l3-autonomy - SKILL.md Agent Skill

name: l3-autonomy description: "L3 agent autonomy primitives (Rust) — envelope tracking, context scoping, messaging, agent factory, plan DAG, enforcement pipeline. For governed multi-agent systems."

31-l3-autonomy: L3 Agent Autonomy Primitives

Feature flags: l3-core (sync types, WASM-compatible), l3 (full runtime, requires tokio) Module root: crates/kailash-kaizen/src/l3/

Architecture

Two layers, mirroring the core/runtime split:

l3::l3_core (l3-core feature) -- sync types, validation, state machines. No tokio. WASM-compatible.
l3::runtime (l3 feature, implies l3-core) -- async execution, channels, spawning. Requires tokio.

Quick Reference

Primitive	Core Type	Runtime Type	Purpose	Skill
Envelope	`EnvelopeTracker`, `EnvelopeSplitter`, `EnvelopeVerdict`	`L3EnforcementPipeline`	Budget monitoring + enforcement	envelope-tracking.md
Context	`ContextScope`, `ScopeProjection`, `DataClassification`	--	Hierarchical data isolation	scoped-context.md
Messaging	`MessageEnvelope`, `L3MessagePayload`, `Priority`	`MessageRouter`, `MessageChannel`, `DeadLetterStore`	Inter-agent communication	messaging.md
Factory	`AgentSpec`, `AgentInstance`, `AgentState`	`AgentFactory`, `AgentInstanceRegistry`, `AgentBuilder`	Agent lifecycle management	agent-factory.md
Plan	`Plan`, `PlanNode`, `PlanEdge`, `PlanModification`	`PlanExecutor`	DAG-based task orchestration	plan-dag.md
Pipeline	--	`L3EnforcementPipeline`, `L3Verdict`, `L3GovernedAgent`	Unified enforcement entry point	envelope-tracking.md

Cross-Cutting

type-convergence.md -- GradientZone, DataClassification, ConstraintEnvelope reuse decisions
invariant-index.md -- All 52 behavioral invariants cross-referenced by primitive

Key Source Directories

crates/kailash-kaizen/src/l3/
  mod.rs                          # Feature-gated l3_core + runtime
  core/
    mod.rs                        # context, envelope, factory, messaging, plan, state_machine, types
    types.rs                      # L3Verdict, PlanGradient (shared types)
    state_machine.rs              # AgentState, WaitReason, TerminationReason
    context/                      # ContextScope, ScopeProjection, DataClassification, ContextValue
    envelope/                     # EnvelopeTracker, EnvelopeSplitter, EnvelopeVerdict
    factory/                      # AgentSpec, AgentInstance, D/T/R address computation
    messaging/                    # MessageEnvelope, L3MessagePayload, Priority, error types
    plan/                         # Plan, PlanNode, PlanEdge, PlanValidator, gradient, modification
  runtime/
    mod.rs                        # agent, factory, messaging, pipeline, plan
    agent.rs                      # L3GovernedAgent
    pipeline.rs                   # L3EnforcementPipeline, L3Verdict, L3VerdictSource
    factory/                      # AgentFactory, AgentInstanceRegistry, AgentBuilder
    messaging/                    # MessageRouter, MessageChannel, DeadLetterStore
    plan/                         # PlanExecutor

Dependencies

L3 depends on types from three other crates:

Crate	Types Used	Notes
`kailash-pact`	`GradientZone`, `ClassificationLevel`	Reused for gradient zones and classification levels
`trust-plane`	`ConstraintEnvelope`, `StrictEnforcer`	Reused for 5-dim envelopes and enforcement
`eatp`	`DataClassification`, `HoldQueue`, `AgentId`, `Capability`	Reused for data classification and human holds