "never calls Z", "asks first") is judged on the trajectory \u2014 what the system actually did. An outcome criterion ("answers the question", "matches the expected answer", "right tone") is judged on the final output. Many criteria need both.',
@@ -38790,6 +38839,7 @@ var EVAL_JUDGE_SYSTEM_PROMPT = [
   "3. Judge whether a required action occurred when it mattered. Never demand a rigid, exact step ordering unless the criterion explicitly requires one; incidental extra steps or interleaved calls are not failures.",
   "4. Grade only the criterion you were given. Do not reward or punish anything outside it: not verbosity or brevity, not style, not parts of the run the criterion doesn't mention, and not whether you would have approached the task differently.",
   "5. When the transcript is truncated or no trajectory was captured, judge what you can see. If what you can see cannot settle the criterion, decline \u2014 do not guess in either direction.",
+  '6. Do not re-litigate the deterministic checks. Whether a tool was or was not called, a step ran, the run completed, or the output contains a required string is already settled by the "Deterministic check results" shown as ground truth \u2014 take those outcomes as given, never re-assert or contradict them, and do not make your verdict hinge on restating one. Your job is the qualitative substance those structural checks cannot decide.',
   "",
   "Reason first, then decide: in one paragraph, walk through the specific evidence for and against the criterion BEFORE settling on a verdict, and never rewrite your reasoning to justify a verdict you already picked.",
   "",
@@ -38798,7 +38848,7 @@ var EVAL_JUDGE_SYSTEM_PROMPT = [
   '- "fail" \u2014 the criterion is clearly not met on the evidence shown.',
   '- "insufficient_evidence" \u2014 what you were shown cannot settle this criterion. This is the honest answer, not a failure; use it instead of guessing.',
   "",
-  'Cite the transcript entries your verdict rests on as an array of {"kind": "message"|"tool_call", "index": N}, using the exact labeled indices. An empty array is correct when the final output alone (which has no labeled index) is the evidence.',
+  'Cite the transcript entries your verdict rests on as an array of {"kind": "message"|"tool_call", "index": N}, using the exact labeled indices. An empty array is correct when the evidence is the final output or the seeded conversation (neither has a labeled index).',
   "",
   'Return ONLY a JSON object of the form {"reasoning": "<one paragraph>", "verdict": "pass"|"fail"|"insufficient_evidence", "evidence": [{"kind": "message"|"tool_call", "index": <integer>}]}. Do not include any text outside the JSON object.',
   "",
@@ -39162,6 +39212,14 @@ var MODEL_FAMILY_PROVIDER_IDS = {
     "google": "gemini-3.5-flash",
     "vercel": "google/gemini-3.5-flash"
   },
+  "gemini-3-5-flash-lite": {
+    "google": "gemini-3.5-flash-lite",
+    "vercel": "google/gemini-3.5-flash-lite"
+  },
+  "gemini-3-6-flash": {
+    "google": "gemini-3.6-flash",
+    "vercel": "google/gemini-3.6-flash"
+  },
   "gemini-3-flash": {
     "google": "gemini-3-flash-preview"
   },
@@ -39196,6 +39254,14 @@ var MODEL_FAMILY_PROVIDER_IDS = {
     "google": "gemini-3.5-flash",
     "vercel": "google/gemini-3.5-flash"
   },
+  "gemini-3.5-flash-lite": {
+    "google": "gemini-3.5-flash-lite",
+    "vercel": "google/gemini-3.5-flash-lite"
+  },
+  "gemini-3.6-flash": {
+    "google": "gemini-3.6-flash",
+    "vercel": "google/gemini-3.6-flash"
+  },
   "gemini-embedding-001": {
     "google": "gemini-embedding-001"
   },
@@ -39532,7 +39598,11 @@ var MODEL_FAMILY_PROVIDER_IDS = {
   "grok-voice-think-fast-1.0": {
     "vercel": "xai/grok-voice-think-fast-1.0"
   },
+  "hy3": {
+    "vercel": "tencent/hy3"
+  },
   "inkling": {
+    "togetherai": "togetherai/thinkingmachines/Inkling",
     "vercel": "thinkingmachines/inkling"
   },
   "interfaze-beta": {
@@ -39595,23 +39665,23 @@ var MODEL_FAMILY_PROVIDER_IDS = {
   "kimi-k3": {
     "vercel": "moonshotai/kimi-k3"
   },
-  "lfm2-24b-a2b": {
-    "togetherai": "togetherai/LiquidAI/LFM2-24B-A2B"
+  "laguna-s-2-1": {
+    "vercel": "poolside/laguna-s-2.1"
   },
-  "LiquidAI/LFM2-24B-A2B": {
-    "togetherai": "togetherai/LiquidAI/LFM2-24B-A2B"
+  "laguna-s-2-1-free": {
+    "vercel": "poolside/laguna-s-2.1-free"
   },
-  "llama-3-2-11b": {
-    "vercel": "meta/llama-3.2-11b"
+  "laguna-s-2.1": {
+    "vercel": "poolside/laguna-s-2.1"
   },
-  "llama-3-2-1b": {
-    "vercel": "meta/llama-3.2-1b"
+  "laguna-s-2.1-free": {
+    "vercel": "poolside/laguna-s-2.1-free"
   },
-  "llama-3-2-3b": {
-    "vercel": "meta/llama-3.2-3b"
+  "lfm2-24b-a2b": {
+    "togetherai": "togetherai/LiquidAI/LFM2-24B-A2B"
   },
-  "llama-3-2-90b": {
-    "vercel": "meta/llama-3.2-90b"
+  "LiquidAI/LFM2-24B-A2B": {
+    "togetherai": "togetherai/LiquidAI/LFM2-24B-A2B"
   },
   "llama-3-3-70b-instruct-turbo": {
     "togetherai": "togetherai/meta-llama/Llama-3.3-70B-Instruct-Turbo"
@@ -39622,18 +39692,6 @@ var MODEL_FAMILY_PROVIDER_IDS = {
   "llama-3.1-8b": {
     "vercel": "meta/llama-3.1-8b"
   },
-  "llama-3.2-11b": {
-    "vercel": "meta/llama-3.2-11b"
-  },
-  "llama-3.2-1b": {
-    "vercel": "meta/llama-3.2-1b"
-  },
-  "llama-3.2-3b": {
-    "vercel": "meta/llama-3.2-3b"
-  },
-  "llama-3.2-90b": {
-    "vercel": "meta/llama-3.2-90b"
-  },
   "llama-3.3-70b": {
     "vercel": "meta/llama-3.3-70b"
   },
@@ -40039,6 +40097,9 @@ var MODEL_FAMILY_PROVIDER_IDS = {
   "text-embedding-ada-002": {
     "openai": "text-embedding-ada-002"
   },
+  "thinkingmachines/Inkling": {
+    "togetherai": "togetherai/thinkingmachines/Inkling"
+  },
   "trinity-large": {
     "vercel": "arcee-ai/trinity-large-thinking"
   },
@@ -45079,6 +45140,70 @@ var MOCK_INVENTORY_LEVELS = MOCK_PRODUCTS.flatMap(
     sku: variant.sku
   }))
 );
+var LOCAL_INFERENCE_PROTOCOL_VERSION = 1;
+var LocalInferenceSelectionSchema = external_exports.object({
+  sessionId: external_exports.string().uuid(),
+  model: external_exports.string().trim().min(1).max(256),
+  backend: external_exports.literal("ollama")
+});
+var LocalInferenceMessageSchema = external_exports.object({
+  role: external_exports.enum(["system", "user", "assistant"]),
+  content: external_exports.string().max(1e6)
+});
+var LocalInferenceInvokeSchema = external_exports.object({
+  type: external_exports.literal("inference.invoke"),
+  protocolVersion: external_exports.literal(LOCAL_INFERENCE_PROTOCOL_VERSION),
+  invocationId: external_exports.string().uuid(),
+  model: external_exports.string().trim().min(1).max(256),
+  messages: external_exports.array(LocalInferenceMessageSchema).min(1).max(500),
+  temperature: external_exports.number().finite().min(0).max(2).optional(),
+  maxTokens: external_exports.number().int().positive().max(131072).optional()
+});
+var LocalInferenceServerFrameSchema = external_exports.discriminatedUnion("type", [
+  LocalInferenceInvokeSchema,
+  external_exports.object({
+    type: external_exports.literal("inference.cancel"),
+    protocolVersion: external_exports.literal(LOCAL_INFERENCE_PROTOCOL_VERSION),
+    invocationId: external_exports.string().uuid()
+  }),
+  external_exports.object({
+    type: external_exports.literal("session.connected"),
+    protocolVersion: external_exports.literal(LOCAL_INFERENCE_PROTOCOL_VERSION),
+    sessionId: external_exports.string().uuid()
+  })
+]);
+var LocalInferenceClientFrameSchema = external_exports.discriminatedUnion("type", [
+  external_exports.object({
+    type: external_exports.literal("session.ready"),
+    protocolVersion: external_exports.literal(LOCAL_INFERENCE_PROTOCOL_VERSION)
+  }),
+  external_exports.object({
+    type: external_exports.literal("inference.chunk"),
+    protocolVersion: external_exports.literal(LOCAL_INFERENCE_PROTOCOL_VERSION),
+    invocationId: external_exports.string().uuid(),
+    sequence: external_exports.number().int().nonnegative(),
+    text: external_exports.string().max(256e3)
+  }),
+  external_exports.object({
+    type: external_exports.literal("inference.complete"),
+    protocolVersion: external_exports.literal(LOCAL_INFERENCE_PROTOCOL_VERSION),
+    invocationId: external_exports.string().uuid(),
+    sequence: external_exports.number().int().nonnegative(),
+    finishReason: external_exports.string().max(64).optional(),
+    usage: external_exports.object({
+      inputTokens: external_exports.number().int().nonnegative().max(1e8).optional(),
+      outputTokens: external_exports.number().int().nonnegative().max(1e8).optional()
+    }).optional()
+  }),
+  external_exports.object({
+    type: external_exports.literal("inference.error"),
+    protocolVersion: external_exports.literal(LOCAL_INFERENCE_PROTOCOL_VERSION),
+    invocationId: external_exports.string().uuid(),
+    sequence: external_exports.number().int().nonnegative(),
+    code: external_exports.enum(["cancelled", "ollama_unavailable", "model_error", "protocol_error"]),
+    message: external_exports.string().min(1).max(2e3)
+  })
+]);
 function flattenTextContent(content) {
   if (typeof content === "string") return content;
   return content.filter((p) => p.type === "text" && typeof p.text === "string").map((p) => p.text).join("");
@@ -45379,7 +45504,11 @@ var DispatchRequestSchema = external_exports.object({
     autoAppendMetadata: external_exports.boolean().optional().default(false),
     debugMode: external_exports.boolean().optional().default(false),
     loggingPolicy: external_exports.enum(["default", "on", "off"]).optional(),
-    useRuntimePackage: external_exports.boolean().optional().default(false),
+    useRuntimePackage: external_exports.boolean().optional().describe(
+      "Selects the execution engine for a streaming FLOW dispatch (two controls, A4-3). `true` forces the @runtypelabs/runtime lane (subject only to eligibility \u2014 an ineligible flow still folds to legacy silently); `false` forces the legacy engine. When omitted, the `use-runtime-package-flows` rollout flag sets the default engine for this caller. An explicit value always wins over the flag. Left undefined on purpose \u2014 do NOT default it, or the flag-as-default case becomes unreachable."
+    ),
+    /** Route prompt model calls through an attached dashboard browser. */
+    localInference: LocalInferenceSelectionSchema.optional(),
     /**
      * Resume a checkpointed agent-loop execution from its latest (or a
      * specific) iteration-boundary snapshot. Runtime-path only: requires
@@ -45400,6 +45529,18 @@ var DispatchRequestSchema = external_exports.object({
       /** Specific boundary to resume from (time-travel); latest when omitted. */
       iteration: external_exports.number().int().nonnegative().optional()
     }).optional(),
+    /**
+     * Client-executed (local / WebMCP) tool outputs for resuming a
+     * client-tool pause on the runtime agent lane. Keyed by the paused
+     * call's per-call id (`toolu_…` / `call_…`) or, as a fallback, the tool
+     * name; the value is the tool's result. Runtime-path only and consumed
+     * ONLY together with `resumeFrom` — the engine resolves each output
+     * against the checkpoint's pending calls, injects the tool-call/result
+     * message pairs so the model sees its prior calls, and re-pauses for any
+     * unresolved parallel siblings. Passthrough-typed so a client can
+     * resolve arbitrary tool return values.
+     */
+    clientToolOutputs: external_exports.record(external_exports.string(), external_exports.unknown()).optional(),
     environment: external_exports.enum(["development", "production"]).optional().default("development"),
     createVersion: external_exports.boolean().optional().default(false),
     versionType: external_exports.enum(["published", "draft", "test", "virtual"]).optional(),
@@ -45431,12 +45572,29 @@ var DispatchRequestSchema = external_exports.object({
     storeResults: true,
     autoAppendMetadata: false,
     debugMode: false,
-    useRuntimePackage: false,
+    // useRuntimePackage intentionally omitted — absent = silent request, so
+    // the `use-runtime-package-flows` flag decides the default engine (A4-3).
     environment: "development",
     createVersion: false
   })
 }).refine((data) => data.flow || data.agent, {
   message: "Either 'flow' or 'agent' configuration must be provided"
+}).superRefine((data, ctx) => {
+  if (!data.options.localInference) return;
+  if (!data.options.useRuntimePackage) {
+    ctx.addIssue({
+      code: "custom",
+      path: ["options", "useRuntimePackage"],
+      message: "Local inference requires useRuntimePackage to be enabled"
+    });
+  }
+  if (data.options.modelOverride) {
+    ctx.addIssue({
+      code: "custom",
+      path: ["options", "modelOverride"],
+      message: "Local inference cannot be combined with modelOverride"
+    });
+  }
 }).superRefine((data, ctx) => {
   if (data.agent?.agentType !== "claude_managed") return;
   if (!hasUsableClaudeManagedSeed(data.messages ?? [])) {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@runtypelabs/cli",
-  "version": "2.31.1",
+  "version": "2.31.3",
   "description": "Command-line interface for Runtype AI platform",
   "type": "module",
   "main": "dist/index.js",
@@ -25,9 +25,9 @@
     "react": "^19.2.4",
     "rosie-skills": "0.8.1",
     "yaml": "^2.9.0",
-    "@runtypelabs/ink-components": "0.3.5",
-    "@runtypelabs/sdk": "6.2.0",
-    "@runtypelabs/terminal-animations": "0.2.2"
+    "@runtypelabs/terminal-animations": "0.2.2",
+    "@runtypelabs/sdk": "6.3.0",
+    "@runtypelabs/ink-components": "0.3.5"
   },
   "devDependencies": {
     "@runtypelabs/persona": "4.8.0",
@@ -41,7 +41,7 @@
     "tsx": "^4.7.1",
     "typescript": "^6.0.3",
     "vitest": "^4.1.0",
-    "@runtypelabs/shared": "3.6.0"
+    "@runtypelabs/shared": "3.8.0"
   },
   "engines": {
     "node": ">=22.12.0"