@mcpjam/inspector 3.7.0 → 3.7.1

package/dist/server/index.js

@@ -35,7 +35,7 @@ import {
   supportsOwnershipProof,
   terminateOwnedProcess,
   terminateOwnedProcessGroup
-} from "./chunk-YCXJ5YSM.js";
+} from "./chunk-XTCUQXBF.js";
 import {
   ALLOWED_HOSTS,
   BROWSER_CONSENT_HEADER,
@@ -241,7 +241,7 @@ function classifyToolFailurePart(part) {
   if (part.type !== "tool-result") return null;
   const toolName3 = typeof part.toolName === "string" ? part.toolName : void 0;
   const toolCallId = typeof part.toolCallId === "string" ? part.toolCallId : void 0;
-  const record4 = (kind, message) => ({
+  const record5 = (kind, message) => ({
     kind,
     ...toolName3 ? { toolName: toolName3 } : {},
     ...toolCallId ? { toolCallId } : {},
@@ -253,20 +253,20 @@ function classifyToolFailurePart(part) {
     ...message && message.trim() ? { message: message.trim() } : {}
   });
   if (typeof part.error === "string" && part.error.trim())
-    return record4("protocol-error", part.error);
+    return record5("protocol-error", part.error);
   if (isRecord(part.error) && typeof part.error.message === "string" && part.error.message.trim()) {
-    return record4("protocol-error", part.error.message);
+    return record5("protocol-error", part.error.message);
   }
   const output = part.output;
   if (isRecord(output) && output.type === "error-text")
-    return record4("protocol-error", errorTextOf(output.value));
+    return record5("protocol-error", errorTextOf(output.value));
   if (isRecord(part.result) && part.result.isError === true)
-    return record4("content-error", callToolResultText(part.result));
+    return record5("content-error", callToolResultText(part.result));
   const unwrapped = unwrapToolOutput(output);
   if (isRecord(unwrapped) && unwrapped.isError === true)
-    return record4("content-error", callToolResultText(unwrapped));
+    return record5("content-error", callToolResultText(unwrapped));
   if (part.isError === true)
-    return record4("content-error", callToolResultText(part));
+    return record5("content-error", callToolResultText(part));
   return null;
 }
 function traceMessagePartIndicatesToolFailure(part) {
@@ -849,10 +849,10 @@ function write(value, ancestors) {
         "Cannot canonicalize a non-plain object (only plain objects, arrays and JSON primitives are part of the contract)"
       );
     }
-    const record4 = value;
+    const record5 = value;
     const parts = [];
-    for (const key of Object.keys(record4).sort()) {
-      const entry3 = record4[key];
+    for (const key of Object.keys(record5).sort()) {
+      const entry3 = record5[key];
       if (entry3 === void 0) continue;
       parts.push(`${JSON.stringify(key)}:${write(entry3, ancestors)}`);
     }
@@ -2243,18 +2243,6 @@ __export(evaluate_exports, {
   finalMessageEndsWithQuestion: () => finalMessageEndsWithQuestion
 });
 
-// ../evaluators/dist/chunk-I4CRERJV.js
-var PROBE = "__predicate_arg_probe__";
-function argMatch(matcher, actualArgs) {
-  const argumentMatching = matcher.argumentMatching ?? "partial";
-  const result = evaluateToolCalls(
-    [{ toolName: PROBE, arguments: matcher.args ?? {} }],
-    [{ toolName: PROBE, arguments: actualArgs ?? {} }],
-    { argumentMatching }
-  );
-  return result.passed;
-}
-
 // ../evaluators/dist/chunk-PYTZA5U2.js
 import { Validator } from "@cfworker/json-schema";
 var DRAFT_2020_12_URIS = /* @__PURE__ */ new Set([
@@ -2329,7 +2317,19 @@ function validateAgainstSchema(schema, instance) {
   }
 }
 
-// ../evaluators/dist/chunk-BIKNSDDD.js
+// ../evaluators/dist/chunk-I4CRERJV.js
+var PROBE = "__predicate_arg_probe__";
+function argMatch(matcher, actualArgs) {
+  const argumentMatching = matcher.argumentMatching ?? "partial";
+  const result = evaluateToolCalls(
+    [{ toolName: PROBE, arguments: matcher.args ?? {} }],
+    [{ toolName: PROBE, arguments: actualArgs ?? {} }],
+    { argumentMatching }
+  );
+  return result.passed;
+}
+
+// ../evaluators/dist/chunk-YNFR3VPW.js
 var MAX_VALUE_CHARS = 200;
 var MAX_ERROR_MSG_CHARS = 200;
 var MAX_ITEMS_SHOWN = 3;
@@ -3683,11 +3683,11 @@ function messagesOf(trace) {
 }
 function countUserTurns(messages) {
   if (!Array.isArray(messages)) return void 0;
-  let count = 0;
+  let count4 = 0;
   for (const message of messages) {
-    if (isRecord3(message) && message.role === "user") count += 1;
+    if (isRecord3(message) && message.role === "user") count4 += 1;
   }
-  return count;
+  return count4;
 }
 function buildTurnTranscript(input) {
   return {
@@ -3800,7 +3800,7 @@ __export(predicate_scorer_exports, {
   predicateScorer: () => predicateScorer
 });
 
-// ../evaluators/dist/chunk-DOG7TBJQ.js
+// ../evaluators/dist/chunk-LICJIDQY.js
 function assertLocallyEvaluable(predicate) {
   if (!requiresRenderObservations(predicate.type)) return;
   throw new Error(
@@ -3878,7 +3878,7 @@ function toEvaluatorRawOutcome(outcome) {
   };
 }
 
-// ../evaluators/dist/chunk-5CFTXLA2.js
+// ../evaluators/dist/chunk-K67IGSPF.js
 function assertLocallyEvaluable2(rule) {
   if (!requiresRenderObservations(rule.type)) return;
   throw new Error(
@@ -8377,7 +8377,7 @@ var init_HostRunner = __esm({
         this.mcpjamLeaseScope = config.mcpjamLeaseScope;
         this.mcpjamProject = config.mcpjamProject;
         this.baseUrls = config.baseUrls;
-        this.systemPrompt = config.systemPrompt ?? (this.hostSnapshot?.systemPrompt && this.hostSnapshot.systemPrompt !== "" ? this.hostSnapshot.systemPrompt : "You are a helpful assistant.");
+        this.systemPrompt = (config.systemPrompt ? config.systemPrompt : void 0) ?? (this.hostSnapshot?.systemPrompt && this.hostSnapshot.systemPrompt !== "" ? this.hostSnapshot.systemPrompt : "You are a helpful assistant.");
         this.temperature = config.temperature ?? this.hostSnapshot?.temperature;
         this.maxSteps = config.maxSteps ?? 10;
         this.customProviders = config.customProviders;
@@ -21155,8 +21155,8 @@ function errorCheck(detail) {
 function skippedCheck(detail) {
   return { status: "skipped", detail };
 }
-function describeCount(count, label) {
-  return `${count} ${label}${count === 1 ? "" : "s"} discovered.`;
+function describeCount(count22, label) {
+  return `${count22} ${label}${count22 === 1 ? "" : "s"} discovered.`;
 }
 function isServerDoctorError(error) {
   return !!error && typeof error === "object" && typeof error.code === "string" && typeof error.message === "string";
@@ -23037,8 +23037,8 @@ var EVAL_VERDICT_DECISION_REASONS = [
   ...EVAL_VALIDITY_DECISION_REASONS,
   ...EVAL_TASK_DECISION_REASONS
 ];
-var evalVerdictDecisionReasonSchema = z2.enum(
-  EVAL_VERDICT_DECISION_REASONS
+var evalVerdictDecisionReasonSchema = z2.enum(EVAL_VERDICT_DECISION_REASONS).describe(
+  "Why the verdict is what it is. The validity reasons are evaluated FIRST and make a run `inconclusive`; the task reasons decide `passed` / `failed` once validity holds. `configuredTrialsNotAttempted` \u2014 some configured trial never ran. `noGradeableTrials` \u2014 nothing in the suite was gradeable. `eligibleTrialsBelowMinimum` \u2014 the explicit `minEligibleTrials` was not reached. `completionRateBelowMinimum` \u2014 measured, and under the floor. `completionRateNotMeasured` \u2014 nothing was attempted, so the floor is unsatisfiable; a not-measured rate never passes one. `evaluatorErrorRateAboveMaximum` \u2014 the grader failed too often for the run to describe the server. `evaluatorErrorRateNotMeasured` \u2014 the same unsatisfiable case for the ceiling. `caseHasNoEligibleTrials` \u2014 a case graded nothing, which is inconclusive even at `passThreshold: 0`. `casePassRateMetThreshold` \u2014 a case's own passing reason. `casePassRateBelowThreshold` \u2014 a case failed its threshold, and so therefore did the suite. `allMeasuredCasesMetThreshold` \u2014 the suite's only passing reason."
 );
 function isEvalValidityDecisionReason(value) {
   return typeof value === "string" && EVAL_VALIDITY_DECISION_REASONS.includes(value);
@@ -25683,6 +25683,416 @@ var evalBacktestContinuationSchema = z2.object({
 evalBacktestDraftSchema.extend({
   continuation: evalBacktestContinuationSchema.optional()
 });
+var SWARM_SESSION_VERDICT_CONTRACT_VERSION = 1;
+var SWARM_ATTEMPT_STATUSES = [
+  "pending",
+  "running",
+  "succeeded",
+  "failed",
+  "rate_limited"
+];
+var SWARM_SESSION_LIFECYCLES = [
+  "pending",
+  "running",
+  "ran",
+  "broke",
+  "limited",
+  "withdrawn"
+];
+var SWARM_SESSION_VERDICTS = [
+  "passed",
+  "failed",
+  "inconclusive",
+  "notEstablished"
+];
+var SWARM_GRADING_STATES = [
+  "notRequested",
+  "queued",
+  "running",
+  "settled",
+  "unavailable"
+];
+var SWARM_SESSION_VERDICT_OF_REASON = {
+  attemptPending: "notEstablished",
+  attemptRunning: "notEstablished",
+  withdrawn: "notEstablished",
+  spendCapReached: "notEstablished",
+  notRun: "notEstablished",
+  executionFailed: "notEstablished",
+  ungraded: "notEstablished",
+  gradingNotClaimed: "notEstablished",
+  criteriaPending: "notEstablished",
+  judgePending: "notEstablished",
+  gatingCriterionFailed: "failed",
+  judgeFailed: "failed",
+  criteriaGradingErrored: "inconclusive",
+  gatingCriterionUnmeasured: "inconclusive",
+  judgeErrored: "inconclusive",
+  gradingUnavailable: "inconclusive",
+  allGatingCriteriaPassed: "passed",
+  judgePassed: "passed",
+  allGradersPassed: "passed"
+};
+var SWARM_SESSION_VERDICT_REASONS = Object.keys(
+  SWARM_SESSION_VERDICT_OF_REASON
+);
+var swarmSessionLifecycleSchema = z2.enum(SWARM_SESSION_LIFECYCLES);
+var swarmSessionVerdictValueSchema = z2.enum(SWARM_SESSION_VERDICTS);
+var swarmGraderRoleSchema = z2.enum(["advisory", "required"]);
+var swarmSessionAttemptInputSchema = z2.object({
+  status: z2.enum(SWARM_ATTEMPT_STATUSES),
+  errorCode: z2.string().nullable().optional()
+}).strict();
+var swarmGradingReadinessSchema = z2.object({
+  /** Durable readiness for decisive grading only, not background observations. */
+  state: z2.enum(SWARM_GRADING_STATES),
+  reasonCode: z2.string().min(1).optional()
+}).strict();
+var criterionResultSchema = z2.object({
+  criterionId: z2.string().min(1),
+  passed: z2.boolean(),
+  status: z2.enum(["scored", "error"]).optional()
+}).strict();
+z2.object({
+  attempt: swarmSessionAttemptInputSchema.nullable(),
+  hasTranscript: z2.boolean(),
+  rubric: z2.array(
+    z2.object({
+      id: z2.string().min(1),
+      role: swarmGraderRoleSchema,
+      predicateType: z2.string().min(1).optional()
+    }).strict()
+  ),
+  criteria: z2.object({
+    status: z2.enum(["pending", "completed", "failed"]),
+    criterionIds: z2.array(z2.string().min(1)).optional(),
+    results: z2.array(criterionResultSchema).optional()
+  }).strict().nullable(),
+  goalScore: z2.discriminatedUnion("status", [
+    z2.object({ status: z2.literal("completed"), passed: z2.boolean() }).strict(),
+    z2.object({ status: z2.literal("running") }).strict(),
+    z2.object({ status: z2.literal("failed") }).strict()
+  ]).nullable(),
+  judge: z2.object({
+    automatic: z2.boolean(),
+    role: swarmGraderRoleSchema,
+    /** Explicit on-demand intent; automatic:false alone does not mean a judge was requested. */
+    requested: z2.boolean().optional()
+  }).strict(),
+  grading: swarmGradingReadinessSchema
+}).strict().superRefine((value, ctx) => {
+  const definitions = new Set(value.rubric.map((entry22) => entry22.id));
+  if (definitions.size !== value.rubric.length) {
+    ctx.addIssue({
+      code: "custom",
+      path: ["rubric"],
+      message: "Duplicate criterion definition"
+    });
+  }
+  for (const [field2, ids] of [
+    ["criterionIds", value.criteria?.criterionIds ?? []],
+    ["results", value.criteria?.results?.map((row2) => row2.criterionId) ?? []]
+  ]) {
+    if (new Set(ids).size !== ids.length || ids.some((id3) => !definitions.has(id3))) {
+      ctx.addIssue({
+        code: "custom",
+        path: ["criteria", field2],
+        message: "Criterion IDs must be unique and defined in the snapshot"
+      });
+    }
+  }
+  const claimed = value.criteria?.criterionIds;
+  if (claimed && value.criteria?.results?.some((row2) => !claimed.includes(row2.criterionId))) {
+    ctx.addIssue({
+      code: "custom",
+      path: ["criteria", "results"],
+      message: "Result outside claimed scope"
+    });
+  }
+});
+var SWARM_SESSION_TRIAL_STATUSES = [
+  "pending",
+  "running",
+  "completed",
+  "failed",
+  "setup_failed",
+  "cancelled"
+];
+var swarmSessionTrialSchema = z2.object({
+  status: z2.enum(SWARM_SESSION_TRIAL_STATUSES),
+  taskVerdict: z2.enum(["passed", "failed"]).optional(),
+  evaluatorError: z2.literal(true).optional()
+}).strict().superRefine((value, ctx) => {
+  const verdict = value.taskVerdict !== void 0;
+  const error = value.evaluatorError === true;
+  if (value.status === "completed" ? verdict === error : verdict || error) {
+    ctx.addIssue({
+      code: "custom",
+      message: "Only completed trials carry exactly one verdict or evaluator error"
+    });
+  }
+});
+var swarmSessionGraderCountsSchema = z2.object({
+  gating: z2.number().int().nonnegative(),
+  gatingPassed: z2.number().int().nonnegative(),
+  gatingFailed: z2.number().int().nonnegative(),
+  gatingUnmeasured: z2.number().int().nonnegative(),
+  advisoryFailed: z2.number().int().nonnegative()
+}).strict();
+var swarmSessionVerdictSchema = z2.object({
+  contractVersion: z2.literal(SWARM_SESSION_VERDICT_CONTRACT_VERSION),
+  lifecycle: swarmSessionLifecycleSchema,
+  verdict: swarmSessionVerdictValueSchema,
+  reason: z2.enum(SWARM_SESSION_VERDICT_REASONS),
+  verdictSource: z2.enum([
+    "goalJudge",
+    "requiredAssertions",
+    "combined",
+    "none"
+  ]),
+  grading: swarmGradingReadinessSchema,
+  graders: z2.object({
+    criteria: z2.enum([
+      "notConfigured",
+      "notClaimed",
+      "pending",
+      "scored",
+      "errored"
+    ]),
+    judge: z2.enum([
+      "notConfigured",
+      "silent",
+      "pending",
+      "scored",
+      "errored"
+    ])
+  }).strict(),
+  counts: swarmSessionGraderCountsSchema,
+  trial: swarmSessionTrialSchema.nullable()
+}).strict().superRefine((value, ctx) => {
+  const fail3 = (message) => ctx.addIssue({ code: "custom", message });
+  if (SWARM_SESSION_VERDICT_OF_REASON[value.reason] !== value.verdict)
+    fail3("Reason contradicts verdict");
+  const c = value.counts;
+  if (c.gating !== c.gatingPassed + c.gatingFailed + c.gatingUnmeasured)
+    fail3("Required measurement counts do not add up");
+  const trial = value.trial;
+  if (trial?.status === "completed") {
+    if (value.lifecycle !== "ran")
+      fail3("Only a completed execution can be a completed trial");
+    if (trial.taskVerdict !== void 0 && trial.taskVerdict !== value.verdict)
+      fail3("Trial contradicts goal result");
+    if (trial.evaluatorError && value.verdict !== "inconclusive")
+      fail3("Evaluator error requires inconclusive grading");
+  }
+  const expected = {
+    pending: "pending",
+    running: "running",
+    broke: "failed",
+    limited: "setup_failed",
+    withdrawn: "cancelled"
+  };
+  if (value.lifecycle !== "ran" && trial?.status !== expected[value.lifecycle])
+    fail3("Trial must preserve execution lifecycle");
+  if (value.lifecycle === "ran" && trial !== null && trial.status !== "completed")
+    fail3("Completed execution must not become skipped or running");
+  if ((value.verdict === "passed" || value.verdict === "failed") && value.verdictSource === "none")
+    fail3("A measured goal verdict needs a source");
+  if (value.lifecycle === "ran" && value.verdict !== "notEstablished" && trial === null)
+    fail3("Settled grading on completed execution needs an observation");
+  if (value.verdict === "notEstablished" && value.verdictSource !== "none")
+    fail3("An undecided goal has no verdict source");
+});
+var SWARM_REPORT_CONTRACT_VERSION = 1;
+var count = z2.number().int().nonnegative();
+var journeyRunVerdictSummarySchema = z2.discriminatedUnion("status", [
+  z2.object({
+    status: z2.literal("pending"),
+    pendingSessions: count,
+    updatedAt: count
+  }).strict(),
+  z2.object({
+    status: z2.literal("decided"),
+    decision: evalVerdictDecisionSchema,
+    updatedAt: count
+  }).strict(),
+  z2.object({
+    status: z2.literal("notEstablished"),
+    reason: z2.literal("gradingNotConfigured"),
+    updatedAt: count
+  }).strict(),
+  z2.object({
+    status: z2.literal("integrityFailed"),
+    reason: z2.string().min(1),
+    updatedAt: count
+  }).strict()
+]);
+var swarmObservationInputSchema = z2.object({
+  evaluatorId: z2.string().min(1),
+  predicateType: z2.enum(
+    Object.keys(ASSERTION_STAGE)
+  ),
+  role: z2.enum(["advisory", "required"]),
+  /** Missing capture/error is unavailable; it must not use the compatibility passed:false. */
+  status: z2.enum(["passed", "failed", "pending", "unavailable"])
+}).strict();
+var swarmReportSessionSchema = z2.object({
+  /** Stable attempt/slot identity, including slots with no session document. */
+  id: z2.string().min(1),
+  /** A producer fact. Grade coverage and lifecycle alone cannot establish a non-start. */
+  startEvidence: z2.enum(["started", "notStarted", "unknown"]),
+  verdict: swarmSessionVerdictSchema,
+  observations: z2.array(swarmObservationInputSchema)
+}).strict().superRefine((row2, ctx) => {
+  if (new Set(row2.observations.map((o) => o.evaluatorId)).size !== row2.observations.length) {
+    ctx.addIssue({
+      code: "custom",
+      path: ["observations"],
+      message: "Duplicate evaluator in one session"
+    });
+  }
+  if (row2.startEvidence === "notStarted" && ["running", "ran"].includes(row2.verdict.lifecycle)) {
+    ctx.addIssue({
+      code: "custom",
+      path: ["startEvidence"],
+      message: "Execution contradicts explicit non-start"
+    });
+  }
+});
+z2.object({
+  runId: z2.string().min(1),
+  executionComplete: z2.boolean(),
+  configuredSessions: count,
+  verdictSummary: journeyRunVerdictSummarySchema.nullable(),
+  /** Frozen run-wide definitions: absent measurement rows still count as unavailable. */
+  evaluatorDefinitions: z2.array(
+    swarmObservationInputSchema.omit({ status: true })
+  ),
+  sessions: z2.array(swarmReportSessionSchema)
+}).strict().superRefine((input, ctx) => {
+  if (input.sessions.length > input.configuredSessions) {
+    ctx.addIssue({
+      code: "custom",
+      path: ["sessions"],
+      message: "Reported sessions exceed configured population"
+    });
+  }
+  if (new Set(input.sessions.map((row2) => row2.id)).size !== input.sessions.length) {
+    ctx.addIssue({
+      code: "custom",
+      path: ["sessions"],
+      message: "Duplicate session/attempt identity"
+    });
+  }
+  const definitions = new Map(
+    input.evaluatorDefinitions.map((row2) => [row2.evaluatorId, row2])
+  );
+  if (definitions.size !== input.evaluatorDefinitions.length) {
+    ctx.addIssue({
+      code: "custom",
+      path: ["evaluatorDefinitions"],
+      message: "Duplicate evaluator definition"
+    });
+  }
+  for (const session of input.sessions) {
+    for (const observation of session.observations) {
+      const definition = definitions.get(observation.evaluatorId);
+      if (!definition || definition.role !== observation.role || definition.predicateType !== observation.predicateType) {
+        ctx.addIssue({
+          code: "custom",
+          path: ["sessions"],
+          message: "Observation contradicts the frozen evaluator definition"
+        });
+      }
+    }
+  }
+});
+var swarmObservationCoverageSchema = z2.object({
+  evaluatorId: z2.string().min(1),
+  predicateType: z2.enum(
+    Object.keys(ASSERTION_STAGE)
+  ),
+  role: z2.enum(["advisory", "required"]),
+  stage: z2.enum(USER_VALUE_STAGES).describe(
+    "The stage this observation was read from, and the only stage it is evidence ABOUT. `connection` \u2014 the server was reachable and the session initialized. `discovery` \u2014 its tools and resources were listed and readable. `selection` \u2014 the model chose the right tool for the request. `call` \u2014 the call was made with usable arguments. `response` \u2014 the server returned data the model could use. `userValue` \u2014 the user's actual request was satisfied."
+  ),
+  unit: z2.literal("sessions"),
+  total: count,
+  passed: count,
+  failed: count,
+  pending: count,
+  unavailable: count
+}).strict().superRefine((value, ctx) => {
+  if (value.total !== value.passed + value.failed + value.pending + value.unavailable) {
+    ctx.addIssue({
+      code: "custom",
+      message: "Observation coverage must partition its population"
+    });
+  }
+  if (value.stage !== ASSERTION_STAGE[value.predicateType]) {
+    ctx.addIssue({
+      code: "custom",
+      message: "Observation stage contradicts the shared evaluator map"
+    });
+  }
+});
+z2.object({
+  contractVersion: z2.literal(SWARM_REPORT_CONTRACT_VERSION),
+  runId: z2.string().min(1),
+  verdict: swarmSessionVerdictValueSchema,
+  verdictSource: z2.enum(["policyV2", "none"]),
+  decision: evalVerdictDecisionSchema.optional(),
+  undecidedReason: z2.enum([
+    "executionPending",
+    "gradingPending",
+    "gradingNotConfigured",
+    "verdictSummaryUnavailable",
+    "integrityFailed"
+  ]).optional(),
+  execution: z2.object({
+    unit: z2.literal("sessions"),
+    configured: count,
+    reported: count,
+    started: count,
+    notStarted: count,
+    unknown: count,
+    completed: count,
+    interrupted: count,
+    /** Only explicit complete execution evidence can establish this. */
+    neverLaunched: z2.boolean()
+  }).strict(),
+  goalGrading: z2.object({
+    unit: z2.literal("sessions"),
+    reported: count,
+    passed: count,
+    failed: count,
+    pending: count,
+    unavailable: count,
+    notRequested: count,
+    /** Readiness is separate: a proven failure can exist while other required grading is pending. */
+    waitingForDecisiveGrading: count
+  }).strict(),
+  observations: z2.array(swarmObservationCoverageSchema)
+}).strict().superRefine((report, ctx) => {
+  const fail3 = (message) => ctx.addIssue({ code: "custom", message });
+  if (report.verdictSource === "policyV2") {
+    if (!report.decision || report.verdict !== report.decision.verdict || report.undecidedReason !== void 0)
+      fail3("Report must copy its authoritative decision");
+  } else if (report.verdict !== "notEstablished" || report.decision !== void 0 || !report.undecidedReason)
+    fail3("Missing authority must have an undecided reason");
+  const e = report.execution;
+  if (e.started + e.notStarted + e.unknown !== e.configured || e.reported > e.configured)
+    fail3("Execution evidence does not cover its configured population");
+  if (e.started + e.notStarted > e.reported || e.completed + e.interrupted > e.reported)
+    fail3("Execution counts exceed reported evidence");
+  if (e.neverLaunched !== (e.configured > 0 && e.reported === e.configured && e.notStarted === e.configured))
+    fail3("Not run requires explicit evidence for the entire population");
+  const g = report.goalGrading;
+  if (g.reported !== e.reported || g.passed + g.failed + g.pending + g.unavailable + g.notRequested !== g.reported || g.waitingForDecisiveGrading > g.reported)
+    fail3("Goal coverage must partition reported sessions");
+  if (new Set(report.observations.map((o) => o.evaluatorId)).size !== report.observations.length || report.observations.some((o) => o.total > e.reported))
+    fail3("Invalid observation population");
+});
 var COMPARATIVE_GATE_FIELDS = [
   "noDeterministicRegressions",
   "maximumP95LatencyIncreaseMs",
@@ -25717,7 +26127,7 @@ var noopPosthog = {
 var posthog = isTelemetryDisabled ? noopPosthog : new PostHog("phc_dTOPniyUNU2kD8Jx8yHMXSqiZHM8I91uWopTMX6EBE9", {
   host: "https://us.i.posthog.com"
 });
-var SDK_VERSION = "8.9.0";
+var SDK_VERSION = "8.10.0";
 var SDK_RELEASE = `@mcpjam/sdk@${SDK_VERSION}`;
 var CAPTURED_ERROR_SYMBOL = Symbol.for("@mcpjam/sdk/captured-eval-error");
 init_internal();
@@ -25728,7 +26138,7 @@ init_eval_tool_execution();
 init_internal();
 init_internal();
 function readSdkVersion() {
-  return "8.9.0";
+  return "8.10.0";
 }
 var DEFAULT_PLATFORM_USER_AGENT = `mcpjam-sdk/${readSdkVersion()}`;
 init_HostRunner();
@@ -40733,7 +41143,7 @@ async function runTransportChecks(ctx, selectedCheckIds) {
           const readErrors = sseResponses.map(
             ({ response }) => response.bodyError
           );
-          const allStreamsReadable = requestErrors.every((error) => error === void 0) && !responseFailures && readErrors.every((error) => error === void 0) && eventCounts.every((count) => count > 0);
+          const allStreamsReadable = requestErrors.every((error) => error === void 0) && !responseFailures && readErrors.every((error) => error === void 0) && eventCounts.every((count22) => count22 > 0);
           results.push(
             allStreamsReadable ? passedResult2(
               TRANSPORT_CHECK_METADATA["server-sse-streams-functional"],
@@ -74624,11 +75034,11 @@ function canonicalizeBooleanLeaves2(value, keys, fieldName) {
   if (!isPlainObject32(value)) {
     throw new Error(`hostConfigV2: ${fieldName} must be a plain object`);
   }
-  const record4 = value;
-  assertOnlyKnownKeys2(record4, new Set(keys), fieldName);
+  const record5 = value;
+  assertOnlyKnownKeys2(record5, new Set(keys), fieldName);
   const out = {};
   for (const key of keys) {
-    const bool = readOptionalBoolean2(record4, key, fieldName);
+    const bool = readOptionalBoolean2(record5, key, fieldName);
     if (bool !== void 0) out[key] = bool;
   }
   return Object.keys(out).length > 0 ? out : void 0;
@@ -74641,10 +75051,10 @@ function canonicalizeResourceVisibility2(value, fieldName) {
   if (!isPlainObject32(value)) {
     throw new Error(`hostConfigV2: ${fieldName} must be a plain object`);
   }
-  const record4 = value;
-  assertOnlyKnownKeys2(record4, new Set(RESOURCE_VISIBILITY_KEYS2), fieldName);
-  const text2 = readOptionalBoolean2(record4, "text", fieldName);
-  const blob = canonicalizeBlobVisibility2(record4.blob, `${fieldName}.blob`);
+  const record5 = value;
+  assertOnlyKnownKeys2(record5, new Set(RESOURCE_VISIBILITY_KEYS2), fieldName);
+  const text2 = readOptionalBoolean2(record5, "text", fieldName);
+  const blob = canonicalizeBlobVisibility2(record5.blob, `${fieldName}.blob`);
   const out = {
     ...text2 !== void 0 ? { text: text2 } : {},
     ...blob !== void 0 ? { blob } : {}
@@ -74658,23 +75068,23 @@ function canonicalizeModelVisibleMcpToolResults2(value) {
       "hostConfigV2: modelVisibleMcpToolResults must be a plain object"
     );
   }
-  const record4 = value;
+  const record5 = value;
   assertOnlyKnownKeys2(
-    record4,
+    record5,
     /* @__PURE__ */ new Set(["directContent", "embeddedResources", "linkedResources"]),
     "modelVisibleMcpToolResults"
   );
   const directContent = canonicalizeBooleanLeaves2(
-    record4.directContent,
+    record5.directContent,
     DIRECT_CONTENT_VISIBILITY_KEYS2,
     "modelVisibleMcpToolResults.directContent"
   );
   const embeddedResources = canonicalizeResourceVisibility2(
-    record4.embeddedResources,
+    record5.embeddedResources,
     "modelVisibleMcpToolResults.embeddedResources"
   );
   const linkedResources = canonicalizeResourceVisibility2(
-    record4.linkedResources,
+    record5.linkedResources,
     "modelVisibleMcpToolResults.linkedResources"
   );
   const out = {
@@ -74699,14 +75109,14 @@ function canonicalizeImageRenderingResource2(value, fieldName) {
   if (!isPlainObject32(value)) {
     throw new Error(`hostConfigV2: ${fieldName} must be a plain object`);
   }
-  const record4 = value;
+  const record5 = value;
   assertOnlyKnownKeys2(
-    record4,
+    record5,
     new Set(MCP_TOOL_RESULT_IMAGE_RENDERING_RESOURCE_KEYS2),
     fieldName
   );
   const blob = canonicalizeBooleanLeaves2(
-    record4.blob,
+    record5.blob,
     MCP_TOOL_RESULT_IMAGE_RENDERING_BLOB_KEYS2,
     `${fieldName}.blob`
   );
@@ -74722,28 +75132,28 @@ function canonicalizeMcpToolResultImageRenderingPolicy2(value) {
       "hostConfigV2: mcpToolResultImageRendering must be a plain object"
     );
   }
-  const record4 = value;
+  const record5 = value;
   assertOnlyKnownKeys2(
-    record4,
+    record5,
     new Set(MCP_TOOL_RESULT_IMAGE_RENDERING_KEYS2),
     "mcpToolResultImageRendering"
   );
   const placement = readOptionalMcpToolResultImageRenderPlacement2(
-    record4,
+    record5,
     "placement",
     "mcpToolResultImageRendering"
   );
   const directContent = canonicalizeBooleanLeaves2(
-    record4.directContent,
+    record5.directContent,
     MCP_TOOL_RESULT_IMAGE_RENDERING_DIRECT_CONTENT_KEYS2,
     "mcpToolResultImageRendering.directContent"
   );
   const embeddedResources = canonicalizeImageRenderingResource2(
-    record4.embeddedResources,
+    record5.embeddedResources,
     "mcpToolResultImageRendering.embeddedResources"
   );
   const linkedResources = canonicalizeImageRenderingResource2(
-    record4.linkedResources,
+    record5.linkedResources,
     "mcpToolResultImageRendering.linkedResources"
   );
   const out = {
@@ -74868,11 +75278,11 @@ function canonicalizeSkillSelection2(value) {
   if (!isPlainObject32(value)) {
     throw new Error("hostConfigV2: skillSelection must be a plain object");
   }
-  const record4 = value;
-  const mode = record4.mode;
+  const record5 = value;
+  const mode = record5.mode;
   if (mode === "all-visible") {
     assertOnlyKnownKeys2(
-      record4,
+      record5,
       SKILL_SELECTION_ALL_VISIBLE_KEYS2,
       "skillSelection"
     );
@@ -74880,11 +75290,11 @@ function canonicalizeSkillSelection2(value) {
   }
   if (mode === "explicit") {
     assertOnlyKnownKeys2(
-      record4,
+      record5,
       SKILL_SELECTION_EXPLICIT_KEYS2,
       "skillSelection"
     );
-    const skillIds = record4.skillIds;
+    const skillIds = record5.skillIds;
     if (!Array.isArray(skillIds)) {
       throw new Error(
         'hostConfigV2: skillSelection.skillIds must be a string[] when mode is "explicit"'
@@ -77605,8 +78015,8 @@ var init_result_guards2 = __esm2({
 });
 function ensureJsonSchemaObject2(schema) {
   if (schema && typeof schema === "object") {
-    const record4 = schema;
-    const base = record4.jsonSchema ? ensureJsonSchemaObject2(record4.jsonSchema) : record4;
+    const record5 = schema;
+    const base = record5.jsonSchema ? ensureJsonSchemaObject2(record5.jsonSchema) : record5;
     if (!("type" in base) || base.type === void 0) {
       base.type = "object";
     }
@@ -78211,7 +78621,7 @@ var init_HostRunner2 = __esm2({
         this.mcpjamLeaseScope = config.mcpjamLeaseScope;
         this.mcpjamProject = config.mcpjamProject;
         this.baseUrls = config.baseUrls;
-        this.systemPrompt = config.systemPrompt ?? (this.hostSnapshot?.systemPrompt && this.hostSnapshot.systemPrompt !== "" ? this.hostSnapshot.systemPrompt : "You are a helpful assistant.");
+        this.systemPrompt = (config.systemPrompt ? config.systemPrompt : void 0) ?? (this.hostSnapshot?.systemPrompt && this.hostSnapshot.systemPrompt !== "" ? this.hostSnapshot.systemPrompt : "You are a helpful assistant.");
         this.temperature = config.temperature ?? this.hostSnapshot?.temperature;
         this.maxSteps = config.maxSteps ?? 10;
         this.customProviders = config.customProviders;
@@ -82457,7 +82867,7 @@ var NEGATIVE_TEST_MODE_DETAILS = {
   }
 };
 function readSdkVersion2() {
-  return "8.9.0";
+  return "8.10.0";
 }
 var CONFORMANCE_CHECKER_VERSION2 = readSdkVersion2();
 var MUST_POINTS2 = 95;
@@ -88699,8 +89109,8 @@ var CommandQueue = class {
   }
   /** Whether every per-tab FIFO is drained. Used for profile snapshots. */
   isIdle() {
-    for (const count of this.depth.values()) {
-      if (count > 0) return false;
+    for (const count4 of this.depth.values()) {
+      if (count4 > 0) return false;
     }
     return true;
   }
@@ -89175,7 +89585,7 @@ function redactAction(action, options = {}) {
       return { kind: action.kind };
     case "act": {
       const target = action.target;
-      const record4 = {
+      const record5 = {
         kind: "act",
         verb: action.verb,
         ...target ? {
@@ -89183,16 +89593,16 @@ function redactAction(action, options = {}) {
         } : {}
       };
       if (typeof action.value === "string") {
-        Object.assign(record4, redactTypedValue(action.value, action.verb, options));
+        Object.assign(record5, redactTypedValue(action.value, action.verb, options));
       }
       if (action.fields?.length) {
-        record4.fields = action.fields.map((field) => ({
+        record5.fields = action.fields.map((field) => ({
           ...field.selector ? { selector: field.selector } : {},
           ...redactTypedValue(field.value, "type", options)
         }));
       }
-      if (action.submit !== void 0) record4.submit = action.submit;
-      return record4;
+      if (action.submit !== void 0) record5.submit = action.submit;
+      return record5;
     }
     case "observe":
       return { kind: "observe", mode: action.mode };
@@ -91137,26 +91547,26 @@ var FRAME_STREAM_MAX_PAYLOAD_BY_KIND = {
   [5]: 512 * 1024
   // video_delta
 };
-function encodeFrameStreamRecord(record4) {
-  const video = isVideoRecord(record4);
-  const payload = record4.kind === FRAME_STREAM_KIND.frame ? record4.jpeg : video ? record4.au : record4.kind === FRAME_STREAM_KIND.end ? new TextEncoder().encode(record4.reason) : record4.stats ? new TextEncoder().encode(JSON.stringify(record4.stats)) : new Uint8Array(0);
+function encodeFrameStreamRecord(record5) {
+  const video = isVideoRecord(record5);
+  const payload = record5.kind === FRAME_STREAM_KIND.frame ? record5.jpeg : video ? record5.au : record5.kind === FRAME_STREAM_KIND.end ? new TextEncoder().encode(record5.reason) : record5.stats ? new TextEncoder().encode(JSON.stringify(record5.stats)) : new Uint8Array(0);
   const bytes2 = new Uint8Array(FRAME_STREAM_HEADER_BYTES + payload.byteLength);
   const view = new DataView(bytes2.buffer);
   view.setUint8(0, video ? FRAME_STREAM_VERSION_VIDEO : FRAME_STREAM_VERSION);
-  view.setUint8(1, record4.kind);
-  if (record4.kind === FRAME_STREAM_KIND.frame || video) {
-    view.setUint16(2, clampU16(record4.deviceWidth), true);
-    view.setUint16(4, clampU16(record4.deviceHeight), true);
-    view.setUint16(6, clampU16(Math.round(record4.scale * 1e3)), true);
-    view.setFloat64(8, record4.ts, true);
-    view.setUint32(16, record4.seq >>> 0, true);
+  view.setUint8(1, record5.kind);
+  if (record5.kind === FRAME_STREAM_KIND.frame || video) {
+    view.setUint16(2, clampU16(record5.deviceWidth), true);
+    view.setUint16(4, clampU16(record5.deviceHeight), true);
+    view.setUint16(6, clampU16(Math.round(record5.scale * 1e3)), true);
+    view.setFloat64(8, record5.ts, true);
+    view.setUint32(16, record5.seq >>> 0, true);
   }
   view.setUint32(20, payload.byteLength, true);
   bytes2.set(payload, FRAME_STREAM_HEADER_BYTES);
   return bytes2;
 }
-function isVideoRecord(record4) {
-  return record4.kind === FRAME_STREAM_KIND.video_key || record4.kind === FRAME_STREAM_KIND.video_delta;
+function isVideoRecord(record5) {
+  return record5.kind === FRAME_STREAM_KIND.video_key || record5.kind === FRAME_STREAM_KIND.video_delta;
 }
 function clampU16(value) {
   if (!Number.isFinite(value) || value <= 0) return 0;
@@ -91272,17 +91682,17 @@ function createFramePacer(sink, onDrop) {
     });
   };
   return {
-    push(bytes2, record4 = {}) {
+    push(bytes2, record5 = {}) {
       if (closed) return;
       if (inFlight3) {
         if (pending !== void 0) {
-          if (pending.record.essential && !record4.essential) {
-            if (record4.counts !== false) onDrop?.();
+          if (pending.record.essential && !record5.essential) {
+            if (record5.counts !== false) onDrop?.();
             return;
           }
           if (pending.record.counts !== false) onDrop?.();
         }
-        pending = { bytes: bytes2, record: record4 };
+        pending = { bytes: bytes2, record: record5 };
         return;
       }
       ship(bytes2);
@@ -91818,17 +92228,17 @@ function withSecretScrub(registry3, executor, documentKeyFor) {
     const result = await executor(command2, context);
     const scrubber = registry3.scrubber();
     const output = result.output;
-    const record4 = typeof output === "object" && output !== null ? output : void 0;
+    const record5 = typeof output === "object" && output !== null ? output : void 0;
     let suppress = false;
-    if (record4?.screenshot !== void 0 && registry3.hasExposure()) {
+    if (record5?.screenshot !== void 0 && registry3.hasExposure()) {
       const tabId = command2.tabId ?? result.stateToken?.tabId;
       const keyAfter = await readKey(tabId);
       suppress = !keyAfter || registry3.exposedAt(keyAfter) || typedBefore && (!keyBefore || registry3.exposedAt(keyBefore));
     }
     if (!scrubber && !suppress) return result;
     let scrubbedOutput = output;
-    if (record4 !== void 0) {
-      const { screenshot, ...rest } = record4;
+    if (record5 !== void 0) {
+      const { screenshot, ...rest } = record5;
       if (suppress) {
         scrubbedOutput = {
           ...scrubber ? scrubber.scrubDeep(rest) : rest,
@@ -93710,14 +94120,14 @@ function assignRefs(root) {
   const entries = /* @__PURE__ */ new Map();
   if (!root) return entries;
   const seen = /* @__PURE__ */ new Map();
-  const count = (node) => {
+  const count4 = (node) => {
     if (isRefWorthy(node)) {
       const key = `${node.role}:${node.name ?? ""}`;
       seen.set(key, (seen.get(key) ?? 0) + 1);
     }
-    for (const child of node.children ?? []) count(child);
+    for (const child of node.children ?? []) count4(child);
   };
-  count(root);
+  count4(root);
   const position = /* @__PURE__ */ new Map();
   let next = 1;
   const visit2 = (node) => {
@@ -93995,7 +94405,7 @@ var WebMcpBridge = class {
     }
     return false;
   }
-  allowChanges(count) {
+  allowChanges(count4) {
     if (!this.localSecurity) return true;
     if (this.discoveryBlocked) return false;
     const now = Date.now();
@@ -94005,8 +94415,8 @@ var WebMcpBridge = class {
       rate.tokens + Math.max(0, now - rate.at) * 1.024
     );
     rate.at = now;
-    if (count > rate.tokens) return this.blockDiscovery();
-    rate.tokens -= count;
+    if (count4 > rate.tokens) return this.blockDiscovery();
+    rate.tokens -= count4;
     this.changeTokens = rate.tokens;
     this.changeAt = now;
     return true;
@@ -94742,9 +95152,9 @@ function canonicalJson3(value, depth = 0, budget = { left: CANONICAL_JSON_BUDGET
     budget.left -= parts2.length + 1;
     return `[${parts2.join(",")}]`;
   }
-  const record4 = value;
+  const record5 = value;
   const parts = [];
-  for (const key of Object.keys(record4).sort()) {
+  for (const key of Object.keys(record5).sort()) {
     if (budget.left <= 0) {
       parts.push('"[budget]":0');
       break;
@@ -94753,7 +95163,7 @@ function canonicalJson3(value, depth = 0, budget = { left: CANONICAL_JSON_BUDGET
       key.length > budget.left ? `${key.slice(0, budget.left)}\u2026` : key
     );
     budget.left -= encodedKey.length + 1;
-    parts.push(`${encodedKey}:${canonicalJson3(record4[key], depth + 1, budget)}`);
+    parts.push(`${encodedKey}:${canonicalJson3(record5[key], depth + 1, budget)}`);
   }
   budget.left -= parts.length + 1;
   return `{${parts.join(",")}}`;
@@ -102003,8 +102413,8 @@ function createFrameRelayStats(options) {
     countDrop(n = 1) {
       dropped += Math.max(0, Math.round(n));
     },
-    setSubscribers(count) {
-      subscribers3 = Math.max(0, Math.round(count));
+    setSubscribers(count4) {
+      subscribers3 = Math.max(0, Math.round(count4));
     },
     mergeDaemon(counters) {
       daemon = counters;
@@ -107596,6 +108006,14 @@ async function fetchRuntimeServerSecrets(args) {
     );
   }
   const RUNTIME_REVEAL_TIMEOUT_MS = 1e4;
+  const scenarioServiceToken = args.scenarioId ? process.env.INSPECTOR_SERVICE_TOKEN : void 0;
+  if (args.scenarioId && !scenarioServiceToken) {
+    throw new WebRouteError(
+      500,
+      ErrorCode.INTERNAL_ERROR,
+      "Server missing INSPECTOR_SERVICE_TOKEN for scenario secret delivery"
+    );
+  }
   const controller = new AbortController();
   const timeoutId = setTimeout(
     () => controller.abort(),
@@ -107604,7 +108022,8 @@ async function fetchRuntimeServerSecrets(args) {
   let response;
   try {
     const headers = {
-      "Content-Type": "application/json"
+      "Content-Type": "application/json",
+      ...scenarioServiceToken ? { "x-inspector-service-token": scenarioServiceToken } : {}
     };
     if (args.workosApiKeyActingAs) {
       const serviceToken3 = process.env.INSPECTOR_SERVICE_TOKEN;
@@ -109074,9 +109493,9 @@ function sanitizeValue(value, keyPath, depth, args, state) {
   }
   return DROP;
 }
-function sanitizeUnknownRecord(record4, args) {
+function sanitizeUnknownRecord(record5, args) {
   const state = { reportedTooDeep: false };
-  return sanitizeValue(record4, "", 0, args, state);
+  return sanitizeValue(record5, "", 0, args, state);
 }
 var APP_ID_KEYS = ["app_id", "appId", "id"];
 var SERVER_KEYS = ["server", "mcp_server", "mcpServer", "serverKey"];
@@ -109092,11 +109511,11 @@ function parsePluginAppConfig(args) {
     );
     return null;
   }
-  const record4 = raw;
+  const record5 = raw;
   let appId;
   let appIdKey;
   for (const key of APP_ID_KEYS) {
-    const value = record4[key];
+    const value = record5[key];
     if (typeof value === "string" && value.length > 0) {
       appId = value;
       appIdKey = key;
@@ -109114,7 +109533,7 @@ function parsePluginAppConfig(args) {
   let declaredServer;
   let serverField;
   for (const key of SERVER_KEYS) {
-    const value = record4[key];
+    const value = record5[key];
     if (typeof value === "string" && value.length > 0) {
       declaredServer = value;
       serverField = key;
@@ -109122,7 +109541,7 @@ function parsePluginAppConfig(args) {
     }
   }
   const unknownFields = {};
-  for (const [key, value] of Object.entries(record4)) {
+  for (const [key, value] of Object.entries(record5)) {
     if (key === appIdKey || key === serverField) continue;
     unknownFields[key] = value;
   }
@@ -109661,8 +110080,8 @@ function applyMcpjamExtension(manifest, rawExtensions, filePaths, issues) {
     );
     return;
   }
-  const record4 = raw;
-  const displayName = record4.displayName ?? record4.display_name;
+  const record5 = raw;
+  const displayName = record5.displayName ?? record5.display_name;
   if (displayName !== void 0) {
     if (typeof displayName !== "string") {
       issues.warn(
@@ -109674,7 +110093,7 @@ function applyMcpjamExtension(manifest, rawExtensions, filePaths, issues) {
     }
   }
   for (const key of ["icon", "logo"]) {
-    const value = record4[key];
+    const value = record5[key];
     if (value === void 0) continue;
     if (typeof value !== "string") {
       issues.warn(
@@ -109708,8 +110127,8 @@ function normalizePluginManifest(raw, context) {
     );
     return { manifest: null };
   }
-  const record4 = raw;
-  const schema = record4.$schema;
+  const record5 = raw;
+  const schema = record5.$schema;
   const schemaVersion = typeof schema === "string" ? Object.keys(PLUGIN_MANIFEST_SCHEMAS).find(
     (version) => PLUGIN_MANIFEST_SCHEMAS[version] === schema
   ) : void 0;
@@ -109722,7 +110141,7 @@ function normalizePluginManifest(raw, context) {
     );
     return { manifest: null };
   }
-  const name15 = record4.name;
+  const name15 = record5.name;
   if (typeof name15 !== "string" || name15.length === 0 || name15.length > MAX_NAME_LENGTH || !AP_NAME.test(name15)) {
     issues.error(
       "MANIFEST_INVALID_NAME",
@@ -109735,18 +110154,18 @@ function normalizePluginManifest(raw, context) {
     name: name15,
     extensions: {}
   };
-  if (record4.version !== void 0) {
-    if (typeof record4.version !== "string" || record4.version.length === 0) {
+  if (record5.version !== void 0) {
+    if (typeof record5.version !== "string" || record5.version.length === 0) {
       issues.error(
         "MANIFEST_INVALID_VERSION",
         `manifest "version" must be a non-empty string when present`
       );
     } else {
-      manifest.version = record4.version;
+      manifest.version = record5.version;
     }
   }
   for (const key of STRING_FIELDS) {
-    const value = record4[key];
+    const value = record5[key];
     if (value === void 0) continue;
     if (typeof value !== "string") {
       issues.error(
@@ -109758,11 +110177,11 @@ function normalizePluginManifest(raw, context) {
     manifest[key] = value;
   }
   for (const key of URL_FIELDS) {
-    if (record4[key] === void 0) continue;
-    const url = readHttpsUrl(key, record4[key], issues);
+    if (record5[key] === void 0) continue;
+    const url = readHttpsUrl(key, record5[key], issues);
     if (url !== void 0) manifest[key] = url;
   }
-  const author = record4.author;
+  const author = record5.author;
   if (author !== void 0) {
     if (author && typeof author === "object" && !Array.isArray(author)) {
       const authorRecord = author;
@@ -109785,7 +110204,7 @@ function normalizePluginManifest(raw, context) {
       );
     }
   }
-  const keywords = record4.keywords;
+  const keywords = record5.keywords;
   if (keywords !== void 0) {
     if (!Array.isArray(keywords) || keywords.some((keyword) => typeof keyword !== "string")) {
       issues.error(
@@ -109796,7 +110215,7 @@ function normalizePluginManifest(raw, context) {
       manifest.keywords = keywords;
     }
   }
-  const extensions = record4.extensions;
+  const extensions = record5.extensions;
   if (extensions !== void 0) {
     if (extensions === null || typeof extensions !== "object" || Array.isArray(extensions)) {
       issues.warn(
@@ -109820,7 +110239,7 @@ function normalizePluginManifest(raw, context) {
       );
     }
   }
-  for (const key of Object.keys(record4)) {
+  for (const key of Object.keys(record5)) {
     if (HANDLED_FIELDS.has(key)) continue;
     if (EXECUTION_AMBIGUOUS_FIELDS.has(key)) {
       issues.error(
@@ -110010,8 +110429,8 @@ function normalizeOAuthHint(serverKey, componentKey, raw, issues) {
   }
   return hint;
 }
-function readDeclaredTransport(serverKey, componentKey, record4, issues) {
-  const declared = record4.type;
+function readDeclaredTransport(serverKey, componentKey, record5, issues) {
+  const declared = record5.type;
   if (declared === "stdio" || declared === "streamable-http" || declared === "sse") {
     return declared;
   }
@@ -110032,16 +110451,16 @@ function normalizeServer(serverKey, sourcePath, raw, issues) {
     );
     return null;
   }
-  const record4 = raw;
+  const record5 = raw;
   const declared = readDeclaredTransport(
     serverKey,
     componentKey,
-    record4,
+    record5,
     issues
   );
   if (declared === null) return null;
   const knownFields = declared === "stdio" ? STDIO_KNOWN_FIELDS : HTTP_KNOWN_FIELDS;
-  for (const key of Object.keys(record4)) {
+  for (const key of Object.keys(record5)) {
     if (!knownFields.has(key)) {
       issues.error(
         "MCP_UNKNOWN_FIELD",
@@ -110052,7 +110471,7 @@ function normalizeServer(serverKey, sourcePath, raw, issues) {
     }
   }
   if (declared === "stdio") {
-    const declaredCommand = record4.command;
+    const declaredCommand = record5.command;
     if (typeof declaredCommand !== "string" || declaredCommand.length === 0) {
       issues.error(
         "MCP_MISSING_COMMAND",
@@ -110081,8 +110500,8 @@ function normalizeServer(serverKey, sourcePath, raw, issues) {
       command2 = `${PLUGIN_ROOT_PLACEHOLDER}/${resolved.path}`;
     }
     let args = [];
-    if (record4.args !== void 0) {
-      if (!Array.isArray(record4.args) || record4.args.some((arg) => typeof arg !== "string")) {
+    if (record5.args !== void 0) {
+      if (!Array.isArray(record5.args) || record5.args.some((arg) => typeof arg !== "string")) {
         issues.error(
           "MCP_INVALID_SERVER",
           `server "${serverKey}": "args" must be an array of strings`,
@@ -110090,18 +110509,18 @@ function normalizeServer(serverKey, sourcePath, raw, issues) {
         );
         return null;
       }
-      args = record4.args;
+      args = record5.args;
     }
     const envRequirements = normalizeEnv(
       serverKey,
       componentKey,
-      record4.env,
+      record5.env,
       issues
     );
     if (envRequirements === null) return null;
     let workingDirectory;
-    if (record4.cwd !== void 0) {
-      const cwd = record4.cwd;
+    if (record5.cwd !== void 0) {
+      const cwd = record5.cwd;
       if (typeof cwd !== "string") {
         issues.error(
           "MCP_INVALID_SERVER",
@@ -110150,7 +110569,7 @@ function normalizeServer(serverKey, sourcePath, raw, issues) {
       }
     };
   }
-  const url = record4.url;
+  const url = record5.url;
   if (typeof url !== "string" || url.length === 0) {
     issues.error(
       "MCP_MISSING_URL",
@@ -110190,12 +110609,12 @@ function normalizeServer(serverKey, sourcePath, raw, issues) {
   const headerRequirements = normalizeHeaders4(
     serverKey,
     componentKey,
-    record4.headers,
+    record5.headers,
     issues
   );
   if (headerRequirements === null) return null;
-  const hasOAuthHint = record4.oauth !== void 0 || record4.authentication !== void 0;
-  const oauth3 = hasOAuthHint ? normalizeOAuthHint(serverKey, componentKey, record4, issues) : void 0;
+  const hasOAuthHint = record5.oauth !== void 0 || record5.authentication !== void 0;
+  const oauth3 = hasOAuthHint ? normalizeOAuthHint(serverKey, componentKey, record5, issues) : void 0;
   return {
     componentKey,
     key: serverKey,
@@ -110227,8 +110646,8 @@ function normalizePluginMcpConfig(raw, context) {
   if (raw === null || typeof raw !== "object" || Array.isArray(raw)) {
     return disabled("MCP_INVALID_CONFIG", "mcp.json must be a JSON object");
   }
-  const record4 = raw;
-  const schema = record4.$schema;
+  const record5 = raw;
+  const schema = record5.$schema;
   const documentVersion = typeof schema === "string" ? Object.keys(PLUGIN_MCP_SCHEMAS).find(
     (version) => PLUGIN_MCP_SCHEMAS[version] === schema
   ) : void 0;
@@ -110244,7 +110663,7 @@ function normalizePluginMcpConfig(raw, context) {
       `mcp.json targets Agent Plugins ${documentVersion} but plugin.json targets ${manifestSchemaVersion}; versions must match`
     );
   }
-  for (const key of Object.keys(record4)) {
+  for (const key of Object.keys(record5)) {
     if (key !== "$schema" && key !== "mcpServers") {
       return disabled(
         "MCP_INVALID_CONFIG",
@@ -110252,7 +110671,7 @@ function normalizePluginMcpConfig(raw, context) {
       );
     }
   }
-  const serverMap = record4.mcpServers;
+  const serverMap = record5.mcpServers;
   if (serverMap === null || serverMap === void 0 || typeof serverMap !== "object" || Array.isArray(serverMap)) {
     return disabled(
       "MCP_INVALID_CONFIG",
@@ -112477,8 +112896,8 @@ function redactServerConfig(config) {
   if (!config || typeof config !== "object") {
     return config;
   }
-  const record4 = config;
-  const redacted = { ...record4 };
+  const record5 = config;
+  const redacted = { ...record5 };
   if ("env" in redacted) {
     redacted.hasEnv = !!redacted.env && typeof redacted.env === "object" && Object.keys(redacted.env).length > 0;
     delete redacted.env;
@@ -114285,12 +114704,12 @@ function pageToolFrom(value) {
   return tool19;
 }
 function pageToolsFromObservation(output) {
-  const record4 = isRecord15(output) ? output : {};
-  const tools5 = Array.isArray(record4.tools) ? record4.tools.map(pageToolFrom).filter((tool19) => tool19 !== null) : [];
+  const record5 = isRecord15(output) ? output : {};
+  const tools5 = Array.isArray(record5.tools) ? record5.tools.map(pageToolFrom).filter((tool19) => tool19 !== null) : [];
   return {
     ok: true,
-    url: typeof record4.url === "string" ? record4.url : "",
-    webmcpSupported: typeof record4.webmcpSupported === "boolean" ? record4.webmcpSupported : tools5.length > 0,
+    url: typeof record5.url === "string" ? record5.url : "",
+    webmcpSupported: typeof record5.webmcpSupported === "boolean" ? record5.webmcpSupported : tools5.length > 0,
     tools: tools5
   };
 }
@@ -114770,11 +115189,11 @@ function bindingRecordedFor(messages, toolCallId) {
     if (role !== "assistant" || !Array.isArray(content)) continue;
     for (const part of content) {
       if (!part || typeof part !== "object") continue;
-      const record4 = part;
-      if (record4.type !== "tool-call" || record4.toolCallId !== toolCallId) {
+      const record5 = part;
+      if (record5.type !== "tool-call" || record5.toolCallId !== toolCallId) {
         continue;
       }
-      return readPageToolBinding(record4.providerOptions);
+      return readPageToolBinding(record5.providerOptions);
     }
   }
   return void 0;
@@ -115514,7 +115933,7 @@ async function ensureOnSandbox(deps, target, contextMode, args) {
           handle2 = booted2;
           await recordLogicalBoot(deps, args, booted2.bootId);
           stream = target.watched ? await sandbox.ensureStream() : void 0;
-          const record4 = {
+          const record5 = {
             sandboxRowId: target.sandboxRowId,
             bootId: booted2.bootId,
             browserdToken: booted2.bearer,
@@ -115531,13 +115950,13 @@ async function ensureOnSandbox(deps, target, contextMode, args) {
             ...args.signal ? { signal: args.signal } : {}
           };
           return stream ? deps.store.record({
-            ...record4,
+            ...record5,
             watched: true,
             stream: {
               url: stream.streamUrl,
               password: stream.streamPassword
             }
-          }) : deps.store.record({ ...record4, watched: false });
+          }) : deps.store.record({ ...record5, watched: false });
         },
         onAdopt: () => {
           handle2 = void 0;
@@ -115840,18 +116259,18 @@ function parseReadyLine(line2) {
     return null;
   }
   if (typeof parsed !== "object" || parsed === null) return null;
-  const record4 = parsed;
-  if (record4.event !== "listening") return null;
-  if (typeof record4.port !== "number") return null;
-  if (typeof record4.bootId !== "string" || record4.bootId.length === 0)
+  const record5 = parsed;
+  if (record5.event !== "listening") return null;
+  if (typeof record5.port !== "number") return null;
+  if (typeof record5.bootId !== "string" || record5.bootId.length === 0)
     return null;
-  const protocolVersion2 = record4.protocolVersion;
+  const protocolVersion2 = record5.protocolVersion;
   if (protocolVersion2 !== void 0 && (typeof protocolVersion2 !== "number" || !Number.isSafeInteger(protocolVersion2) || protocolVersion2 < 1)) {
     return null;
   }
   return {
-    port: record4.port,
-    bootId: record4.bootId,
+    port: record5.port,
+    bootId: record5.bootId,
     ...typeof protocolVersion2 === "number" ? { protocolVersion: protocolVersion2 } : {}
   };
 }
@@ -116352,14 +116771,14 @@ var BrowserdClient = class {
           reason = void 0;
           break;
         }
-        for (const record4 of decoded.records) {
-          if (record4.kind === FRAME_STREAM_KIND.frame) args.onFrame(record4);
-          else if (record4.kind === FRAME_STREAM_KIND.video_key || record4.kind === FRAME_STREAM_KIND.video_delta) {
-            args.onVideo?.(record4);
-          } else if (record4.kind === FRAME_STREAM_KIND.heartbeat) {
-            if (record4.stats) args.onStats?.(record4.stats);
-          } else if (record4.kind === FRAME_STREAM_KIND.end)
-            reason = record4.reason;
+        for (const record5 of decoded.records) {
+          if (record5.kind === FRAME_STREAM_KIND.frame) args.onFrame(record5);
+          else if (record5.kind === FRAME_STREAM_KIND.video_key || record5.kind === FRAME_STREAM_KIND.video_delta) {
+            args.onVideo?.(record5);
+          } else if (record5.kind === FRAME_STREAM_KIND.heartbeat) {
+            if (record5.stats) args.onStats?.(record5.stats);
+          } else if (record5.kind === FRAME_STREAM_KIND.end)
+            reason = record5.reason;
         }
         if (reason !== void 0) break;
       }
@@ -116712,8 +117131,8 @@ async function startHostedRecording(handle2, deps) {
   const { client: client3, sandboxRowId, sandboxId, sessionId, bootId } = handle2;
   const existing = active.get(sandboxRowId);
   if (existing?.bootId === bootId) return;
-  const record4 = client3.record;
-  if (!record4) return;
+  const record5 = client3.record;
+  if (!record5) return;
   const now = deps.now ?? Date.now;
   const id3 = recordingIdFor(sessionId);
   let abandoned = false;
@@ -116730,7 +117149,7 @@ async function startHostedRecording(handle2, deps) {
       (async () => {
         const status = await client3.status();
         if (status.kind !== "ok" || !status.features?.includes("record")) return;
-        const started2 = await record4({
+        const started2 = await record5({
           action: "start",
           id: id3,
           fps: HOSTED_RECORDING_FPS
@@ -116760,7 +117179,7 @@ async function startHostedRecording(handle2, deps) {
           sandboxId,
           bootId,
           connect: () => deps.connect(sandboxId),
-          stop: () => record4({ action: "stop" }),
+          stop: () => record5({ action: "stop" }),
           fps: HOSTED_RECORDING_FPS,
           startedAtMs: now()
         });
@@ -124082,9 +124501,9 @@ function toAcpMcpServers(json) {
 }
 function attributeCursorToolCall(args) {
   const { rawToolName, input, keyToServerId } = args;
-  const record4 = input && typeof input === "object" ? input : {};
-  const provider = record4.providerIdentifier;
-  const inner = record4.toolName;
+  const record5 = input && typeof input === "object" ? input : {};
+  const provider = record5.providerIdentifier;
+  const inner = record5.toolName;
   if (typeof provider !== "string" || typeof inner !== "string" || !inner) {
     return { toolName: rawToolName };
   }
@@ -125177,7 +125596,7 @@ function cloneTraceValue(value) {
 }
 function getPromptIndex(messageHistory) {
   const userCount = messageHistory.reduce(
-    (count, message) => count + (message?.role === "user" ? 1 : 0),
+    (count4, message) => count4 + (message?.role === "user" ? 1 : 0),
     0
   );
   return Math.max(0, userCount - 1);
@@ -127007,13 +127426,13 @@ function stageDerivationToMetadata(derivation) {
 }
 function projectStageDerivation(metadata) {
   if (!metadata || typeof metadata !== "object") return {};
-  const record4 = metadata;
-  if (!("stageResults" in record4)) return {};
+  const record5 = metadata;
+  if (!("stageResults" in record5)) return {};
   const derivation = stageDerivationSchema2.safeParse({
-    stageResults: record4.stageResults,
-    ...record4.firstFailedStage !== void 0 ? { firstFailedStage: record4.firstFailedStage } : {},
-    ...record4.failureCategory !== void 0 ? { failureCategory: record4.failureCategory } : {},
-    stageAnalyzerVersion: record4.stageAnalyzerVersion
+    stageResults: record5.stageResults,
+    ...record5.firstFailedStage !== void 0 ? { firstFailedStage: record5.firstFailedStage } : {},
+    ...record5.failureCategory !== void 0 ? { failureCategory: record5.failureCategory } : {},
+    stageAnalyzerVersion: record5.stageAnalyzerVersion
   });
   if (derivation.success) {
     return {
@@ -127023,7 +127442,7 @@ function projectStageDerivation(metadata) {
       stageAnalyzerVersion: derivation.data.stageAnalyzerVersion
     };
   }
-  const version = record4.stageAnalyzerVersion;
+  const version = record5.stageAnalyzerVersion;
   return {
     stageResultsUnverified: true,
     ...typeof version === "number" && Number.isInteger(version) && version >= 0 ? { stageAnalyzerVersion: version } : {}
@@ -128212,8 +128631,8 @@ var EVAL_VERDICT_DECISION_REASONS2 = [
   ...EVAL_VALIDITY_DECISION_REASONS2,
   ...EVAL_TASK_DECISION_REASONS2
 ];
-var evalVerdictDecisionReasonSchema2 = z14.enum(
-  EVAL_VERDICT_DECISION_REASONS2
+var evalVerdictDecisionReasonSchema2 = z14.enum(EVAL_VERDICT_DECISION_REASONS2).describe(
+  "Why the verdict is what it is. The validity reasons are evaluated FIRST and make a run `inconclusive`; the task reasons decide `passed` / `failed` once validity holds. `configuredTrialsNotAttempted` \u2014 some configured trial never ran. `noGradeableTrials` \u2014 nothing in the suite was gradeable. `eligibleTrialsBelowMinimum` \u2014 the explicit `minEligibleTrials` was not reached. `completionRateBelowMinimum` \u2014 measured, and under the floor. `completionRateNotMeasured` \u2014 nothing was attempted, so the floor is unsatisfiable; a not-measured rate never passes one. `evaluatorErrorRateAboveMaximum` \u2014 the grader failed too often for the run to describe the server. `evaluatorErrorRateNotMeasured` \u2014 the same unsatisfiable case for the ceiling. `caseHasNoEligibleTrials` \u2014 a case graded nothing, which is inconclusive even at `passThreshold: 0`. `casePassRateMetThreshold` \u2014 a case's own passing reason. `casePassRateBelowThreshold` \u2014 a case failed its threshold, and so therefore did the suite. `allMeasuredCasesMetThreshold` \u2014 the suite's only passing reason."
 );
 function isEvalValidityDecisionReason2(value) {
   return typeof value === "string" && EVAL_VALIDITY_DECISION_REASONS2.includes(value);
@@ -128984,10 +129403,17 @@ function predicateResultsFrom(criteria) {
   if (criteria?.status !== "completed") return void 0;
   const results = criteria.results ?? [];
   if (results.length === 0) return void 0;
-  return results.map((entry3) => ({ passed: entry3.passed }));
+  return results.map((entry3) => ({
+    passed: entry3.passed,
+    ...entry3.status ? { status: entry3.status } : {},
+    ...entry3.predicate ? { predicate: entry3.predicate } : {}
+  }));
 }
 function judgeEvidenceFrom(input) {
-  const { criteria, goalJudge, hasUserAsk } = input;
+  const { goalJudge, hasUserAsk } = input;
+  const criteria = input.source === "swarm" && input.swarmPolicy?.judgeDecisive ? void 0 : input.criteria;
+  if (input.source === "swarm" && input.swarmPolicy && !input.swarmPolicy.judgeDecisive)
+    return void 0;
   if (criteria?.status === "failed") return void 0;
   if (criteria?.status === "completed") {
     if ((criteria.results ?? []).length > 0) return void 0;
@@ -129035,9 +129461,40 @@ function buildChatSessionStageInput(input) {
   if (toolSignals) evidence.toolSignals = toolSignals;
   const predicateResults = predicateResultsFrom(input.criteria);
   if (predicateResults) evidence.predicateResults = predicateResults;
-  if (input.criteria?.status === "failed") evidence.evaluatorErrored = true;
+  if (input.source === "swarm" && input.swarmPolicy?.judgeDecisive) {
+    evidence.predicateResults = predicateResults?.filter(
+      (r) => r.passed === false || r.status === "error" || r.predicate?.type && ASSERTION_STAGE2[r.predicate.type] !== "userValue"
+    );
+  }
+  if (input.criteria?.status === "failed" && (input.source !== "swarm" || !input.swarmPolicy || input.swarmPolicy.requiredCriteria > 0))
+    evidence.evaluatorErrored = true;
   const judgeEvidence = judgeEvidenceFrom(input);
   if (judgeEvidence) evidence.judgeEvidence = judgeEvidence;
+  if (input.source === "swarm" && input.swarmPolicy?.requiredCriteria) {
+    const requiredResults = input.criteria?.results?.filter(
+      (r) => r.predicate?.role !== "advisory"
+    ) ?? [];
+    const requiredFailure = requiredResults.some(
+      (r) => r.status !== "error" && !r.passed
+    );
+    if (!requiredFailure && input.criteria?.status !== "failed") {
+      if (!input.criteria || input.criteria.status === "pending") {
+        evidence.judgeEvidence = {
+          status: "pending",
+          pendingKind: "scheduled"
+        };
+        evidence.predicateResults = evidence.predicateResults?.filter(
+          (r) => r.passed === false
+        );
+      } else if (requiredResults.length < input.swarmPolicy.requiredCriteria || requiredResults.some((r) => r.status === "error")) {
+        evidence.evaluatorErrored = true;
+      }
+    }
+  }
+  if (input.source === "swarm" && input.swarmPolicy?.judgeDecisive && input.goalJudge?.status === "completed" && input.goalJudge.passed === false) {
+    delete evidence.evaluatorErrored;
+    if (judgeEvidence) evidence.judgeEvidence = judgeEvidence;
+  }
   return {
     authored: buildChatSessionAuthoredCase(input),
     evidence,
@@ -129682,16 +130139,16 @@ function buildFrictionCallRecords(args) {
   const results = args.resultsByToolCallId;
   const records = [];
   for (const [index, call] of args.toolsCalled.entries()) {
-    const record4 = isRecord21(call) ? call : {};
-    const toolName22 = typeof record4.toolName === "string" && record4.toolName.length > 0 ? record4.toolName : typeof record4.tool === "string" && record4.tool.length > 0 ? record4.tool : typeof record4.name === "string" && record4.name.length > 0 ? record4.name : "";
-    const toolCallId = typeof record4.toolCallId === "string" && record4.toolCallId.length > 0 ? record4.toolCallId : void 0;
+    const record5 = isRecord21(call) ? call : {};
+    const toolName22 = typeof record5.toolName === "string" && record5.toolName.length > 0 ? record5.toolName : typeof record5.tool === "string" && record5.tool.length > 0 ? record5.tool : typeof record5.name === "string" && record5.name.length > 0 ? record5.name : "";
+    const toolCallId = typeof record5.toolCallId === "string" && record5.toolCallId.length > 0 ? record5.toolCallId : void 0;
     const entry3 = toolCallId ? results?.get(toolCallId) : void 0;
     const timed = entry3 !== void 0 && typeof entry3.startedAtMs === "number" && typeof entry3.settledAtMs === "number";
     records.push({
       index,
       toolName: toolName22,
       ...toolCallId ? { toolCallId } : {},
-      arguments: record4.arguments,
+      arguments: record5.arguments,
       ...entry3 ? {
         result: {
           ...normalizeFrictionResult(entry3.raw),
@@ -129746,18 +130203,18 @@ function compareSignals(a, b) {
 function deriveTrialFrictionSignals(records) {
   const callCount = records.length;
   const resultAvailableCount = records.filter(
-    (record4) => record4.resultAvailable
+    (record5) => record5.resultAvailable
   ).length;
   const timedCallCount = records.filter(
-    (record4) => record4.ordering === "timed"
+    (record5) => record5.ordering === "timed"
   ).length;
   const counts = { callCount, resultAvailableCount, timedCallCount };
   if (callCount === 0) return notMeasured("noToolCalls", counts);
   if (callCount > MAX_FRICTION_CALLS) return notMeasured("truncated", counts);
   const canonicalArguments = [];
-  for (const record4 of records) {
+  for (const record5 of records) {
     try {
-      canonicalArguments.push((0, canonical_exports3.canonicalJson)(record4.arguments ?? null));
+      canonicalArguments.push((0, canonical_exports3.canonicalJson)(record5.arguments ?? null));
     } catch (error) {
       if (error instanceof canonical_exports3.CanonicalJsonError) {
         return notMeasured("evidenceIncomplete", counts);
@@ -129810,7 +130267,7 @@ function deriveTrialFrictionSignals(records) {
   if (resultAvailableCount !== callCount) {
     identifierSignals = { state: "notMeasured", reason: "resultsUnavailable" };
   } else {
-    const modes = new Set(records.map((record4) => record4.ordering));
+    const modes = new Set(records.map((record5) => record5.ordering));
     if (modes.size !== 1 || modes.has("unknown")) {
       identifierSignals = { state: "notMeasured", reason: "orderingUnknown" };
     }
@@ -129924,18 +130381,18 @@ function deriveTrialFrictionSignalsFromCalls(args) {
   if (args.evidenceHole) {
     return notMeasured(args.evidenceHole, {
       callCount: records.length,
-      resultAvailableCount: records.filter((record4) => record4.resultAvailable).length,
-      timedCallCount: records.filter((record4) => record4.ordering === "timed").length
+      resultAvailableCount: records.filter((record5) => record5.resultAvailable).length,
+      timedCallCount: records.filter((record5) => record5.ordering === "timed").length
     });
   }
   return deriveTrialFrictionSignals(records);
 }
 function projectFrictionSignals(metadata) {
   if (!metadata || typeof metadata !== "object") return {};
-  const record4 = metadata;
-  if (!("frictionSignals" in record4)) return {};
+  const record5 = metadata;
+  if (!("frictionSignals" in record5)) return {};
   const parsed = evalTrialFrictionSignalsSchema.safeParse(
-    record4.frictionSignals
+    record5.frictionSignals
   );
   if (parsed.success) return { frictionSignals: parsed.data };
   return { frictionSignalsUnverified: true };
@@ -130006,10 +130463,10 @@ var suspectedConditionVerdictSchema = z14.discriminatedUnion("status", [
 ]);
 function projectSuspectedConditionVerdict(metadata) {
   if (!metadata || typeof metadata !== "object") return {};
-  const record4 = metadata;
-  if (!("suspectedConditionVerdict" in record4)) return {};
+  const record5 = metadata;
+  if (!("suspectedConditionVerdict" in record5)) return {};
   const parsed = suspectedConditionVerdictSchema.safeParse(
-    record4.suspectedConditionVerdict
+    record5.suspectedConditionVerdict
   );
   if (parsed.success) return { suspectedConditionVerdict: parsed.data };
   return { suspectedConditionUnverified: true };
@@ -130633,10 +131090,10 @@ function assembleEvalRunDecisionSummary(input) {
 }
 function legacyTrialCounts(summary) {
   if (!summary || typeof summary !== "object") return void 0;
-  const count = (value) => typeof value === "number" && Number.isInteger(value) && value >= 0 ? value : void 0;
-  const total = count(summary.total);
-  const passed2 = count(summary.passed);
-  const failed3 = count(summary.failed);
+  const count22 = (value) => typeof value === "number" && Number.isInteger(value) && value >= 0 ? value : void 0;
+  const total = count22(summary.total);
+  const passed2 = count22(summary.passed);
+  const failed3 = count22(summary.failed);
   if (total === void 0 && passed2 === void 0 && failed3 === void 0) {
     return void 0;
   }
@@ -130762,8 +131219,8 @@ function nonEmptyString(value) {
 }
 function toolName(value) {
   if (!value || typeof value !== "object") return void 0;
-  const record4 = value;
-  return nonEmptyString(record4.toolName) ?? nonEmptyString(record4.tool) ?? nonEmptyString(record4.name);
+  const record5 = value;
+  return nonEmptyString(record5.toolName) ?? nonEmptyString(record5.tool) ?? nonEmptyString(record5.name);
 }
 function toolNames(calls) {
   const names = (calls ?? []).map(toolName).filter((name15) => name15 !== void 0);
@@ -132050,6 +132507,417 @@ var evalBacktestContinuationSchema2 = z14.object({
 var evalBacktestRequestSchema = evalBacktestDraftSchema2.extend({
   continuation: evalBacktestContinuationSchema2.optional()
 });
+var SWARM_SESSION_VERDICT_CONTRACT_VERSION2 = 1;
+var SWARM_STAGE_EVIDENCE_VERSION = 2;
+var SWARM_ATTEMPT_STATUSES2 = [
+  "pending",
+  "running",
+  "succeeded",
+  "failed",
+  "rate_limited"
+];
+var SWARM_SESSION_LIFECYCLES2 = [
+  "pending",
+  "running",
+  "ran",
+  "broke",
+  "limited",
+  "withdrawn"
+];
+var SWARM_SESSION_VERDICTS2 = [
+  "passed",
+  "failed",
+  "inconclusive",
+  "notEstablished"
+];
+var SWARM_GRADING_STATES2 = [
+  "notRequested",
+  "queued",
+  "running",
+  "settled",
+  "unavailable"
+];
+var SWARM_SESSION_VERDICT_OF_REASON2 = {
+  attemptPending: "notEstablished",
+  attemptRunning: "notEstablished",
+  withdrawn: "notEstablished",
+  spendCapReached: "notEstablished",
+  notRun: "notEstablished",
+  executionFailed: "notEstablished",
+  ungraded: "notEstablished",
+  gradingNotClaimed: "notEstablished",
+  criteriaPending: "notEstablished",
+  judgePending: "notEstablished",
+  gatingCriterionFailed: "failed",
+  judgeFailed: "failed",
+  criteriaGradingErrored: "inconclusive",
+  gatingCriterionUnmeasured: "inconclusive",
+  judgeErrored: "inconclusive",
+  gradingUnavailable: "inconclusive",
+  allGatingCriteriaPassed: "passed",
+  judgePassed: "passed",
+  allGradersPassed: "passed"
+};
+var SWARM_SESSION_VERDICT_REASONS2 = Object.keys(
+  SWARM_SESSION_VERDICT_OF_REASON2
+);
+var swarmSessionLifecycleSchema2 = z14.enum(SWARM_SESSION_LIFECYCLES2);
+var swarmSessionVerdictValueSchema2 = z14.enum(SWARM_SESSION_VERDICTS2);
+var swarmGraderRoleSchema2 = z14.enum(["advisory", "required"]);
+var swarmSessionAttemptInputSchema2 = z14.object({
+  status: z14.enum(SWARM_ATTEMPT_STATUSES2),
+  errorCode: z14.string().nullable().optional()
+}).strict();
+var swarmGradingReadinessSchema2 = z14.object({
+  /** Durable readiness for decisive grading only, not background observations. */
+  state: z14.enum(SWARM_GRADING_STATES2),
+  reasonCode: z14.string().min(1).optional()
+}).strict();
+var criterionResultSchema2 = z14.object({
+  criterionId: z14.string().min(1),
+  passed: z14.boolean(),
+  status: z14.enum(["scored", "error"]).optional()
+}).strict();
+var swarmSessionVerdictInputSchema = z14.object({
+  attempt: swarmSessionAttemptInputSchema2.nullable(),
+  hasTranscript: z14.boolean(),
+  rubric: z14.array(
+    z14.object({
+      id: z14.string().min(1),
+      role: swarmGraderRoleSchema2,
+      predicateType: z14.string().min(1).optional()
+    }).strict()
+  ),
+  criteria: z14.object({
+    status: z14.enum(["pending", "completed", "failed"]),
+    criterionIds: z14.array(z14.string().min(1)).optional(),
+    results: z14.array(criterionResultSchema2).optional()
+  }).strict().nullable(),
+  goalScore: z14.discriminatedUnion("status", [
+    z14.object({ status: z14.literal("completed"), passed: z14.boolean() }).strict(),
+    z14.object({ status: z14.literal("running") }).strict(),
+    z14.object({ status: z14.literal("failed") }).strict()
+  ]).nullable(),
+  judge: z14.object({
+    automatic: z14.boolean(),
+    role: swarmGraderRoleSchema2,
+    /** Explicit on-demand intent; automatic:false alone does not mean a judge was requested. */
+    requested: z14.boolean().optional()
+  }).strict(),
+  grading: swarmGradingReadinessSchema2
+}).strict().superRefine((value, ctx) => {
+  const definitions = new Set(value.rubric.map((entry3) => entry3.id));
+  if (definitions.size !== value.rubric.length) {
+    ctx.addIssue({
+      code: "custom",
+      path: ["rubric"],
+      message: "Duplicate criterion definition"
+    });
+  }
+  for (const [field, ids] of [
+    ["criterionIds", value.criteria?.criterionIds ?? []],
+    ["results", value.criteria?.results?.map((row2) => row2.criterionId) ?? []]
+  ]) {
+    if (new Set(ids).size !== ids.length || ids.some((id3) => !definitions.has(id3))) {
+      ctx.addIssue({
+        code: "custom",
+        path: ["criteria", field],
+        message: "Criterion IDs must be unique and defined in the snapshot"
+      });
+    }
+  }
+  const claimed = value.criteria?.criterionIds;
+  if (claimed && value.criteria?.results?.some((row2) => !claimed.includes(row2.criterionId))) {
+    ctx.addIssue({
+      code: "custom",
+      path: ["criteria", "results"],
+      message: "Result outside claimed scope"
+    });
+  }
+});
+var SWARM_SESSION_TRIAL_STATUSES2 = [
+  "pending",
+  "running",
+  "completed",
+  "failed",
+  "setup_failed",
+  "cancelled"
+];
+var swarmSessionTrialSchema2 = z14.object({
+  status: z14.enum(SWARM_SESSION_TRIAL_STATUSES2),
+  taskVerdict: z14.enum(["passed", "failed"]).optional(),
+  evaluatorError: z14.literal(true).optional()
+}).strict().superRefine((value, ctx) => {
+  const verdict = value.taskVerdict !== void 0;
+  const error = value.evaluatorError === true;
+  if (value.status === "completed" ? verdict === error : verdict || error) {
+    ctx.addIssue({
+      code: "custom",
+      message: "Only completed trials carry exactly one verdict or evaluator error"
+    });
+  }
+});
+var swarmSessionGraderCountsSchema2 = z14.object({
+  gating: z14.number().int().nonnegative(),
+  gatingPassed: z14.number().int().nonnegative(),
+  gatingFailed: z14.number().int().nonnegative(),
+  gatingUnmeasured: z14.number().int().nonnegative(),
+  advisoryFailed: z14.number().int().nonnegative()
+}).strict();
+var swarmSessionVerdictSchema2 = z14.object({
+  contractVersion: z14.literal(SWARM_SESSION_VERDICT_CONTRACT_VERSION2),
+  lifecycle: swarmSessionLifecycleSchema2,
+  verdict: swarmSessionVerdictValueSchema2,
+  reason: z14.enum(SWARM_SESSION_VERDICT_REASONS2),
+  verdictSource: z14.enum([
+    "goalJudge",
+    "requiredAssertions",
+    "combined",
+    "none"
+  ]),
+  grading: swarmGradingReadinessSchema2,
+  graders: z14.object({
+    criteria: z14.enum([
+      "notConfigured",
+      "notClaimed",
+      "pending",
+      "scored",
+      "errored"
+    ]),
+    judge: z14.enum([
+      "notConfigured",
+      "silent",
+      "pending",
+      "scored",
+      "errored"
+    ])
+  }).strict(),
+  counts: swarmSessionGraderCountsSchema2,
+  trial: swarmSessionTrialSchema2.nullable()
+}).strict().superRefine((value, ctx) => {
+  const fail3 = (message) => ctx.addIssue({ code: "custom", message });
+  if (SWARM_SESSION_VERDICT_OF_REASON2[value.reason] !== value.verdict)
+    fail3("Reason contradicts verdict");
+  const c = value.counts;
+  if (c.gating !== c.gatingPassed + c.gatingFailed + c.gatingUnmeasured)
+    fail3("Required measurement counts do not add up");
+  const trial = value.trial;
+  if (trial?.status === "completed") {
+    if (value.lifecycle !== "ran")
+      fail3("Only a completed execution can be a completed trial");
+    if (trial.taskVerdict !== void 0 && trial.taskVerdict !== value.verdict)
+      fail3("Trial contradicts goal result");
+    if (trial.evaluatorError && value.verdict !== "inconclusive")
+      fail3("Evaluator error requires inconclusive grading");
+  }
+  const expected = {
+    pending: "pending",
+    running: "running",
+    broke: "failed",
+    limited: "setup_failed",
+    withdrawn: "cancelled"
+  };
+  if (value.lifecycle !== "ran" && trial?.status !== expected[value.lifecycle])
+    fail3("Trial must preserve execution lifecycle");
+  if (value.lifecycle === "ran" && trial !== null && trial.status !== "completed")
+    fail3("Completed execution must not become skipped or running");
+  if ((value.verdict === "passed" || value.verdict === "failed") && value.verdictSource === "none")
+    fail3("A measured goal verdict needs a source");
+  if (value.lifecycle === "ran" && value.verdict !== "notEstablished" && trial === null)
+    fail3("Settled grading on completed execution needs an observation");
+  if (value.verdict === "notEstablished" && value.verdictSource !== "none")
+    fail3("An undecided goal has no verdict source");
+});
+var SWARM_REPORT_CONTRACT_VERSION2 = 1;
+var count2 = z14.number().int().nonnegative();
+var journeyRunVerdictSummarySchema2 = z14.discriminatedUnion("status", [
+  z14.object({
+    status: z14.literal("pending"),
+    pendingSessions: count2,
+    updatedAt: count2
+  }).strict(),
+  z14.object({
+    status: z14.literal("decided"),
+    decision: evalVerdictDecisionSchema2,
+    updatedAt: count2
+  }).strict(),
+  z14.object({
+    status: z14.literal("notEstablished"),
+    reason: z14.literal("gradingNotConfigured"),
+    updatedAt: count2
+  }).strict(),
+  z14.object({
+    status: z14.literal("integrityFailed"),
+    reason: z14.string().min(1),
+    updatedAt: count2
+  }).strict()
+]);
+var swarmObservationInputSchema2 = z14.object({
+  evaluatorId: z14.string().min(1),
+  predicateType: z14.enum(
+    Object.keys(ASSERTION_STAGE2)
+  ),
+  role: z14.enum(["advisory", "required"]),
+  /** Missing capture/error is unavailable; it must not use the compatibility passed:false. */
+  status: z14.enum(["passed", "failed", "pending", "unavailable"])
+}).strict();
+var swarmReportSessionSchema2 = z14.object({
+  /** Stable attempt/slot identity, including slots with no session document. */
+  id: z14.string().min(1),
+  /** A producer fact. Grade coverage and lifecycle alone cannot establish a non-start. */
+  startEvidence: z14.enum(["started", "notStarted", "unknown"]),
+  verdict: swarmSessionVerdictSchema2,
+  observations: z14.array(swarmObservationInputSchema2)
+}).strict().superRefine((row2, ctx) => {
+  if (new Set(row2.observations.map((o) => o.evaluatorId)).size !== row2.observations.length) {
+    ctx.addIssue({
+      code: "custom",
+      path: ["observations"],
+      message: "Duplicate evaluator in one session"
+    });
+  }
+  if (row2.startEvidence === "notStarted" && ["running", "ran"].includes(row2.verdict.lifecycle)) {
+    ctx.addIssue({
+      code: "custom",
+      path: ["startEvidence"],
+      message: "Execution contradicts explicit non-start"
+    });
+  }
+});
+var swarmReportInputSchema = z14.object({
+  runId: z14.string().min(1),
+  executionComplete: z14.boolean(),
+  configuredSessions: count2,
+  verdictSummary: journeyRunVerdictSummarySchema2.nullable(),
+  /** Frozen run-wide definitions: absent measurement rows still count as unavailable. */
+  evaluatorDefinitions: z14.array(
+    swarmObservationInputSchema2.omit({ status: true })
+  ),
+  sessions: z14.array(swarmReportSessionSchema2)
+}).strict().superRefine((input, ctx) => {
+  if (input.sessions.length > input.configuredSessions) {
+    ctx.addIssue({
+      code: "custom",
+      path: ["sessions"],
+      message: "Reported sessions exceed configured population"
+    });
+  }
+  if (new Set(input.sessions.map((row2) => row2.id)).size !== input.sessions.length) {
+    ctx.addIssue({
+      code: "custom",
+      path: ["sessions"],
+      message: "Duplicate session/attempt identity"
+    });
+  }
+  const definitions = new Map(
+    input.evaluatorDefinitions.map((row2) => [row2.evaluatorId, row2])
+  );
+  if (definitions.size !== input.evaluatorDefinitions.length) {
+    ctx.addIssue({
+      code: "custom",
+      path: ["evaluatorDefinitions"],
+      message: "Duplicate evaluator definition"
+    });
+  }
+  for (const session of input.sessions) {
+    for (const observation of session.observations) {
+      const definition = definitions.get(observation.evaluatorId);
+      if (!definition || definition.role !== observation.role || definition.predicateType !== observation.predicateType) {
+        ctx.addIssue({
+          code: "custom",
+          path: ["sessions"],
+          message: "Observation contradicts the frozen evaluator definition"
+        });
+      }
+    }
+  }
+});
+var swarmObservationCoverageSchema2 = z14.object({
+  evaluatorId: z14.string().min(1),
+  predicateType: z14.enum(
+    Object.keys(ASSERTION_STAGE2)
+  ),
+  role: z14.enum(["advisory", "required"]),
+  stage: z14.enum(USER_VALUE_STAGES2).describe(
+    "The stage this observation was read from, and the only stage it is evidence ABOUT. `connection` \u2014 the server was reachable and the session initialized. `discovery` \u2014 its tools and resources were listed and readable. `selection` \u2014 the model chose the right tool for the request. `call` \u2014 the call was made with usable arguments. `response` \u2014 the server returned data the model could use. `userValue` \u2014 the user's actual request was satisfied."
+  ),
+  unit: z14.literal("sessions"),
+  total: count2,
+  passed: count2,
+  failed: count2,
+  pending: count2,
+  unavailable: count2
+}).strict().superRefine((value, ctx) => {
+  if (value.total !== value.passed + value.failed + value.pending + value.unavailable) {
+    ctx.addIssue({
+      code: "custom",
+      message: "Observation coverage must partition its population"
+    });
+  }
+  if (value.stage !== ASSERTION_STAGE2[value.predicateType]) {
+    ctx.addIssue({
+      code: "custom",
+      message: "Observation stage contradicts the shared evaluator map"
+    });
+  }
+});
+var swarmReportSchema = z14.object({
+  contractVersion: z14.literal(SWARM_REPORT_CONTRACT_VERSION2),
+  runId: z14.string().min(1),
+  verdict: swarmSessionVerdictValueSchema2,
+  verdictSource: z14.enum(["policyV2", "none"]),
+  decision: evalVerdictDecisionSchema2.optional(),
+  undecidedReason: z14.enum([
+    "executionPending",
+    "gradingPending",
+    "gradingNotConfigured",
+    "verdictSummaryUnavailable",
+    "integrityFailed"
+  ]).optional(),
+  execution: z14.object({
+    unit: z14.literal("sessions"),
+    configured: count2,
+    reported: count2,
+    started: count2,
+    notStarted: count2,
+    unknown: count2,
+    completed: count2,
+    interrupted: count2,
+    /** Only explicit complete execution evidence can establish this. */
+    neverLaunched: z14.boolean()
+  }).strict(),
+  goalGrading: z14.object({
+    unit: z14.literal("sessions"),
+    reported: count2,
+    passed: count2,
+    failed: count2,
+    pending: count2,
+    unavailable: count2,
+    notRequested: count2,
+    /** Readiness is separate: a proven failure can exist while other required grading is pending. */
+    waitingForDecisiveGrading: count2
+  }).strict(),
+  observations: z14.array(swarmObservationCoverageSchema2)
+}).strict().superRefine((report, ctx) => {
+  const fail3 = (message) => ctx.addIssue({ code: "custom", message });
+  if (report.verdictSource === "policyV2") {
+    if (!report.decision || report.verdict !== report.decision.verdict || report.undecidedReason !== void 0)
+      fail3("Report must copy its authoritative decision");
+  } else if (report.verdict !== "notEstablished" || report.decision !== void 0 || !report.undecidedReason)
+    fail3("Missing authority must have an undecided reason");
+  const e = report.execution;
+  if (e.started + e.notStarted + e.unknown !== e.configured || e.reported > e.configured)
+    fail3("Execution evidence does not cover its configured population");
+  if (e.started + e.notStarted > e.reported || e.completed + e.interrupted > e.reported)
+    fail3("Execution counts exceed reported evidence");
+  if (e.neverLaunched !== (e.configured > 0 && e.reported === e.configured && e.notStarted === e.configured))
+    fail3("Not run requires explicit evidence for the entire population");
+  const g = report.goalGrading;
+  if (g.reported !== e.reported || g.passed + g.failed + g.pending + g.unavailable + g.notRequested !== g.reported || g.waitingForDecisiveGrading > g.reported)
+    fail3("Goal coverage must partition reported sessions");
+  if (new Set(report.observations.map((o) => o.evaluatorId)).size !== report.observations.length || report.observations.some((o) => o.total > e.reported))
+    fail3("Invalid observation population");
+});
 var export_CanonicalJsonError = canonical_exports3.CanonicalJsonError;
 var export_EMIT_CANONICAL_ROLE = policy_spelling_exports3.EMIT_CANONICAL_ROLE;
 var export_EVALUATOR_KINDS2 = evaluator_types_exports3.EVALUATOR_KINDS;
@@ -132318,7 +133186,7 @@ function collectResultText(output, depth = 0) {
   if (Array.isArray(output)) {
     return output.flatMap((entry3) => collectResultText(entry3, depth + 1));
   }
-  const record4 = output;
+  const record5 = output;
   const texts = [];
   for (const key of [
     "text",
@@ -132328,7 +133196,7 @@ function collectResultText(output, depth = 0) {
     "value",
     "stdout"
   ]) {
-    const nested = record4[key];
+    const nested = record5[key];
     if (nested !== void 0) {
       texts.push(...collectResultText(nested, depth + 1));
     }
@@ -132877,29 +133745,29 @@ function stableJson(value) {
   if (Array.isArray(value)) {
     return `[${value.map(stableJson).join(",")}]`;
   }
-  const record4 = value;
-  return `{${Object.keys(record4).sort().map((key) => `${JSON.stringify(key)}:${stableJson(record4[key])}`).join(",")}}`;
+  const record5 = value;
+  return `{${Object.keys(record5).sort().map((key) => `${JSON.stringify(key)}:${stableJson(record5[key])}`).join(",")}}`;
 }
 function scopeStepUpInputHash(input) {
   return createHash14("sha256").update(stableJson(input)).digest("base64url");
 }
-function expireIfNeeded(record4) {
-  if ((record4.status === "pending" || record4.status === "retrying") && record4.expiresAt <= Date.now()) {
-    record4.status = "expired";
-    record4.input = void 0;
-    record4.inputPresent = false;
-    record4.terminalReason = "continuation expired";
+function expireIfNeeded(record5) {
+  if ((record5.status === "pending" || record5.status === "retrying") && record5.expiresAt <= Date.now()) {
+    record5.status = "expired";
+    record5.input = void 0;
+    record5.inputPresent = false;
+    record5.terminalReason = "continuation expired";
   }
-  return record4;
+  return record5;
 }
-function scrubTerminal(record4, status, reason) {
-  record4.status = status;
-  record4.input = void 0;
-  record4.inputPresent = false;
-  record4.terminalReason = reason;
+function scrubTerminal(record5, status, reason) {
+  record5.status = status;
+  record5.input = void 0;
+  record5.inputPresent = false;
+  record5.terminalReason = reason;
   const timer = setTimeout(() => {
-    if (localContinuations.get(record4.continuationId) === record4) {
-      localContinuations.delete(record4.continuationId);
+    if (localContinuations.get(record5.continuationId) === record5) {
+      localContinuations.delete(record5.continuationId);
     }
   }, SCOPE_STEP_UP_LIVE_TTL_MS);
   timer.unref?.();
@@ -132949,71 +133817,71 @@ function createLocalScopeStepUpContinuation(input) {
   };
 }
 function claimLocalScopeStepUpContinuation(input) {
-  const record4 = localContinuations.get(input.continuationId);
-  if (!record4) throw new Error("scope_step_up_continuation_not_found");
-  expireIfNeeded(record4);
-  if (record4.bindingKey !== input.bindingKey || record4.toolCallId !== input.toolCallId) {
+  const record5 = localContinuations.get(input.continuationId);
+  if (!record5) throw new Error("scope_step_up_continuation_not_found");
+  expireIfNeeded(record5);
+  if (record5.bindingKey !== input.bindingKey || record5.toolCallId !== input.toolCallId) {
     throw new Error("scope_step_up_continuation_not_found");
   }
-  if (record4.status === "retrying") {
+  if (record5.status === "retrying") {
     throw new Error("scope_step_up_retry_in_progress");
   }
-  if (record4.status !== "pending" || !record4.inputPresent) {
-    throw new Error(`scope_step_up_continuation_${record4.status}`);
+  if (record5.status !== "pending" || !record5.inputPresent) {
+    throw new Error(`scope_step_up_continuation_${record5.status}`);
   }
-  record4.status = "retrying";
+  record5.status = "retrying";
   return {
-    continuationId: record4.continuationId,
-    bindingKey: record4.bindingKey,
-    serverId: record4.serverId,
-    ...record4.resourceUrl ? { resourceUrl: record4.resourceUrl } : {},
-    toolCallId: record4.toolCallId,
-    toolName: record4.toolName,
-    input: record4.input,
-    inputHash: record4.inputHash,
-    challenge: record4.challenge,
-    expiresAt: record4.expiresAt
+    continuationId: record5.continuationId,
+    bindingKey: record5.bindingKey,
+    serverId: record5.serverId,
+    ...record5.resourceUrl ? { resourceUrl: record5.resourceUrl } : {},
+    toolCallId: record5.toolCallId,
+    toolName: record5.toolName,
+    input: record5.input,
+    inputHash: record5.inputHash,
+    challenge: record5.challenge,
+    expiresAt: record5.expiresAt
   };
 }
 function markLocalScopeStepUpWireStarted(continuationId) {
-  const record4 = localContinuations.get(continuationId);
-  if (!record4 || record4.status !== "retrying") {
+  const record5 = localContinuations.get(continuationId);
+  if (!record5 || record5.status !== "retrying") {
     throw new Error("scope_step_up_continuation_not_retrying");
   }
-  record4.wireStartedAt = Date.now();
+  record5.wireStartedAt = Date.now();
 }
 function completeLocalScopeStepUpContinuation(continuationId) {
-  const record4 = localContinuations.get(continuationId);
-  if (!record4) return;
-  scrubTerminal(record4, "completed");
+  const record5 = localContinuations.get(continuationId);
+  if (!record5) return;
+  scrubTerminal(record5, "completed");
 }
 function failLocalScopeStepUpContinuation(continuationId, reason) {
-  const record4 = localContinuations.get(continuationId);
-  if (!record4) return;
-  scrubTerminal(record4, "failed", reason);
+  const record5 = localContinuations.get(continuationId);
+  if (!record5) return;
+  scrubTerminal(record5, "failed", reason);
 }
 function cancelLocalScopeStepUpContinuation(continuationId, reason) {
-  const record4 = localContinuations.get(continuationId);
-  if (!record4) return;
-  if (record4.status === "retrying" && record4.wireStartedAt !== void 0) {
-    scrubTerminal(record4, "indeterminate", reason);
+  const record5 = localContinuations.get(continuationId);
+  if (!record5) return;
+  if (record5.status === "retrying" && record5.wireStartedAt !== void 0) {
+    scrubTerminal(record5, "indeterminate", reason);
     return;
   }
-  scrubTerminal(record4, "cancelled", reason);
+  scrubTerminal(record5, "cancelled", reason);
 }
 function cancelLocalScopeStepUpContinuationForRequest(input) {
-  const record4 = localContinuations.get(input.continuationId);
-  if (!record4) throw new Error("scope_step_up_continuation_not_found");
-  expireIfNeeded(record4);
-  if (record4.bindingKey !== input.bindingKey || record4.toolCallId !== input.toolCallId) {
+  const record5 = localContinuations.get(input.continuationId);
+  if (!record5) throw new Error("scope_step_up_continuation_not_found");
+  expireIfNeeded(record5);
+  if (record5.bindingKey !== input.bindingKey || record5.toolCallId !== input.toolCallId) {
     throw new Error("scope_step_up_continuation_not_found");
   }
-  if (record4.status !== "pending") {
-    throw new Error(`scope_step_up_continuation_${record4.status}`);
+  if (record5.status !== "pending") {
+    throw new Error(`scope_step_up_continuation_${record5.status}`);
   }
-  const toolName3 = record4.toolName;
-  const serverId = record4.serverId;
-  scrubTerminal(record4, "cancelled", input.reason);
+  const toolName3 = record5.toolName;
+  const serverId = record5.serverId;
+  scrubTerminal(record5, "cancelled", input.reason);
   return { serverId, toolName: toolName3 };
 }
 
@@ -134284,16 +135152,16 @@ async function readState() {
       "the local harness grant store is not an object"
     );
   }
-  const record4 = parsed;
-  if (record4.version !== 1 || !Array.isArray(record4.workspaces) || !Array.isArray(record4.harnessGrants) || !record4.workspaces.every(isWorkspaceGrant) || !record4.harnessGrants.every(isPersistedHarnessGrant)) {
+  const record5 = parsed;
+  if (record5.version !== 1 || !Array.isArray(record5.workspaces) || !Array.isArray(record5.harnessGrants) || !record5.workspaces.every(isWorkspaceGrant) || !record5.harnessGrants.every(isPersistedHarnessGrant)) {
     throw new GrantStateUnreadableError(
       "the local harness grant store has an unrecognized or malformed shape"
     );
   }
   return {
     version: 1,
-    workspaces: [...record4.workspaces],
-    harnessGrants: [...record4.harnessGrants]
+    workspaces: [...record5.workspaces],
+    harnessGrants: [...record5.harnessGrants]
   };
 }
 async function writeState(state) {
@@ -137073,13 +137941,13 @@ async function readRegistry() {
       "the local harness process registry is not an object"
     );
   }
-  const record4 = parsed;
-  if (record4.version !== 1 || !Array.isArray(record4.records)) {
+  const record5 = parsed;
+  if (record5.version !== 1 || !Array.isArray(record5.records)) {
     throw new RegistryUnreadableError(
       "the local harness process registry has an unrecognized shape"
     );
   }
-  return { version: 1, records: record4.records };
+  return { version: 1, records: record5.records };
 }
 async function writeRegistry(registry3) {
   const dir = localHarnessStateRoot();
@@ -137094,22 +137962,22 @@ async function writeRegistry(registry3) {
 function mintSupervisorNonce() {
   return `sup_${randomUUID21()}`;
 }
-function recordProcess(record4) {
+function recordProcess(record5) {
   return withRegistryLock(async () => {
     const registry3 = await readRegistry();
     registry3.records = registry3.records.filter(
-      (r) => r.sessionId !== record4.sessionId
+      (r) => r.sessionId !== record5.sessionId
     );
-    registry3.records.push(record4);
+    registry3.records.push(record5);
     await writeRegistry(registry3);
   });
 }
 function updateLifecycleState(sessionId, lifecycleState) {
   return withRegistryLock(async () => {
     const registry3 = await readRegistry();
-    const record4 = registry3.records.find((r) => r.sessionId === sessionId);
-    if (!record4) return;
-    record4.lifecycleState = lifecycleState;
+    const record5 = registry3.records.find((r) => r.sessionId === sessionId);
+    if (!record5) return;
+    record5.lifecycleState = lifecycleState;
     await writeRegistry(registry3);
   });
 }
@@ -137123,14 +137991,14 @@ function forgetProcess(sessionId) {
     if (registry3.records.length !== before) await writeRegistry(registry3);
   });
 }
-async function owningSupervisorProvablyExited(record4, platform) {
-  if (record4.supervisorPid === void 0 || record4.supervisorBirthIdentity === void 0) {
+async function owningSupervisorProvablyExited(record5, platform) {
+  if (record5.supervisorPid === void 0 || record5.supervisorBirthIdentity === void 0) {
     return false;
   }
-  const probe = await probeProcess(record4.supervisorPid, platform);
+  const probe = await probeProcess(record5.supervisorPid, platform);
   if (probe.state === "gone") return true;
   if (probe.state === "unknown") return false;
-  return probe.identity !== record4.supervisorBirthIdentity;
+  return probe.identity !== record5.supervisorBirthIdentity;
 }
 function janitorOutcomeForUnsettled(outcome) {
   if (outcome === "escaped") return "escaped";
@@ -137155,90 +138023,90 @@ function reclaimAbandonedProcesses(args) {
     const registry3 = await readRegistry();
     const results = [];
     const survivors = [];
-    for (const record4 of registry3.records) {
-      if (record4.supervisorNonce === args.liveNonce) {
-        survivors.push(record4);
+    for (const record5 of registry3.records) {
+      if (record5.supervisorNonce === args.liveNonce) {
+        survivors.push(record5);
         results.push({
-          sessionId: record4.sessionId,
+          sessionId: record5.sessionId,
           outcome: "skipped-live-supervisor"
         });
         continue;
       }
       if (!supportsOwnershipProof(platform)) {
-        survivors.push(record4);
+        survivors.push(record5);
         results.push({
-          sessionId: record4.sessionId,
+          sessionId: record5.sessionId,
           outcome: "skipped-unprovable"
         });
         continue;
       }
-      if (!await owningSupervisorProvablyExited(record4, platform)) {
-        survivors.push(record4);
+      if (!await owningSupervisorProvablyExited(record5, platform)) {
+        survivors.push(record5);
         results.push({
-          sessionId: record4.sessionId,
+          sessionId: record5.sessionId,
           outcome: "skipped-live-supervisor"
         });
         continue;
       }
-      const rootProbe = await probeProcess(record4.rootPid, platform);
+      const rootProbe = await probeProcess(record5.rootPid, platform);
       if (rootProbe.state === "unknown") {
-        survivors.push(record4);
+        survivors.push(record5);
         results.push({
-          sessionId: record4.sessionId,
+          sessionId: record5.sessionId,
           outcome: "skipped-unprovable"
         });
         continue;
       }
       if (rootProbe.state === "gone") {
-        const group = await probeProcessGroup(record4.rootPid, platform);
+        const group = await probeProcessGroup(record5.rootPid, platform);
         if (group === "live") {
-          survivors.push(record4);
-          results.push({ sessionId: record4.sessionId, outcome: "escaped" });
+          survivors.push(record5);
+          results.push({ sessionId: record5.sessionId, outcome: "escaped" });
           logger.warn(
             "[local-harness] descendants outlived their root; record kept",
-            { sessionId: record4.sessionId }
+            { sessionId: record5.sessionId }
           );
           continue;
         }
         if (group === "unknown") {
-          survivors.push(record4);
+          survivors.push(record5);
           results.push({
-            sessionId: record4.sessionId,
+            sessionId: record5.sessionId,
             outcome: "skipped-unprovable"
           });
           continue;
         }
-        await removeSessionState(record4.sessionStateDir);
-        results.push({ sessionId: record4.sessionId, outcome: "already-gone" });
+        await removeSessionState(record5.sessionStateDir);
+        results.push({ sessionId: record5.sessionId, outcome: "already-gone" });
         continue;
       }
-      if (rootProbe.identity !== record4.processBirthIdentity) {
-        survivors.push(record4);
-        results.push({ sessionId: record4.sessionId, outcome: "not-owned" });
+      if (rootProbe.identity !== record5.processBirthIdentity) {
+        survivors.push(record5);
+        results.push({ sessionId: record5.sessionId, outcome: "not-owned" });
         logger.warn("[local-harness] refusing to reclaim a reused pid", {
-          sessionId: record4.sessionId
+          sessionId: record5.sessionId
         });
         continue;
       }
       const outcome = await terminateOwnedProcessGroup({
-        pid: record4.rootPid,
-        birthIdentity: record4.processBirthIdentity,
+        pid: record5.rootPid,
+        birthIdentity: record5.processBirthIdentity,
         graceMs: args.graceMs ?? 5e3,
         platform
       });
       if (outcome.outcome === "graceful" || outcome.outcome === "forced") {
-        await removeSessionState(record4.sessionStateDir);
-        results.push({ sessionId: record4.sessionId, outcome: "terminated" });
+        await removeSessionState(record5.sessionStateDir);
+        results.push({ sessionId: record5.sessionId, outcome: "terminated" });
         continue;
       }
       if (outcome.outcome === "already-gone") {
-        await removeSessionState(record4.sessionStateDir);
-        results.push({ sessionId: record4.sessionId, outcome: "already-gone" });
+        await removeSessionState(record5.sessionStateDir);
+        results.push({ sessionId: record5.sessionId, outcome: "already-gone" });
         continue;
       }
-      survivors.push(record4);
+      survivors.push(record5);
       results.push({
-        sessionId: record4.sessionId,
+        sessionId: record5.sessionId,
         outcome: janitorOutcomeForUnsettled(outcome.outcome)
       });
     }
@@ -137569,7 +138437,7 @@ var LocalHarnessSupervisor = class {
       );
     }
     if (request3.role === "root") {
-      const record4 = {
+      const record5 = {
         sessionId: request3.sessionId,
         supervisorNonce: this.nonce,
         supervisorPid: process.pid,
@@ -137585,7 +138453,7 @@ var LocalHarnessSupervisor = class {
         sessionStateDir: request3.sessionStateDir
       };
       try {
-        await recordProcess(record4);
+        await recordProcess(record5);
         await updateLifecycleState(request3.sessionId, "running");
       } catch (error) {
         await abandon();
@@ -137832,48 +138700,48 @@ var LocalHarnessSupervisor = class {
 // server/utils/harness/local/session-registry.ts
 var sessions2 = /* @__PURE__ */ new Map();
 var unstopped = /* @__PURE__ */ new Set();
-function registerLocalHarnessSession(record4) {
-  sessions2.set(record4.sessionId, record4);
+function registerLocalHarnessSession(record5) {
+  sessions2.set(record5.sessionId, record5);
 }
 function forgetLocalHarnessSession(sessionId) {
   sessions2.delete(sessionId);
 }
-function forgetLocalHarnessSessionRecord(record4) {
-  unstopped.delete(record4);
-  if (sessions2.get(record4.sessionId) !== record4) return false;
-  sessions2.delete(record4.sessionId);
+function forgetLocalHarnessSessionRecord(record5) {
+  unstopped.delete(record5);
+  if (sessions2.get(record5.sessionId) !== record5) return false;
+  sessions2.delete(record5.sessionId);
   return true;
 }
-function claimRecord(record4) {
-  const claimed = sessions2.get(record4.sessionId) === record4 || unstopped.has(record4);
+function claimRecord(record5) {
+  const claimed = sessions2.get(record5.sessionId) === record5 || unstopped.has(record5);
   if (!claimed) return false;
-  if (sessions2.get(record4.sessionId) === record4) {
-    sessions2.delete(record4.sessionId);
+  if (sessions2.get(record5.sessionId) === record5) {
+    sessions2.delete(record5.sessionId);
   }
-  unstopped.delete(record4);
+  unstopped.delete(record5);
   return true;
 }
-async function endRecord(record4) {
-  const sessionId = record4.sessionId;
+async function endRecord(record5) {
+  const sessionId = record5.sessionId;
   const errors = [];
   try {
-    record4.gateway?.revoke();
+    record5.gateway?.revoke();
   } catch (error) {
     errors.push(`gateway revoke: ${messageOf(error)}`);
   }
   try {
-    await record4.gateway?.close();
+    await record5.gateway?.close();
   } catch (error) {
     errors.push(`gateway close: ${messageOf(error)}`);
   }
   try {
-    await record4.revokeLease?.();
+    await record5.revokeLease?.();
   } catch (error) {
     errors.push(`lease revoke: ${messageOf(error)}`);
   }
   let stopped = true;
   try {
-    const outcome = await record4.stop();
+    const outcome = await record5.stop();
     if (outcome?.stopped !== true) {
       stopped = false;
       errors.push(
@@ -137885,15 +138753,15 @@ async function endRecord(record4) {
     errors.push(`stop: ${messageOf(error)}`);
   }
   if (stopped) {
-    unstopped.delete(record4);
+    unstopped.delete(record5);
     try {
-      await record4.releaseRuntime?.();
+      await record5.releaseRuntime?.();
     } catch (error) {
       errors.push(`runtime release: ${messageOf(error)}`);
     }
   } else {
-    unstopped.add(record4);
-    if (!sessions2.has(sessionId)) sessions2.set(sessionId, record4);
+    unstopped.add(record5);
+    if (!sessions2.has(sessionId)) sessions2.set(sessionId, record5);
   }
   if (errors.length > 0) {
     logger.warn("[local-harness] session teardown had failures", {
@@ -137908,7 +138776,7 @@ async function stopAllLocalHarnessSessions() {
   const mine = candidates.filter(claimRecord);
   const results = await Promise.all(
     mine.map(
-      (record4) => endRecord(record4).catch(() => ({
+      (record5) => endRecord(record5).catch(() => ({
         stopped: false,
         errors: ["unexpected"]
       }))
@@ -140193,12 +141061,12 @@ async function fetchHarnessProxyTokens(args) {
 }
 function readHarnessEvidenceDecision(raw) {
   if (!raw || typeof raw !== "object" || Array.isArray(raw)) return void 0;
-  const record4 = raw;
-  if (typeof record4.captureEnabled !== "boolean") return void 0;
-  const gradingSource = record4.gradingSource === "evidence" ? "evidence" : "narration";
+  const record5 = raw;
+  if (typeof record5.captureEnabled !== "boolean") return void 0;
+  const gradingSource = record5.gradingSource === "evidence" ? "evidence" : "narration";
   return {
-    captureEnabled: record4.captureEnabled,
-    gradingSource: record4.captureEnabled ? gradingSource : "narration"
+    captureEnabled: record5.captureEnabled,
+    gradingSource: record5.captureEnabled ? gradingSource : "narration"
   };
 }
 
@@ -140351,8 +141219,8 @@ function stableHarnessValue(value) {
   if (Array.isArray(value)) {
     return `[${value.map(stableHarnessValue).join(",")}]`;
   }
-  const record4 = value;
-  return `{${Object.keys(record4).sort().map((key) => `${JSON.stringify(key)}:${stableHarnessValue(record4[key])}`).join(",")}}`;
+  const record5 = value;
+  return `{${Object.keys(record5).sort().map((key) => `${JSON.stringify(key)}:${stableHarnessValue(record5[key])}`).join(",")}}`;
 }
 function matchHarnessScopeStepUpToolCall(args) {
   const { observed, challenge } = args;
@@ -145595,11 +146463,11 @@ function unwrapJsonEnvelope(value) {
     if (!current || typeof current !== "object" || Array.isArray(current)) {
       return current;
     }
-    const record4 = current;
-    if (record4.type !== "json" || !("value" in record4)) {
+    const record5 = current;
+    if (record5.type !== "json" || !("value" in record5)) {
       return current;
     }
-    current = record4.value;
+    current = record5.value;
   }
   return current;
 }
@@ -145608,11 +146476,11 @@ function isModelVisibleImageOutput(value) {
   if (!output || typeof output !== "object" || Array.isArray(output)) {
     return false;
   }
-  const record4 = output;
-  if (record4.type !== "content" || !Array.isArray(record4.value)) {
+  const record5 = output;
+  if (record5.type !== "content" || !Array.isArray(record5.value)) {
     return false;
   }
-  return record4.value.some((part) => {
+  return record5.value.some((part) => {
     if (!part || typeof part !== "object" || Array.isArray(part)) {
       return false;
     }
@@ -146707,6 +147575,10 @@ async function processOneStep(ctx) {
     }
   }
   Object.assign(convexHeaders2, guestIpForwardHeaders(ipHash));
+  const scenarioServiceToken = process.env.INSPECTOR_SERVICE_TOKEN?.trim();
+  if (scenarioId && scenarioServiceToken) {
+    convexHeaders2["x-inspector-service-token"] = scenarioServiceToken;
+  }
   let res;
   onModelHandover?.();
   try {
@@ -148402,23 +149274,23 @@ function stampMcpToolOriginProviderOptions(messages, tools5) {
         if (!part || typeof part !== "object" || Array.isArray(part)) {
           return part;
         }
-        const record4 = part;
-        if (record4.type !== "tool-call" && record4.type !== "tool-result") {
+        const record5 = part;
+        if (record5.type !== "tool-call" && record5.type !== "tool-result") {
           return part;
         }
-        const toolName3 = record4.toolName;
+        const toolName3 = record5.toolName;
         if (typeof toolName3 !== "string") return part;
         const serverId = readToolServerId(tools5, toolName3);
         const providerOptions = mergePageToolBindingMetadata(
           withPageToolAttributionMetadata(
-            mergeMcpToolOriginMetadata(record4.providerOptions, serverId),
+            mergeMcpToolOriginMetadata(record5.providerOptions, serverId),
             tools5[toolName3]
           ),
-          record4.type === "tool-call" ? pageToolBindingOf(tools5[toolName3]) : void 0
+          record5.type === "tool-call" ? pageToolBindingOf(tools5[toolName3]) : void 0
         );
         if (!providerOptions) return part;
         messageChanged = true;
-        return { ...record4, providerOptions };
+        return { ...record5, providerOptions };
       }
     );
     if (!messageChanged) return message;
@@ -152397,8 +153269,37 @@ var PlatformApiError = class extends SdkError3 {
 function isPlatformApiError(error) {
   return error instanceof PlatformApiError;
 }
+var REFUSAL_REASON_PATTERN = /^[a-z0-9_]{1,64}$/;
+function describePlatformRefusal(error) {
+  if (!isPlatformApiError(error)) return void 0;
+  if (error.status !== 429 && error.code !== "RATE_LIMITED") return void 0;
+  const details = error.details ?? {};
+  const text2 = (key) => {
+    const value = details[key];
+    return typeof value === "string" && REFUSAL_REASON_PATTERN.test(value) ? value : void 0;
+  };
+  const flag = (key) => typeof details[key] === "boolean" ? details[key] : void 0;
+  const retryAfterMs = details.retryAfterMs;
+  const retryAfterSeconds = error.retryAfter !== void 0 ? error.retryAfter : typeof retryAfterMs === "number" && Number.isFinite(retryAfterMs) && retryAfterMs >= 0 ? Math.ceil(retryAfterMs / 1e3) : void 0;
+  const refusal = { status: error.status, code: error.code };
+  const reason = text2("code");
+  if (reason) refusal.reason = reason;
+  const gatedBy = text2("gatedBy");
+  if (gatedBy) refusal.gatedBy = gatedBy;
+  const canTopUp = flag("canTopUp");
+  if (canTopUp !== void 0) refusal.canTopUp = canTopUp;
+  const retryable = flag("isRetryable");
+  if (retryable !== void 0) refusal.retryable = retryable;
+  if (retryAfterSeconds !== void 0)
+    refusal.retryAfterSeconds = retryAfterSeconds;
+  return refusal;
+}
+function platformRefusalHint(refusal) {
+  const when = refusal.retryAfterSeconds !== void 0 ? `Retry after ${refusal.retryAfterSeconds}s, not sooner.` : "Wait before retrying; do not retry in a loop.";
+  return refusal.canTopUp === false ? `${when} This is a usage limit: topping up credits does not lift it.` : when;
+}
 function readSdkVersion3() {
-  return "8.9.0";
+  return "8.10.0";
 }
 var DEFAULT_PLATFORM_API_BASE_URL = "https://app.mcpjam.com/api/v1";
 function resolvePlatformRequestUrl(spec) {
@@ -157633,8 +158534,8 @@ var EVAL_VERDICT_DECISION_REASONS3 = [
   ...EVAL_VALIDITY_DECISION_REASONS3,
   ...EVAL_TASK_DECISION_REASONS3
 ];
-var evalVerdictDecisionReasonSchema3 = z24.enum(
-  EVAL_VERDICT_DECISION_REASONS3
+var evalVerdictDecisionReasonSchema3 = z24.enum(EVAL_VERDICT_DECISION_REASONS3).describe(
+  "Why the verdict is what it is. The validity reasons are evaluated FIRST and make a run `inconclusive`; the task reasons decide `passed` / `failed` once validity holds. `configuredTrialsNotAttempted` \u2014 some configured trial never ran. `noGradeableTrials` \u2014 nothing in the suite was gradeable. `eligibleTrialsBelowMinimum` \u2014 the explicit `minEligibleTrials` was not reached. `completionRateBelowMinimum` \u2014 measured, and under the floor. `completionRateNotMeasured` \u2014 nothing was attempted, so the floor is unsatisfiable; a not-measured rate never passes one. `evaluatorErrorRateAboveMaximum` \u2014 the grader failed too often for the run to describe the server. `evaluatorErrorRateNotMeasured` \u2014 the same unsatisfiable case for the ceiling. `caseHasNoEligibleTrials` \u2014 a case graded nothing, which is inconclusive even at `passThreshold: 0`. `casePassRateMetThreshold` \u2014 a case's own passing reason. `casePassRateBelowThreshold` \u2014 a case failed its threshold, and so therefore did the suite. `allMeasuredCasesMetThreshold` \u2014 the suite's only passing reason."
 );
 function isEvalValidityDecisionReason3(value) {
   return typeof value === "string" && EVAL_VALIDITY_DECISION_REASONS3.includes(value);
@@ -159241,10 +160142,10 @@ function assembleEvalRunDecisionSummary2(input) {
 }
 function legacyTrialCounts2(summary) {
   if (!summary || typeof summary !== "object") return void 0;
-  const count = (value) => typeof value === "number" && Number.isInteger(value) && value >= 0 ? value : void 0;
-  const total = count(summary.total);
-  const passed2 = count(summary.passed);
-  const failed3 = count(summary.failed);
+  const count22 = (value) => typeof value === "number" && Number.isInteger(value) && value >= 0 ? value : void 0;
+  const total = count22(summary.total);
+  const passed2 = count22(summary.passed);
+  const failed3 = count22(summary.failed);
   if (total === void 0 && passed2 === void 0 && failed3 === void 0) {
     return void 0;
   }
@@ -159370,8 +160271,8 @@ function nonEmptyString2(value) {
 }
 function toolName2(value) {
   if (!value || typeof value !== "object") return void 0;
-  const record4 = value;
-  return nonEmptyString2(record4.toolName) ?? nonEmptyString2(record4.tool) ?? nonEmptyString2(record4.name);
+  const record5 = value;
+  return nonEmptyString2(record5.toolName) ?? nonEmptyString2(record5.tool) ?? nonEmptyString2(record5.name);
 }
 function toolNames2(calls) {
   const names = (calls ?? []).map(toolName2).filter((name15) => name15 !== void 0);
@@ -160503,6 +161404,416 @@ var evalBacktestContinuationSchema3 = z24.object({
 evalBacktestDraftSchema3.extend({
   continuation: evalBacktestContinuationSchema3.optional()
 });
+var SWARM_SESSION_VERDICT_CONTRACT_VERSION3 = 1;
+var SWARM_ATTEMPT_STATUSES3 = [
+  "pending",
+  "running",
+  "succeeded",
+  "failed",
+  "rate_limited"
+];
+var SWARM_SESSION_LIFECYCLES3 = [
+  "pending",
+  "running",
+  "ran",
+  "broke",
+  "limited",
+  "withdrawn"
+];
+var SWARM_SESSION_VERDICTS3 = [
+  "passed",
+  "failed",
+  "inconclusive",
+  "notEstablished"
+];
+var SWARM_GRADING_STATES3 = [
+  "notRequested",
+  "queued",
+  "running",
+  "settled",
+  "unavailable"
+];
+var SWARM_SESSION_VERDICT_OF_REASON3 = {
+  attemptPending: "notEstablished",
+  attemptRunning: "notEstablished",
+  withdrawn: "notEstablished",
+  spendCapReached: "notEstablished",
+  notRun: "notEstablished",
+  executionFailed: "notEstablished",
+  ungraded: "notEstablished",
+  gradingNotClaimed: "notEstablished",
+  criteriaPending: "notEstablished",
+  judgePending: "notEstablished",
+  gatingCriterionFailed: "failed",
+  judgeFailed: "failed",
+  criteriaGradingErrored: "inconclusive",
+  gatingCriterionUnmeasured: "inconclusive",
+  judgeErrored: "inconclusive",
+  gradingUnavailable: "inconclusive",
+  allGatingCriteriaPassed: "passed",
+  judgePassed: "passed",
+  allGradersPassed: "passed"
+};
+var SWARM_SESSION_VERDICT_REASONS3 = Object.keys(
+  SWARM_SESSION_VERDICT_OF_REASON3
+);
+var swarmSessionLifecycleSchema3 = z24.enum(SWARM_SESSION_LIFECYCLES3);
+var swarmSessionVerdictValueSchema3 = z24.enum(SWARM_SESSION_VERDICTS3);
+var swarmGraderRoleSchema3 = z24.enum(["advisory", "required"]);
+var swarmSessionAttemptInputSchema3 = z24.object({
+  status: z24.enum(SWARM_ATTEMPT_STATUSES3),
+  errorCode: z24.string().nullable().optional()
+}).strict();
+var swarmGradingReadinessSchema3 = z24.object({
+  /** Durable readiness for decisive grading only, not background observations. */
+  state: z24.enum(SWARM_GRADING_STATES3),
+  reasonCode: z24.string().min(1).optional()
+}).strict();
+var criterionResultSchema3 = z24.object({
+  criterionId: z24.string().min(1),
+  passed: z24.boolean(),
+  status: z24.enum(["scored", "error"]).optional()
+}).strict();
+z24.object({
+  attempt: swarmSessionAttemptInputSchema3.nullable(),
+  hasTranscript: z24.boolean(),
+  rubric: z24.array(
+    z24.object({
+      id: z24.string().min(1),
+      role: swarmGraderRoleSchema3,
+      predicateType: z24.string().min(1).optional()
+    }).strict()
+  ),
+  criteria: z24.object({
+    status: z24.enum(["pending", "completed", "failed"]),
+    criterionIds: z24.array(z24.string().min(1)).optional(),
+    results: z24.array(criterionResultSchema3).optional()
+  }).strict().nullable(),
+  goalScore: z24.discriminatedUnion("status", [
+    z24.object({ status: z24.literal("completed"), passed: z24.boolean() }).strict(),
+    z24.object({ status: z24.literal("running") }).strict(),
+    z24.object({ status: z24.literal("failed") }).strict()
+  ]).nullable(),
+  judge: z24.object({
+    automatic: z24.boolean(),
+    role: swarmGraderRoleSchema3,
+    /** Explicit on-demand intent; automatic:false alone does not mean a judge was requested. */
+    requested: z24.boolean().optional()
+  }).strict(),
+  grading: swarmGradingReadinessSchema3
+}).strict().superRefine((value, ctx) => {
+  const definitions = new Set(value.rubric.map((entry3) => entry3.id));
+  if (definitions.size !== value.rubric.length) {
+    ctx.addIssue({
+      code: "custom",
+      path: ["rubric"],
+      message: "Duplicate criterion definition"
+    });
+  }
+  for (const [field, ids] of [
+    ["criterionIds", value.criteria?.criterionIds ?? []],
+    ["results", value.criteria?.results?.map((row2) => row2.criterionId) ?? []]
+  ]) {
+    if (new Set(ids).size !== ids.length || ids.some((id3) => !definitions.has(id3))) {
+      ctx.addIssue({
+        code: "custom",
+        path: ["criteria", field],
+        message: "Criterion IDs must be unique and defined in the snapshot"
+      });
+    }
+  }
+  const claimed = value.criteria?.criterionIds;
+  if (claimed && value.criteria?.results?.some((row2) => !claimed.includes(row2.criterionId))) {
+    ctx.addIssue({
+      code: "custom",
+      path: ["criteria", "results"],
+      message: "Result outside claimed scope"
+    });
+  }
+});
+var SWARM_SESSION_TRIAL_STATUSES3 = [
+  "pending",
+  "running",
+  "completed",
+  "failed",
+  "setup_failed",
+  "cancelled"
+];
+var swarmSessionTrialSchema3 = z24.object({
+  status: z24.enum(SWARM_SESSION_TRIAL_STATUSES3),
+  taskVerdict: z24.enum(["passed", "failed"]).optional(),
+  evaluatorError: z24.literal(true).optional()
+}).strict().superRefine((value, ctx) => {
+  const verdict = value.taskVerdict !== void 0;
+  const error = value.evaluatorError === true;
+  if (value.status === "completed" ? verdict === error : verdict || error) {
+    ctx.addIssue({
+      code: "custom",
+      message: "Only completed trials carry exactly one verdict or evaluator error"
+    });
+  }
+});
+var swarmSessionGraderCountsSchema3 = z24.object({
+  gating: z24.number().int().nonnegative(),
+  gatingPassed: z24.number().int().nonnegative(),
+  gatingFailed: z24.number().int().nonnegative(),
+  gatingUnmeasured: z24.number().int().nonnegative(),
+  advisoryFailed: z24.number().int().nonnegative()
+}).strict();
+var swarmSessionVerdictSchema3 = z24.object({
+  contractVersion: z24.literal(SWARM_SESSION_VERDICT_CONTRACT_VERSION3),
+  lifecycle: swarmSessionLifecycleSchema3,
+  verdict: swarmSessionVerdictValueSchema3,
+  reason: z24.enum(SWARM_SESSION_VERDICT_REASONS3),
+  verdictSource: z24.enum([
+    "goalJudge",
+    "requiredAssertions",
+    "combined",
+    "none"
+  ]),
+  grading: swarmGradingReadinessSchema3,
+  graders: z24.object({
+    criteria: z24.enum([
+      "notConfigured",
+      "notClaimed",
+      "pending",
+      "scored",
+      "errored"
+    ]),
+    judge: z24.enum([
+      "notConfigured",
+      "silent",
+      "pending",
+      "scored",
+      "errored"
+    ])
+  }).strict(),
+  counts: swarmSessionGraderCountsSchema3,
+  trial: swarmSessionTrialSchema3.nullable()
+}).strict().superRefine((value, ctx) => {
+  const fail3 = (message) => ctx.addIssue({ code: "custom", message });
+  if (SWARM_SESSION_VERDICT_OF_REASON3[value.reason] !== value.verdict)
+    fail3("Reason contradicts verdict");
+  const c = value.counts;
+  if (c.gating !== c.gatingPassed + c.gatingFailed + c.gatingUnmeasured)
+    fail3("Required measurement counts do not add up");
+  const trial = value.trial;
+  if (trial?.status === "completed") {
+    if (value.lifecycle !== "ran")
+      fail3("Only a completed execution can be a completed trial");
+    if (trial.taskVerdict !== void 0 && trial.taskVerdict !== value.verdict)
+      fail3("Trial contradicts goal result");
+    if (trial.evaluatorError && value.verdict !== "inconclusive")
+      fail3("Evaluator error requires inconclusive grading");
+  }
+  const expected = {
+    pending: "pending",
+    running: "running",
+    broke: "failed",
+    limited: "setup_failed",
+    withdrawn: "cancelled"
+  };
+  if (value.lifecycle !== "ran" && trial?.status !== expected[value.lifecycle])
+    fail3("Trial must preserve execution lifecycle");
+  if (value.lifecycle === "ran" && trial !== null && trial.status !== "completed")
+    fail3("Completed execution must not become skipped or running");
+  if ((value.verdict === "passed" || value.verdict === "failed") && value.verdictSource === "none")
+    fail3("A measured goal verdict needs a source");
+  if (value.lifecycle === "ran" && value.verdict !== "notEstablished" && trial === null)
+    fail3("Settled grading on completed execution needs an observation");
+  if (value.verdict === "notEstablished" && value.verdictSource !== "none")
+    fail3("An undecided goal has no verdict source");
+});
+var SWARM_REPORT_CONTRACT_VERSION3 = 1;
+var count3 = z24.number().int().nonnegative();
+var journeyRunVerdictSummarySchema3 = z24.discriminatedUnion("status", [
+  z24.object({
+    status: z24.literal("pending"),
+    pendingSessions: count3,
+    updatedAt: count3
+  }).strict(),
+  z24.object({
+    status: z24.literal("decided"),
+    decision: evalVerdictDecisionSchema3,
+    updatedAt: count3
+  }).strict(),
+  z24.object({
+    status: z24.literal("notEstablished"),
+    reason: z24.literal("gradingNotConfigured"),
+    updatedAt: count3
+  }).strict(),
+  z24.object({
+    status: z24.literal("integrityFailed"),
+    reason: z24.string().min(1),
+    updatedAt: count3
+  }).strict()
+]);
+var swarmObservationInputSchema3 = z24.object({
+  evaluatorId: z24.string().min(1),
+  predicateType: z24.enum(
+    Object.keys(ASSERTION_STAGE3)
+  ),
+  role: z24.enum(["advisory", "required"]),
+  /** Missing capture/error is unavailable; it must not use the compatibility passed:false. */
+  status: z24.enum(["passed", "failed", "pending", "unavailable"])
+}).strict();
+var swarmReportSessionSchema3 = z24.object({
+  /** Stable attempt/slot identity, including slots with no session document. */
+  id: z24.string().min(1),
+  /** A producer fact. Grade coverage and lifecycle alone cannot establish a non-start. */
+  startEvidence: z24.enum(["started", "notStarted", "unknown"]),
+  verdict: swarmSessionVerdictSchema3,
+  observations: z24.array(swarmObservationInputSchema3)
+}).strict().superRefine((row2, ctx) => {
+  if (new Set(row2.observations.map((o) => o.evaluatorId)).size !== row2.observations.length) {
+    ctx.addIssue({
+      code: "custom",
+      path: ["observations"],
+      message: "Duplicate evaluator in one session"
+    });
+  }
+  if (row2.startEvidence === "notStarted" && ["running", "ran"].includes(row2.verdict.lifecycle)) {
+    ctx.addIssue({
+      code: "custom",
+      path: ["startEvidence"],
+      message: "Execution contradicts explicit non-start"
+    });
+  }
+});
+z24.object({
+  runId: z24.string().min(1),
+  executionComplete: z24.boolean(),
+  configuredSessions: count3,
+  verdictSummary: journeyRunVerdictSummarySchema3.nullable(),
+  /** Frozen run-wide definitions: absent measurement rows still count as unavailable. */
+  evaluatorDefinitions: z24.array(
+    swarmObservationInputSchema3.omit({ status: true })
+  ),
+  sessions: z24.array(swarmReportSessionSchema3)
+}).strict().superRefine((input, ctx) => {
+  if (input.sessions.length > input.configuredSessions) {
+    ctx.addIssue({
+      code: "custom",
+      path: ["sessions"],
+      message: "Reported sessions exceed configured population"
+    });
+  }
+  if (new Set(input.sessions.map((row2) => row2.id)).size !== input.sessions.length) {
+    ctx.addIssue({
+      code: "custom",
+      path: ["sessions"],
+      message: "Duplicate session/attempt identity"
+    });
+  }
+  const definitions = new Map(
+    input.evaluatorDefinitions.map((row2) => [row2.evaluatorId, row2])
+  );
+  if (definitions.size !== input.evaluatorDefinitions.length) {
+    ctx.addIssue({
+      code: "custom",
+      path: ["evaluatorDefinitions"],
+      message: "Duplicate evaluator definition"
+    });
+  }
+  for (const session of input.sessions) {
+    for (const observation of session.observations) {
+      const definition = definitions.get(observation.evaluatorId);
+      if (!definition || definition.role !== observation.role || definition.predicateType !== observation.predicateType) {
+        ctx.addIssue({
+          code: "custom",
+          path: ["sessions"],
+          message: "Observation contradicts the frozen evaluator definition"
+        });
+      }
+    }
+  }
+});
+var swarmObservationCoverageSchema3 = z24.object({
+  evaluatorId: z24.string().min(1),
+  predicateType: z24.enum(
+    Object.keys(ASSERTION_STAGE3)
+  ),
+  role: z24.enum(["advisory", "required"]),
+  stage: z24.enum(USER_VALUE_STAGES3).describe(
+    "The stage this observation was read from, and the only stage it is evidence ABOUT. `connection` \u2014 the server was reachable and the session initialized. `discovery` \u2014 its tools and resources were listed and readable. `selection` \u2014 the model chose the right tool for the request. `call` \u2014 the call was made with usable arguments. `response` \u2014 the server returned data the model could use. `userValue` \u2014 the user's actual request was satisfied."
+  ),
+  unit: z24.literal("sessions"),
+  total: count3,
+  passed: count3,
+  failed: count3,
+  pending: count3,
+  unavailable: count3
+}).strict().superRefine((value, ctx) => {
+  if (value.total !== value.passed + value.failed + value.pending + value.unavailable) {
+    ctx.addIssue({
+      code: "custom",
+      message: "Observation coverage must partition its population"
+    });
+  }
+  if (value.stage !== ASSERTION_STAGE3[value.predicateType]) {
+    ctx.addIssue({
+      code: "custom",
+      message: "Observation stage contradicts the shared evaluator map"
+    });
+  }
+});
+z24.object({
+  contractVersion: z24.literal(SWARM_REPORT_CONTRACT_VERSION3),
+  runId: z24.string().min(1),
+  verdict: swarmSessionVerdictValueSchema3,
+  verdictSource: z24.enum(["policyV2", "none"]),
+  decision: evalVerdictDecisionSchema3.optional(),
+  undecidedReason: z24.enum([
+    "executionPending",
+    "gradingPending",
+    "gradingNotConfigured",
+    "verdictSummaryUnavailable",
+    "integrityFailed"
+  ]).optional(),
+  execution: z24.object({
+    unit: z24.literal("sessions"),
+    configured: count3,
+    reported: count3,
+    started: count3,
+    notStarted: count3,
+    unknown: count3,
+    completed: count3,
+    interrupted: count3,
+    /** Only explicit complete execution evidence can establish this. */
+    neverLaunched: z24.boolean()
+  }).strict(),
+  goalGrading: z24.object({
+    unit: z24.literal("sessions"),
+    reported: count3,
+    passed: count3,
+    failed: count3,
+    pending: count3,
+    unavailable: count3,
+    notRequested: count3,
+    /** Readiness is separate: a proven failure can exist while other required grading is pending. */
+    waitingForDecisiveGrading: count3
+  }).strict(),
+  observations: z24.array(swarmObservationCoverageSchema3)
+}).strict().superRefine((report, ctx) => {
+  const fail3 = (message) => ctx.addIssue({ code: "custom", message });
+  if (report.verdictSource === "policyV2") {
+    if (!report.decision || report.verdict !== report.decision.verdict || report.undecidedReason !== void 0)
+      fail3("Report must copy its authoritative decision");
+  } else if (report.verdict !== "notEstablished" || report.decision !== void 0 || !report.undecidedReason)
+    fail3("Missing authority must have an undecided reason");
+  const e = report.execution;
+  if (e.started + e.notStarted + e.unknown !== e.configured || e.reported > e.configured)
+    fail3("Execution evidence does not cover its configured population");
+  if (e.started + e.notStarted > e.reported || e.completed + e.interrupted > e.reported)
+    fail3("Execution counts exceed reported evidence");
+  if (e.neverLaunched !== (e.configured > 0 && e.reported === e.configured && e.notStarted === e.configured))
+    fail3("Not run requires explicit evidence for the entire population");
+  const g = report.goalGrading;
+  if (g.reported !== e.reported || g.passed + g.failed + g.pending + g.unavailable + g.notRequested !== g.reported || g.waitingForDecisiveGrading > g.reported)
+    fail3("Goal coverage must partition reported sessions");
+  if (new Set(report.observations.map((o) => o.evaluatorId)).size !== report.observations.length || report.observations.some((o) => o.total > e.reported))
+    fail3("Invalid observation population");
+});
 var DECISION_SUMMARY_FALLBACK_PAGE_LIMIT = 200;
 var DECISION_SUMMARY_FALLBACK_MAX_PAGES = 100;
 function buildEvalRunDecisionSummary(input) {
@@ -160869,13 +162180,13 @@ function primitiveStatusFromCheck(check) {
   return "loaded";
 }
 function coerceToolInfo(value) {
-  const record4 = objectRecord(value);
-  const name15 = record4 ? optionalString2(record4, "name") : void 0;
-  if (!record4 || !name15) {
+  const record5 = objectRecord(value);
+  const name15 = record5 ? optionalString2(record5, "name") : void 0;
+  if (!record5 || !name15) {
     return void 0;
   }
-  const title = optionalString2(record4, "title");
-  const description = optionalDescription(record4);
+  const title = optionalString2(record5, "title");
+  const description = optionalDescription(record5);
   return {
     name: name15,
     ...title ? { title } : {},
@@ -160883,15 +162194,15 @@ function coerceToolInfo(value) {
   };
 }
 function coerceResourceInfo(value) {
-  const record4 = objectRecord(value);
-  const uri = record4 ? optionalString2(record4, "uri") : void 0;
-  if (!record4 || !uri) {
+  const record5 = objectRecord(value);
+  const uri = record5 ? optionalString2(record5, "uri") : void 0;
+  if (!record5 || !uri) {
     return void 0;
   }
-  const name15 = optionalString2(record4, "name");
-  const title = optionalString2(record4, "title");
-  const description = optionalDescription(record4);
-  const mimeType = optionalString2(record4, "mimeType");
+  const name15 = optionalString2(record5, "name");
+  const title = optionalString2(record5, "title");
+  const description = optionalDescription(record5);
+  const mimeType = optionalString2(record5, "mimeType");
   return {
     uri,
     ...name15 ? { name: name15 } : {},
@@ -160901,14 +162212,14 @@ function coerceResourceInfo(value) {
   };
 }
 function coercePromptInfo(value) {
-  const record4 = objectRecord(value);
-  const name15 = record4 ? optionalString2(record4, "name") : void 0;
-  if (!record4 || !name15) {
+  const record5 = objectRecord(value);
+  const name15 = record5 ? optionalString2(record5, "name") : void 0;
+  if (!record5 || !name15) {
     return void 0;
   }
-  const promptArguments = Array.isArray(record4.arguments) ? record4.arguments.map(coercePromptArgumentInfo).filter(isPresent) : void 0;
-  const title = optionalString2(record4, "title");
-  const description = optionalDescription(record4);
+  const promptArguments = Array.isArray(record5.arguments) ? record5.arguments.map(coercePromptArgumentInfo).filter(isPresent) : void 0;
+  const title = optionalString2(record5, "title");
+  const description = optionalDescription(record5);
   return {
     name: name15,
     ...title ? { title } : {},
@@ -160917,25 +162228,25 @@ function coercePromptInfo(value) {
   };
 }
 function coercePromptArgumentInfo(value) {
-  const record4 = objectRecord(value);
-  const name15 = record4 ? optionalString2(record4, "name") : void 0;
-  if (!record4 || !name15) {
+  const record5 = objectRecord(value);
+  const name15 = record5 ? optionalString2(record5, "name") : void 0;
+  if (!record5 || !name15) {
     return void 0;
   }
-  const description = optionalDescription(record4);
+  const description = optionalDescription(record5);
   return {
     name: name15,
     ...description ? { description } : {},
-    ...typeof record4.required === "boolean" ? { required: record4.required } : {}
+    ...typeof record5.required === "boolean" ? { required: record5.required } : {}
   };
 }
 function coerceServerInfo(value) {
-  const record4 = objectRecord(value);
-  if (!record4) {
+  const record5 = objectRecord(value);
+  if (!record5) {
     return void 0;
   }
-  const name15 = optionalString2(record4, "name");
-  const version = optionalString2(record4, "version");
+  const name15 = optionalString2(record5, "name");
+  const version = optionalString2(record5, "version");
   if (!name15 && !version) {
     return void 0;
   }
@@ -160947,12 +162258,12 @@ function coerceServerInfo(value) {
 function objectRecord(value) {
   return value && typeof value === "object" && !Array.isArray(value) ? value : void 0;
 }
-function optionalString2(record4, key) {
-  const value = record4[key];
+function optionalString2(record5, key) {
+  const value = record5[key];
   return typeof value === "string" && value.trim().length > 0 ? value.trim() : void 0;
 }
-function optionalDescription(record4) {
-  const description = optionalString2(record4, "description");
+function optionalDescription(record5) {
+  const description = optionalString2(record5, "description");
   return description ? truncateText(description, MAX_PRIMITIVE_DESCRIPTION_LENGTH) : void 0;
 }
 function truncateText(value, maxLength) {
@@ -161310,27 +162621,27 @@ function evaluateHostCompat(requirements, profile) {
       (name15) => !requirements.appOnlyWidgets.includes(name15)
     );
     if (blockedAppOnly.length > 0) {
-      const count = blockedAppOnly.length;
+      const count22 = blockedAppOnly.length;
       findings.push({
         lane: "apps",
         severity: "blocker",
         code: "app_only_unrenderable",
         tools: blockedAppOnly,
-        title: `${count === 1 ? "Interactive tool" : `${count} interactive tools`} unavailable`,
-        detail: `${formatToolNames(blockedAppOnly)} only ${count === 1 ? "works" : "work"} inside a widget. ${profile.label} does not render ${count === 1 ? "it" : "them"}.`,
+        title: `${count22 === 1 ? "Interactive tool" : `${count22} interactive tools`} unavailable`,
+        detail: `${formatToolNames(blockedAppOnly)} only ${count22 === 1 ? "works" : "work"} inside a widget. ${profile.label} does not render ${count22 === 1 ? "it" : "them"}.`,
         remediation,
         provenance: profile.provenance
       });
     }
     if (degradedFallback.length > 0) {
-      const count = degradedFallback.length;
+      const count22 = degradedFallback.length;
       findings.push({
         lane: "apps",
         severity: "degraded",
         code: "widget_text_fallback",
         tools: degradedFallback,
-        title: `${count === 1 ? "Interactive view" : `${count} interactive views`} unavailable`,
-        detail: `${formatToolNames(degradedFallback)} provide${count === 1 ? "s" : ""} an interactive view. ${profile.label} shows the plain-text result instead.`,
+        title: `${count22 === 1 ? "Interactive view" : `${count22} interactive views`} unavailable`,
+        detail: `${formatToolNames(degradedFallback)} provide${count22 === 1 ? "s" : ""} an interactive view. ${profile.label} shows the plain-text result instead.`,
         remediation,
         provenance: profile.provenance
       });
@@ -167412,7 +168723,7 @@ var readinessRunScopedInput = z24.object({
 });
 var startReadinessInput = serverScopedInput.extend({
   includeLlmObservations: z24.boolean().optional().describe(
-    "Ask a model for optional experience observations. Included with MCPJam \u2014 no credits are consumed. Defaults false; the deterministic grade is complete without it."
+    "Ask a model for optional experience observations. Included with MCPJam; no customer credits consumed; subject to MCPJam's daily analysis budget. Defaults false; the deterministic grade is complete without it."
   ),
   idempotencyKey: z24.string().trim().min(1).max(200).optional().describe(
     "Replay guard. A retry carrying the same key returns the run it already started rather than dialling the target twice."
@@ -167426,7 +168737,7 @@ var startOpenAIReadinessInput = startReadinessInput.extend({
 var startClaudeReadinessRunOperation = {
   name: "start_claude_readiness_run",
   title: "Start a Claude directory readiness run",
-  description: "Grade a saved MCP server against Anthropic's connector-directory rules. Starts a durable run and returns its id \u2014 poll `get_readiness_run` for the verdict, which is NOT in this response. `includeLlmObservations` adds an optional model pass: included with MCPJam, so no credits are consumed either way.",
+  description: "Grade a saved MCP server against Anthropic's connector-directory rules. Starts a durable run and returns its id \u2014 poll `get_readiness_run` for the verdict, which is NOT in this response. `includeLlmObservations` adds an optional model pass: included with MCPJam, so no customer credits are consumed either way; it is subject to MCPJam's daily analysis budget.",
   readOnly: false,
   risk: "none",
   permalink: noPermalink(
@@ -167464,7 +168775,7 @@ var startClaudeReadinessRunOperation = {
 var startOpenAIReadinessRunOperation = {
   name: "start_openai_readiness_run",
   title: "Start an OpenAI directory readiness run",
-  description: "Grade a saved MCP server against OpenAI's app-directory rules. Requires an explicit `submissionMode` \u2014 it is never inferred, because guessing turns a missing input into a clean bill of health. Starts a durable run and returns its id; poll `get_readiness_run` for the verdict. `includeLlmObservations` adds an optional model pass: included with MCPJam, so no credits are consumed either way.",
+  description: "Grade a saved MCP server against OpenAI's app-directory rules. Requires an explicit `submissionMode` \u2014 it is never inferred, because guessing turns a missing input into a clean bill of health. Starts a durable run and returns its id; poll `get_readiness_run` for the verdict. `includeLlmObservations` adds an optional model pass: included with MCPJam, so no customer credits are consumed either way; it is subject to MCPJam's daily analysis budget.",
   readOnly: false,
   risk: "none",
   permalink: noPermalink(
@@ -169901,14 +171212,14 @@ var generateEvalCasesInput = z24.object({
     "Condition generated cases on a realistic range of user styles so the queries read like different users wrote them."
   ),
   idempotencyKey: z24.string().trim().min(1).max(256).optional().describe(
-    "Retry-safety key: pass one, because a retry must not take a second slice of the project's daily generation quota. Repeating a call with the same key replays the first attempt's drafts and returns the cases it already created."
+    "Retry-safety key: pass one, because a retry must not take a second slice of the organization's daily generation quota. Repeating a call with the same key replays the first attempt's drafts and returns the cases it already created."
   )
 });
 var generateEvalCasesOperation = {
   name: "generate_eval_cases",
   risk: "none",
   title: "Generate MCPJam eval cases",
-  description: "AI-generate test cases from the suite's server tools and persist them into the suite. Connects the servers to discover tools. Included with MCPJam \u2014 no credits are consumed; it counts against the project's daily generation quota. For a suite with attached project environments, tools are discovered from the environment's closed server set \u2014 pass environment to choose which one. The authoring model is platform-controlled; set caseModels to choose the generated cases' execution models. IDEMPOTENT on idempotencyKey: pass one, because a retry must not take a second slice of that quota.",
+  description: "AI-generate test cases from the suite's server tools and persist them into the suite. Connects the servers to discover tools. Included with MCPJam; no customer credits consumed; subject to usage limits: a per-minute burst limit and the organization's daily generation quota. A refusal is RATE_LIMITED with a retry time \u2014 wait until then; topping up credits does not lift it. For a suite with attached project environments, tools are discovered from the environment's closed server set \u2014 pass environment to choose which one. The authoring model is platform-controlled; set caseModels to choose the generated cases' execution models. IDEMPOTENT on idempotencyKey: pass one, because a retry must not take a second slice of that quota.",
   readOnly: false,
   permalink: derivePermalinks(
     (result) => result.created.flatMap(
@@ -170389,7 +171700,7 @@ var proposeEvalDescriptionRewriteInput = evalRunScopedInput.extend({
 var proposeEvalDescriptionRewriteOperation = {
   name: "propose_eval_description_rewrite",
   title: "Propose an eval description rewrite",
-  description: "Draft a rewritten tool description from a finished eval run's failed trials: the tool's current description and input schema, sibling tool names, the expected vs observed calls, and the failing prompts. Included with MCPJam \u2014 no credits are consumed; it counts against the project's daily generation quota. Returns immediately with a proposing receipt \u2014 poll get_eval_description_experiment until status is proposed (or failed). Does not launch runs, write a verdict, or change a gate. Report-only throughout.",
+  description: "Draft a rewritten tool description from a finished eval run's failed trials: the tool's current description and input schema, sibling tool names, the expected vs observed calls, and the failing prompts. Included with MCPJam; no customer credits consumed; subject to MCPJam's daily analysis budget. Returns immediately with a proposing receipt \u2014 poll get_eval_description_experiment until status is proposed (or failed). Does not launch runs, write a verdict, or change a gate. Report-only throughout.",
   readOnly: false,
   risk: "none",
   permalink: noPermalink("mutation-only"),
@@ -173343,7 +174654,7 @@ var journeyRunSessionsInput = journeyRunSelectorInput.extend({
 var listJourneyRunSessionsOperation = {
   name: "list_journey_run_sessions",
   title: "List the sessions a journey run produced",
-  description: "The chat sessions a journey run produced \u2014 one per persona attempt against each target \u2014 with readiness, goal scores and a first-message preview. Transcript bodies are not on this API yet; use the returned `id` in the app to open a session.",
+  description: "The chat sessions a journey run produced \u2014 one per persona attempt against each target \u2014 with graded verdicts, check observations, readiness, goal scores and a first-message preview. `verdict` is the graded goal result; `outcome` is execution lifecycle. A broken execution may have met its goal. Transcript bodies are not on this API yet; use the returned `id` in the app to open a session.",
   readOnly: true,
   permalink: derivePermalinks(
     (result) => result.items.map((session) => ({
@@ -174385,7 +175696,7 @@ var generatePersonasInput = generationGroundingInput.extend({
 var generatePersonasOperation = {
   name: "generate_personas",
   title: "Draft MCPJam personas with a model",
-  description: "Draft candidate personas grounded in what the project's servers actually do. NOTHING IS SAVED \u2014 pick what you want and pass it to create_persona. Runs a model, included with MCPJam \u2014 no credits are consumed; it counts against the project's daily generation quota. Exactly one of environmentId or serverAttachmentId.",
+  description: "Draft candidate personas grounded in what the project's servers actually do. NOTHING IS SAVED \u2014 pick what you want and pass it to create_persona. Runs a model. Included with MCPJam; no customer credits consumed; subject to usage limits: a per-minute burst limit and the organization's daily generation quota. A refusal is RATE_LIMITED with a retry time \u2014 wait until then; topping up credits does not lift it. Exactly one of environmentId or serverAttachmentId.",
   readOnly: false,
   risk: "none",
   permalink: noPermalink(
@@ -174426,7 +175737,7 @@ var generateJourneysInput = generationGroundingInput.extend({
 var generateJourneysOperation = {
   name: "generate_journeys",
   title: "Draft MCPJam journeys with a model",
-  description: "Draft candidate journeys for a persona, grounded in the project's servers. NOTHING IS SAVED \u2014 pass what you want to create_journey. Included with MCPJam \u2014 no credits are consumed; it counts against the project's daily generation quota. Exactly one of environmentId or serverAttachmentId.",
+  description: "Draft candidate journeys for a persona, grounded in the project's servers. NOTHING IS SAVED \u2014 pass what you want to create_journey. Included with MCPJam; no customer credits consumed; subject to usage limits: a per-minute burst limit and the organization's daily generation quota. A refusal is RATE_LIMITED with a retry time \u2014 wait until then; topping up credits does not lift it. Exactly one of environmentId or serverAttachmentId.",
   readOnly: false,
   risk: "none",
   permalink: noPermalink(
@@ -174571,10 +175882,11 @@ var waveSelectorInput = z24.object({
   project: z24.string().trim().min(1).optional().describe(PROJECT_SELECTOR_DESCRIPTION),
   wave: z24.string().trim().min(1).describe("Wave id \u2014 the `waveId` on a journey run.")
 });
+var INCLUDED_ANALYSIS_FAILURE_NOTE = "A failed analysis carries errorCode: `platform_cap_exceeded` means MCPJam's own daily budget for this analysis is used up \u2014 nothing was charged, it resets at 00:00 UTC, and neither re-requesting nor topping up credits helps before then; `platform_unavailable` means MCPJam could not reserve capacity \u2014 try again later, not in a loop.";
 var getWaveInsightsOperation = {
   name: "get_wave_insights",
   title: "Get an MCPJam wave's insights",
-  description: "The model's analysis of a whole wave, if one has been requested. Poll this after request_wave_insights \u2014 status goes pending \u2192 completed. Not-found means nobody has requested it, which is different from 'requested and still working'.",
+  description: "The model's analysis of a whole wave, if one has been requested. Poll this after request_wave_insights \u2014 status goes pending \u2192 completed. Not-found means nobody has requested it, which is different from 'requested and still working'. " + INCLUDED_ANALYSIS_FAILURE_NOTE,
   readOnly: true,
   permalink: derivePermalinks((result) => [
     // A wave IS a journey run on the Swarms surface: `/swarms/<waveId>`.
@@ -174605,7 +175917,7 @@ var requestWaveInsightsInput = waveSelectorInput.extend({
 var requestWaveInsightsOperation = {
   name: "request_wave_insights",
   title: "Request MCPJam wave insights",
-  description: "Ask a model to analyze a whole wave. Returns immediately with status pending; poll get_wave_insights. Included with MCPJam \u2014 no credits are consumed; it COUNTS against the organization's daily insight quota, which is SHARED with user-testing insights, so a request here takes one from there. Read the run scorecards first; they cost no quota and usually explain the failure.",
+  description: "Ask a model to analyze a whole wave. Returns immediately with status pending; poll get_wave_insights. Included with MCPJam; no customer credits consumed; subject to usage limits: it COUNTS against the organization's daily insight quota, which is SHARED with user-testing insights, so a request here takes one from there. Read the run scorecards first; they cost no quota and usually explain the failure.",
   readOnly: false,
   risk: "none",
   permalink: noPermalink("mutation-only"),
@@ -174918,7 +176230,7 @@ var userTestingWindowInput = userTestingScenarioSelectorInput.extend({
 var getUserTestingInsightsOperation = {
   name: "get_user_testing_insights",
   title: "Get a user-testing window's insights",
-  description: "The model's analysis of one analysis window, if one has been requested. Not-found means nobody has requested it, which is different from requested-and-still-working.",
+  description: "The model's analysis of one analysis window, if one has been requested. Not-found means nobody has requested it, which is different from requested-and-still-working. " + INCLUDED_ANALYSIS_FAILURE_NOTE,
   readOnly: true,
   permalink: noPermalink(
     "no-addressable-resource",
@@ -174951,7 +176263,7 @@ var requestUserTestingInsightsInput = userTestingScenarioSelectorInput.extend(
 var requestUserTestingInsightsOperation = {
   name: "request_user_testing_insights",
   title: "Request insights for a user-testing scenario",
-  description: "Ask a model to analyze the scenario's current window. Returns immediately with the windowId and status pending; poll get_user_testing_insights. Included with MCPJam \u2014 no credits are consumed; it COUNTS against the organization's daily insight quota, which is SHARED with swarm wave insights. A 409 means the window has not been mined yet \u2014 wait, do not retry in a loop.",
+  description: "Ask a model to analyze the scenario's current window. Returns immediately with the windowId and status pending; poll get_user_testing_insights. Included with MCPJam; no customer credits consumed; subject to usage limits: it COUNTS against the organization's daily insight quota, which is SHARED with swarm wave insights. A 409 means the window has not been mined yet \u2014 wait, do not retry in a loop.",
   readOnly: false,
   risk: "none",
   permalink: noPermalink("mutation-only"),
@@ -175975,6 +177287,13 @@ function capForModel(value) {
 }
 function toToolError(error, fallback) {
   const message = error instanceof Error && error.message.trim() ? error.message : "";
+  const refusal = describePlatformRefusal(error);
+  if (refusal) {
+    return {
+      error: `${message || fallback} ${platformRefusalHint(refusal)}`,
+      refusal
+    };
+  }
   return { error: message || fallback };
 }
 function buildMcpjamTool(id3, opts) {
@@ -176672,12 +177991,12 @@ function linkedResourceReaderForPart(part, fallbackServerId, options) {
 }
 function readServerIdFromToolResultPart(part) {
   if (!part || typeof part !== "object") return void 0;
-  const record4 = part;
-  const legacy = record4.serverId;
+  const record5 = part;
+  const legacy = record5.serverId;
   if (typeof legacy === "string" && legacy.length > 0) {
     return legacy;
   }
-  return readMcpToolOriginServerId(record4.providerOptions);
+  return readMcpToolOriginServerId(record5.providerOptions);
 }
 function readToolCallId(part) {
   if (!part || typeof part !== "object") return void 0;
@@ -176700,11 +178019,11 @@ function unwrapJsonToolOutput(output) {
     if (!current || typeof current !== "object" || Array.isArray(current)) {
       return current;
     }
-    const record4 = current;
-    if (record4.type !== "json" || !Object.prototype.hasOwnProperty.call(record4, "value")) {
+    const record5 = current;
+    if (record5.type !== "json" || !Object.prototype.hasOwnProperty.call(record5, "value")) {
       return current;
     }
-    current = record4.value;
+    current = record5.value;
   }
   return current;
 }
@@ -176731,14 +178050,14 @@ function readReplayableImageModelOutput(output, options) {
   if (!output || typeof output !== "object" || Array.isArray(output)) {
     return void 0;
   }
-  const record4 = output;
-  if (record4.type !== "content" || !Array.isArray(record4.value)) {
+  const record5 = output;
+  if (record5.type !== "content" || !Array.isArray(record5.value)) {
     return void 0;
   }
   let sawImageCandidate = false;
   let sawOmissionMarker = false;
   const value = [];
-  for (const part of record4.value) {
+  for (const part of record5.value) {
     if (!part || typeof part !== "object" || Array.isArray(part)) {
       return void 0;
     }
@@ -176786,10 +178105,10 @@ function readReplayableImageModelOutput(output, options) {
 }
 function stripInternalProviderOptions(part) {
   if (!part || typeof part !== "object" || Array.isArray(part)) return part;
-  const record4 = part;
-  if (!record4.providerOptions) return part;
-  const providerOptions = stripMcpToolOriginMetadata(record4.providerOptions);
-  const { providerOptions: _providerOptions, ...rest } = record4;
+  const record5 = part;
+  if (!record5.providerOptions) return part;
+  const providerOptions = stripMcpToolOriginMetadata(record5.providerOptions);
+  const { providerOptions: _providerOptions, ...rest } = record5;
   return providerOptions ? { ...rest, providerOptions } : rest;
 }
 async function mapMcpImageToolOutputs(messages, options = {}) {
@@ -177348,6 +178667,9 @@ chatV2.post("/", async (c) => {
             }
           );
           const failClosedMessage = `Couldn't load this scenario's settings, so the turn was stopped to avoid running with the wrong configuration. ${runtime.error}`;
+          if (runtime.code === "SCENARIO_SIGN_IN_REQUIRED") {
+            return c.json({ error: runtime.error, code: runtime.code }, 401);
+          }
           if (runtime.code === "SCENARIO_ACCESS_STALE") {
             return c.json(
               { error: failClosedMessage, code: "SCENARIO_ACCESS_STALE" },
@@ -177410,7 +178732,10 @@ chatV2.post("/", async (c) => {
       let enabled = typeof hostRuntimeConfig?.localBrowserEnabled === "boolean" ? hostRuntimeConfig.localBrowserEnabled : void 0;
       if (!hostRuntimeConfig && typeof body.projectId === "string" && body.projectId && c.req.header("authorization") && !isGuestChatRequest(c.req.header("authorization"))) {
         try {
-          enabled = await readLocalBrowserSetting(await getConvexBearerForRequest(c), body.projectId);
+          enabled = await readLocalBrowserSetting(
+            await getConvexBearerForRequest(c),
+            body.projectId
+          );
         } catch {
           localBrowserSettingsUnavailable = true;
         }
@@ -177657,9 +178982,7 @@ chatV2.post("/", async (c) => {
       localConsentValid
     });
     const localBrowserRequested = body.browserEngine === "local";
-    const browserRollout = !localBrowserSettingsUnavailable && resolvedExecution.builtInToolIds?.includes(
-      BROWSER_BUILT_IN_TOOL_ID
-    ) ? await resolveBrowserRollout(c, localBrowserRequested) : { enabled: false, actor: null };
+    const browserRollout = !localBrowserSettingsUnavailable && resolvedExecution.builtInToolIds?.includes(BROWSER_BUILT_IN_TOOL_ID) ? await resolveBrowserRollout(c, localBrowserRequested) : { enabled: false, actor: null };
     const localBrowserGuestId = localBrowserRequested && browserRollout.enabled && browserRollout.actor?.guest ? browserRollout.actor.id : void 0;
     const browserConsentToken = c.req.header(BROWSER_CONSENT_HEADER);
     const browserConsentValid = browserRollout.enabled && (!requestIsGuest || Boolean(localBrowserGuestId)) && !isScenarioSession && await verifyLocalBrowserConsent(browserConsentToken);
@@ -183025,14 +184348,14 @@ function safeHost(url) {
   }
 }
 function headersToRecord(input, init2) {
-  const record4 = {};
+  const record5 = {};
   const raw = init2?.headers ?? input?.headers;
   if (raw) {
     new Headers(raw).forEach((value, key) => {
-      record4[key] = value;
+      record5[key] = value;
     });
   }
-  return record4;
+  return record5;
 }
 function encodeBody(body) {
   if (body === null || body === void 0) return void 0;
@@ -184676,7 +185999,7 @@ async function createJourneyRun(convexHttpUrl3, bearer, args) {
       ...args.environmentIds?.length ? { environmentIds: args.environmentIds } : {},
       // Asserted by this process, never a caller: the runner is the only
       // honest source for what it can execute.
-      runnerCapabilities: [...runnerCapabilities()]
+      runnerCapabilities: [...runnerCapabilities(), "swarm-standard-checks-v1"]
     },
     NON_LLM_TIMEOUT_MS
   );
@@ -184726,7 +186049,8 @@ async function claimSwarmChecks(convexHttpUrl3, bearer, args, signal) {
     {
       projectId: args.projectId,
       runId: args.runId,
-      chatSessionId: args.chatSessionId
+      chatSessionId: args.chatSessionId,
+      checkCapability: "swarm-standard-checks-v1"
     },
     NON_LLM_TIMEOUT_MS,
     signal
@@ -187077,8 +188401,8 @@ var LAUNCH_REJECTION_STATUS = {
 function environmentLaunchRejectionError(error) {
   const data = error?.data;
   if (!data || typeof data !== "object" || Array.isArray(data)) return null;
-  const record4 = data;
-  const code = typeof record4.code === "string" ? record4.code : void 0;
+  const record5 = data;
+  const code = typeof record5.code === "string" ? record5.code : void 0;
   if (!code) return null;
   const status = LAUNCH_REJECTION_STATUS[code];
   if (!status) return null;
@@ -187088,8 +188412,8 @@ function environmentLaunchRejectionError(error) {
       reason: "env_not_found"
     });
   }
-  const message = typeof record4.message === "string" && record4.message.trim() ? record4.message : "This eval run could not be started as requested.";
-  const details = record4.details && typeof record4.details === "object" && !Array.isArray(record4.details) ? record4.details : {};
+  const message = typeof record5.message === "string" && record5.message.trim() ? record5.message : "This eval run could not be started as requested.";
+  const details = record5.details && typeof record5.details === "object" && !Array.isArray(record5.details) ? record5.details : {};
   return new WebRouteError(status, ErrorCode.VALIDATION_ERROR, message, {
     code,
     ...details,
@@ -187099,10 +188423,10 @@ function environmentLaunchRejectionError(error) {
 function environmentModelRequiredError(error) {
   const data = error?.data;
   if (!data || typeof data !== "object" || Array.isArray(data)) return null;
-  const record4 = data;
-  if (record4.code !== "ENV_MODEL_REQUIRED") return null;
-  const message = typeof record4.message === "string" && record4.message.trim() ? record4.message : "This environment has no model to run. Set a model on the environment, or pin one on its client.";
-  const details = record4.details && typeof record4.details === "object" && !Array.isArray(record4.details) ? record4.details : {};
+  const record5 = data;
+  if (record5.code !== "ENV_MODEL_REQUIRED") return null;
+  const message = typeof record5.message === "string" && record5.message.trim() ? record5.message : "This environment has no model to run. Set a model on the environment, or pin one on its client.";
+  const details = record5.details && typeof record5.details === "object" && !Array.isArray(record5.details) ? record5.details : {};
   return new WebRouteError(409, ErrorCode.CONFLICT, message, {
     code: "ENV_MODEL_REQUIRED",
     ...details,
@@ -191893,7 +193217,7 @@ function slimPhaseSignal(signal) {
 }
 var SETUP_AUDIT_METADATA_KEY = "stageSetupAudit";
 function capSetupAuditMetadata(raw, capBytes = SETUP_SIGNALS_METADATA_CAP_BYTES) {
-  const fits = (record4) => Buffer.byteLength(JSON.stringify(record4), "utf8") <= capBytes;
+  const fits = (record5) => Buffer.byteLength(JSON.stringify(record5), "utf8") <= capBytes;
   if (fits(raw)) return raw;
   const withoutReasons = {
     ...raw,
@@ -192144,7 +193468,7 @@ function createRunSetupObserver(options) {
   const lists = /* @__PURE__ */ new Map();
   let canaryResult;
   let canaryPromise;
-  const record4 = (into, serverId, init2) => {
+  const record5 = (into, serverId, init2) => {
     const detail = init2.outcome === "failed" ? describeSetupFailure(init2.error, {
       serverId,
       ...options.context?.(serverId) ?? {},
@@ -192193,8 +193517,8 @@ function createRunSetupObserver(options) {
     };
   };
   return {
-    recordConnect: (serverId, init2) => record4(connects, serverId, init2),
-    recordToolsList: (serverId, init2) => record4(lists, serverId, init2),
+    recordConnect: (serverId, init2) => record5(connects, serverId, init2),
+    recordToolsList: (serverId, init2) => record5(lists, serverId, init2),
     ensureEgressCanary,
     buildSignals,
     failureDetail: (serverId, phase) => (phase === "connection" ? connects : lists).get(serverId)?.detail,
@@ -204783,11 +206107,11 @@ function asRecord11(value) {
   return typeof value === "object" && value !== null && !Array.isArray(value) ? value : void 0;
 }
 function summarizeTool(tool19) {
-  const record4 = asRecord11(tool19);
-  const name15 = typeof record4?.name === "string" ? record4.name : void 0;
+  const record5 = asRecord11(tool19);
+  const name15 = typeof record5?.name === "string" ? record5.name : void 0;
   if (!name15) return void 0;
-  const description = typeof record4?.description === "string" ? record4.description : "";
-  const annotations2 = asRecord11(record4?.annotations);
+  const description = typeof record5?.description === "string" ? record5.description : "";
+  const annotations2 = asRecord11(record5?.annotations);
   const hints = annotations2 ? Object.entries(annotations2).filter(([, value]) => typeof value === "boolean").map(([key, value]) => `${key}=${value}`).join(" ") : "";
   return `- ${name15}${hints ? ` [${hints}]` : ""}: ${description.slice(0, 400)}`;
 }
@@ -209839,13 +211163,13 @@ var SUPPORTED_AUDIO_FORMATS = /* @__PURE__ */ new Set([
 ]);
 function readErrorMessage(payload, fallback) {
   if (!payload || typeof payload !== "object") return fallback;
-  const record4 = payload;
-  if (typeof record4.error === "string") return record4.error;
-  if (record4.error && typeof record4.error === "object") {
-    const error = record4.error;
+  const record5 = payload;
+  if (typeof record5.error === "string") return record5.error;
+  if (record5.error && typeof record5.error === "object") {
+    const error = record5.error;
     if (typeof error.message === "string") return error.message;
   }
-  if (typeof record4.message === "string") return record4.message;
+  if (typeof record5.message === "string") return record5.message;
   return fallback;
 }
 function readUpstreamError(payload, fallback) {
@@ -209853,16 +211177,16 @@ function readUpstreamError(payload, fallback) {
   if (!payload || typeof payload !== "object") {
     return { error: message };
   }
-  const record4 = payload;
-  const code = typeof record4.code === "string" ? record4.code : void 0;
+  const record5 = payload;
+  const code = typeof record5.code === "string" ? record5.code : void 0;
   const isVoiceBudgetError = code === MCPJAM_VOICE_BUDGET_CODE;
   const isVoiceInProgressError = code === MCPJAM_VOICE_IN_PROGRESS_CODE;
   return {
     error: isVoiceBudgetError ? MCPJAM_VOICE_BUDGET_MESSAGE : isVoiceInProgressError ? MCPJAM_VOICE_IN_PROGRESS_MESSAGE : message,
     ...code ? { code } : {},
-    ...typeof record4.isRetryable === "boolean" ? { isRetryable: record4.isRetryable } : {},
-    ...typeof record4.retryAfter === "number" ? { retryAfter: record4.retryAfter } : {},
-    ...isVoiceBudgetError && typeof record4.details === "string" ? { details: record4.details } : {}
+    ...typeof record5.isRetryable === "boolean" ? { isRetryable: record5.isRetryable } : {},
+    ...typeof record5.retryAfter === "number" ? { retryAfter: record5.retryAfter } : {},
+    ...isVoiceBudgetError && typeof record5.details === "string" ? { details: record5.details } : {}
   };
 }
 function validateRequest(body) {
@@ -216486,8 +217810,8 @@ function stableJson2(value) {
   if (Array.isArray(value)) {
     return `[${value.map(stableJson2).join(",")}]`;
   }
-  const record4 = value;
-  return `{${Object.keys(record4).sort().map((key) => `${JSON.stringify(key)}:${stableJson2(record4[key])}`).join(",")}}`;
+  const record5 = value;
+  return `{${Object.keys(record5).sort().map((key) => `${JSON.stringify(key)}:${stableJson2(record5[key])}`).join(",")}}`;
 }
 function hashInput(value) {
   return createHash24("sha256").update(stableJson2(value)).digest("base64url");
@@ -217661,12 +218985,12 @@ function turnSkillProvenance(spec, options) {
     ].filter(isProvenanceRow)
   };
 }
-function readString5(record4, key) {
-  const value = record4[key];
+function readString5(record5, key) {
+  const value = record5[key];
   return typeof value === "string" && value.length > 0 ? value : null;
 }
-function readBoolean3(record4, key) {
-  const value = record4[key];
+function readBoolean3(record5, key) {
+  const value = record5[key];
   return typeof value === "boolean" ? value : null;
 }
 function toEnvironmentPreview(spec, options) {
@@ -217906,12 +219230,30 @@ chatV22.post("/", async (c) => {
     rpcCollector = createHostedRpcLogCollector(rawBody);
     const hostedBody = parseWithSchema(hostedChatSchema, rawBody);
     if (!c.get("guestId") && hostedBody.projectId && hostedBody.chatSessionId) {
-      const allowed = await apiSessionWriteAllowed(rawBody.origin, async (signal) => {
-        const service = new BrowserSessionService();
-        if (!service.enabled) return { writable: true };
-        return service.agentRequest("assert_web_writable", { bearer: bearerToken, projectId: hostedBody.projectId, body: { conversationId: hostedBody.chatSessionId }, signal: AbortSignal.any([signal, c.req.raw.signal]) });
-      });
-      if (!allowed) return c.json({ code: "API_SESSION_READ_ONLY", error: "This API session is view-only in Playground. Continue it through the session API." }, 409);
+      const allowed = await apiSessionWriteAllowed(
+        rawBody.origin,
+        async (signal) => {
+          const service = new BrowserSessionService();
+          if (!service.enabled) return { writable: true };
+          return service.agentRequest(
+            "assert_web_writable",
+            {
+              bearer: bearerToken,
+              projectId: hostedBody.projectId,
+              body: { conversationId: hostedBody.chatSessionId },
+              signal: AbortSignal.any([signal, c.req.raw.signal])
+            }
+          );
+        }
+      );
+      if (!allowed)
+        return c.json(
+          {
+            code: "API_SESSION_READ_ONLY",
+            error: "This API session is view-only in Playground. Continue it through the session API."
+          },
+          409
+        );
     }
     const { initializePins, mcpProtocolVersionsByServerId } = extractMcpInitializeOptions(rawBody);
     const body = rawBody;
@@ -218086,6 +219428,11 @@ chatV22.post("/", async (c) => {
           error: runtime.error
         });
         const failClosedMessage = `Couldn't load this scenario's settings, so the turn was stopped to avoid running with the wrong configuration. ${runtime.error}`;
+        if (runtime.code === "SCENARIO_SIGN_IN_REQUIRED") {
+          throw new WebRouteError(401, ErrorCode.UNAUTHORIZED, runtime.error, {
+            code: runtime.code
+          });
+        }
         if (runtime.code === "SCENARIO_ACCESS_STALE") {
           throw new WebRouteError(
             409,
@@ -220073,8 +221420,8 @@ mcpjamAgent.post("/widget-content", async (c) => {
       if (!content || typeof content !== "object") {
         return webError(c, 404, ErrorCode.NOT_FOUND, "No content in resource");
       }
-      const record4 = content;
-      const contentMimeType = typeof record4.mimeType === "string" ? record4.mimeType : void 0;
+      const record5 = content;
+      const contentMimeType = typeof record5.mimeType === "string" ? record5.mimeType : void 0;
       const mimeTypeValid = contentMimeType !== void 0 && ACCEPTED_WIDGET_MIMETYPES2.has(contentMimeType);
       const mimeTypeWarning = !mimeTypeValid ? contentMimeType ? `Invalid mimetype "${contentMimeType}" - expected one of: ${[
         ...ACCEPTED_WIDGET_MIMETYPES2
@@ -220082,10 +221429,10 @@ mcpjamAgent.post("/widget-content", async (c) => {
         ...ACCEPTED_WIDGET_MIMETYPES2
       ].join(", ")}` : null;
       let html;
-      if (typeof record4.text === "string") {
-        html = record4.text;
-      } else if (typeof record4.blob === "string") {
-        html = Buffer.from(record4.blob, "base64").toString("utf-8");
+      if (typeof record5.text === "string") {
+        html = record5.text;
+      } else if (typeof record5.blob === "string") {
+        html = Buffer.from(record5.blob, "base64").toString("utf-8");
       } else {
         return webError(
           c,
@@ -220094,7 +221441,7 @@ mcpjamAgent.post("/widget-content", async (c) => {
           "No HTML content in resource"
         );
       }
-      const resourceMeta = record4._meta;
+      const resourceMeta = record5._meta;
       const uiMeta = resolveUiResourceMeta({ contentMeta: resourceMeta });
       const effectiveCspMode = body.cspMode ?? "permissive";
       if (body.injectOpenAiCompat === true) {
@@ -220875,6 +222222,7 @@ async function runSyntheticHostSession(adapter) {
     modelDefinition,
     systemPrompt,
     temperature,
+    maxSteps,
     requireToolApproval,
     respectToolVisibility,
     progressiveToolDiscovery,
@@ -221252,6 +222600,7 @@ async function runSyntheticHostSession(adapter) {
         sourceType: persist.sourceType,
         systemPrompt: prepared.enhancedSystemPrompt,
         temperature: prepared.resolvedTemperature,
+        ...maxSteps !== void 0 ? { maxSteps } : {},
         // `computer` / `finish_widget` merge into the advertised set; the
         // prepareAdvertisedTools hook hides them until a widget is mounted.
         tools: { ...prepared.allTools, ...browser.computerWidgetTools },
@@ -222298,10 +223647,37 @@ async function prepareTargetGrounding(args) {
   }
 }
 
+// server/services/checks/swarm-check-evidence.ts
+var record2 = (value) => !!value && typeof value === "object" && !Array.isArray(value);
+function swarmCheckInventory(envelope) {
+  if (!record2(envelope) || !record2(envelope.recordedContext)) return void 0;
+  const snapshots = envelope.recordedContext.toolSnapshots;
+  if (!Array.isArray(snapshots) || snapshots.length !== 1) return void 0;
+  const entry3 = snapshots[0];
+  if (!record2(entry3) || !record2(entry3.snapshot) || !Array.isArray(entry3.snapshot.servers))
+    return void 0;
+  const tools5 = {};
+  for (const server2 of entry3.snapshot.servers) {
+    if (!record2(server2) || server2.captureError || !Array.isArray(server2.tools))
+      return void 0;
+    for (const tool19 of server2.tools) {
+      if (!record2(tool19) || typeof tool19.name !== "string" || Object.hasOwn(tools5, tool19.name))
+        return void 0;
+      tools5[tool19.name] = {
+        inputSchema: tool19.inputSchema,
+        ...typeof tool19.description === "string" ? { description: tool19.description } : {},
+        ...record2(tool19.annotations) ? { annotations: tool19.annotations } : {}
+      };
+    }
+  }
+  return toTranscriptToolInventory(tools5);
+}
+
 // server/services/checks/run-predicates-on-chat-session.ts
 function extractToolCallsFromEnvelopeMessages(messages) {
   return extractToolCallsFromConversation({ messages }).map((toolCall) => ({
     toolName: toolCall.toolName,
+    ...toolCall.toolCallId ? { toolCallId: toolCall.toolCallId } : {},
     arguments: toolCall.arguments ?? {}
   }));
 }
@@ -222421,12 +223797,14 @@ async function runSwarmChecks(args) {
     return { status: "failed", error };
   };
   const messages = Array.isArray(claim.envelope?.messages) ? claim.envelope.messages : null;
-  if (messages === null) {
+  if (messages === null || messages.length === 0 || claim.envelope?.traceComplete === false) {
     return reportFailure("transcript envelope unreadable");
   }
   let criterionResults;
   try {
     const transcript = buildIterationTranscript({
+      ...extractTranscriptEvidence(claim.envelope),
+      toolInventory: swarmCheckInventory(claim.envelope),
       trace: {
         messages,
         ...claim.envelope?.spans ? { spans: claim.envelope.spans } : {}
@@ -222434,8 +223812,8 @@ async function runSwarmChecks(args) {
       // The SHARED walker, not a copy: two extractors would let an
       // envelope-format or dedupe fix land on one grading path and not the
       // other, so the same session could grade differently depending on who
-      // asked. (Its identity dedupe — same tool + same args collapses to one
-      // entry — is a known limitation, now a single known limitation.)
+      // asked. Distinct call IDs preserve repeated calls; only duplicate
+      // representations of the same call are collapsed.
       toolCalls: extractToolCallsFromEnvelopeMessages(messages),
       // Session-level token totals, materialized backend-side from turn-trace
       // usage and returned on the claim. `null`/absent means no turn reported
@@ -222456,6 +223834,9 @@ async function runSwarmChecks(args) {
     criterionResults = claim.criteria.map((entry3, index) => ({
       criterionId: entry3.id,
       passed: results[index]?.passed ?? false,
+      // Older evaluators omit status on scored rows; only a missing result
+      // or an explicit evaluator error is unmeasured.
+      status: results[index] ? results[index].status ?? "scored" : "error",
       reason: results[index]?.reason ?? "evaluator returned no verdict"
     }));
   } catch (error) {
@@ -223261,6 +224642,7 @@ async function withArtifactFlushDeadline(work) {
   }
 }
 var MAX_CONCURRENT_TARGETS = 3;
+var SWARM_PERSONA_TURN_MAX_STEPS = 10;
 function targetSessionIdentity(target) {
   return {
     hostId: target.hostId,
@@ -223755,6 +225137,7 @@ async function runJourneyFanOut(opts) {
               modelDefinition,
               systemPrompt: target.systemPrompt,
               temperature: target.temperature,
+              maxSteps: SWARM_PERSONA_TURN_MAX_STEPS,
               requireToolApproval: target.requireToolApproval,
               respectToolVisibility: target.respectToolVisibility,
               progressiveToolDiscovery: target.progressiveToolDiscovery,
@@ -223917,7 +225300,7 @@ async function runJourneyFanOut(opts) {
               error: err instanceof Error ? err.message : String(err)
             });
           }
-          if (hasRubric && terminal.status !== "rate_limited") {
+          if (hasRubric) {
             try {
               const graded = await runSwarmChecks({
                 convexHttpUrl: convexHttpUrl3,
@@ -224499,7 +225882,7 @@ async function launchJourneyRun(deps, input) {
       // the yes/no — the criteria themselves come back from the claim, so
       // the authoritative list is always the backend's pinned copy and
       // never a value that rode along in process memory.
-      hasRubric: (snapshot2.rubric?.length ?? 0) > 0,
+      hasRubric: ((snapshot2.standardCheckProfile?.criteria ?? snapshot2.rubric)?.length ?? 0) > 0,
       convexHttpUrl: convexHttpUrl3,
       getBearer: deps.getRunBearer,
       // Host-aware: each host connects ONLY its own pinned required servers
@@ -225616,10 +226999,10 @@ var ACCEPTED_WIDGET_MIMETYPES3 = /* @__PURE__ */ new Set([
 ]);
 function extractHtmlFromResourceContent4(content) {
   if (!content || typeof content !== "object") return "";
-  const record4 = content;
-  if (typeof record4.text === "string") return record4.text;
-  if (typeof record4.blob === "string") {
-    return Buffer.from(record4.blob, "base64").toString("utf-8");
+  const record5 = content;
+  if (typeof record5.text === "string") return record5.text;
+  if (typeof record5.blob === "string") {
+    return Buffer.from(record5.blob, "base64").toString("utf-8");
   }
   return "";
 }
@@ -231749,8 +233132,8 @@ function asArray(value) {
   return Array.isArray(value) ? value : void 0;
 }
 function isPredicateRow(value) {
-  const record4 = asRecord13(value);
-  return typeof record4?.passed === "boolean" && asRecord13(record4.predicate) !== void 0;
+  const record5 = asRecord13(value);
+  return typeof record5?.passed === "boolean" && asRecord13(record5.predicate) !== void 0;
 }
 function stepErrorFromStoredChain(stageResults) {
   const rows = asArray(stageResults);
@@ -232119,7 +233502,8 @@ async function postServiceRoute(path8, body) {
 }
 async function claimNextStageDerivation(claimedBy) {
   const { status, body } = await postServiceRoute(`${BASE_PATH2}/claim`, {
-    claimedBy
+    claimedBy,
+    evidenceVersion: SWARM_STAGE_EVIDENCE_VERSION
   });
   if (status === 404) return { kind: "disabled" };
   if (status !== 200 || !body?.ok) {
@@ -232253,7 +233637,13 @@ function normalizeCriteria(value) {
     if (!isRecord37(entry3)) return [];
     const criterionId = nonEmptyString4(entry3.criterionId);
     if (!criterionId || typeof entry3.passed !== "boolean") return [];
-    return [{ criterionId, passed: entry3.passed }];
+    return [
+      {
+        criterionId,
+        passed: entry3.passed,
+        ...entry3.status === "scored" || entry3.status === "error" ? { status: entry3.status } : {}
+      }
+    ];
   });
   const droppedRows = results.length !== value.results.length;
   const covered = new Set(results.map((row2) => row2.criterionId));
@@ -232296,6 +233686,22 @@ function normalizeClaimedEvidence(claim) {
   const spans = normalizeSpans(envelope.spans);
   const readiness2 = normalizeReadiness(claim.evidence.readiness);
   const criteria = normalizeCriteria(claim.evidence.criteria);
+  const definitions = claim.evidence.criterionDefinitions;
+  if (criteria?.results && Array.isArray(definitions)) {
+    criteria.results = criteria.results.map((row2) => {
+      const definition = definitions.find(
+        (d) => isRecord37(d) && d.id === row2.criterionId
+      );
+      return isRecord37(definition) && isRecord37(definition.predicate) && typeof definition.predicate.type === "string" ? {
+        ...row2,
+        predicate: {
+          type: definition.predicate.type,
+          role: definition.predicate.role === "advisory" ? "advisory" : "required"
+        }
+      } : row2;
+    });
+  }
+  const swarmPolicy = claim.evidence.swarmPolicy;
   const goalJudge = normalizeGoalJudge(claim.evidence.goalScore);
   return {
     ok: true,
@@ -232306,7 +233712,13 @@ function normalizeClaimedEvidence(claim) {
       ...spans.length > 0 ? { spans } : {},
       ...readiness2 ? { readiness: readiness2 } : {},
       ...criteria ? { criteria } : {},
-      ...goalJudge ? { goalJudge } : {}
+      ...goalJudge ? { goalJudge } : {},
+      ...source === "swarm" && isRecord37(swarmPolicy) && typeof swarmPolicy.judgeDecisive === "boolean" && typeof swarmPolicy.requiredCriteria === "number" ? {
+        swarmPolicy: {
+          judgeDecisive: swarmPolicy.judgeDecisive,
+          requiredCriteria: swarmPolicy.requiredCriteria
+        }
+      } : {}
     }
   };
 }
@@ -233375,8 +234787,8 @@ function measureClientMessage(buffer2, holdsInput) {
       return decide(20, true);
     case RFB_CLIENT_MESSAGE.SET_ENCODINGS: {
       if (buffer2.length < 4) return { kind: "incomplete" };
-      const count = readU16(buffer2, 2);
-      return decide(4 + count * 4, true);
+      const count4 = readU16(buffer2, 2);
+      return decide(4 + count4 * 4, true);
     }
     case RFB_CLIENT_MESSAGE.FRAMEBUFFER_UPDATE_REQUEST:
       return decide(10, true);
@@ -233486,10 +234898,10 @@ function parseProtocolVersion(bytes2) {
 }
 function parseSecurityTypes(bytes2) {
   if (bytes2.length < 1) return null;
-  const count = bytes2[0];
-  if (count === 0) return [];
-  if (bytes2.length < 1 + count) return null;
-  return Array.from(bytes2.subarray(1, 1 + count));
+  const count4 = bytes2[0];
+  if (count4 === 0) return [];
+  if (bytes2.length < 1 + count4) return null;
+  return Array.from(bytes2.subarray(1, 1 + count4));
 }
 function vncPasswordKey(password) {
   const key = Buffer.alloc(8, 0);
@@ -233895,14 +235307,14 @@ function createComputerBrowserFramesWsHandler(upgradeWebSocket2, deps = {}) {
     ...args.codec ? { codec: args.codec } : {},
     ...args.sharp ? { sharp: true } : {},
     ...args.onVideo ? {
-      onVideo: (record4) => args.onVideo?.({
-        key: record4.kind === FRAME_STREAM_KIND.video_key,
-        au: record4.au,
-        deviceWidth: record4.deviceWidth,
-        deviceHeight: record4.deviceHeight,
-        scale: record4.scale,
-        ts: record4.ts,
-        seq: record4.seq
+      onVideo: (record5) => args.onVideo?.({
+        key: record5.kind === FRAME_STREAM_KIND.video_key,
+        au: record5.au,
+        deviceWidth: record5.deviceWidth,
+        deviceHeight: record5.deviceHeight,
+        scale: record5.scale,
+        ts: record5.ts,
+        seq: record5.seq
       })
     } : {},
     signal: args.signal,
@@ -234112,20 +235524,20 @@ function createComputerBrowserFramesWsHandler(upgradeWebSocket2, deps = {}) {
           ...videoAgreed ? { codec: "h264" } : {},
           ...sharpAgreed ? { sharp: true } : {},
           ...videoAgreed ? {
-            onVideo: (record4) => {
+            onVideo: (record5) => {
               if (closed) {
                 stats3?.countDrop();
                 return;
               }
               const bytes2 = new Uint8Array(
                 encodeFrameStreamRecord({
-                  kind: record4.key ? FRAME_STREAM_KIND.video_key : FRAME_STREAM_KIND.video_delta,
-                  deviceWidth: record4.deviceWidth,
-                  deviceHeight: record4.deviceHeight,
-                  scale: record4.scale,
+                  kind: record5.key ? FRAME_STREAM_KIND.video_key : FRAME_STREAM_KIND.video_delta,
+                  deviceWidth: record5.deviceWidth,
+                  deviceHeight: record5.deviceHeight,
+                  scale: record5.scale,
                   ts: Date.now(),
-                  seq: record4.seq,
-                  au: record4.au
+                  seq: record5.seq,
+                  au: record5.au
                 })
               );
               stats3?.offer(bytes2.byteLength, () => ws.send(bytes2));
@@ -234732,6 +236144,13 @@ function translateConvexWriteError(error, options) {
       structuredMessage ?? "This feature is not available for your organization."
     );
   }
+  if (code === "SIGN_IN_REQUIRED") {
+    return new WebRouteError(
+      401,
+      ErrorCode.UNAUTHORIZED,
+      structuredMessage ?? "Sign in to use this feature."
+    );
+  }
   if (code === "EVAL_COMPARE_BASELINE_CONFLICT" || code === "EVAL_COMPARE_BASELINE_INVALID") {
     return new WebRouteError(
       400,
@@ -235465,8 +236884,8 @@ tools4.post(
       );
       const durationMs = Math.max(0, Date.now() - startedAt);
       if (result && typeof result === "object" && !Array.isArray(result)) {
-        const record4 = result;
-        return "durationMs" in record4 ? record4 : { ...record4, durationMs };
+        const record5 = result;
+        return "durationMs" in record5 ? record5 : { ...record5, durationMs };
       }
       return result;
     },
@@ -235750,12 +237169,12 @@ var ITERATION_SCORE_INTEGRITY = /* @__PURE__ */ new Set(["score_integrity_invali
 var RUN_SCORE_INTEGRITY = /* @__PURE__ */ new Set(["valid", "invalid"]);
 function toScoreProjection(metadata, vocabulary = 1) {
   if (!metadata || typeof metadata !== "object") return {};
-  const record4 = metadata;
-  const integrity = typeof record4.scoreIntegrity === "string" && ITERATION_SCORE_INTEGRITY.has(record4.scoreIntegrity) ? { scoreIntegrity: record4.scoreIntegrity } : {};
+  const record5 = metadata;
+  const integrity = typeof record5.scoreIntegrity === "string" && ITERATION_SCORE_INTEGRITY.has(record5.scoreIntegrity) ? { scoreIntegrity: record5.scoreIntegrity } : {};
   if ("scoreIntegrity" in integrity) return integrity;
-  const scores = scoreResultArraySchema.safeParse(record4.scores);
+  const scores = scoreResultArraySchema.safeParse(record5.scores);
   const config = evaluationConfigSnapshotSchema2.safeParse(
-    record4.evaluationConfig
+    record5.evaluationConfig
   );
   if (!scores.success || !config.success) return {};
   return {
@@ -235944,12 +237363,12 @@ function costCoverage(value) {
   const source = value;
   const side = (raw) => {
     const inner = isRecord38(raw) ? raw : {};
-    const count = (raw2) => {
+    const count4 = (raw2) => {
       const n = numOrNull(raw2);
       return typeof n === "number" && Number.isSafeInteger(n) && n >= 0 ? n : 0;
     };
-    const costed = count(inner.costed);
-    const total = count(inner.total);
+    const costed = count4(inner.costed);
+    const total = count4(inner.total);
     return { costed: costed <= total ? costed : 0, total };
   };
   return { base: side(source.base), compare: side(source.compare) };
@@ -237706,15 +239125,15 @@ function internalCaseToSteps(testCase) {
 }
 function toPublicCaseImportClaim(raw) {
   if (!raw || typeof raw !== "object") return void 0;
-  const record4 = raw;
-  const status = record4.status;
+  const record5 = raw;
+  const status = record5.status;
   if (typeof status !== "string" || !IMPORT_MAPPING_STATUSES2.includes(status)) {
     return void 0;
   }
   return {
     status,
-    ...typeof record4.sourceCaseKey === "string" && record4.sourceCaseKey ? { sourceCaseKey: record4.sourceCaseKey } : {},
-    ...typeof record4.note === "string" && record4.note ? { note: record4.note } : {}
+    ...typeof record5.sourceCaseKey === "string" && record5.sourceCaseKey ? { sourceCaseKey: record5.sourceCaseKey } : {},
+    ...typeof record5.note === "string" && record5.note ? { note: record5.note } : {}
   };
 }
 function toCaseDto(testCase, vocabulary = 1) {
@@ -249040,6 +250459,8 @@ function toJourneyRunDto(row2) {
     id: row2._id,
     projectId: row2.projectId,
     journeyId: row2.journeyRefId,
+    ...row2.verdictSummary ? { verdictSummary: row2.verdictSummary } : {},
+    ...row2.report ? { report: row2.report } : {},
     // `swarmRunGroupId` upstream. Renamed because the public meaning is "the
     // batch this run was launched with", and every run of a solo relaunch is
     // a wave of one.
@@ -249096,6 +250517,9 @@ function toJourneySessionDto(row2, outcome) {
      * A caller that listed a run's sessions could not then look one of them up.
      */
     id: row2.id,
+    ...row2.verdict ? { verdict: row2.verdict } : {},
+    ...row2.observations ? { observations: row2.observations } : {},
+    ...row2.criteria ? { criteria: row2.criteria } : {},
     /**
      * The runtime key, kept but named for what it is. It is what the chat
      * transport and the app's own deep links use, so dropping it would strand
@@ -252150,7 +253574,7 @@ var evalBacktestRequestSchema2 = evalBacktestDraftSchema4.extend({
 });
 
 // server/services/evals/assertion-backtest.ts
-var record2 = (value) => !!value && typeof value === "object" && !Array.isArray(value);
+var record3 = (value) => !!value && typeof value === "object" && !Array.isArray(value);
 function originalResults(row2) {
   return Array.isArray(row2.results) ? row2.results.flatMap((value) => {
     const parsed = scoreResultSchema2.safeParse(value);
@@ -252159,7 +253583,7 @@ function originalResults(row2) {
 }
 function rulesFor(row2, draft) {
   if (draft.assertions.mode === "replace") return draft.assertions.list;
-  const original = Array.isArray(row2.predicates) ? row2.predicates : record2(row2.predicates) && Array.isArray(row2.predicates.list) ? row2.predicates.list : null;
+  const original = Array.isArray(row2.predicates) ? row2.predicates : record3(row2.predicates) && Array.isArray(row2.predicates.list) ? row2.predicates.list : null;
   if (!original || original.length + draft.assertions.list.length > 100)
     return null;
   const rules = [];
@@ -252191,7 +253615,7 @@ function backtestIteration(row2, draft) {
     trace: { messages: evidence?.messages ?? [] }
   });
   const evaluated = missing ? [] : evaluatePredicates(transcript, rules);
-  const definitions = record2(row2.evaluationConfig) && Array.isArray(row2.evaluationConfig.definitions) ? row2.evaluationConfig.definitions.flatMap((value) => {
+  const definitions = record3(row2.evaluationConfig) && Array.isArray(row2.evaluationConfig.definitions) ? row2.evaluationConfig.definitions.flatMap((value) => {
     const parsed = resolvedScoreDefinitionSchema2.safeParse(value);
     return parsed.success ? [parsed.data] : [];
   }) : [];
@@ -252306,7 +253730,7 @@ async function runAssertionBacktest(input) {
   const differences = [];
   const seen = /* @__PURE__ */ new Set();
   let page3;
-  let count = 0;
+  let count4 = 0;
   for (let n = 0; n < 10; n++) {
     if (input.signal?.aborted) throw new Error("Backtest cancelled");
     const priorPage = page3;
@@ -252355,7 +253779,7 @@ async function runAssertionBacktest(input) {
         throw new Error("Duplicate or invalid backtest iteration identity");
       seen.add(row2.iterationId);
       differences.push(...backtestIteration(row2, draft));
-      count++;
+      count4++;
       if (Date.now() > deadline2) throw new Error("Backtest deadline exceeded");
     }
     if (page3.isDone) break;
@@ -252368,7 +253792,7 @@ async function runAssertionBacktest(input) {
     sourceHash: page3.sourceHash,
     draftHash,
     configRevision: page3.configRevision,
-    complete: page3.isDone && count > 0 && differences.length > 0 && comparable === differences.length,
+    complete: page3.isDone && count4 > 0 && differences.length > 0 && comparable === differences.length,
     continuationAvailable: !page3.isDone,
     ...!page3.isDone ? {
       continuation: {
@@ -252379,7 +253803,7 @@ async function runAssertionBacktest(input) {
       }
     } : {},
     counts: {
-      iterations: count,
+      iterations: count4,
       comparable,
       ungradable: differences.length - comparable,
       flipped: differences.filter((row2) => row2.flipped).length
@@ -253083,8 +254507,8 @@ function describeClientImpact(input) {
     [n("scenarioAttachmentCount"), "scenario attachment"],
     [n("activeLegacyJourneyCount"), "active legacy journey"]
   ];
-  const total = parts.reduce((sum2, [count]) => sum2 + count, 0);
-  const listed = parts.map(([count, noun]) => `${count} ${noun}${count === 1 ? "" : "s"}`).join(", ");
+  const total = parts.reduce((sum2, [count4]) => sum2 + count4, 0);
+  const listed = parts.map(([count4, noun]) => `${count4} ${noun}${count4 === 1 ? "" : "s"}`).join(", ");
   const affected = total === 0 ? "Nothing durable currently uses this client" : `This will affect ${listed}`;
   return `${affected}; future direct client and playground use also follows the edit. ${unchanged}`;
 }
@@ -253837,7 +255261,7 @@ var AGENT_OP_REGISTRY = [
       confirmSeverity: "none"
     },
     promptNotes: [
-      "- `propose_eval_description_rewrite` returns a proposing receipt, not a finished rewrite. Poll `get_eval_description_experiment` until status is proposed (or failed). Requesting again takes another slice of the daily generation quota."
+      "- `propose_eval_description_rewrite` returns a proposing receipt, not a finished rewrite. Poll `get_eval_description_experiment` until status is proposed (or failed). Requesting again runs another analysis against MCPJam's daily analysis budget."
     ]
   },
   {
@@ -254147,7 +255571,8 @@ var AGENT_OP_REGISTRY = [
       confirmSeverity: "none"
     },
     promptNotes: [
-      "- `request_wave_insights` consumes no credits, but it counts against a daily insight QUOTA shared with user-testing insights \u2014 a request here takes one from there. Read the run scorecards first; they cost no quota and usually explain the failure without a model pass."
+      "- `request_wave_insights` consumes no credits, but it counts against a daily insight QUOTA shared with user-testing insights \u2014 a request here takes one from there. Read the run scorecards first; they cost no quota and usually explain the failure without a model pass.",
+      "- Included operations (generation and insights) can be refused with `RATE_LIMITED`. `canTopUp` is false on those refusals: tell the user when it lifts (`retryAfterSeconds`, or 00:00 UTC for a daily budget), and do not retry sooner, suggest topping up credits, or switch identities to get around it."
     ]
   },
   // ── USER TESTING ──────────────────────────────────────────────────────
@@ -255420,9 +256845,9 @@ proposedActions.post(
         "Approved actions can only be executed from the surface that collected the approval."
       );
     }
-    let record4;
+    let record5;
     try {
-      record4 = await getProposedAction(actionId);
+      record5 = await getProposedAction(actionId);
     } catch (error) {
       logger.error("[v1/proposed-actions] could not read the proposal", {
         error: error instanceof Error ? error.message : String(error),
@@ -255434,20 +256859,20 @@ proposedActions.post(
         "Could not reach MCPJam to check that approval. Try again in a moment."
       );
     }
-    const recordTenantId = record4?.surfaceTenantId ?? record4?.teamId ?? null;
-    if (!record4 || (record4.surface ?? "slack") !== actor.surfaceKind || recordTenantId !== actor.tenantId) {
+    const recordTenantId = record5?.surfaceTenantId ?? record5?.teamId ?? null;
+    if (!record5 || (record5.surface ?? "slack") !== actor.surfaceKind || recordTenantId !== actor.tenantId) {
       return v1Error(c, "NOT_FOUND", "That approval is no longer available.");
     }
-    if (record4.organizationId !== actor.organizationId) {
+    if (record5.organizationId !== actor.organizationId) {
       return v1Error(c, "NOT_FOUND", "That approval is no longer available.");
     }
-    if (record4.projectId !== projectId) {
+    if (record5.projectId !== projectId) {
       return v1Error(c, "NOT_FOUND", "That approval is no longer available.");
     }
-    const operation = GATED_BY_NAME2.get(record4.operation);
+    const operation = GATED_BY_NAME2.get(record5.operation);
     if (!operation) {
       logger.warn("[v1/proposed-actions] proposal names an unknown operation", {
-        operation: record4.operation.slice(0, 100).replace(/[\r\n]/g, "")
+        operation: record5.operation.slice(0, 100).replace(/[\r\n]/g, "")
       });
       return v1Error(
         c,
@@ -255468,7 +256893,7 @@ proposedActions.post(
         "Could not check your organization's settings right now. Try again in a moment."
       );
     }
-    if (disabledOperations.has(record4.operation)) {
+    if (disabledOperations.has(record5.operation)) {
       return v1Error(
         c,
         "FORBIDDEN",
@@ -255510,7 +256935,7 @@ proposedActions.post(
       }
       return v1Error(c, "NOT_FOUND", "That approval is no longer available.");
     }
-    if (claim.operation !== record4.operation || claim.projectId !== record4.projectId) {
+    if (claim.operation !== record5.operation || claim.projectId !== record5.projectId) {
       logger.error(
         "[v1/proposed-actions] claim diverged from the pre-claim read",
         { actionId }
@@ -256240,15 +257665,15 @@ var commandSchema2 = z84.discriminatedUnion("op", [
 
 // server/services/browser-tool-evidence.ts
 var MAX_SCREENSHOT_CHARS = 512e3;
-var record3 = (value) => value !== null && typeof value === "object" && !Array.isArray(value) ? value : void 0;
+var record4 = (value) => value !== null && typeof value === "object" && !Array.isArray(value) ? value : void 0;
 function screenshotOf(value) {
-  const obj = record3(value);
+  const obj = record4(value);
   if (!obj) return void 0;
   if (typeof obj.screenshot === "string") return obj.screenshot;
   for (const key of ["page", "output", "result"]) {
-    const nested = record3(obj[key]);
+    const nested = record4(obj[key]);
     if (typeof nested?.screenshot === "string") return nested.screenshot;
-    const page3 = record3(nested?.page);
+    const page3 = record4(nested?.page);
     if (typeof page3?.screenshot === "string") return page3.screenshot;
   }
   return void 0;
@@ -256264,7 +257689,7 @@ function capture(raw) {
 function redactBrowserScreenshots(value, pointer) {
   if (Array.isArray(value))
     return value.map((item) => redactBrowserScreenshots(item, pointer));
-  const obj = record3(value);
+  const obj = record4(value);
   if (!obj) return value;
   if (obj.type === "image-data" || obj.type === "image" || obj.type === "image_url")
     return {
@@ -256320,7 +257745,7 @@ function wrapBrowserToolsForEvidence(tools5, args) {
                   source: "browser_tool",
                   ts: started2,
                   elapsedMs: Date.now() - started2,
-                  ok: result !== void 0 && !record3(result)?.error && record3(result)?.ok !== false,
+                  ok: result !== void 0 && !record4(result)?.error && record4(result)?.ok !== false,
                   ...image ? { screenshotBase64: image.base64 } : {}
                 });
                 if (image && url) {
@@ -256340,7 +257765,7 @@ function wrapBrowserToolsForEvidence(tools5, args) {
 function redactBrowserEvidenceTree(value, evidence) {
   if (Array.isArray(value))
     return value.map((item2) => redactBrowserEvidenceTree(item2, evidence));
-  const obj = record3(value);
+  const obj = record4(value);
   if (!obj) return value;
   const item = evidence.find((entry3) => entry3.toolCallId === obj.toolCallId);
   if (item) return redactBrowserScreenshots(obj, item);
@@ -256937,31 +258362,31 @@ function readToolOutput2(output) {
   if (!output || typeof output !== "object" || Array.isArray(output)) {
     return output;
   }
-  const record4 = output;
-  if (typeof record4.type === "string" && "value" in record4) {
-    return record4.value;
+  const record5 = output;
+  if (typeof record5.type === "string" && "value" in record5) {
+    return record5.value;
   }
   return output;
 }
 function isErrorOutput(output) {
   if (!output || typeof output !== "object") return false;
-  const record4 = output;
-  if (record4.type === "error-json" || record4.type === "error-text") return true;
-  const value = "value" in record4 ? record4.value : output;
+  const record5 = output;
+  if (record5.type === "error-json" || record5.type === "error-text") return true;
+  const value = "value" in record5 ? record5.value : output;
   return Boolean(
     value && typeof value === "object" && value.isError === true
   );
 }
 function projectMessages(messages, offset) {
   return messages.map((message, position) => {
-    const record4 = message && typeof message === "object" ? message : {};
+    const record5 = message && typeof message === "object" ? message : {};
     const bounded = boundPayload(
-      record4.content ?? record4.parts ?? message,
+      record5.content ?? record5.parts ?? message,
       MAX_MESSAGE_PAYLOAD_CHARS
     );
     return {
       index: offset + position,
-      role: typeof record4.role === "string" ? record4.role : "unknown",
+      role: typeof record5.role === "string" ? record5.role : "unknown",
       content: bounded.value,
       ...bounded.truncated ? { truncated: true } : {}
     };
@@ -262041,12 +263466,12 @@ function verifySlackLinkState(token, secret, now = Date.now()) {
   if (!payload || typeof payload !== "object" || Array.isArray(payload)) {
     return null;
   }
-  const record4 = payload;
-  if (typeof record4.linkSessionId !== "string" || record4.linkSessionId.length === 0 || typeof record4.exp !== "number") {
+  const record5 = payload;
+  if (typeof record5.linkSessionId !== "string" || record5.linkSessionId.length === 0 || typeof record5.exp !== "number") {
     return null;
   }
-  if (record4.exp <= now) return null;
-  return { linkSessionId: record4.linkSessionId, exp: record4.exp };
+  if (record5.exp <= now) return null;
+  return { linkSessionId: record5.linkSessionId, exp: record5.exp };
 }
 function hashOauthState(state) {
   return createHash30("sha256").update(state).digest("hex");
@@ -263169,20 +264594,20 @@ function verifyCliAuthState(token, secret, now = Date.now()) {
   if (!payload || typeof payload !== "object" || Array.isArray(payload)) {
     return null;
   }
-  const record4 = payload;
-  if (typeof record4.cliRedirectUri !== "string" || typeof record4.cliState !== "string" || typeof record4.exp !== "number") {
+  const record5 = payload;
+  if (typeof record5.cliRedirectUri !== "string" || typeof record5.cliState !== "string" || typeof record5.exp !== "number") {
     return null;
   }
-  if (record4.exp <= now) {
+  if (record5.exp <= now) {
     return null;
   }
-  if (!isAllowedLoopbackRedirect(record4.cliRedirectUri)) {
+  if (!isAllowedLoopbackRedirect(record5.cliRedirectUri)) {
     return null;
   }
   return {
-    cliRedirectUri: record4.cliRedirectUri,
-    cliState: record4.cliState,
-    exp: record4.exp
+    cliRedirectUri: record5.cliRedirectUri,
+    cliState: record5.cliState,
+    exp: record5.exp
   };
 }
 function isAllowedLoopbackRedirect(uri) {
@@ -264534,8 +265959,8 @@ async function startNetworkMonitor(sandbox, context) {
         }
         continue;
       }
-      const record4 = parseConnectionRecord(line2);
-      if (!record4 || records >= RECORD_LIMIT_PER_MINUTE || total >= TOTAL_RECORD_LIMIT) {
+      const record5 = parseConnectionRecord(line2);
+      if (!record5 || records >= RECORD_LIMIT_PER_MINUTE || total >= TOTAL_RECORD_LIMIT) {
         dropped++;
         continue;
       }
@@ -264543,7 +265968,7 @@ async function startNetworkMonitor(sandbox, context) {
       records++;
       total++;
       logger.info("[github-checks] network connection", {
-        ...record4,
+        ...record5,
         ...base,
         observedAt: now
       });
@@ -265389,10 +266814,10 @@ function matchesServerCapabilitiesSchema(value) {
     if (capability !== void 0 && !isPlainObject11(capability)) return false;
   }
   for (const field of ["experimental", "extensions"]) {
-    const record4 = capabilities2[field];
-    if (record4 === void 0) continue;
-    if (!isPlainObject11(record4)) return false;
-    const entries = Object.values(record4);
+    const record5 = capabilities2[field];
+    if (record5 === void 0) continue;
+    if (!isPlainObject11(record5)) return false;
+    const entries = Object.values(record5);
     if (!entries.every((entry3) => isPlainObject11(entry3))) return false;
   }
   for (const [field, flags] of CAPABILITY_BOOLEAN_FLAGS) {
@@ -266447,13 +267872,13 @@ function kindForMode(mode) {
 }
 function parseLsFilesEntries(stdout) {
   const out = [];
-  for (const record4 of stdout.split("\0")) {
-    if (record4.length === 0) continue;
-    const tab = record4.indexOf("	");
+  for (const record5 of stdout.split("\0")) {
+    if (record5.length === 0) continue;
+    const tab = record5.indexOf("	");
     if (tab === -1) continue;
-    const path8 = record4.slice(tab + 1);
+    const path8 = record5.slice(tab + 1);
     if (path8.length === 0) continue;
-    const mode = record4.slice(0, tab).split(" ")[0] ?? "";
+    const mode = record5.slice(0, tab).split(" ")[0] ?? "";
     out.push({ path: path8, kind: kindForMode(mode) });
     if (out.length >= MAX_REPO_FILES) break;
   }
@@ -266635,8 +268060,8 @@ function parseListenerReport(stdout) {
     try {
       const parsed = JSON.parse(line2.slice(LISTENER_MARKER.length + 1));
       if (typeof parsed !== "object" || parsed === null) return null;
-      const record4 = parsed;
-      const processes = Array.isArray(record4.processes) ? record4.processes : [];
+      const record5 = parsed;
+      const processes = Array.isArray(record5.processes) ? record5.processes : [];
       return { processes };
     } catch {
       return null;