Review policies

The classifier recommends; the consumer-owned gate.PermissionReviewPolicy decides whether that recommendation is eligible for automatic one-shot approval. A policy is sealed by its constructor and is deliberately at least as restrictive as Harness’s hard review ceiling.

Policy shape

PermissionReviewPolicy has these exported fields:

NewPermissionReviewPolicy(revision, maximum, minimum, absoluteHuman, material) validates and seals the value. DefaultPermissionReviewPolicy(revision) permits low and medium with unknown authorization and requires medium authorization for high, with ReviewRiskHigh as the maximum automatic risk. Sealed() reports whether a constructor seal is present; rig.WithPermissionReviewPolicy rejects an unsealed literal.

Risk constants are ReviewRiskLow, ReviewRiskMedium, ReviewRiskHigh, ReviewRiskCritical. Authorization constants are ReviewAuthorizationUnknown, ReviewAuthorizationLow, ReviewAuthorizationMedium, and ReviewAuthorizationHigh. Recommendation constants are ReviewAllow and ReviewNeedsHuman.

Assessment

PermissionAssessment carries Basis ReviewBasis, Risk ReviewRisk, Authorization ReviewAuthorization, Categories []ReviewRiskCategory, Recommendation ReviewRecommendation, and Rationale string. Rationale is valid UTF-8, required for non-low risk, and bounded by MaxPermissionReviewRationaleBytes = 2048. Categories are validated by ValidateReviewCategories: only the 14 closed categories are accepted, with no duplicates and no more than MaxReviewCategories = 14.

ReviewContext.Truncation.Material and the policy’s MaterialTruncation mask are both checked. A classifier cannot make material context loss disappear by recommending allow. The assessment basis must exactly equal the subject basis.

Decision

EvaluatePermissionAssessment(policy, subject, assessment) returns ReviewDecision{Eligible bool; Reason ReviewDecisionReason}. It validates the subject digest, sealed policy, policy revision, basis, assessment, recommendation, truncation, risk ceiling, absolute-human categories, and authorization in that order. Reasons are closed: ReviewDecisionEligible, ReviewDecisionInvalidPolicy, ReviewDecisionInvalidAssessment, ReviewDecisionBasisMismatch, ReviewDecisionRecommendation, ReviewDecisionRiskCeiling, ReviewDecisionAuthorization, ReviewDecisionAbsoluteHuman, ReviewDecisionMaterialTruncation, ReviewDecisionNoApplicableClassifier, and ReviewDecisionClassifierStatus.

package main

import "github.com/looprig/harness/pkg/gate"

func policy() (gate.PermissionReviewPolicy, error) {
    return gate.NewPermissionReviewPolicy(
        "permission-review/v1",
        gate.ReviewRiskHigh,
        map[gate.ReviewRisk]gate.ReviewAuthorization{
            gate.ReviewRiskLow: gate.ReviewAuthorizationUnknown,
            gate.ReviewRiskMedium: gate.ReviewAuthorizationUnknown,
            gate.ReviewRiskHigh: gate.ReviewAuthorizationMedium,
        },
        nil,
        0,
    )
}

An eligible decision is not a gate response. It is an input to the private session review path, which still rechecks observations and live basis before one-shot approval.

Combine

CombinePermissionAssessments(policy, classifiers, outcomes) applies conjunctive semantics. Every ordered classifier contributes one outcome. Outcomes must carry the classifier’s exact revision, a valid subject digest, the policy revision, and a unique common subject identity. An applicable non-allowed status returns ReviewDecisionClassifierStatus; an applicable allowed assessment is evaluated locally; a non-applicable outcome is neutral only with ReviewStatusNotApplicable; no applicable outcome returns ReviewDecisionNoApplicableClassifier.

This prevents one permissive classifier from overriding another applicable classifier’s failure. It also prevents combining assessments produced for different tool executions, context revisions, security ceilings, or policy revisions.

Configuration

The rig boundary wires this policy with rig.WithPermissionReviewPolicy(policy). It must be paired with rig.WithPermissionClassifiers(classifiers); either alone is an incomplete permission-review configuration. Evidence and security-ceiling options are required according to classifier definitions, while rig.WithPermissionReviewObservations is optional. rig.Define returns typed *rig.DefinitionError values for invalid, incomplete, missing, unused, or duplicate options.

Source: pkg/gate/review.go, pkg/gate/review_policy.go, and option validation in pkg/rig/definition.go. Policy and subject validation are covered by pkg/rig/hustle_fingerprint_test.go and the permission-review option tests in pkg/rig.

Source and proof

• PermissionReviewPolicy
• assessment combination
• policy and fingerprint tests