Overview

The Agent Workspace is the client-side runtime that gives every Allternit agent a persistent, inspectable, and reproducible home on disk. It is implemented by the allternit-agent-workspace crate and provides the 5-layer living workspace architecture that sits between the user, the agent, and the kernel’s deterministic enforcement systems. The workspace is deliberately not the kernel. It is the bridge:
  • Users edit markdown files to express identity, policy, and goals.
  • Agents read those markdown files to understand constraints and context.
  • The kernel enforces policy, maintains the ledger, and executes tool calls.
The workspace keeps the agent grounded across sessions, crashes, and compaction boundaries by rehydrating state from markdown and checkpoints on every boot.

Key principle

The Kernel enforces. The Markdown distills. The Agent consumes.

The 5 layers

The workspace is organized into five layers, each represented by markdown files under .allternit/.

Layer 1: Cognitive Persistence

Cognitive files keep the agent’s working memory human-readable and recoverable.
  • BRAIN.md — Human view of the active task graph.
  • MEMORY.md — Index into daily session logs.
  • memory/YYYY-MM-DD.md — Daily logs synced from the kernel ledger.
  • active-tasks.md — Current task queue.
  • lessons.md — Curated learnings.
  • self-review.md — Introspection logs.

Layer 2: Identity Stabilization

Identity files stabilize who the agent is and how it should behave.
  • IDENTITY.md — Name, nature, vibe.
  • SOUL.md — Behavior guidelines.
  • USER.md — User preferences.
  • VOICE.md — Voice and tone configuration.
  • POLICY.md — Runtime policy overrides that take effect immediately.

Layer 3: Governance & Decision

Governance files are the constitution and rulebook.
  • AGENTS.md — Supreme law. All agents must read this first.
  • PLAYBOOK.md — Operational procedures distilled from cookbooks.
  • TOOLS.md — Tool configuration for this workspace.
  • SYSTEM.md — Environment constraints.
  • CHANNELS.md — MCP and external channel configuration.

Layer 4: Modular Skills

Skills are reusable capabilities kept in skills/{name}/.
Each skill has a human-readable SKILL.md procedure and a machine-readable contract.json with triggers, permissions, and input/output schemas.

Layer 5: Business Topology

Business files enable multi-tenancy and project management.
  • CLIENTS.md — Client registry.
  • business/crm/ — Contact management.
  • business/Projects/ — Project workspaces.
  • business/content/ — Content templates.

Boot sequence

When an agent starts, the workspace runs a deterministic 21-phase boot sequence grouped into six phases.
The boot sequence produces a BootContext that the agent uses for the rest of the session.

Components

The allternit-agent-workspace crate exposes the following primary components.

Quick start

Add the crate

Initialize a workspace

Check policy

Build a context pack

Save and restore checkpoints

CLI usage

The crate also ships a workspace-cli binary.

Workspace layout

A fully populated .allternit/ directory looks like this:

Kernel sync

The workspace is designed to sync bidirectionally with the authoritative kernel.

Sync directions

When the kernel is unavailable, the workspace falls back to offline mode using cached context packs and local policy.

Deterministic rehydration

At every context boundary the agent is rehydrated from the workspace:
  1. New session → Full boot sequence.
  2. Post-compaction → Load checkpoint + recent memory.
  3. Subagent spawn → Inherit parent context pack.
  4. Tool call boundary → Policy check + receipt logging.
  5. Crash recovery → Restore from checkpoint.

WASM build for Shell UI

The crate can be compiled to WebAssembly for browser-based interfaces such as the Shell UI.

Configuration

Kernel sync and workspace behavior can be configured in .allternit/config.toml.

Concepts

  • Workspace — A directory on disk that contains the 5-layer markdown structure under .allternit/.
  • Boot context — The assembled state produced by the 21-phase boot sequence.
  • Context pack — A deterministic bundle of all workspace layers prepared for the agent’s working memory.
  • Policy engine — Client-side parser for AGENTS.md and POLICY.md that gates tool and file operations.
  • Skill — A reusable capability made of SKILL.md plus contract.json.
  • Checkpoint — A snapshot of agent and task state used for crash recovery and session continuity.
  • Rehydration — The process of restoring agent context from markdown and checkpoints at a boundary.
  • Kernel sync — Bidirectional flow of receipts, policy, skills, and checkpoints between the workspace and the authoritative kernel.