Stack Layer Map — Atlas Graph to Package Mapping

The 14 Layers

The atlas graph defines 14 stack layers (stack-layers cluster). Each layer represents an architectural concern in the a5c agent stack. This document maps each layer to its implementing packages, primary graph node kinds, and current implementation maturity.

Layer Mapping

L1: Model

**Scope:** Trained artifact and declared capability surface

**Gap:** Model capabilities are cataloged but not executable. No package owns model selection, routing, or capability negotiation at this layer.

---

L2: Provider

**Scope:** Hosted or self-hosted serving boundary

**Gap:** Provider concepts are split between transport-adapter (proxy) and adapters-cli (provider translation). No unified provider abstraction package.

---

L3: Transport

**Scope:** Wire and client path between agent core and provider

**Gap:** Codec capabilities (tool schema translation, cost normalization) are planned but not fully implemented. The v6.0 plan had a codec architecture refactor that was started but not completed.

---

L4: Agent-Core

**Scope:** Inner loop and graph/turn semantics for one agent brain

**Note:** Two packages implement this layer for different concerns: genty-core (unified babysitter loop) and agent-comm-adapter (harness-neutral dispatch). This is intentional per v6.0 spec.

---

L5: Agent-Runtime

**Scope:** Host process and operational runtime for an agent core

---

L6: Agent-Platform

**Scope:** Extension, distribution, launch, and ecosystem surface

**Gap:** AgentPlatformImpl is defined in graph but has no standalone implementation package. Platform concerns are scattered across agent-platform, extensions-adapter, and agent-catalog.

---

L7: Workspace

**Scope:** Materialized working context the agent reads/writes/indexes

**Gap:** No dedicated workspace abstraction. The concept of "what the agent can see/modify" is implicit in the SDK's run directory layout rather than an explicit workspace contract.

---

L8: Execution

**Scope:** Invocation environment for agent-driven tools and commands

---

L9: Sandbox

**Scope:** Policy-enforcement perimeter around execution and side effects

**Gap:** Sandbox as a policy enforcement layer (filesystem restrictions, network controls, resource limits) is aspirational. Only human approval gates are implemented.

---

L10: Interaction

**Scope:** User-facing action primitives (commands, keybindings, controls)

---

L11: Presentation

**Scope:** Outermost rendering surface (TUI, CLI, web, IDE, API)

---

L12: Knowledge Fabric

**Scope:** Durable organizational knowledge, memory, retrieval, indexing

**Gap:** The graph defines rich knowledge concepts (MemorySystem, RetrievalPipeline, VectorStore) but no package implements durable agent memory or semantic retrieval. agent-catalog is metadata-only.

---

L13: Orchestration

**Scope:** Control-plane primitives for agent work coordination

**Note:** babysitter-sdk is intentionally the monolithic center of orchestration per v6.0 decision. Splitting into smaller packages was evaluated and rejected for pragmatic reasons.

---

L14: Governance

**Scope:** Policy, risk, compliance, audit, approval controls

**Gap:** AgentGovernanceImpl is defined but implementation examples are sparse. The governance module in agent-platform exists but covers a narrow surface (plugin governance, hook validation).

Maturity Summary