Jarvis Docs
Architecture

Workmode and Authority Boundaries

Canonical workmode contracts, lease semantics, and authority-chain enforcement for workflow dispatch

Workmode and Authority Boundaries

Phase 5.1 establishes the foundational runtime contracts for workflow dispatch and admission in Nous-OSS. This document describes workmode contracts, lease semantics, and authority-chain guardrails that enforce deterministic, fail-closed admission before higher-level control-plane capabilities are layered on.

Overview

  • Workmode — Execution context that defines entrypoint, SOP path, allowed artifact surfaces, and scope boundaries.
  • Lease — Time-bound, non-transferable authorization for an orchestration agent to execute within a workmode.
  • Authority chain — Strict hierarchy: nous_cortexorchestration_agentworker_agent. Authority narrows down-chain only.

All dispatch admissions validate target identity, action admissibility, and authority-chain constraints. Invalid attempts are blocked and evidence-logged.

Chat is intent, not authority (PCP-001): Chat turns are proposals interpreted by the runtime. Chat cannot directly mutate execution state; control intents route through operator-control. See Operator Control and Chat.

Canonical Authority Chain

AuthorityRoleMay dispatch to
nous_cortexExecutive authorityorchestration_agent, worker_agent
orchestration_agentScoped orchestrationworker_agent only
worker_agentScoped execution— (cannot dispatch)

Invariants:

  • WMODE-002: Authority must narrow down-chain. An orchestration agent cannot dispatch to cortex; a worker cannot dispatch to anyone.
  • WMODE-003: Nested orchestration is forbidden. An orchestration agent cannot dispatch to another orchestration agent.
  • WMODE-010: A worker agent cannot escalate authority or dispatch authoritative agents.

System Workmodes

Canonical system workmodes shipped with Nous:

WorkmodePurposeAllowed artifact surfaces
system:implementationImplementation workflow.worklog/, self/, docs/, packages/, scripts/
system:architectureArchitecture ideation.architecture/ only
system:skill_authoringSkill authoring.skills/ only

Skill-installed workmodes follow the pattern skill:<slug>:<entrypoint>.

Skill Creator Admission (Phase 7.5)

system:skill_authoring governs creation and revision of skill artifacts only. It does not grant admission or promotion authority.

Admission and promotion are control-plane decisions with additional hard gates:

  1. Skill contract validation must pass (runtime artifacts present and valid).
  2. SkillBench evidence must be complete and fixed-model-locked.
  3. Attribution thesis evidence must classify uplift source and recommendation.
  4. Final admission/promotion decision must be recorded by nous_cortex.

Blocked or invalid states return deterministic reason codes and evidence refs.

Examples:

  • SCM-003-WORKER-SELF-PROMOTION: worker lane attempted self-promotion.
  • SCM-005-MODEL-DRIFT: fixed-model drift detected in benchmark evidence.
  • SKADM-002-CORTEX-AUTH-REQUIRED: non-cortex actor attempted final decision.
  • EVID-001-MISSING-WITNESS: required witness linkage missing.

Lease Contract

A lease is issued by nous_cortex and grants an orchestration agent the right to execute within a workmode for a bounded time.

Lease rules:

  • Non-transferable
  • Revocable by cortex at any time
  • Cannot widen scope or capabilities after issuance without re-authorization
  • Context-profile switches require explicit transition event before downstream dispatch

Lifecycle: Lease exits when task objectives complete, TTL expires, budget guard trips, policy violation requires stop, or cortex revokes.

Lifecycle Admission

Lifecycle actions (start, pause, resume, stop, recover) are gated by:

  • Control state — Valid transitions only (e.g. pause requires running).

  • Confirmation proof — Required per ADR-004. Runtime-issued, scope-bound, action-bound. Cannot use stale or UI-only approval.

Unknown control state or missing confirmation proof → fail-closed (deny).

Recovery Governance (Phase 5.4)

Lifecycle action recover and resume authorization are governed by the failure-recovery architecture (AA-011):

  • Checkpoint protocol — Only committed checkpoints are resumable; two-phase prepare/commit ensures consistency.
  • Retry policy — Side-effecting operations require idempotency proof before retry; unbounded retry is impossible by contract.
  • Rollback policyunknown_external_effect blocks blind auto-resume; explicit review-gated decision evidence required.
  • Terminal states — Every recovery flow ends in recovery_completed, recovery_blocked_review_required, or recovery_failed_hard_stop.

FR invariants (witnessd): FR-001 (committed checkpoints only), FR-004/FR-005 (retry budget, idempotency), FR-006/FR-007 (rollback, domain boundary), FR-008 (no orphan recovery).

See Recovery Governance for operator guidance.

Blocked Actions and Reason Codes

When a dispatch or lifecycle action is blocked, the admission guard returns a reasonCode and evidenceRefs.

CodeMeaning
WMODE-002Authority widening blocked — e.g. worker attempting to dispatch, or orchestrator dispatching to cortex
WMODE-003Nested orchestration blocked — orchestrator attempting to dispatch to orchestrator
WMODE-010Worker escalation blocked — worker cannot dispatch to any authoritative agent
POL-CONTROL-STATE-BLOCKEDControl state undefined or invalid for admission
OPCTL-CONFIRMATION-REQUIREDLifecycle action requires confirmation proof
OPCTL-INVALID-STATEControl state does not allow the requested action

All blocked outcomes include evidenceRefs (min 1) for audit linkage per ADR-002.

On this page