SDK changelogs

​

TypeScript

​

2.21.2

• Fixed protectStream() to send the request body shape the streaming evaluate endpoint expects: a single-item items array ({action_type, actor_id, context}). The API key is now sent in the Authorization header and no longer in the request body.
• parseSseStream() now treats the event: complete frame as end-of-stream, in addition to event: done.
• parseSseStream() error frames now read the error_code field, falling back to code.
• Decision frames now prefer the permit_token field over the legacy decision_id.

​

2.21.1

• Fixed evaluate() silently dropping evaluation_profile, override, and completion_proofs from the request. These fields are now sent when set. Callers not using them are unaffected.

​

2.21.0

• New AtlaSentClient.explainAuthority({ principalId, requestedScope, resourceId? }) read-only method. It returns paths (each matching authority mechanism: direct_grant, delegation, or role_capability) and unresolved findings.
• New client.authorityIntelligence.integrityAudit(query?) method, with an optional decisionWindowDays option (integer, 1–3650). Requires an API key with the authority_intelligence:read scope. Adds the countFindingsByClassification() helper and IntegrityReport and related types.
• EvaluateResponse gains optional humanApprovalRequired and humanApprovalStatus fields (not_required, pending, satisfied, rejected, expired, revoked). They are undefined on runtimes that do not emit them.
• CrossOrgPermissionCheckResult gains trust_precheck_passed, authorizes_execution (always false), requires_local_authority_evaluation (always true), and conditions_evaluated. New summarizeTrustPrecheck() helper.
• Deprecated: allowed and summarizeCrossOrgPermission() on cross-org permission checks, and CrossOrgPermissionCheckRequest.identity_id, which the server does not read.
• Added NO_EXECUTION_WINDOW and OUTSIDE_EXECUTION_WINDOW to DENY_CODES. Fixed checkCrossOrgPermission() and listCrossOrgPermissionChecks(), which called a path that returned 404. They now use /v1/federation/permission-check[s].

​

2.20.0

• EvaluateRequest now uses the canonical actor_id and action_type field names.
• Deprecated: agent and action on EvaluateRequest. They are still accepted as aliases and emit a one-time console.warn. They will be removed in a future major release. The request sent over the wire is unchanged.
• New resolveEvaluateIdentity(input) helper resolves { action_type, actor_id } from either the canonical or the legacy input shape. evaluate(), evaluatePreflight(), and protectStream() all accept both shapes.

​

2.19.0

• New client.clinicalTrials sub-client for the clinical trial unblinding gate. It provides list, get, and history (require clinical:read), plus blind, requestUnblind, emergencyUnblind, and verifyPermit (require clinical:manage).

​

2.18.0

• New AtlaSentClient.complianceControls({ framework?, from?, to? }) method. It reads the compliance control catalog with a live enforcement status for each clause (enforced, partial, not_enforced, no_data, attested). Requires the compliance:read scope.
• New AtlaSentClient.complianceEvidencePack({ framework, from?, to? }) method. It fetches a signed compliance evidence pack for one framework. The returned bundle can be hashed offline against sha256, and signature, signingStatus, and keyId carry the signing state.

​

2.17.0 — 2026-06-10

• New client.smsOtp sub-client with send() and verify() for one-time-passcode step-up authentication (break_glass, api_key_create, governance_hold_approve). It requires JWT session auth and does not accept API keys.
• New client.usageMetering sub-client with list() for paginated billable evaluation records and summary() for usage aggregated by billing period. Requires the usage:read scope.

​

2.16.0 — 2026-06-04

• New optional evaluation_profile field (EvaluationProfile) on EvaluateRequest controls which evaluation layers run. Pass "basic" to skip snapshot enforcement. Unknown values fall back to "standard" on the server.
• New optional override field (EmergencyOverrideV1) on EvaluateRequest for an emergency override that clears snapshot hard blocks. It is only evaluated when evaluation_profile is "advanced" or "enterprise", and authority_actor_id must differ from actor_id.

​

2.15.0 — 2026-06-04

• New client.submitAssertion() submits a point-in-time fact from an external source (e.g. GitHub, Stripe, Slack) so that policy rules can gate on it during evaluation. It returns assertion_id, payload_hash, and reused. Requires the assertions:write scope.

​

2.14.0 — 2026-06-03

• New optional state_snapshot field on EvaluateRequest. Action classes that require a state snapshot return decision: "deny" with deny_code: "SNAPSHOT_REQUIRED" when it is omitted. The snapshot is recorded in the audit chain alongside the permit.

​

2.13.0 — 2026-06-03

• New client.getLicense() and client.verifyLicense(blob) methods for self-hosted and air-gapped deployments, with new LicenseStatus and LicenseVerifyResult types. A valid: false verification result is returned, not thrown.

​

2.12.0 — 2026-05-28

• The SDK now includes a trust-root snapshot and a TrustRootManager that refreshes it in the background every 4 hours. getGlobalTrustRootManager() returns a shared instance, and checkExpiry() reports "ok", "half_life", or "expired".
• verifyBundle() uses the global trust-root snapshot automatically when no trustRoot option is passed.
• Breaking: verifyAuditBundle() now throws BundleVerificationError when the trust snapshot has expired, instead of returning { verified: false }. Pass allowExpiredSnapshot: true to opt out, e.g. in air-gapped environments.
• Revoked signing keys and keys with the wrong role are now rejected with BundleVerificationError (reason: "key_revoked" or "key_role_mismatch"). Adds the "permit_signing_key_revoked" PermitOutcome and the AtlaSentDeniedError.isSigningKeyRevoked getter.

​

2.11.0 — 2026-05-27

• New sub-clients: client.auth (refresh, refreshWithIdp, listIdpConnections), client.scim (SCIM 2.0 users and groups), and client.evidenceBundles (create, get, download).
• New exported Decision type ("allow" | "deny" | "hold" | "escalate"), Permit.permitExpiresAt, and EvaluateResult.reasons (a string[] alongside reason).
• New verifyEvidenceBundle(bundle) function verifies an evidence bundle offline.
• Retries on transient errors now use full-jitter exponential back-off.
• The SDK now warns when it runs in a browser, because API keys should not be used client-side. Suppress the warning with suppressBrowserWarning: true.

​

2.9.0 — 2026-05-25

• AtlaSentDeniedError now carries an outcome field (PermitOutcome) with the isRevoked, isExpired, isConsumed, and isNotFound predicates. Also adds the normalizePermitOutcome(raw) helper.
• protect() and protectWithEvidence() now validate the action format before making a network call, and throw AtlaSentError with code: "bad_request" if it is invalid.
• Evaluate responses now include a riskEnvelope with the composite risk score. Pass explain: true to get the per-factor breakdown in riskEnvelope.factors.

​

2.8.0 — 2026-05-24

• New client.replay({ evaluationId }) re-evaluates a recorded decision against its originally pinned policy bundle. It reports variance kinds such as POLICY_DRIFT, ENGINE_DRIFT, and BUNDLE_MISSING without throwing.

​

2.7.0 — 2026-05-24

• New client.replayDecision(decisionId) re-evaluates a recorded decision against its originally pinned policy bundle and engine version. It reports a variance of NONE, DECISION_CHANGED, or ENVELOPE_DRIFT. No permit is issued. The underlying endpoint is alpha, so its shapes may change.

​

Python

​

2.20.0

• Added the AtlaSentClient.clinical_trials sub-client (ClinicalTrialsClient, in the new atlasent.clinical_client module) for the clinical trial unblinding gate.
• Read methods list, get, and history require the clinical:read scope.
• Write methods blind, request_unblind, emergency_unblind, and verify_permit require the clinical:manage scope.

​

2.19.0

• Added clinical trial blinding and unblinding wire types in atlasent.clinical: ClinicalTrialBlind, ClinicalUnblindingEvent, ClinicalBlindRequest, ClinicalBlindResponse, ClinicalUnblindRequest, ClinicalEmergencyRequest, ClinicalMutationResponse, ClinicalTrialListResponse, ClinicalTrialGetResponse, and ClinicalHistoryResponse.
• Added the ClinicalBlindingStatus and ClinicalUnblindingEventType literals.
• Added the helpers is_unblinded() and latest_unblinding_event().
• All of these are exported from the top-level atlasent package.

​

2.18.0

• Added AtlaSentClient.compliance_controls(*, framework=None, from_=None, to=None) and its AsyncAtlaSentClient counterpart. They read the compliance control catalog, with live enforcement status per regulatory clause (enforced, partial, not_enforced, no_data, or attested). Requires the compliance:read scope.
• Added AtlaSentClient.compliance_evidence_pack(*, framework, from_=None, to=None) and its async counterpart, which fetch a signed compliance evidence pack for one framework. You can hash the returned bundle offline and compare it with sha256. The signature, signing_status, and key_id fields report the signing state.
• Added the result models ComplianceControlsResult, ComplianceEvidencePackResult, ComplianceControl, ComplianceEvidenceControl, ComplianceEvidenceBundle, ComplianceSummary, and ComplianceWindow, plus the ComplianceControlStatus literal.

​

2.17.0 — 2026-06-10

• Added SmsOtpClient (atlasent.sms_otp) for SMS OTP step-up authentication, with send() and verify() methods. It uses JWT session auth only, not API keys. Supported action_context values are "break_glass", "api_key_create", and "governance_hold_approve".
• Added UsageMeteringClient (atlasent.usage_metering) for listing billable evaluation records (list()) and fetching aggregated usage summaries (summary()). Requires the usage:read scope.
• Both clients can be imported directly with from atlasent import SmsOtpClient, UsageMeteringClient.

​

2.16.0 — 2026-06-04

• Added the optional EvaluateRequest.evaluation_profile field ("basic", "standard", "advanced", or "enterprise"). With "basic", snapshot enforcement is skipped, while policy, risk envelope, and audit still run. The server treats unknown values as "standard".
• Added the optional EvaluateRequest.override field for emergency overrides of snapshot hard blocks. It is only evaluated when evaluation_profile is "advanced" or "enterprise". The authority_actor_id must hold the override:execute scope and must be different from actor_id.

​

2.15.0 — 2026-06-03

• Added the optional EvaluateRequest.state_snapshot field for attaching a system state snapshot to evaluate calls.
• The snapshot is required when the action class has requires_state_snapshot=True. If you omit it for such a class, the call returns decision="deny" with the deny code SNAPSHOT_REQUIRED.

​

2.14.0 — 2026-06-03

• Added AtlaSentClient.get_license(), which returns a LicenseStatus for self-hosted and air-gapped deployments. It reports status, org_slug, posture, expires_at, grace_until, features, and the capacity limits eval_limit and seat_limit.
• Added AtlaSentClient.verify_license(blob), which returns a LicenseVerifyResult. A result with valid=False is returned rather than raised.
• LicenseStatus and LicenseVerifyResult are exported from the top-level atlasent package.

​

2.13.0 — 2026-05-28

• Added TrustRootManager (atlasent.trust_root). It loads a bundled trust-root snapshot at startup and refreshes it in the background from https://keys.atlasent.io/.well-known/ every 4 hours. get_global_trust_root_manager() returns the process-wide instance, and check_expiry() returns "ok", "half_life", or "expired".
• verify_bundle() now uses the global trust-root snapshot automatically when you don’t pass a trust_root argument.
• Added the BundleVerificationError exception, the "permit_signing_key_revoked" value for PermitOutcome, and the AtlaSentDeniedError.is_signing_key_revoked property.
• Breaking: verify_audit_bundle() now raises BundleVerificationError on trust failures instead of returning a falsy result. In air-gapped environments, pass allow_expired_snapshot=True to opt out of expiry enforcement.

​

2.12.0 — 2026-05-28

• Added the client.sso sub-client (SsoClient) for managing SSO connections, JIT rules, SSO enforcement (enforce), and readiness status (get_status).
• Added the client.access_governance_log sub-client (AccessGovernanceLogClient), whose paginated list() accepts limit, cursor, event_type, actor_id, from_, and to filters.
• Added the client.evidence_bundles sub-client (EvidenceBundlesClient) with list, create, get, and download methods.
• Added the read-only client.api_key and client.base_url properties on AtlaSentClient.

​

2.11.0 — 2026-05-27

• Added the atlasent.auth module with refresh_token(), refresh_with_idp(), and list_idp_connections().
• Added the atlasent.scim_client module (ScimClient, with .users and .groups sub-clients) for SCIM 2.0 provisioning.
• Added the atlasent.evidence_bundle module with create_evidence_bundle(), get_evidence_bundle(), and download_evidence_bundle().
• Added verify_evidence_bundle() and EvidenceVerificationResult (in atlasent.replay) for verifying evidence bundles offline, without a call to the backend.
• Added the DecisionValue type alias, Permit.permit_expires_at, and EvaluateResult.reasons. When the server returns only a single reason, reasons holds it as a one-element list.

​

2.7.0 — 2026-05-24

• Added EvaluateResult.risk_envelope (EvaluateRiskEnvelope) with weighted_score, engine_decision, envelope_decision, promoted, hard_blocks, and factors. It is None on responses from older servers.
• Added EvaluateRequest.explain. Set it to True to have the server populate risk_envelope.factors with EvaluateRiskEnvelopeFactor entries.
• EvaluateRiskEnvelope and EvaluateRiskEnvelopeFactor are exported from atlasent.

​

2.6.0 — 2026-05-24

• Added AtlaSentClient.replay(*, evaluation_id), which re-evaluates a recorded decision against its originally pinned policy bundle and engine version. Replay writes no audit record and mints no permit.
• Added the ReplayResponse model and the ReplayVarianceKind literal (NONE, DECISION_CHANGED, ENVELOPE_DRIFT, POLICY_DRIFT, CHAIN_TAMPER, ENGINE_DRIFT, and BUNDLE_MISSING).
• Fixed an ImportError when importing atlasent.async_client.

​

2.5.0 — 2026-05-22

• Added read methods for advisory governance agents on AtlaSentClient and AsyncAtlaSentClient: list_governance_agents(), list_governance_findings(*, change_id, agent_slug=None), and list_governance_evaluations(*, change_id, agent_slug=None).
• Added the GovernanceAgent, GovernanceAgentFinding, GovernanceAgentEvaluation, and AgentEvidenceRef models, plus the highest_agent_finding_severity() helper.

​

2.4.0 — 2026-05-14

• Added the PRODUCTION_DEPLOY_ACTION constant ("production.deploy"), the canonical Deploy Gate action.
• Examples and docstrings now use "production.deploy". The legacy "deployment.production" action string still works.

​

2.2.0 — 2026-05-05

• Added identity assertion models: IdentityAssertionV1, IdentityAssertionBinding, IdentityIssuer, IdentitySubject, IdentityIssuerKey, and IdentityTrustedIssuersConfig.
• ApprovalArtifactV1 gained an optional identity_assertion field. The server requires it whenever human approval is required.

​

2.1.0 — 2026-05-05

• Added approval artifact models, including ApprovalArtifactV1, ApprovalReference, and PermitApprovalBinding.
• evaluate() gained the resource_id, amount, approval, and require_approval keyword arguments, and verify() gained require_approval.
• Added EvaluateResult.permit_approval, VerifyResult.consumed, and VerifyResult.approval. If consumed is True on an APPROVAL_LINKAGE_MISSING result, the permit has been used up, so don’t retry.

​

2.0.0 — 2026-04-30

• Breaking: The SDK now sends the canonical wire format (action_type, actor_id, context) and no longer includes api_key in request bodies. This requires an AtlaSent API deployment that accepts the canonical format.
• Breaking: EvaluateResult.decision is now a string ("allow", "deny", "hold", or "escalate") instead of a bool. The boolean is still available as result.permitted.
• Added canonical result attributes: permit_token, request_id, expires_at, and denial on EvaluateResult, and valid, outcome, and verify_error_code on VerifyResult. Legacy result attributes are still populated. Building EvaluateRequest or VerifyRequest with the legacy keyword names still works but emits a DeprecationWarning.

​

Go

Go SDK release notes will appear here.