OpenSpec Workbench for VS Code

OpenSpec Workbench brings the complete OpenSpec change workflow into VS Code. It uses native tree views, markdown editors, diffs, Source Control, Chat, and notifications while keeping lifecycle and security behavior in @openspec-ui/core.

Project site: https://openspec-ui.dev.

Screenshots

Workbench overview

Every view, with something in it

The OpenSpec Workbench's seven views: Changes with what the sweep cleared and a change's row saying what blocks it, the Archive, Specs with their requirement counts, Processes, the template catalogue, the Change Graph with a landed branch folded away, and the Human-Only Inbox

Compact repository overview

Compact OpenSpec Workbench with the Changes, Archive, Specs, Processes, Templates, Change Graph and Human-Only Inbox views

Changes and archive

Individual checklist entries remain nested under their tasks.md artifact, while archived changes expose lifecycle and template actions from the native context menu.

Nested tasks

Completed checklist items nested under a change Tasks artifact

Archive actions

Archived change context menu: unarchive, copy tasks as a template, rollback, show the change timeline, reveal it in the change graph, show what it follows, show what it cost, explain and recommend harness settings, and delete

Archived task history

Expanded archived change tasks beside the archive context menu

Where each change stands, and the Pipeline

The Changes tree states each change's standing beside its name, in the same word every surface uses. The words are:

Running, or Running in another working directory;
Waiting for you, or Waiting in another working directory;
Failed or Stopped at a stage;
Blocked, Ready or Done;
Further along in another copy;
Merged in #N, Archived on main, or Deleted on main.

Hovering a change shows where each part of that word was read from.

The word follows a run: it says Running within a few seconds of a run starting on the change, and stops saying it when the run ends, with no Refresh. Only the run records are read for that, never git.

Changes tree with each change's state and its standing word, such as in-progress — Ready

OpenSpec Workbench: Open Pipeline, also in the Changes view's title bar, opens the Pipeline in a panel of its own. It is the same picture the standalone app draws: every change as a card in the order the changes declare, with its state, its tasks, what a live run last said, and what can start alongside what.

The Pipeline panel in VS Code with two ready changes as cards, each with Start and Show tasks, and the worktree command suggested for each

A card's Start opens the run dialog for that change. On a run this editor holds, the card answers a checkpoint or a permission, and Stop asks the run to stop, with a reason. See stopping a run.

Finding one change, and hiding what has landed

The Archive, Specs and Change Graph views each take a filter from their title bar. Every word typed has to appear somewhere in the row, so "pipeline blocked" finds the blocked pipeline change rather than everything called pipeline. A narrowed view says what it is narrowed by and how much of itself it is showing, and a view that matches nothing says so with the words it was given.

The Archive view narrowed by a filter, reading: Filtered by "shipped" - showing 1 of 3

The Change Graph folds away every branch whose root and every change under it are archived, and ends its rows with the count it is hiding. One press shows them again; a filter that finds something inside a folded branch opens that branch for the reading. A branch a live change follows, or is waiting on, is never folded.

The Change Graph drawing a change waiting on another, and a row reading "1 landed relation hidden - Press to show"

Stating a relation without opening a file

Add Relation and Remove Relation sit on a change's row, in the Changes view and in the Change Graph. Adding asks which relation - Follows, Supersedes or Blocked by, each with the sentence that says what it means - and then which change, from the ones the workspace has. Removing offers only the relations that change actually states.

An edit that names a change the workspace does not have, a change naming itself, or one that would close a cycle is refused before anything is written, naming the changes in the cycle.

The relation picker open over the Changes view, offering Follows, Supersedes and Blocked by with what each one means

What the archive left behind

Archiving can leave a directory holding nothing but a file this product wrote, and that directory used to be listed as a change with no tasks. It is cleared now, on activation and on an interval, and the Changes view says what went. A directory holding something this product did not write, or one whose name is nowhere in the archive, is shown rather than removed: it may be a change you have not written yet.

The Changes view reading "Cleared 1 directory the archive left behind", with a directory it kept beneath it

Specs, templates, and repository setup

Canonical spec editor

Canonical spec selected in the Specs tree and open in the VS Code editor

Requirement counts

Specs tree listing capabilities and their requirement counts

Template actions

Template context menu with customize and insert actions

Repository bootstrap

Repository Setup tree with agent instructions, Dependabot, and Copilot instruction generators

Features

Navigate config.yaml, active changes, archived changes, canonical specs, and every proposal, design, task list, and delta spec.
Create, validate, archive, unarchive, and delete changes from contextual actions with confirmation for destructive operations. Expand a change, then expand its collapsible "Tasks" entry to see the individual tasks.md checklist items nested under it; select one to jump to it in the editor, or delete it (active changes only, and only if not yet done) with confirmation.
See queued, running, completed, failed, cancelled, interrupted, and rolled-back operations in the Processes view. Read-only work can run in parallel; workspace mutations are serialized to keep checkpoints isolated.
See what a run spent, where the agent reports it. A recorded cost appears beside the run, and a chain stops at a stage boundary when a configured ceiling is reached — naming the budget rather than failing a stage. A run whose agent reported nothing says so; it never shows $0.00, because unreported is not the same as free. Which agents report what, and which ceiling can act on each, is in LIMITS.md.
Read the Change Graph view: each change nested under the ones it follows, archived ones marked, and any waiting on a change that has not landed shown as waiting — so what can be started now is answerable without opening a file. It is read-only on purpose; a change with two parents appears under each, and every action stays in the Changes and Archive trees where a change appears exactly once.
Open the Human-Only Inbox to see every open item that is waiting on a person, across all active changes, each naming the change it belongs to and opening it when selected. When nothing is waiting the view says so rather than showing an empty list. It offers no way to mark an item done: those items are reported by a person after observing the thing they name, and a button on a surface that cannot observe it would turn that rule into a formality.
Open the Pipeline in a panel ("OpenSpec Workbench: Open Pipeline"): every active change as a card with its state, its tasks and what a live run last said, and every other working directory beneath. A card starts its change, answers a run's checkpoint or permission, and asks a run to stop with a reason. A run held elsewhere offers Stop only when it is signed by the same enrolled person as this machine's key.
See each change's standing word beside it in the Changes tree — the same word the Pipeline, the standalone app and openspec-ui-cli ready use — with where each part of it was read from on hover.
Say that a run signed by an unenrolled key was yours with It was me in the Human-Only Inbox; its runs then read as signed by you.
Ask a change what it follows — "Show What This Change Follows" walks back to the changes it grew out of and opens any of them. This is how a decision is traced to the reason for it once the change that made it is archived.
Move between the Changes/Archive trees and the Change Graph on request. "Reveal in Change Graph", on a change in either tree, expands and selects every row it occupies (a change following more than one other change has one row per parent) and reports the count when there is more than one; if the change states no relation it says so rather than appearing to do nothing — most changes are not in the graph, since it shows only what states a relation. "Reveal in Changes", on a graph row, opens the Archive tree for an archived change and Changes otherwise. Both are off unless invoked: openspec-ui.followSelectionInChangeGraph (default false) turns on following the Changes/Archive selection into the graph automatically, without stealing focus from the list being browsed.
Use @openspec in VS Code Chat with /plan, /implement, /review, /status, and /validate.
Or run plan/implement/review directly from the Process Dashboard's own agent picker (Claude CLI, GitHub Copilot CLI, Codex CLI, Gemini CLI, or a local OpenAI-compatible LLM) — a separate mechanism from VS Code Chat, see "Agents" below.
Recover process history and checkpointed runs after extension reload. Start VS Code Agent implementation sessions, finish them for review, and roll back only files changed by a run. Rollback refuses to overwrite later edits and discloses files omitted by checkpoint size limits.
"Rollback Change" on a Change item (in either the Changes or Archive tree) rolls back every process ever run against that Change, restoring every touched file to its state before the earliest of those runs — works identically for active and archived changes. Same fail-closed behavior as single-process rollback: any file changed outside what the system knows about refuses the entire restore, not just that one file.
Open the Process Dashboard with Workspace root and Change directory filled from the current VS Code workspace. The default message-bridge dashboard follows the active VS Code color theme, including dark and high-contrast themes.
Open native markdown and diff editors instead of custom replacements.
Browse built-in and project-level change templates in the Templates view, grouped first by origin (Built-in / Project) and then by category. Customize a built-in template into your project (keeping a backlink to the version it was forked from), insert a rendered template into any active change, or delete a project-level template with confirmation — built-in templates ship as part of the extension and are never deletable through the UI. 17 built-in templates across 10 categories (data-layer, framework-migration, architecture-migration, testing, ci-cd, auth, containerization, observability, configuration, release-management), covering Python, Node.js/TypeScript, ASP.NET Core, and language-agnostic patterns.
Bootstrap repository files for a picked project type (Node.js/ TypeScript or Python): agent instructions (CLAUDE.md/AGENTS.md), .github/dependabot.yml, and path-scoped .github/instructions/ <subtype>.instructions.md. Reachable either from the Command Palette or from the "Repository Setup" node at the top of the Changes tree. Any of these commands leaves an existing file alone (and tells you so) if it wasn't created by this extension — never overwrites hand-written content.

Requirements

VS Code 1.90 or newer.
OpenSpec CLI available as openspec on PATH.
A workspace containing openspec/config.yaml, or initialize OpenSpec from the CLI before using the Workbench.
VS Code Chat and an available language model for /plan, /implement, and /review. Deterministic lifecycle, status, and validation actions do not require AI.

Workflow

Open the OpenSpec activity-bar container.
Create or expand a change in Changes.
Edit Proposal, Design, Tasks, and delta Specs in native editors.
Run Validate, then choose Implement with VS Code Agent — this runs through VS Code's own Copilot Chat, using whatever model is selected in Chat's own model picker. If you specifically want a particular CLI agent (e.g. the real Anthropic Claude Code CLI, not Claude via Copilot) to make the change instead, use OpenSpec Workbench: Open Process Dashboard and its own agent picker instead — see "Agents" below.
Follow the run in Processes. When Agent work is done, choose Finish Implementation & Review.
Review native diffs, roll back the checkpoint if needed, or archive the completed change. Archiving creates or updates canonical specs.

Chat

Enter @openspec and choose a slash command. The first argument must be an active change id:

@openspec /review openspec-workbench focus on rollback safety

Agents

To run plan/implement/review through a specific CLI agent (Claude CLI, GitHub Copilot CLI, Codex CLI, Gemini CLI, or a local OpenAI-compatible LLM) instead of VS Code's native Chat/Agent handoff:

Run OpenSpec Workbench: Open Process Dashboard from the Command Palette (or the toolbar button on a change in Changes).
In the panel, pick implement (or plan/review) from the command dropdown.
Pick the agent you want from the agent dropdown next to it — each option shows a best-effort detected/not-detected badge.
Click Run.

This picker is independent of VS Code Chat: it runs the command through this extension's own CLI-agent protocol (@openspec-ui/core's buildDefaultAgentRunners), resolved at activation from the open workspace. Each listed CLI tool (claude/copilot/codex/gemini) must already be installed and authenticated on the machine separately — the extension never handles API keys itself. If a tool is missing, the run fails immediately with a clear error instead of hanging. This works the same way in both the default message-bridge dashboard and the optional local-server mode (openspec-ui.transport.localServer.enabled). None of these tools need a VS Code extension or any VS Code-specific setup — a plain CLI login is enough, the same as using it from a terminal. See the root repository README.md's "Agent Selection" section for the full picture, including how this differs from the @openspec Chat Participant above.

Agentic Harness

OpenSpec Workbench: Configure Harness Settings and OpenSpec Workbench: Configure Harness for this Change set the per-stage agent, autonomy level, and review gate the commands above (and a chain run) use by default. Each opens a panel of its own: the first for the workspace's global file, the second for the one change it was run on, titled Harness: <change>, already loaded; OpenSpec Workbench: Run with Agentic Harness starts a single-stage run or a chain depending on the resolved autonomy level. See the root repository's HARNESS.md for every setting these commands can and cannot reach, and LIMITS.md for what actually caps a run's spending.

A chain run can be answered mid-stage: when a stage's agent asks for permission before a sensitive action, the chain view offers an Allow/Deny control naming the request, and the answer reaches that stage's own run. Under autonomous there is no confirmation channel, so a permission request there ends the run with a stated reason instead of waiting.

Settings

openspec-ui.transport.localServer.enabled: use the optional standalone REST/WS shell inside the dashboard, for both the AI panel and the Pipeline panel. The default extension path imports core directly and uses an in-process message bridge. Because the optional localhost shell is a cross-origin iframe, it retains the standalone palette instead of inheriting VS Code theme variables. With this on, the Pipeline panel's cards are read by the server's own process instead of the editor's, so they keep answering while a harness chain runs in this editor.
openspec-ui.checkpointRetentionDays: days to keep process/checkpoint history before it's pruned, once, on the next window reload. Defaults to 0, which keeps everything forever — unchanged from every prior version. A positive number prunes anything older than that many days.

Warning: once a process is pruned this way, both single-process Rollback and "Rollback Change" become permanently unavailable for it — pruning itself cannot be undone. Leave this at 0 (the default) unless you specifically want bounded history growth and have accepted that trade-off.

Development

npm run typecheck --workspace openspec-ui-vscode
npm run lint --workspace openspec-ui-vscode
npm run test --workspace openspec-ui-vscode
npm run build --workspace openspec-ui-vscode
npm run test:integration --workspace openspec-ui-vscode

To try a local change in your own VS Code instead of just building:

npm run reinstall:local --workspace openspec-ui-vscode

Unlike npm run package (which only produces the .vsix), this also force-installs it — the extension's version in package.json often does not change between local iterations, and code --install-extension silently does nothing when the target version is already installed, so a plain reinstall without --force can leave you testing stale code. You still need to run "Developer: Reload Window" yourself afterward — that step is not automated.