Stores: Plan in Its Own Repo
Beta. Stores, references, working context, and worksets are new. Command names, flags, file formats, and JSON output may still change shape between releases. Every walkthrough below was run against the current build, but re-read this guide after upgrading.
The problem this solves
OpenSpec normally lives inside one code repo: an openspec/ folder next to
your code, holding specs and changes for that repo.
That stops fitting the moment your planning is bigger than one repo:
- Your work spans several repos — one feature touches the API server, the
web app, and a shared library. Whose
openspec/folder does the plan live in? - Your team plans before code exists, or plans things that never become code in this repo.
- Requirements are owned by one team and consumed by others. The wiki version drifts, and your coding agent can't read it anyway.
A store is the answer: a standalone repo whose whole job is planning.
It has the same openspec/ shape you already know — specs and changes —
plus a small identity file. You register it on your machine once, by name,
and then every normal OpenSpec command can work in it from anywhere.
The shape
team-plans (a store: planning in its own repo)
├── .openspec-store/store.yaml identity: "I am team-plans"
└── openspec/
├── specs/ what is true
└── changes/ what is in motion
▲
│ registered on each machine by name;
│ shared by pushing/cloning like any repo
┌─────────────┼─────────────┐
│ │ │
web-app api-server mobile-app
(code repo) (code repo) (code repo)Two rules keep this simple:
- A store is just a git repo. You commit, push, pull, and review it yourself. OpenSpec never clones, syncs, or pushes anything on its own.
- Declarations, not machinery. Repos can declare how they relate to stores (shown below). Declarations change what OpenSpec can tell you — never where your commands act.
Five minutes to your first store
Two commands take you from nothing to a working, store-scoped change:
openspec store setup team-plans --path ~/openspec/team-plansStore ready: team-plans
Location: /Users/you/openspec/team-plans
OpenSpec root: ready
Registry: registered
Next: run normal OpenSpec commands against this store, for example:
openspec new change <change-id> --store team-plans
Share this store by committing and pushing it like any Git repo.openspec new change add-login --store team-plansUsing OpenSpec root: team-plans (/Users/you/openspec/team-plans)
Created change 'add-login' at /Users/you/openspec/team-plans/openspec/changes/add-login/
Schema: spec-driven
Next: openspec status --change add-login --store team-plansThat's the whole model. From here the lifecycle is exactly what you know —
status, instructions, validate, archive — with --store team-plans
on each command, and every printed hint carries the flag for you. The
Using OpenSpec root: line always tells you where a command is acting.
Story: one team, one planning repo
A team keeps its specs and changes in team-plans instead of scattering
them across code repos.
Day one (whoever sets it up):
openspec store setup team-plans --path ~/openspec/team-plans \
--remote git@github.com:acme/team-plans.git
git -C ~/openspec/team-plans push -u origin mainPassing --remote records the clone URL inside the store's own identity
file (.openspec-store/store.yaml), in the initial commit. Every future
clone is born knowing where it came from, so health checks and error
messages can print a complete, pasteable fix for teammates who don't have
it yet.
Every teammate (once per machine):
git clone git@github.com:acme/team-plans.git ~/openspec/team-plans
openspec store register ~/openspec/team-plansFrom then on, everyone works in the same planning repo by name:
openspec status --store team-plans --change add-login
openspec show add-login --store team-plansSharing work is git, on purpose. A change you create exists only in your checkout until you commit and push it — same as code. Plans get branches, pull requests, and review for free, because a store is an ordinary repo.
Connecting the team's code repos. A code repo whose planning is fully
externalized needs exactly one line, in openspec/config.yaml:
# web-app/openspec/config.yaml
store: team-plansNow every OpenSpec command run inside web-app acts on team-plans with
no flags at all:
cd ~/src/web-app
openspec status --change add-loginUsing OpenSpec root: team-plans (/Users/you/openspec/team-plans)
...The pointer is a fallback, never an override: an explicit --store always
wins, and if the repo grows real planning folders of its own, those win
(with a warning to remove the stale pointer).
One default for every repo on your machine. If you work across many
code repos that all plan into the same store, set it once, globally,
instead of adding the store: line to each repo:
openspec config set defaultStore team-plansNow any command run outside a planning root — and with no --store and no
project pointer — resolves to team-plans. It sits at the bottom of the
precedence list, so --store, a local root, and a project store: pointer
all still win. The root banner and JSON root block report
source: "global_default" with the store id, so you can always tell a
machine-wide default from a repo's own pointer. Clear it with
openspec config unset defaultStore. If the id is not registered, commands
error and tell you to register it or clear the stale default.
Example: one feature, two component repos
Suppose add-checkout-promo changes both checkout-api and
checkout-web. The team wants one shared product contract, while each code
repo still needs its own implementation tasks, branch, and review.
Use two layers:
- Keep the shared behavior in
team-plans. - Keep implementation plans in each component repo and reference the store as read-only upstream context.
First, plan the shared contract in the store:
openspec new change add-checkout-promo --store team-plans
openspec status --change add-checkout-promo --store team-plansThe proposal and specs should describe the behavior at the boundary between the components — for example, the promotion fields returned by the service and how the frontend handles an ineligible checkout. Review this change in the store repo like any other branch and pull request.
What context does planning see?
Selecting a store changes the OpenSpec root; it does not discover or read every code repo that uses that store. Store instructions see the artifacts and configured context in the store. They see component code only when those folders are also available to the agent or editor and the agent reads them.
A workset is a convenient way to open the planning store and both code repos together:
openspec workset create checkout-promo \
--member ~/openspec/team-plans \
--member ~/src/checkout-api \
--member ~/src/checkout-web \
--tool code
openspec workset open checkout-promoThis makes the folders visible in one IDE workspace. It does not copy source context into the store, select affected repos, or grant an agent permission to edit them. Put durable cross-component facts in the shared specs; do not rely on a planner remembering source it happened to inspect.
How does implementation start in each repo?
When no explicit --store or nearer openspec/ root applies, a
store: team-plans pointer routes commands to that store. It does not split
one store task list by the directory from which apply was invoked. OpenSpec
currently does not route tasks to repos.
When each component needs an independently scoped apply/review cycle, give it a local OpenSpec root and reference the central store instead of pointing at it:
# checkout-api/openspec/config.yaml (and likewise in checkout-web)
schema: spec-driven
references:
- team-plansAfter the shared contract is approved and available in the store's main specs, create a small local change for the component's part:
cd ~/src/checkout-api
openspec new change implement-checkout-promo-api
cd ~/src/checkout-web
openspec new change implement-checkout-promo-uiThe reference index in each repo's instructions supplies the store spec's
summary and exact openspec show ... --store team-plans fetch command. Each
local proposal cites that shared contract, and its tasks describe only work
in that component. Then run /opsx:apply in each repo separately; root
resolution keeps the artifacts and implementation edits scoped to that repo.
The service and frontend changes can now be tested, reviewed, merged, and
archived independently.
If implementation must begin while the shared store change is still active,
fetch it explicitly with
openspec show add-checkout-promo --store team-plans; reference indexes list
canonical store specs, not active store changes. Keep the store branch and
component branches linked in their pull-request descriptions so reviewers
can see which version of the contract each implementation follows.
Story: requirements that cross team lines
A platform team owns the requirements. Product teams build against them, in their own repos, with their own designs. A reference describes that relationship without moving anyone's work.
platform-reqs (store) api-server (code repo)
owned by the platform team owned by a product team
┌──────────────────────────┐ ┌──────────────────────────┐
│ openspec/specs/ │ ◀────────│ openspec/config.yaml │
│ payments/spec.md │ reads │ references: │
│ auth/spec.md │ │ - platform-reqs │
│ │ │ openspec/specs/ │
│ openspec/changes/ │ │ (their own designs) │
│ platform work │ │ openspec/changes/ │
│ │ │ (their own work) │
│ │ └──────────────────────────┘
└──────────────────────────┘The product team declares what it draws on in its repo's
openspec/config.yaml:
references:
- platform-reqsReferences are read-only context. The repo keeps its own openspec/ root;
work stays there. What changes: openspec instructions in that repo now
includes an index of the referenced store's specs — each with a one-line
summary and the exact fetch command (openspec show <spec-id> --type spec --store platform-reqs). An agent working in api-server can find the
upstream payment requirements, cite them, and write its low-level design in
the repo's own root — without anyone pasting context around.
A reference can carry its clone source, so teammates who don't have the store yet get a complete fix instead of a dead end:
references:
- { id: platform-reqs, remote: "git@github.com:acme/platform-reqs.git" }When you want the plan and code open together, make a workset. This is personal and explicit: each person chooses the folders they actually work with on their machine. Nothing about those local checkout paths is committed to the shared planning repo.
openspec workset create platform \
--member ~/openspec/platform-reqs \
--member ~/src/api-server \
--member ~/src/web-appTwo questions you can always ask
"Is my setup healthy?" — openspec doctor checks the current root and
its referenced stores, read-only, with a pasteable fix per finding:
Doctor
Root
Location: /Users/you/src/api-server
OpenSpec root: ok
References
- platform-reqs: ok (/Users/you/openspec/platform-reqs)
- design-system: Referenced store 'design-system' is not registered on this machine.
Fix: git clone -- git@github.com:acme/design-system.git '/Users/you/openspec/design-system' && openspec store register '/Users/you/openspec/design-system' --id design-system
"What am I working with?" — openspec context assembles the working
set from OpenSpec declarations: the root and the stores it references.
Working context for api-server (/Users/you/src/api-server)
OpenSpec root
api-server /Users/you/src/api-server
Referenced stores
platform-reqs /Users/you/openspec/platform-reqs
Fetch: openspec show <spec-id> --type spec --store platform-reqsBoth support --json for agents. openspec context --code-workspace <path> additionally writes a VS Code workspace file containing the whole
set — the only write this command performs.
Worksets: reopen the folders you work on together
Separate from all of the above: most people open the same few folders together every session — the planning repo plus two or three code repos. A workset is a personal, named view of exactly that, reopened with one command in your tool of choice.
workset "platform" openspec workset open platform
├── team-plans ~/openspec/team-plans │
├── api-server ~/src/api-server ▼
└── web-app ~/src/web-app all three open in your toolopenspec workset create platform \
--member ~/openspec/team-plans --member ~/src/api-server \
--tool code
openspec workset listplatform (opens in VS Code)
team-plans /Users/you/openspec/team-plans
api-server /Users/you/src/api-serveropenspec workset open platform then launches the saved tool: editors
(VS Code, Cursor) open one window with every member and return. The first
member is the primary. Override the tool any time with --tool <id>.
Worksets are deliberately not shared state. They live on your machine,
are never committed, and make no claims about the work — they only record
what you like open together. Removing one never touches the member
folders. New tools are configuration, not code: anything launched via a
workspace file or per-folder attach flags can be added under the openers
key in the global config (openspec config edit).
How commands decide where to act
Every normal command resolves its root the same way, in this order:
1. --store <id> you said so explicitly → that store
2. nearest openspec/ a real planning root here → this repo
(walking up from cwd)
3. store: pointer config.yaml declares a store → that store
4. defaultStore global config sets a machine → that store
default
5. none of the above stores registered on this → error with a
machine? selection hint
no stores registered? → the current
directory
(classic behavior)The Using OpenSpec root: line (and the root block in --json output)
tells you which case you're in.
Known limitations
- Beta shape. Everything on this page may change between releases — names, flags, file formats, JSON keys.
- One checkout per store id per machine. Registering a second checkout
under the same id fails with a hint to
store unregisterfirst. - No sync, ever — by design. OpenSpec never clones, pulls, or pushes. A stale checkout shows stale specs until you pull; references are indexed live from whatever is on disk.
- Empty planning folders can be absent. A new store may not have
openspec/changes/,openspec/specs/, oropenspec/changes/archive/in Git yet. That is accepted during the beta; those folders appear once normal commands create files for them. - Pointer repos stay pointers. A config-only repo whose
openspec/config.yamldeclaresstore: <id>is treated as externalized planning, not as a store checkout to register. Remove thestore:line first if you intentionally want to convert that repo into a local store root. - Some commands stay where they are.
view,templates,schemas, and the deprecated noun forms (openspec change show, ...) act on the current directory only — no--store. - Per-machine state is per-machine. The store registry and worksets are local settings. Nothing about your machine's layout is ever committed to shared planning.
- Two launch styles for worksets. A tool that can't be launched with a workspace file or per-folder attach flags can't be added as an opener.
- Agent JSON has a known casing split (store-family keys are snake_case, workflow-family camelCase). Documented in the agent contract; unifying it is deferred to a versioned release.
Where things live
| What | Where | Shared? |
|---|---|---|
| A store's planning | <store>/openspec/ (specs, changes) | Yes — commit and push it |
| A store's identity | <store>/.openspec-store/store.yaml | Yes — committed with the store |
| The store registry | <data dir>/openspec/stores/registry.yaml | No — this machine only |
| Worksets | <data dir>/openspec/worksets/ | No — this machine only |
<data dir> is ~/.local/share/openspec on macOS and Linux (or
$XDG_DATA_HOME/openspec when set), and %LOCALAPPDATA%\openspec on
Windows.
Reference
Exact flags and JSON shapes for every command on this page: CLI reference (Stores, Doctor, Working context, Personal worksets) and the agent contract.
OpenSpec on a Team
Everything in the other guides works the same whether you're solo or on a team of twenty. What changes on a team is the questions around the edges: where do the specs live, how do teammates review a…
Commands
This is the reference for OpenSpec's slash commands. These commands are invoked in your AI coding assistant's chat interface (e.g., Claude Code, Cursor, Devin Desktop).