Role and Capability Matrix

Canonical markdown page for this topic. Rendered reference version: role-and-capability-matrix.

Last Updated: 2026-03-22 Status: Active reference

This document defines the current role/capability boundaries for Jorvis agents and OpenClaw.

It exists to make one thing explicit:

For the concrete development/operations usage model, use:

Human Governance Roles

Role	Primary use	Can do	Must not do
Architect	planning, sequencing, ADRs	define bounded next step and acceptance criteria	hide major scope in a “small” fix
Executor	implementation, tests, deploy work	implement approved focused scope	self-gate or self-merge
Gatekeeper	review, GO/NO-GO, merge constraints	evaluate evidence, CI, risk, scope discipline	rewrite implementation during review
Tester	E2E and regression validation	produce evidence-first repro and verdict input	blend testing with undocumented fixes
Universal	safe bootstrap across roles	read current truth, then assume one assigned role	silently combine multiple authorities at once
Coordinator	orchestration-only mode	route work and keep shared context aligned	quietly become implementer or gatekeeper

Current OpenClaw behavior is bounded to advisory and plan-only modes.

OpenClaw mode	Current state	Notes
User Assistant	Active	help/reference, explainer, personal/work assistant
Admin Observer	Active	monitoring/improvement advisory behavior with non-mutating admin snapshot enrichment
Admin Planner	Partial / advisory only	planning intent exists, but not yet backed by a guarded execution gateway
Admin Operator	Not delegated	approval gateway and rollback model remain separate human-governed concerns

Capability	User Assistant	Admin Observer	Admin Planner	Admin Operator
help/reference	✅	✅	✅	✅
Jorvis explainer	✅	✅	✅	✅
personal/work memory assistant	✅	✅	✅	✅
read runtime/admin snapshot	❌	✅	✅	✅
improvement-oriented advisory	❌	✅	✅	✅
plan-only change guidance	❌	advisory only	advisory only	not delegated
dry-run internal tooling	❌	❌	bounded advisory	not delegated
approved mutation tooling	❌	❌	❌	not delegated
merge/release/governance authority	❌	❌	❌	❌

Current canonical typed capability IDs (Jorvis typed surface, PR #407):

Current canonical execution modes: chat, advisory, plan_only Current canonical memory scopes: reference, user, admin

Important current truth:

canonical capability routing is now Jorvis-side via POST /api/v1/internal/openclaw-capabilities/resolve
the bridge consumes this typed surface (PR #410) with a local keyword-heuristic fallback
the old bridge-only labels (admin_monitoring_assistant, admin_improvement_advisor, admin_change_assistant) are superseded by the canonical typed IDs above

OpenClaw must not be treated as:

OpenClaw may:

It may not silently cross from advisor to authority.

The bounded mode ladder remains:

Each later stage requires stronger controls than the previous one and does not transfer governance authority to OpenClaw.

Boundary requirements for any separately authorized expansion: