Skip to main content

Context System Overview

The aibox context system controls what project-level instructions and structured working memory are available to AI harnesses. It has two modes: processkit, which installs the full structured context layer, and harness-only, which keeps only the generated container and harness setup.

As of v0.16.0, the system is split across two cleanly separated projects:

  • aibox owns the container — devcontainers, addons, the CLI, the install/apply/migrate machinery, and the project skeleton (aibox.lock, .gitignore, provider pointer files, and, in harness-only mode, a minimal AGENTS.md).
  • processkit owns the content — every skill, every primitive schema, every state machine, the canonical AGENTS.md template, the processes, and the package YAMLs that compose them.
  • The user-side context/ directory is shared territory in processkit mode. aibox creates it, processkit fills it, and the user edits in place. An immutable upstream snapshot is kept under context/templates/processkit/<version>/ for the three-way diff that aibox apply uses to detect upstream changes versus local edits.

The Problem​

AI coding agents like Claude operate best when they understand not just the code, but the project's goals, decisions, and current state. Without structure, this information ends up scattered across chat histories, stale comments, and the developer's memory.

A single root-level instructions file is not enough for non-trivial projects. It works well for instructions and preferences, but it does not provide a standard place for decisions, backlog, progress tracking, or team conventions.

How Context Files Work​

With the default context.mode = "processkit" and a real [processkit].version pinned, your project looks something like this after aibox init and aibox apply:

my-project/
├── AGENTS.md                       # Canonical agent entry — rendered from processkit scaffolding
├── CLAUDE.md                       # Thin pointer to AGENTS.md (provider entry file)
├── aibox.toml
├── .devcontainer/
└── context/
    ├── skills/                     # Editable skill copies
    ├── processes/                  # release, code-review, feature-development, bug-fix
    ├── schemas/                    # primitive schemas
    ├── state-machines/             # state machine definitions
    └── templates/
        └── processkit/
            └── v0.27.4/            # Immutable upstream snapshot, base of three-way diffs

With context.mode = "harness-only", aibox writes only the container and harness surfaces:

my-project/
├── AGENTS.md                       # Minimal project instructions generated by aibox
├── CLAUDE.md                       # Thin pointer to AGENTS.md when Claude is enabled
├── aibox.toml
├── .aibox-home/
└── .devcontainer/

Harness-only mode intentionally omits processkit content, processkit MCP gateway config, processkit hooks/preauth, processkit command adapters, and processkit Migration entities. Generated text surfaces contain no processkit references.

AGENTS.md, CLAUDE.md, and provider files​

AGENTS.md at the project root is the canonical agent entry document. It is either rendered from the processkit template (context.mode = "processkit") or generated as a minimal aibox-owned file (context.mode = "harness-only"). It is write-if-missing during aibox init; aibox does not overwrite local edits. The agents.md ecosystem convention is to read this file from any AI harness.

When Claude is enabled in [ai].harnesses, aibox also writes a thin CLAUDE.md at the project root that points at AGENTS.md. In processkit mode, command adapters are projected into enabled harness surfaces, including .claude/skills/<name>/SKILL.md for Claude Code. Canonical skill content still lives under context/skills/; provider-specific files are generated shims. In harness-only mode, only the pointer is written.

OWNER.md — Developer Identity​

OWNER.md captures the developer's identity and preferences. It is created during aibox init (or by the owner-profile skill the first time the agent asks), with fields that help AI agents understand who they are working with:

  • Name — how the developer prefers to be addressed
  • Domain expertise — areas of knowledge and experience
  • Primary languages — programming languages used most often
  • Communication language — natural language for responses (e.g., English, German)
  • Timezone — for scheduling and availability context
  • Working hours — typical availability window
  • Current focus — what the developer is currently working on or learning
  • Communication preferences — style and conventions for AI interactions

Skills and Processes​

Skills and processes are owned by processkit and are only installed when context.mode = "processkit". For full documentation on what's available, how skills are organised, and which packages to use, see:

Version Tracking​

Resolved versions are split between desired and applied state:

[context]
mode = "processkit"
packages = ["product"]

[processkit]
version = "v0.27.4"

[processkit.context]
schema_version = "1.0.0"

aibox.toml declares the desired image and processkit versions. aibox.lock records the CLI version and exact resolved processkit release, checksum, addon selection, and managed runtime-home state last applied to the project.

When the schema evolves, aibox doctor flags version mismatches and aibox apply runs the relevant migrations. See Migration for details.

Relationship to aibox.toml​

The [context] section in aibox.toml selects the context mode. The [processkit] section is used only in processkit mode and pins which version of the content repository this project consumes:

[context]
mode = "processkit"
packages = ["product"]

[processkit]
source  = "https://github.com/projectious-work/processkit.git"
version = "v0.27.4"

Run aibox apply after editing [processkit].version to pull a new release. For harness-only projects, set only:

[context]
mode = "harness-only"

Design Principles​

Convention over configuration. File names and locations are standardised so AI agents can find them without special instructions.

Human-readable first. Context files and root instructions are Markdown. They are useful without any tooling.

Editable in place. Everything under context/skills/, context/processes/, context/schemas/, and context/state-machines/ is yours to edit. The immutable snapshot under context/templates/processkit/<version>/ exists only as the base of aibox apply's three-way diff.

No lock-in. Context files are plain Markdown and YAML in a context/ directory, and harness-only instructions are plain root-level Markdown. Stop using aibox and the files remain useful.

Clean boundary between container and content. aibox owns the box; processkit owns what goes in it. Each ships on its own cadence.