openmerit 0.1.4 → 0.1.5

package/dist/protocol/src/schemas.d.ts

@@ -0,0 +1,450 @@
+import { Type, type TSchema } from "typebox";
+import type { IntentKind } from "./index.ts";
+export declare const OPENMERIT_SCHEMA_DIALECT: "https://json-schema.org/draft/2020-12/schema";
+export declare const OPENMERIT_RESULT_SCHEMA_VERSION: "0.2";
+export declare const EvidenceReferenceSchema: Type.TObject<{
+    id: Type.TString;
+    source: Type.TUnion<Type.TLiteral<string>[]>;
+    uri: Type.TString;
+    mediaType: Type.TOptional<Type.TString>;
+    digest: Type.TOptional<Type.TString>;
+}>;
+export declare const MetricObjectiveSchema: Type.TObject<{
+    metricId: Type.TUnion<Type.TLiteral<string>[]>;
+    required: Type.TBoolean;
+    minimumSamples: Type.TInteger;
+    tolerance: Type.TNumber;
+    constraint: Type.TOptional<Type.TObject<{
+        operator: Type.TUnion<Type.TLiteral<string>[]>;
+        value: Type.TNumber;
+    }>>;
+}>;
+export declare const TaskProfileSchema: Type.TObject<{
+    id: Type.TString;
+    revision: Type.TInteger;
+    name: Type.TString;
+    goal: Type.TString;
+    objectives: Type.TArray<Type.TObject<{
+        metricId: Type.TUnion<Type.TLiteral<string>[]>;
+        required: Type.TBoolean;
+        minimumSamples: Type.TInteger;
+        tolerance: Type.TNumber;
+        constraint: Type.TOptional<Type.TObject<{
+            operator: Type.TUnion<Type.TLiteral<string>[]>;
+            value: Type.TNumber;
+        }>>;
+    }>>;
+    createdAt: Type.TString;
+    confirmedAt: Type.TOptional<Type.TString>;
+}>;
+export declare const ObservabilityCoverageSchema: Type.TObject<{
+    status: Type.TUnion<Type.TLiteral<string>[]>;
+    requiredMetricIds: Type.TArray<Type.TUnion<Type.TLiteral<string>[]>>;
+    coveredMetricIds: Type.TArray<Type.TUnion<Type.TLiteral<string>[]>>;
+    missingMetricIds: Type.TArray<Type.TUnion<Type.TLiteral<string>[]>>;
+    checkedAt: Type.TString;
+}>;
+export declare const SupervisedGraduationPolicySchema: Type.TObject<{
+    mode: Type.TLiteral<"supervised_graduation">;
+    evaluationBudget: Type.TObject<{
+        currency: Type.TString;
+        maximumSpend: Type.TNumber;
+        maximumCandidateCount: Type.TOptional<Type.TInteger>;
+        maximumRunsPerCandidate: Type.TOptional<Type.TInteger>;
+        expiresAt: Type.TOptional<Type.TString>;
+    }>;
+    automaticSwapsEnabled: Type.TBoolean;
+    confirmationRequiredUntilVerifiedSwaps: Type.TInteger;
+    requirePostSwapVerification: Type.TLiteral<true>;
+    rollbackOnRegression: Type.TBoolean;
+    checkPolicy: Type.TObject<{
+        baselineAssessment: Type.TObject<{
+            afterCompletedTasks: Type.TOptional<Type.TInteger>;
+            afterElapsedSeconds: Type.TOptional<Type.TInteger>;
+        }>;
+        frontierReassessment: Type.TObject<{
+            afterCompletedTasks: Type.TOptional<Type.TInteger>;
+            afterElapsedSeconds: Type.TOptional<Type.TInteger>;
+            onModelCatalogChange: Type.TBoolean;
+        }>;
+        regressionCheck: Type.TObject<{
+            afterCompletedTasks: Type.TOptional<Type.TInteger>;
+            afterElapsedSeconds: Type.TOptional<Type.TInteger>;
+            thresholds: Type.TArray<Type.TObject<{
+                metricId: Type.TUnion<Type.TLiteral<string>[]>;
+                relativeChangeAtLeast: Type.TNumber;
+            }>>;
+        }>;
+        postSwapVerification: Type.TObject<{
+            afterCompletedTasks: Type.TOptional<Type.TInteger>;
+            afterElapsedSeconds: Type.TOptional<Type.TInteger>;
+        }>;
+        cooldownSeconds: Type.TInteger;
+        execution: Type.TObject<{
+            mode: Type.TUnion<Type.TLiteral<string>[]>;
+            infrastructureChangesAllowed: Type.TBoolean;
+        }>;
+    }>;
+}>;
+export declare const MetricRecordSchema: Type.TObject<{
+    metricId: Type.TUnion<Type.TLiteral<string>[]>;
+    state: Type.TUnion<Type.TLiteral<string>[]>;
+    scope: Type.TUnion<Type.TLiteral<string>[]>;
+    direction: Type.TUnion<Type.TLiteral<string>[]>;
+    method: Type.TUnion<Type.TLiteral<string>[]>;
+    value: Type.TOptional<Type.TNumber>;
+    unit: Type.TString;
+    sampleCount: Type.TInteger;
+    interval: Type.TOptional<Type.TObject<{
+        lower: Type.TNumber;
+        upper: Type.TNumber;
+        confidenceLevel: Type.TOptional<Type.TNumber>;
+    }>>;
+    window: Type.TOptional<Type.TObject<{
+        startedAt: Type.TString;
+        endedAt: Type.TString;
+    }>>;
+    evidence: Type.TArray<Type.TObject<{
+        id: Type.TString;
+        source: Type.TUnion<Type.TLiteral<string>[]>;
+        uri: Type.TString;
+        mediaType: Type.TOptional<Type.TString>;
+        digest: Type.TOptional<Type.TString>;
+    }>>;
+    note: Type.TOptional<Type.TString>;
+}>;
+export declare const CandidateDescriptorSchema: Type.TObject<{
+    candidateId: Type.TString;
+    modelId: Type.TString;
+    modelVersion: Type.TOptional<Type.TString>;
+    reasons: Type.TArray<Type.TString>;
+}>;
+export declare const CandidateAssessmentSchema: Type.TObject<{
+    id: Type.TString;
+    taskProfileId: Type.TString;
+    taskProfileRevision: Type.TInteger;
+    candidateId: Type.TString;
+    modelId: Type.TString;
+    modelVersion: Type.TOptional<Type.TString>;
+    productRevision: Type.TString;
+    harnessId: Type.TString;
+    startedAt: Type.TString;
+    completedAt: Type.TString;
+    metrics: Type.TArray<Type.TObject<{
+        metricId: Type.TUnion<Type.TLiteral<string>[]>;
+        state: Type.TUnion<Type.TLiteral<string>[]>;
+        scope: Type.TUnion<Type.TLiteral<string>[]>;
+        direction: Type.TUnion<Type.TLiteral<string>[]>;
+        method: Type.TUnion<Type.TLiteral<string>[]>;
+        value: Type.TOptional<Type.TNumber>;
+        unit: Type.TString;
+        sampleCount: Type.TInteger;
+        interval: Type.TOptional<Type.TObject<{
+            lower: Type.TNumber;
+            upper: Type.TNumber;
+            confidenceLevel: Type.TOptional<Type.TNumber>;
+        }>>;
+        window: Type.TOptional<Type.TObject<{
+            startedAt: Type.TString;
+            endedAt: Type.TString;
+        }>>;
+        evidence: Type.TArray<Type.TObject<{
+            id: Type.TString;
+            source: Type.TUnion<Type.TLiteral<string>[]>;
+            uri: Type.TString;
+            mediaType: Type.TOptional<Type.TString>;
+            digest: Type.TOptional<Type.TString>;
+        }>>;
+        note: Type.TOptional<Type.TString>;
+    }>>;
+}>;
+export declare const FrontierSnapshotSchema: Type.TObject<{
+    id: Type.TString;
+    taskProfileId: Type.TString;
+    taskProfileRevision: Type.TInteger;
+    productRevision: Type.TString;
+    createdAt: Type.TString;
+    frontierCandidateIds: Type.TArray<Type.TString>;
+    dominatedCandidateIds: Type.TArray<Type.TString>;
+    ineligibleCandidateIds: Type.TArray<Type.TString>;
+    unresolvedCandidateIds: Type.TArray<Type.TString>;
+    comparisons: Type.TArray<Type.TObject<{
+        candidateId: Type.TString;
+        otherCandidateId: Type.TString;
+        state: Type.TUnion<Type.TLiteral<string>[]>;
+        reasons: Type.TArray<Type.TString>;
+    }>>;
+    assessmentIds: Type.TArray<Type.TString>;
+    calculationEvidence: Type.TArray<Type.TObject<{
+        id: Type.TString;
+        source: Type.TUnion<Type.TLiteral<string>[]>;
+        uri: Type.TString;
+        mediaType: Type.TOptional<Type.TString>;
+        digest: Type.TOptional<Type.TString>;
+    }>>;
+}>;
+export declare const FrontierRecommendationSchema: Type.TObject<{
+    frontierSnapshotId: Type.TString;
+    selectedCandidateId: Type.TString;
+    currentCandidateId: Type.TString;
+    reasons: Type.TArray<Type.TString>;
+    preferenceEvidence: Type.TArray<Type.TObject<{
+        id: Type.TString;
+        source: Type.TUnion<Type.TLiteral<string>[]>;
+        uri: Type.TString;
+        mediaType: Type.TOptional<Type.TString>;
+        digest: Type.TOptional<Type.TString>;
+    }>>;
+}>;
+export declare const INTENT_OUTPUT_SCHEMAS: {
+    readonly establish_evals: Type.TObject<{
+        taskProfile: Type.TObject<{
+            confirmedAt: Type.TString;
+            id: Type.TString;
+            revision: Type.TInteger;
+            name: Type.TString;
+            goal: Type.TString;
+            objectives: Type.TArray<Type.TObject<{
+                metricId: Type.TUnion<Type.TLiteral<string>[]>;
+                required: Type.TBoolean;
+                minimumSamples: Type.TInteger;
+                tolerance: Type.TNumber;
+                constraint: Type.TOptional<Type.TObject<{
+                    operator: Type.TUnion<Type.TLiteral<string>[]>;
+                    value: Type.TNumber;
+                }>>;
+            }>>;
+            createdAt: Type.TString;
+        }>;
+        policy: Type.TObject<{
+            mode: Type.TLiteral<"supervised_graduation">;
+            evaluationBudget: Type.TObject<{
+                currency: Type.TString;
+                maximumSpend: Type.TNumber;
+                maximumCandidateCount: Type.TOptional<Type.TInteger>;
+                maximumRunsPerCandidate: Type.TOptional<Type.TInteger>;
+                expiresAt: Type.TOptional<Type.TString>;
+            }>;
+            automaticSwapsEnabled: Type.TBoolean;
+            confirmationRequiredUntilVerifiedSwaps: Type.TInteger;
+            requirePostSwapVerification: Type.TLiteral<true>;
+            rollbackOnRegression: Type.TBoolean;
+            checkPolicy: Type.TObject<{
+                baselineAssessment: Type.TObject<{
+                    afterCompletedTasks: Type.TOptional<Type.TInteger>;
+                    afterElapsedSeconds: Type.TOptional<Type.TInteger>;
+                }>;
+                frontierReassessment: Type.TObject<{
+                    afterCompletedTasks: Type.TOptional<Type.TInteger>;
+                    afterElapsedSeconds: Type.TOptional<Type.TInteger>;
+                    onModelCatalogChange: Type.TBoolean;
+                }>;
+                regressionCheck: Type.TObject<{
+                    afterCompletedTasks: Type.TOptional<Type.TInteger>;
+                    afterElapsedSeconds: Type.TOptional<Type.TInteger>;
+                    thresholds: Type.TArray<Type.TObject<{
+                        metricId: Type.TUnion<Type.TLiteral<string>[]>;
+                        relativeChangeAtLeast: Type.TNumber;
+                    }>>;
+                }>;
+                postSwapVerification: Type.TObject<{
+                    afterCompletedTasks: Type.TOptional<Type.TInteger>;
+                    afterElapsedSeconds: Type.TOptional<Type.TInteger>;
+                }>;
+                cooldownSeconds: Type.TInteger;
+                execution: Type.TObject<{
+                    mode: Type.TUnion<Type.TLiteral<string>[]>;
+                    infrastructureChangesAllowed: Type.TBoolean;
+                }>;
+            }>;
+        }>;
+        observability: Type.TObject<{
+            status: Type.TUnion<Type.TLiteral<string>[]>;
+            requiredMetricIds: Type.TArray<Type.TUnion<Type.TLiteral<string>[]>>;
+            coveredMetricIds: Type.TArray<Type.TUnion<Type.TLiteral<string>[]>>;
+            missingMetricIds: Type.TArray<Type.TUnion<Type.TLiteral<string>[]>>;
+            checkedAt: Type.TString;
+        }>;
+    }>;
+    readonly instrument_observability: Type.TObject<{
+        coveredMetricIds: Type.TArray<Type.TUnion<Type.TLiteral<string>[]>>;
+        missingMetricIds: Type.TArray<Type.TUnion<Type.TLiteral<string>[]>>;
+    }>;
+    readonly run_assessment: Type.TObject<{
+        baselineEvidenceSufficient: Type.TBoolean;
+        metrics: Type.TOptional<Type.TArray<Type.TObject<{
+            metricId: Type.TUnion<Type.TLiteral<string>[]>;
+            state: Type.TUnion<Type.TLiteral<string>[]>;
+            scope: Type.TUnion<Type.TLiteral<string>[]>;
+            direction: Type.TUnion<Type.TLiteral<string>[]>;
+            method: Type.TUnion<Type.TLiteral<string>[]>;
+            value: Type.TOptional<Type.TNumber>;
+            unit: Type.TString;
+            sampleCount: Type.TInteger;
+            interval: Type.TOptional<Type.TObject<{
+                lower: Type.TNumber;
+                upper: Type.TNumber;
+                confidenceLevel: Type.TOptional<Type.TNumber>;
+            }>>;
+            window: Type.TOptional<Type.TObject<{
+                startedAt: Type.TString;
+                endedAt: Type.TString;
+            }>>;
+            evidence: Type.TArray<Type.TObject<{
+                id: Type.TString;
+                source: Type.TUnion<Type.TLiteral<string>[]>;
+                uri: Type.TString;
+                mediaType: Type.TOptional<Type.TString>;
+                digest: Type.TOptional<Type.TString>;
+            }>>;
+            note: Type.TOptional<Type.TString>;
+        }>>>;
+    }>;
+    readonly discover_candidates: Type.TObject<{
+        candidates: Type.TArray<Type.TObject<{
+            candidateId: Type.TString;
+            modelId: Type.TString;
+            modelVersion: Type.TOptional<Type.TString>;
+            reasons: Type.TArray<Type.TString>;
+        }>>;
+    }>;
+    readonly run_challenger_trials: Type.TObject<{
+        assessments: Type.TArray<Type.TObject<{
+            id: Type.TString;
+            taskProfileId: Type.TString;
+            taskProfileRevision: Type.TInteger;
+            candidateId: Type.TString;
+            modelId: Type.TString;
+            modelVersion: Type.TOptional<Type.TString>;
+            productRevision: Type.TString;
+            harnessId: Type.TString;
+            startedAt: Type.TString;
+            completedAt: Type.TString;
+            metrics: Type.TArray<Type.TObject<{
+                metricId: Type.TUnion<Type.TLiteral<string>[]>;
+                state: Type.TUnion<Type.TLiteral<string>[]>;
+                scope: Type.TUnion<Type.TLiteral<string>[]>;
+                direction: Type.TUnion<Type.TLiteral<string>[]>;
+                method: Type.TUnion<Type.TLiteral<string>[]>;
+                value: Type.TOptional<Type.TNumber>;
+                unit: Type.TString;
+                sampleCount: Type.TInteger;
+                interval: Type.TOptional<Type.TObject<{
+                    lower: Type.TNumber;
+                    upper: Type.TNumber;
+                    confidenceLevel: Type.TOptional<Type.TNumber>;
+                }>>;
+                window: Type.TOptional<Type.TObject<{
+                    startedAt: Type.TString;
+                    endedAt: Type.TString;
+                }>>;
+                evidence: Type.TArray<Type.TObject<{
+                    id: Type.TString;
+                    source: Type.TUnion<Type.TLiteral<string>[]>;
+                    uri: Type.TString;
+                    mediaType: Type.TOptional<Type.TString>;
+                    digest: Type.TOptional<Type.TString>;
+                }>>;
+                note: Type.TOptional<Type.TString>;
+            }>>;
+        }>>;
+    }>;
+    readonly calculate_frontier: Type.TObject<{
+        assessments: Type.TArray<Type.TObject<{
+            id: Type.TString;
+            taskProfileId: Type.TString;
+            taskProfileRevision: Type.TInteger;
+            candidateId: Type.TString;
+            modelId: Type.TString;
+            modelVersion: Type.TOptional<Type.TString>;
+            productRevision: Type.TString;
+            harnessId: Type.TString;
+            startedAt: Type.TString;
+            completedAt: Type.TString;
+            metrics: Type.TArray<Type.TObject<{
+                metricId: Type.TUnion<Type.TLiteral<string>[]>;
+                state: Type.TUnion<Type.TLiteral<string>[]>;
+                scope: Type.TUnion<Type.TLiteral<string>[]>;
+                direction: Type.TUnion<Type.TLiteral<string>[]>;
+                method: Type.TUnion<Type.TLiteral<string>[]>;
+                value: Type.TOptional<Type.TNumber>;
+                unit: Type.TString;
+                sampleCount: Type.TInteger;
+                interval: Type.TOptional<Type.TObject<{
+                    lower: Type.TNumber;
+                    upper: Type.TNumber;
+                    confidenceLevel: Type.TOptional<Type.TNumber>;
+                }>>;
+                window: Type.TOptional<Type.TObject<{
+                    startedAt: Type.TString;
+                    endedAt: Type.TString;
+                }>>;
+                evidence: Type.TArray<Type.TObject<{
+                    id: Type.TString;
+                    source: Type.TUnion<Type.TLiteral<string>[]>;
+                    uri: Type.TString;
+                    mediaType: Type.TOptional<Type.TString>;
+                    digest: Type.TOptional<Type.TString>;
+                }>>;
+                note: Type.TOptional<Type.TString>;
+            }>>;
+        }>>;
+        frontierSnapshot: Type.TObject<{
+            id: Type.TString;
+            taskProfileId: Type.TString;
+            taskProfileRevision: Type.TInteger;
+            productRevision: Type.TString;
+            createdAt: Type.TString;
+            frontierCandidateIds: Type.TArray<Type.TString>;
+            dominatedCandidateIds: Type.TArray<Type.TString>;
+            ineligibleCandidateIds: Type.TArray<Type.TString>;
+            unresolvedCandidateIds: Type.TArray<Type.TString>;
+            comparisons: Type.TArray<Type.TObject<{
+                candidateId: Type.TString;
+                otherCandidateId: Type.TString;
+                state: Type.TUnion<Type.TLiteral<string>[]>;
+                reasons: Type.TArray<Type.TString>;
+            }>>;
+            assessmentIds: Type.TArray<Type.TString>;
+            calculationEvidence: Type.TArray<Type.TObject<{
+                id: Type.TString;
+                source: Type.TUnion<Type.TLiteral<string>[]>;
+                uri: Type.TString;
+                mediaType: Type.TOptional<Type.TString>;
+                digest: Type.TOptional<Type.TString>;
+            }>>;
+        }>;
+        recommendation: Type.TOptional<Type.TObject<{
+            frontierSnapshotId: Type.TString;
+            selectedCandidateId: Type.TString;
+            currentCandidateId: Type.TString;
+            reasons: Type.TArray<Type.TString>;
+            preferenceEvidence: Type.TArray<Type.TObject<{
+                id: Type.TString;
+                source: Type.TUnion<Type.TLiteral<string>[]>;
+                uri: Type.TString;
+                mediaType: Type.TOptional<Type.TString>;
+                digest: Type.TOptional<Type.TString>;
+            }>>;
+        }>>;
+    }>;
+    readonly investigate_regression: Type.TObject<{
+        regressionDetected: Type.TBoolean;
+        affectedMetricIds: Type.TArray<Type.TUnion<Type.TLiteral<string>[]>>;
+        likelyCause: Type.TOptional<Type.TString>;
+    }>;
+    readonly apply_model_swap: Type.TObject<{
+        appliedCandidateId: Type.TString;
+        previousCandidateId: Type.TOptional<Type.TString>;
+    }>;
+    readonly verify_model_swap: Type.TObject<{
+        verified: Type.TBoolean;
+        regressionDetected: Type.TBoolean;
+    }>;
+    readonly rollback_model_swap: Type.TObject<{
+        restoredCandidateId: Type.TString;
+    }>;
+};
+export declare function intentOutputSchemaFor(kind: IntentKind): TSchema;
+export declare function validateIntentOutputSchema(kind: IntentKind, output: unknown): readonly string[];
+export declare function completionToolSchemaFor(kind: IntentKind): TSchema;

package/dist/protocol/src/schemas.js

@@ -0,0 +1,224 @@
+import { Type } from "typebox";
+import { Check, Errors } from "typebox/value";
+export const OPENMERIT_SCHEMA_DIALECT = "https://json-schema.org/draft/2020-12/schema";
+export const OPENMERIT_RESULT_SCHEMA_VERSION = "0.2";
+const strictObject = (properties, options = {}) => Type.Object(properties, { additionalProperties: false, ...options });
+const metricIds = [
+    "task_quality", "task_success", "consistency", "instruction_following",
+    "tool_use_performance", "structured_output_reliability", "hallucination_rate",
+    "reasoning_efficiency", "input_token_consumption", "output_token_consumption",
+    "total_task_cost", "cost_per_successful_task", "time_to_first_useful_output",
+    "end_to_end_task_latency", "throughput", "retry_rate", "agent_steps",
+    "recovery_ability", "context_handling", "retrieval_use_quality",
+    "long_horizon_performance", "human_intervention_rate", "preference_fit",
+];
+const evidenceSources = [
+    "harness_trace", "evaluation_artifact", "observability_record", "external_source", "user_feedback",
+];
+const stringEnum = (values) => Type.Union(values.map((value) => Type.Literal(value)));
+const cadenceSchema = () => strictObject({
+    afterCompletedTasks: Type.Optional(Type.Integer({ minimum: 1 })),
+    afterElapsedSeconds: Type.Optional(Type.Integer({ minimum: 1 })),
+});
+export const EvidenceReferenceSchema = strictObject({
+    id: Type.String({ minLength: 1 }),
+    source: stringEnum(evidenceSources),
+    uri: Type.String({ minLength: 1 }),
+    mediaType: Type.Optional(Type.String({ minLength: 1 })),
+    digest: Type.Optional(Type.String({ minLength: 1 })),
+});
+export const MetricObjectiveSchema = strictObject({
+    metricId: stringEnum(metricIds),
+    required: Type.Boolean(),
+    minimumSamples: Type.Integer({ minimum: 1 }),
+    tolerance: Type.Number({ minimum: 0 }),
+    constraint: Type.Optional(strictObject({
+        operator: stringEnum(["at_least", "at_most"]),
+        value: Type.Number(),
+    })),
+});
+export const TaskProfileSchema = strictObject({
+    id: Type.String({ minLength: 1 }),
+    revision: Type.Integer({ minimum: 1 }),
+    name: Type.String({ minLength: 1 }),
+    goal: Type.String({ minLength: 1 }),
+    objectives: Type.Array(MetricObjectiveSchema, { minItems: 1 }),
+    createdAt: Type.String({ minLength: 1 }),
+    confirmedAt: Type.Optional(Type.String({ minLength: 1 })),
+});
+export const ObservabilityCoverageSchema = strictObject({
+    status: stringEnum(["ready", "incomplete"]),
+    requiredMetricIds: Type.Array(stringEnum(metricIds)),
+    coveredMetricIds: Type.Array(stringEnum(metricIds)),
+    missingMetricIds: Type.Array(stringEnum(metricIds)),
+    checkedAt: Type.String({ minLength: 1 }),
+});
+const ConfirmedTaskProfileSchema = strictObject({
+    ...TaskProfileSchema.properties,
+    confirmedAt: Type.String({ minLength: 1 }),
+});
+export const SupervisedGraduationPolicySchema = strictObject({
+    mode: Type.Literal("supervised_graduation"),
+    evaluationBudget: strictObject({
+        currency: Type.String({ minLength: 1 }),
+        maximumSpend: Type.Number({ minimum: 0 }),
+        maximumCandidateCount: Type.Optional(Type.Integer({ minimum: 1 })),
+        maximumRunsPerCandidate: Type.Optional(Type.Integer({ minimum: 1 })),
+        expiresAt: Type.Optional(Type.String({ minLength: 1 })),
+    }),
+    automaticSwapsEnabled: Type.Boolean(),
+    confirmationRequiredUntilVerifiedSwaps: Type.Integer({ minimum: 0 }),
+    requirePostSwapVerification: Type.Literal(true),
+    rollbackOnRegression: Type.Boolean(),
+    checkPolicy: strictObject({
+        baselineAssessment: cadenceSchema(),
+        frontierReassessment: strictObject({
+            afterCompletedTasks: Type.Optional(Type.Integer({ minimum: 1 })),
+            afterElapsedSeconds: Type.Optional(Type.Integer({ minimum: 1 })),
+            onModelCatalogChange: Type.Boolean(),
+        }),
+        regressionCheck: strictObject({
+            afterCompletedTasks: Type.Optional(Type.Integer({ minimum: 1 })),
+            afterElapsedSeconds: Type.Optional(Type.Integer({ minimum: 1 })),
+            thresholds: Type.Array(strictObject({
+                metricId: stringEnum(metricIds),
+                relativeChangeAtLeast: Type.Number({ exclusiveMinimum: 0 }),
+            })),
+        }),
+        postSwapVerification: cadenceSchema(),
+        cooldownSeconds: Type.Integer({ minimum: 0 }),
+        execution: strictObject({
+            mode: stringEnum(["active_session_only", "persistent"]),
+            infrastructureChangesAllowed: Type.Boolean(),
+        }),
+    }),
+});
+export const MetricRecordSchema = strictObject({
+    metricId: stringEnum(metricIds),
+    state: stringEnum(["measured", "not_applicable", "not_configured", "insufficient_evidence"]),
+    scope: stringEnum(["single_run", "sample_window"]),
+    direction: stringEnum(["maximize", "minimize"]),
+    method: stringEnum(["observed", "derived", "evaluator", "human", "external"]),
+    value: Type.Optional(Type.Number()),
+    unit: Type.String({ minLength: 1 }),
+    sampleCount: Type.Integer({ minimum: 0 }),
+    interval: Type.Optional(strictObject({
+        lower: Type.Number(),
+        upper: Type.Number(),
+        confidenceLevel: Type.Optional(Type.Number({ minimum: 0, maximum: 1 })),
+    })),
+    window: Type.Optional(strictObject({
+        startedAt: Type.String({ minLength: 1 }),
+        endedAt: Type.String({ minLength: 1 }),
+    })),
+    evidence: Type.Array(EvidenceReferenceSchema),
+    note: Type.Optional(Type.String()),
+});
+export const CandidateDescriptorSchema = strictObject({
+    candidateId: Type.String({ minLength: 1 }),
+    modelId: Type.String({ minLength: 1 }),
+    modelVersion: Type.Optional(Type.String({ minLength: 1 })),
+    reasons: Type.Array(Type.String()),
+});
+export const CandidateAssessmentSchema = strictObject({
+    id: Type.String({ minLength: 1 }),
+    taskProfileId: Type.String({ minLength: 1 }),
+    taskProfileRevision: Type.Integer({ minimum: 1 }),
+    candidateId: Type.String({ minLength: 1 }),
+    modelId: Type.String({ minLength: 1 }),
+    modelVersion: Type.Optional(Type.String({ minLength: 1 })),
+    productRevision: Type.String({ minLength: 1 }),
+    harnessId: Type.String({ minLength: 1 }),
+    startedAt: Type.String({ minLength: 1 }),
+    completedAt: Type.String({ minLength: 1 }),
+    metrics: Type.Array(MetricRecordSchema),
+});
+export const FrontierSnapshotSchema = strictObject({
+    id: Type.String({ minLength: 1 }),
+    taskProfileId: Type.String({ minLength: 1 }),
+    taskProfileRevision: Type.Integer({ minimum: 1 }),
+    productRevision: Type.String({ minLength: 1 }),
+    createdAt: Type.String({ minLength: 1 }),
+    frontierCandidateIds: Type.Array(Type.String()),
+    dominatedCandidateIds: Type.Array(Type.String()),
+    ineligibleCandidateIds: Type.Array(Type.String()),
+    unresolvedCandidateIds: Type.Array(Type.String()),
+    comparisons: Type.Array(strictObject({
+        candidateId: Type.String({ minLength: 1 }),
+        otherCandidateId: Type.String({ minLength: 1 }),
+        state: stringEnum(["dominates", "does_not_dominate", "unresolved"]),
+        reasons: Type.Array(Type.String()),
+    })),
+    assessmentIds: Type.Array(Type.String()),
+    calculationEvidence: Type.Array(EvidenceReferenceSchema, { minItems: 1 }),
+});
+export const FrontierRecommendationSchema = strictObject({
+    frontierSnapshotId: Type.String({ minLength: 1 }),
+    selectedCandidateId: Type.String({ minLength: 1 }),
+    currentCandidateId: Type.String({ minLength: 1 }),
+    reasons: Type.Array(Type.String()),
+    preferenceEvidence: Type.Array(EvidenceReferenceSchema),
+});
+export const INTENT_OUTPUT_SCHEMAS = {
+    establish_evals: strictObject({
+        taskProfile: ConfirmedTaskProfileSchema,
+        policy: SupervisedGraduationPolicySchema,
+        observability: ObservabilityCoverageSchema,
+    }),
+    instrument_observability: strictObject({
+        coveredMetricIds: Type.Array(stringEnum(metricIds)),
+        missingMetricIds: Type.Array(stringEnum(metricIds)),
+    }),
+    run_assessment: strictObject({
+        baselineEvidenceSufficient: Type.Boolean(),
+        metrics: Type.Optional(Type.Array(MetricRecordSchema)),
+    }),
+    discover_candidates: strictObject({
+        candidates: Type.Array(CandidateDescriptorSchema),
+    }),
+    run_challenger_trials: strictObject({
+        assessments: Type.Array(CandidateAssessmentSchema),
+    }),
+    calculate_frontier: strictObject({
+        assessments: Type.Array(CandidateAssessmentSchema),
+        frontierSnapshot: FrontierSnapshotSchema,
+        recommendation: Type.Optional(FrontierRecommendationSchema),
+    }),
+    investigate_regression: strictObject({
+        regressionDetected: Type.Boolean(),
+        affectedMetricIds: Type.Array(stringEnum(metricIds)),
+        likelyCause: Type.Optional(Type.String()),
+    }),
+    apply_model_swap: strictObject({
+        appliedCandidateId: Type.String({ minLength: 1 }),
+        previousCandidateId: Type.Optional(Type.String({ minLength: 1 })),
+    }),
+    verify_model_swap: strictObject({
+        verified: Type.Boolean(),
+        regressionDetected: Type.Boolean(),
+    }),
+    rollback_model_swap: strictObject({
+        restoredCandidateId: Type.String({ minLength: 1 }),
+    }),
+};
+export function intentOutputSchemaFor(kind) {
+    return INTENT_OUTPUT_SCHEMAS[kind];
+}
+export function validateIntentOutputSchema(kind, output) {
+    const schema = intentOutputSchemaFor(kind);
+    if (Check(schema, output))
+        return [];
+    return [...Errors(schema, output)].map((error) => `${error.instancePath || "/"} ${error.message}`);
+}
+export function completionToolSchemaFor(kind) {
+    return strictObject({
+        intentId: Type.String({ minLength: 1 }),
+        status: stringEnum(["succeeded", "failed", "cancelled"]),
+        summary: Type.String({ minLength: 1 }),
+        evidence: Type.Array(EvidenceReferenceSchema),
+        outputs: Type.Optional(intentOutputSchemaFor(kind)),
+        errorCode: Type.Optional(Type.String({ minLength: 1 })),
+        errorMessage: Type.Optional(Type.String({ minLength: 1 })),
+        retryable: Type.Optional(Type.Boolean()),
+    });
+}