Documentation
¶
Overview ¶
Package policy holds the gate.command-safety classifier's own deterministic command-safety consistency policy.
This is deliberately NOT Harness's PermissionReviewPolicy (github.com/looprig/harness/pkg/gate.PermissionReviewPolicy): that type is the consumer-owned, local decision ceiling Harness itself applies AFTER a classifier's assessment has been validated, and it independently re-derives eligibility from the assessment's Risk/Authorization/Categories using its own matrix. This package never touches that matrix and never grants, denies, or widens eligibility on its own.
Instead, Policy encodes this classifier's own closed-taxonomy decision structure — which risk floor a reported category implies, which category can never resolve to an eligible allow, and which authorization threshold this classifier itself expects at each risk level — as deterministic Go data rather than leaving that structure purely to prompt-following. Reconcile consumes an already strictly-decoded model assessment (as internal/wire produces it) and, before that assessment ever crosses this module's public boundary, tightens an inconsistent allow recommendation to needs_human. It never lowers a reported risk, never raises a reported authorization, and never turns an existing needs_human back into allow: it only ever moves toward the more conservative outcome.
Index ¶
Constants ¶
const DefaultRevision = "command-safety-policy/v1"
DefaultRevision is the initial command-safety policy revision. Per design §19.2 ("prompt, policy, schema, wire, and corpus revisions are independently explicit and included in classifier identity"), bump this whenever DefaultPolicy's taxonomy content changes meaning.
Variables ¶
This section is empty.
Functions ¶
func Reconcile ¶
func Reconcile(p Policy, assessment gate.PermissionAssessment) gate.PermissionAssessment
Reconcile deterministically re-derives a fail-secure recommendation from an already strictly-decoded model assessment. An assessment that does not recommend allow is already the most conservative outcome and is returned unchanged. Otherwise, Reconcile checks, in order:
- A risk value outside the closed four-value gate.ReviewRisk enum (low/medium/high/critical) — including the zero value — always needs a human. Reconcile is a fail-secure function and must not depend on its caller having already enforced the closed enum (in production, internal/wire's strict decoders do, before Reconcile ever runs, but Reconcile itself must not rely on that): riskRank returns 0 for exactly this case and for no valid enum value, so this check and the critical-risk check below share one comparison.
- Critical risk always needs a human, regardless of category or authorization.
- Any category in p.AbsoluteHumanCategories always needs a human, at any authorization level.
- Any category in p.CategoryMinimumRisk whose floor exceeds the assessment's own reported risk is an internally inconsistent report and needs a human.
- The assessment's reported authorization must meet or exceed p.MinimumAuthorization for its reported risk.
Reconcile never lowers the reported risk, never raises the reported authorization, and never invents or drops a category: only Recommendation may change, and only from allow to needs_human.
Types ¶
type Policy ¶
type Policy struct {
Revision string
// CategoryMinimumRisk floors the risk every listed category implies. A
// decoded assessment that selects a category from this table yet
// reports a risk below that category's floor is internally
// inconsistent — the reported risk understates what the model's own
// category selection already claims — and Reconcile treats that
// inconsistency as unsafe to auto-approve.
CategoryMinimumRisk map[gate.ReviewRiskCategory]gate.ReviewRisk
// AbsoluteHumanCategories can never resolve to an eligible allow from
// this classifier, at any authorization level, including a strong,
// post-warning authorization: a confirmed disclosure to an untrusted
// destination, for example, is not made safe by user approval.
AbsoluteHumanCategories map[gate.ReviewRiskCategory]struct{}
// MinimumAuthorization is this classifier's own default minimum
// authorization required at each risk level before recommending allow,
// independent of category. It mirrors, but does not replace or
// substitute for, Harness's own local ceiling.
MinimumAuthorization map[gate.ReviewRisk]gate.ReviewAuthorization
}
Policy is the classifier's own deterministic command-safety consistency policy. Its zero value (no revision, nil tables) is a structurally valid "no additional tightening" policy; DefaultPolicy returns the populated, named default.
func DefaultPolicy ¶
func DefaultPolicy() Policy
DefaultPolicy returns the initial command-safety taxonomy policy. Tenant or product code may construct a stricter Policy value directly; this package places no floor on how conservative a caller-supplied Policy may be, only on how it is applied (Reconcile only ever tightens).
Source Files