North Star

This is where Exosuit is headed, not a description of every surface as it stands today. The mature experience starts from a workspace. The workspace is the collaboration boundary, and the extension follows the project: orientation, state, controls, and next actions all show up where the work is happening.

It’s worth reading as one picture of how the four pillars (inverted sovereignty, high coherence, phased rigor, and evolutionary intelligence) play out once the workflow is wired all the way through.

Genesis

A new project begins by establishing its mission. The user states what the project is for, and that intent becomes the anchor every later decision is measured against. Exosuit resolves the project from its git repository: the canonical project identity is derived from the git common directory, so linked worktrees share one project, one state store, and one daemon (RFC 10184).

From the first moment, project state is canonical in the Exosuit state store rather than scattered across ad-hoc files. There is one place to read where the project stands and one place progress is written (RFC 10180).

Modes of collaboration

The agent shifts between three modes depending on where the work is in its lifecycle (RFC 10155):

• Thinking Partner — exploration and design. Surfaces tensions, challenges assumptions, asks “why.” Used during planning and design reviews.
• Chief of Staff — organization and cadence. Restores context, tracks obligations, keeps the plan coherent. Used at phase transitions.
• Maker — execution. Follows the plan, works incrementally, keeps scope bounded. Used during implementation.

The same agent moves between modes; the mode is a stance, not a separate persona.

Planning

Work is structured into a hierarchy: epochs group phases, phases hold goals, and goals decompose into tasks. Planning fills in this structure collaboratively, with the user approving the shape before execution begins. Each goal is calibrated to a single Prepare → Execute → Review cycle, which keeps phases plannable and progress legible (RFC 00230).

Research

When a technical unknown blocks progress, the work opens as a research track rather than a guess. The agent investigates by reading code, fetching references, and recording findings, and the conclusion is captured as a decision. Durable decisions become RFCs (documents on disk); the structured backlog and progress stay canonical state (RFC 10180).

Deep work

During implementation the user picks the active task and the agent works it with full project context in view. Steering keeps the agent aligned: progress commands return entity-scoped feedback, so a concern the user raises reaches the agent at the moment it can act on it rather than waiting for an explicit status check (RFC 10182, RFC 10181).

Phase transition

A phase closes only on evidence. Verification runs the project’s checks, and coherence catches drift between what the documents claim and what the code does. When everything is green and the outcome is recorded, the phase finishes and its learning carries forward into the next one.

Evolution

The system improves as it is used. Repeated patterns are promoted into axioms, decisions accumulate as RFCs, and the workflow itself becomes a thing the project can refine.