@runtypelabs/cli 2.25.0 → 2.25.2

package/dist/index.js

@@ -15566,7 +15566,7 @@ function resolveFallback(parsed, context, tracking) {
 }
 var templateEngine = new SimpleTemplateEngine();
 
-// ../shared/dist/chunk-CNO2UJFO.mjs
+// ../shared/dist/chunk-54HZCG2Y.mjs
 var tokensSchema = external_exports.object({ input: external_exports.number(), output: external_exports.number() });
 var wireStopReasonSchema = external_exports.enum([
   "end_turn",
@@ -15916,12 +15916,15 @@ var unifiedAwaitEventSchema = external_exports.object({
   origin: external_exports.enum(["webmcp", "sdk"]).optional(),
   pageOrigin: external_exports.string().optional(),
   // Discriminates an AUTO-RESUMING durable pause (`crawl_pending` /
-  // `durable_poll`, resumed server-side by the crawl / wait-until
-  // Workflow) from a CLIENT-RESOLVABLE local-tool / WebMCP await (absent), so a
-  // streaming consumer does not render a "resume" prompt for a pause that
-  // resumes itself. Mirrors the JSON `pausedReason` discriminator
-  // (execution-pipeline.ts); the server-resumed pauses carry `crawlId` /
-  // `stepId` instead of the tool fields. UX context; never a control signal.
+  // `durable_poll`, resumed server-side by the crawl / wait-until Workflow;
+  // `detached`, a Claude Managed watch-budget detach — the stream ends but
+  // the execution continues server-side, so the consumer should hold its
+  // reconnect handle and render "still working") from a CLIENT-RESOLVABLE
+  // local-tool / WebMCP await (absent), so a streaming consumer does not
+  // render a "resume" prompt for a pause that resumes itself. Mirrors the
+  // JSON `pausedReason` discriminator (execution-pipeline.ts); the
+  // server-resumed pauses carry `crawlId` / `stepId` instead of the tool
+  // fields. UX context; never a control signal.
   awaitReason: external_exports.string().optional(),
   crawlId: external_exports.string().optional(),
   stepId: external_exports.string().optional()
@@ -16428,12 +16431,15 @@ var flowAwaitEventSchema = external_exports.object({
   origin: external_exports.enum(["webmcp", "sdk"]).optional(),
   pageOrigin: external_exports.string().optional(),
   // Discriminates an AUTO-RESUMING durable pause (`crawl_pending` /
-  // `durable_poll`, resumed server-side by the crawl / wait-until
-  // Workflow) from a CLIENT-RESOLVABLE local-tool / WebMCP await (absent). The
-  // server-resumed pauses carry `crawlId` / `stepId` instead of the tool fields.
-  // Open `z.string()` (not an enum) for forward-compat — kept in lockstep with
-  // the JSON `pausedReason` discriminator in execution-pipeline.ts. UX context;
-  // never a control signal.
+  // `durable_poll`, resumed server-side by the crawl / wait-until Workflow;
+  // `detached`, a Claude Managed watch-budget detach — the stream ends but the
+  // execution continues server-side and the headless finalize stamps the real
+  // terminal for `?after=` replay) from a CLIENT-RESOLVABLE local-tool /
+  // WebMCP await (absent). The server-resumed pauses carry `crawlId` /
+  // `stepId` instead of the tool fields. Open `z.string()` (not an enum) for
+  // forward-compat — kept in lockstep with the JSON `pausedReason`
+  // discriminator in execution-pipeline.ts. UX context; never a control
+  // signal.
   awaitReason: external_exports.string().optional(),
   crawlId: external_exports.string().optional(),
   stepId: external_exports.string().optional(),
@@ -16715,6 +16721,117 @@ function neutralizeCsvFormula(value) {
   if (STRICT_NUMBER.test(value)) return value;
   return `'${value}`;
 }
+var agUiRoleSchema = external_exports.enum(["developer", "system", "assistant", "user", "tool"]);
+var runStartedEventSchema = external_exports.object({
+  type: external_exports.literal("RUN_STARTED"),
+  threadId: external_exports.string(),
+  runId: external_exports.string()
+});
+var runFinishedEventSchema = external_exports.object({
+  type: external_exports.literal("RUN_FINISHED"),
+  threadId: external_exports.string(),
+  runId: external_exports.string(),
+  result: external_exports.unknown().optional()
+});
+var runErrorEventSchema = external_exports.object({
+  type: external_exports.literal("RUN_ERROR"),
+  message: external_exports.string(),
+  code: external_exports.string().optional()
+});
+var stepStartedEventSchema = external_exports.object({
+  type: external_exports.literal("STEP_STARTED"),
+  stepName: external_exports.string()
+});
+var stepFinishedEventSchema = external_exports.object({
+  type: external_exports.literal("STEP_FINISHED"),
+  stepName: external_exports.string()
+});
+var textMessageStartEventSchema = external_exports.object({
+  type: external_exports.literal("TEXT_MESSAGE_START"),
+  messageId: external_exports.string(),
+  role: agUiRoleSchema
+});
+var textMessageContentEventSchema = external_exports.object({
+  type: external_exports.literal("TEXT_MESSAGE_CONTENT"),
+  messageId: external_exports.string(),
+  /** Non-empty incremental text. */
+  delta: external_exports.string()
+});
+var textMessageEndEventSchema = external_exports.object({
+  type: external_exports.literal("TEXT_MESSAGE_END"),
+  messageId: external_exports.string()
+});
+var toolCallStartEventSchema = external_exports.object({
+  type: external_exports.literal("TOOL_CALL_START"),
+  toolCallId: external_exports.string(),
+  toolCallName: external_exports.string(),
+  parentMessageId: external_exports.string().optional()
+});
+var toolCallArgsEventSchema = external_exports.object({
+  type: external_exports.literal("TOOL_CALL_ARGS"),
+  toolCallId: external_exports.string(),
+  /** Incremental (or, for non-streaming providers, complete) JSON-text of args. */
+  delta: external_exports.string()
+});
+var toolCallEndEventSchema = external_exports.object({
+  type: external_exports.literal("TOOL_CALL_END"),
+  toolCallId: external_exports.string()
+});
+var toolCallResultEventSchema = external_exports.object({
+  type: external_exports.literal("TOOL_CALL_RESULT"),
+  messageId: external_exports.string(),
+  toolCallId: external_exports.string(),
+  content: external_exports.string(),
+  role: external_exports.literal("tool").optional()
+});
+var reasoningStartEventSchema = external_exports.object({
+  type: external_exports.literal("REASONING_START"),
+  messageId: external_exports.string()
+});
+var reasoningMessageStartEventSchema = external_exports.object({
+  type: external_exports.literal("REASONING_MESSAGE_START"),
+  messageId: external_exports.string(),
+  role: external_exports.literal("reasoning")
+});
+var reasoningMessageContentEventSchema = external_exports.object({
+  type: external_exports.literal("REASONING_MESSAGE_CONTENT"),
+  messageId: external_exports.string(),
+  /** Non-empty incremental reasoning text. */
+  delta: external_exports.string()
+});
+var reasoningMessageEndEventSchema = external_exports.object({
+  type: external_exports.literal("REASONING_MESSAGE_END"),
+  messageId: external_exports.string()
+});
+var reasoningEndEventSchema = external_exports.object({
+  type: external_exports.literal("REASONING_END"),
+  messageId: external_exports.string()
+});
+var customEventSchema = external_exports.object({
+  type: external_exports.literal("CUSTOM"),
+  name: external_exports.string(),
+  value: external_exports.unknown()
+});
+var agUiSSEEventSchema = external_exports.discriminatedUnion("type", [
+  runStartedEventSchema,
+  runFinishedEventSchema,
+  runErrorEventSchema,
+  stepStartedEventSchema,
+  stepFinishedEventSchema,
+  textMessageStartEventSchema,
+  textMessageContentEventSchema,
+  textMessageEndEventSchema,
+  toolCallStartEventSchema,
+  toolCallArgsEventSchema,
+  toolCallEndEventSchema,
+  toolCallResultEventSchema,
+  reasoningStartEventSchema,
+  reasoningMessageStartEventSchema,
+  reasoningMessageContentEventSchema,
+  reasoningMessageEndEventSchema,
+  reasoningEndEventSchema,
+  customEventSchema
+]);
 var AGENT_CONTENT_CONFIG_KEYS = {
   model: true,
   systemPrompt: true,
@@ -21429,13 +21546,13 @@ function collectFlowStructureIssues(flowData, deps, buckets) {
           { stepIndex }
         );
         for (const issue2 of codeResult.issues) {
-          const location = issue2.line ? ` (line ${issue2.line}, column ${issue2.column})` : "";
+          const location2 = issue2.line ? ` (line ${issue2.line}, column ${issue2.column})` : "";
           const issueCode = issue2.code === "SYNTAX_ERROR" ? "CODE_SYNTAX_ERROR" : `CODE_${issue2.code}`;
           addIssue(
             issue2.severity,
             {
               code: issueCode,
-              message: `${issue2.message}${location}`,
+              message: `${issue2.message}${location2}`,
               path: `flowSteps[${stepIndex}].config.script`,
               step: stepRef,
               details: {
@@ -38055,12 +38172,48 @@ var aiGraderSchema = external_exports.object({
   criteria: external_exports.string().min(1),
   /** Reference-guided when the case has `expected`. */
   useExpected: external_exports.boolean().default(false),
-  /** Defaults to a cheap routed model (e.g. claude-haiku-4-5) at execution time. */
+  /** Defaults to the Runtype Eval Judge flow's model pair (`EVAL_JUDGE_MODEL` with a cross-provider fallback) at execution time. Ignored when `judgeFlowId` is set (the flow's own steps pin their models). */
   model: external_exports.string().optional(),
+  /**
+   * Custom judge flow — the escape hatch when the platform judge prompt isn't
+   * enough (extensions plan §1.2.5). The referenced flow (owned by the
+   * evaluating account) is executed through the normal flow pipeline as the
+   * judge for this grader: it receives the four `EvalJudgeVariables`
+   * (`criteria` / `caseName` / `expected` / `transcript`, see
+   * `eval-judge-flow.ts`) as inputs and its final output must be the same
+   * verdict JSON contract the platform judge returns. Validated at run time:
+   * a missing/inaccessible flow, a failed run, or malformed output resolves to
+   * an `insufficient_evidence` outcome with a visible grader error — a
+   * miswired custom judge warns, it never decides cases.
+   */
+  judgeFlowId: external_exports.string().min(1).optional(),
   /** Pass cutoff for the 1-5 scale. */
   threshold: external_exports.number().min(1).max(5).optional(),
+  /**
+   * The platform judge-prompt version that produced this grader's most recent
+   * outcome. Stamped by `grader-executor.ts` at grade time into
+   * `eval_case_scores.grader_snapshot` (design decision: the judge is a
+   * versioned platform prompt, not a persisted flow) — not settable by
+   * callers; any caller-supplied value is overwritten before scoring (or
+   * stripped entirely for `judgeFlowId` graders, whose outcomes the platform
+   * prompt did not produce).
+   */
+  judgePromptVersion: external_exports.number().int().positive().optional(),
   ...severityFields
 });
+var JUDGE_VERDICTS = ["pass", "fail", "insufficient_evidence"];
+var judgeVerdictSchema = external_exports.enum(JUDGE_VERDICTS);
+var JUDGE_EVIDENCE_KINDS = ["message", "tool_call"];
+var judgeEvidenceSchema = external_exports.object({
+  kind: external_exports.enum(JUDGE_EVIDENCE_KINDS),
+  /**
+   * Position in the transcript the judge was shown: an index into
+   * `GradingTrace.steps` for `kind: 'message'`, or `GradingTrace.toolCalls`
+   * for `kind: 'tool_call'` — the same ordering the transcript renderer
+   * numbered when building the judge prompt.
+   */
+  index: external_exports.number().int().nonnegative()
+});
 var graderConfigSchema = external_exports.union([checkGraderSchema, aiGraderSchema]);
 var gradersSchema = external_exports.array(graderConfigSchema);
 var caseExpectedSchema = external_exports.object({
@@ -38068,6 +38221,8 @@ var caseExpectedSchema = external_exports.object({
   json: external_exports.unknown().optional(),
   facts: external_exports.array(external_exports.string()).optional()
 });
+var EVAL_CASE_ORIGINS = ["manual", "saved_from_run", "generated", "imported", "fpo"];
+var evalCaseOriginSchema = external_exports.enum(EVAL_CASE_ORIGINS);
 var DEFAULT_MODELS_FOR_NEW_ACCOUNTS = [
   // Mixlayer models
   {
@@ -38201,6 +38356,45 @@ function renderDefaultModelStrengthsGuide() {
     (m2) => `- \`${m2.modelId}\`${m2.isDefault ? " (account default)" : ""} \u2014 ${m2.strengths}`
   ).join("\n");
 }
+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.',
+  "",
+  "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.',
+  "2. Judge actions, not claims. If the output says a refund was issued but no refund tool call appears in the trajectory, the action did not happen.",
+  "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.",
+  "",
+  "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.",
+  "",
+  "Then return exactly one verdict:",
+  '- "pass" \u2014 the criterion is clearly met on the evidence shown.',
+  '- "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.',
+  "",
+  '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.',
+  "",
+  "Format examples only \u2014 never carry facts from these into a real case:",
+  '{"reasoning": "The agent called verify_identity before issue_refund, satisfying the criterion.", "verdict": "pass", "evidence": [{"kind": "tool_call", "index": 0}, {"kind": "tool_call", "index": 1}]}',
+  '{"reasoning": "The criterion forbids promising a delivery date, but the final output commits to delivery by Friday.", "verdict": "fail", "evidence": []}'
+].join("\n");
+var EVAL_JUDGE_USER_PROMPT_TEMPLATE = [
+  "Case: {{caseName}}",
+  "",
+  "Criterion (what this flow/agent is supposed to do):",
+  "{{criteria}}",
+  "",
+  'Expected answer (reference; the output should convey the same facts and conclusion, wording may differ; "(none provided)" when there is no reference):',
+  "{{expected}}",
+  "",
+  "Run transcript (trajectory and final output):",
+  "{{transcript}}"
+].join("\n");
 var userProfileFeaturesSchema = external_exports.object({
   enableAgentSkills: external_exports.boolean(),
   // Gates the WebMCP-powered dashboard assistant (copilot panel + chat
@@ -41371,6 +41565,7 @@ var surfaceSchema = external_exports.object({
     "messaging",
     "telegram",
     "a2a",
+    "ag-ui",
     "hosted-page",
     "chrome_extension"
   ]),
@@ -42873,7 +43068,8 @@ var PLATFORM_CATALOG = {
     "messaging",
     "telegram",
     "a2a",
-    "chrome_extension"
+    "chrome_extension",
+    "ag-ui"
   ],
   availableFlowPrimitives: FLOW_STEP_TYPES.map((stepType) => {
     const meta3 = FLOW_STEP_TYPE_METADATA[stepType];
@@ -42917,6 +43113,11 @@ var PLATFORM_CATALOG = {
     restBase: "/v1/skills"
   }
 };
+function publicSurfaceTypeMetadata() {
+  return Object.fromEntries(
+    Object.entries(SURFACE_TYPE_METADATA).filter(([, info]) => !info.hiddenFromCatalog)
+  );
+}
 var SURFACE_TYPE_METADATA = {
   chat: {
     description: "Website widget, support bots, conversational interfaces. Uses Persona and can optionally accept WebMCP page tools discovered from document.modelContext.",
@@ -43312,6 +43513,30 @@ var SURFACE_TYPE_METADATA = {
       executionHint: "You are running inside a Chrome extension side panel. You may have browser tools (chrome:*) to read the current page, interact with forms, and manage tabs \u2014 use them when the user asks about or wants to act on the page they are viewing. Mutating actions (clicking, filling, navigating) ask the user for confirmation before running."
     },
     behaviorTypeRef: "runtype://types/surface-configs"
+  },
+  "ag-ui": {
+    description: "AG-UI protocol endpoint (agent-to-user event stream over SSE)",
+    useCases: ["custom agent frontends", "CopilotKit / AG-UI clients", "embedded agent UIs"],
+    examples: ["CopilotKit app", "Custom React agent UI", "AG-UI dojo client"],
+    // Staging-test phase (`enable-ag-ui-surface` flag): keep the type out of
+    // agent-facing catalogs until GA. Runtime trait lookups still resolve.
+    hiddenFromCatalog: true,
+    traits: {
+      streaming: "required",
+      messagesMutable: false,
+      deliveryModel: "real_time",
+      mediaSupport: "none",
+      textMediaDelivery: "none",
+      interactiveUI: "none",
+      inboundMediaSupport: false,
+      consumerType: "machine",
+      reasoningVisibility: "pass_through",
+      markdownDialect: "markdown",
+      threadModel: "flat",
+      senderIdentity: "authenticated",
+      maxResponseLength: null,
+      executionHint: null
+    }
   }
 };
 var EMAIL_HTML_CAPABILITY_HINT = [
@@ -43320,7 +43545,7 @@ var EMAIL_HTML_CAPABILITY_HINT = [
   '- Images: embed an image with `<img src="https://..." alt="...">` and it renders inline in the email. (Some clients ask the recipient to load remote images.)'
 ].join("\n");
 var SURFACE_TYPE_GUIDE = (() => {
-  const lines = Object.entries(SURFACE_TYPE_METADATA).map(([type, info]) => {
+  const lines = Object.entries(publicSurfaceTypeMetadata()).map(([type, info]) => {
     return `- **${type}**: ${info.description}`;
   });
   return `## Surface Type Guide
@@ -49317,6 +49542,9 @@ logsCommand.command("list").description("Query persisted logs with filtering, se
           printJson(res);
           return;
         }
+        if (res.data?.degraded) {
+          console.error(chalk14.yellow("Warning: historical logs are temporarily unavailable; showing recent entries only"));
+        }
         const entries = res.data?.entries ?? [];
         if (entries.length === 0) {
           console.log(chalk14.gray("No logs found"));
@@ -49409,6 +49637,9 @@ logsCommand.command("stats").description("Aggregate log stats: totals, counts by
           return;
         }
         const stats = res.data;
+        if (stats?.degraded) {
+          console.error(chalk14.yellow("Warning: historical log storage is temporarily unavailable; counts cover recent entries only"));
+        }
         printDetail("Log Stats", [
           { label: "Total", value: stats.total },
           { label: "By Level", value: formatCounts(stats.byLevel ?? {}) },
@@ -69676,7 +69907,7 @@ function workerPackageJson(name, runtimeDep) {
         "@runtypelabs/runtime": runtimeDep
       },
       devDependencies: {
-        "@cloudflare/workers-types": "^4.0.0",
+        "@cloudflare/workers-types": "^5.0.0",
         typescript: "^5.3.3",
         wrangler: "^4.0.0"
       }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@runtypelabs/cli",
-  "version": "2.25.0",
+  "version": "2.25.2",
   "description": "Command-line interface for Runtype AI platform",
   "type": "module",
   "main": "dist/index.js",
@@ -24,7 +24,7 @@
     "rosie-skills": "0.8.1",
     "yaml": "^2.9.0",
     "@runtypelabs/ink-components": "0.3.4",
-    "@runtypelabs/sdk": "5.8.0",
+    "@runtypelabs/sdk": "5.9.0",
     "@runtypelabs/terminal-animations": "0.2.1"
   },
   "devDependencies": {
@@ -39,7 +39,7 @@
     "tsx": "^4.7.1",
     "typescript": "^6.0.3",
     "vitest": "^4.1.0",
-    "@runtypelabs/shared": "1.45.0"
+    "@runtypelabs/shared": "1.46.0"
   },
   "engines": {
     "node": ">=22.12.0"