Reading the Workspace Tree

The workspace tree is meant to be legible. Folder names come from human names, sanitized for filesystems, with a suffix only when needed to avoid collision. UUIDs remain in the database as stable identity; paths are for humans and agents.

Source folders

Workspace
├── Agent
│   ├── AGENTS.md
│   ├── CONTEXT.md
│   ├── GUARDRAILS.md
│   ├── TOOLS.md
│   ├── memory/
│   └── skills/
├── Spaces
│   └── board-pack/
│       ├── CONTEXT.md
│       ├── artifacts/
│       ├── docs/
│       ├── goals/
│       │   └── customer-onboarding/
│       └── plans/
└── User
    ├── USER.md
    └── memory/

Settings -> Workspace shows the editable source workspace. The S3 backing keys are tenant-scoped (tenants/acme/agents/..., tenants/acme/spaces/..., tenants/acme/users/...), but operators should think in terms of these three roots.

Runtime workspace

When a turn starts, ThinkWork hydrates a different tree into /workspace:

/workspace
├── AGENTS.md
├── CONTEXT.md
├── skills/
├── workspaces/
├── Spaces/
│   └── board-pack/
│       ├── SPACE.md
│       ├── CONTEXT.md
│       ├── artifacts/
│       ├── docs/
│       ├── goals/
│       ├── plans/
│       └── workflows/
├── User/
│   ├── USER.md
│   └── memory/
└── Thread/
    ├── THREAD.md
    ├── GOAL.md
    ├── PROGRESS.md
    ├── TASKS.md
    └── notes/

The Agent source becomes the root of /workspace. User context stays explicit under User/. The active Space is hydrated under Spaces/<active-space>/. Other authorized Spaces are not hydrated by default; they are listed in the generated Workspace Routing section of AGENTS.md, which the renderer composes from authorization state (active Space, other authorized Spaces, acting user, participants, and Agent Profiles) and which replaces the former Spaces/INDEX.md. The agent reads non-active Spaces read-only through the fetch_workspace_source tool.

Thread runtime folders in S3 hold .hydrate_manifest.json, .rendered_at, generated Thread projection files, and writable Thread/notes/ files. The manifest points back to the Agent, Space, User, and Thread sources used for hydration and write-back.

How to read common files

Path	Meaning	Edit how
AGENTS.md	Agent identity, routing, and high-priority instructions	Edit Agent source
CONTEXT.md	Durable Agent-level context	Edit Agent source
Workspace Routing section of AGENTS.md	Generated routing: authorized Spaces, acting user, participants, Agent Profiles	Do not edit; recomposed on every render
Spaces/<space>/SPACE.md	Active Space manifest, workflow, and policy overview	Edit Space source
Spaces/<space>/CONTEXT.md	Active Space context	Edit Space source
User/USER.md	User/requester context	Edit user/profile source, not the runtime copy
User/memory/*.md	User-scoped working memory notes	Edit User source
skills/<slug>/SKILL.md	Installed skill instructions	Install or update skills
Spaces/<space>/goals/<template>/GOAL.md	Reusable Space Goal template	Edit Space source
Thread/THREAD.md	Generated Thread summary	Use Thread actions; runtime projection is read-only
Thread/GOAL.md	Generated current Goal projection	Use Goal actions; runtime projection is read-only
Thread/PROGRESS.md	Generated progress projection from database task rows	Use set_task_status or UI actions
Thread/TASKS.md	Generated checklist/task projection	Use set_task_status or UI actions
Thread/notes/*.md	Thread-scoped working notes and findings	Agent or operator can edit as narrative work
.hydrate_manifest.json	Runtime source map, provenance, ETags, and read-only mounts	Do not edit; regenerate by rendering the Thread runtime
.rendered_at	Runtime freshness marker	Do not edit

Owner lanes

Reconcile routes changed files back to their owning source:

Runtime path pattern	Owner lane	Backing source
Agent root files and skills/**	Agent	tenants/{tenant}/agents/{agent-folder}/
Spaces/<space>/** docs, artifacts, plans, and workflows	Space	tenants/{tenant}/spaces/{space-folder}/
User/** user context and memory	User	tenants/{tenant}/users/{user-folder}/
Thread/notes/**	Thread notes	tenants/{tenant}/threads/{thread-folder}/
Thread/THREAD.md, Thread/GOAL.md, Thread/PROGRESS.md, Thread/TASKS.md	Status	Generated read-only projection

If a changed file does not map to an allowed lane, reconcile rejects it and records the rejection on the turn. The generated Workspace Routing section of AGENTS.md is likewise render output: the composer strips and rebuilds it on every render, so hand edits to that section do not persist.

Operator checklist

• Use Agent folders for behavior that should apply everywhere.
• Use Space folders for project, customer, workflow, and shared team context.
• Use User folders for requester-specific context.
• Use Thread notes for raw findings that belong to the current thread.
• Use platform actions and tools for task status, review gates, access, and policy.
• Treat runtime metadata as rebuildable output.
• In runtime smoke tests, expect Spaces/, User/, and Thread/; reject top-level Agent/, singular Space/, root USER.md, workspace/, source/, and workspace-archives/.