Admin Console
Trust + Decision plane operator surface for packs, policies, decision specs, intent catalogs, approvals, and the review queue.
The operator surface for every governed artifact — packs, policies, DecisionSpecs, approval gates, memory review.
- Permissioned registry views: pack, policy bundles, Decision Catalog, Intent-Task Catalog, adapter registry
- Live queues: approval gates, memory review, improvement proposals
- Trace bundles + scorecards from Observability
- RBAC and identity assertions for operators
- Validated pack / policy / DecisionSpec promotions, signed when the supply-chain profile requires it
- Approval-gate verdicts (approve / deny with rationale)
- Memory review-queue actions (approve / edit / reject)
- Improvement-proposal verdicts (ship / discard with rationale)
- Audit log entries for every operator action
- ApprovalGate
- DecisionApproval
- DecisionAudit
- ReplayPacket
The Admin Console is the operator surface that spans the Trust and Decision planes. It is where packs are reviewed, policies are linted, Decision Specs are versioned, approval gates are answered, and the memory review queue is worked. Cryptographic signing is an optional supply-chain profile, not a base schema requirement.
This is a component design, not a shipped administration service. Enforce exact control-profile requirements in backend authorization and protected stores. UI visibility, an operator role name or a signed release cannot let an operator demote mandatory controls, approve their own prohibited change, or assert current runtime conformance without evidence.
Definition
A web surface (and equivalent CLI) over the registries and queues that drive the runtime: pack registry, policy bundles, Decision Catalog, Intent-Task Catalog, adapter registry, approval queue, memory review queue, replay harness, scorecard dashboards.
Why it exists
Without a single operator surface, governance work scatters across spreadsheets, ticket systems, and bespoke UIs. The Console is the change-control choke point that makes promotion deliberate.
Inputs
- Least-privilege views of the applicable registries, tenant/owner namespaces and queues; an operator does not automatically read every tenant’s artifacts
- Live queues: approval gates, memory review, improvement proposals
- Trace bundles + scorecards from the Observability component
- RBAC and identity assertions for operators
Outputs
- Validated pack / policy / DecisionSpec promotions, with signatures when required by the deployment profile
- Approval-gate verdicts (
approve/denywith rationale) - Memory review-queue actions (
approve/edit/reject) - Improvement-proposal verdicts (ship / discard with rationale)
- Audit log entries for every operator action
Surfaces
| Surface | Purpose |
|---|---|
| Pack registry | author, lint, publish, deprecate, and optionally sign a Context Pack |
| Policy bundle editor | author JsonLogic rules; lint for unreachable / unsatisfiable rules; optionally sign releases |
| Decision Catalog | author DecisionSpec; bind to policy rules; version |
| Intent-Task Catalog | author Intents and TaskTemplates; bind to DecisionSpec |
| Adapter registry | onboard adapters; declare native ActionRisk, v1 approval_mode projection, schemas, identity, and idempotency per capability |
| Approval queue | review the complete frozen action, permitted evidence view, amount/destination, expiry and policy/profile/release pins; show changed or stale bindings explicitly |
| Memory review queue | promote / edit / reject candidate facts |
| Improvement proposals | review Insight Synthesizer / Strategy Compiler / Autotune outputs |
| Replay harness | re-run past Decision Records against pinned snapshot |
| scorecard dashboards | per intent, per pack version, per tenant |
| Authority and running work | inspect scoped mandates and triggers, revoke grants, pause future work, request stop and inspect confirmed/partial/unknown effect state |
| Recovery and accounting | owned reconciliation queues, checkpoint migration, stale-worker/fencing failures and aggregate budget reservations/usage |
| Control/evidence registry | exact profile/control revisions, release/environment pins, evidence age and useful/denied case results; show insufficient evidence without a passing badge |
How it works
- Authenticate the operator under the deployment’s required method, tenant, audience and current entitlement; high-impact policy may require WebAuthn/hardware-backed credentials.
- RBAC scopes the editable surfaces (e.g.,
policy_writer,approver,correction_writer). - Edits go through review with required approver roles per change class.
- Promotions are recorded in an append-only registry log and signed when the active supply-chain profile requires it.
- Operator actions emit audit records bound to the operator’s identity.
For an effect approval, validate the stored proposal/request and complete ActionBinding, required distinct quorum, current role entitlements, expiry and separation policy. Atomically redeem it for the same stable operation; a stale screen cannot authorize changed arguments or a second effect. Edits, promotion and policy administration also need protected revision checks and independent required review. Re-pin of a release does not undo external effects or permit resuming a checkpoint under changed interpretation.
Failure modes
- Operator action without assertion — refuse.
- Promotion bypassing required reviewer role — refuse.
- Approval-gate answer without frozen evidence snapshot match — refuse and emit security event.
- Stale registry view in the UI causing accidental rollback — mitigated by optimistic concurrency on every edit.
- Changed proposal, expired approval, revoked entitlement or prohibited self-approval — backend rejects even if an old UI still enables the button.
- Cancellation still pending or provider effect uncertain — display it as unresolved and route to its recovery owner; a stop click is not proof of rollback.
Operational concerns
- Separation of duties enforced by RBAC, not by UX.
- Operator throughput on each queue monitored and alerted.
- Strong operator authentication for approval obligations selected by policy; deployments may require WebAuthn or hardware keys for high-impact effects.
- Audit access for tenants on shared environments.
- Minimized permissioned evidence views, secure action rendering and actual telemetry/export processor controls; keep credential and compensation authority out of operator-visible payloads unless explicitly authorized.
- Owned escalation/reconciliation queues, response objectives and tested dependency-outage procedures appropriate to the workflow; no universal availability target is implied.
Evaluation metrics
- Time from change authored → promoted.
- Approval gate p50 / p99 response time.
- Memory review queue throughput and reject rate.
- Operator action audit completeness (target: 100%).