TechNative releases specgetty as open source: a terminal tool for reviewing OpenSpec change proposals and specifications without losing your focus.

Spec-driven development (SDD) for AI coding assistants. - Fission-AI/OpenSpec

OpenSpecはAIを利用したコーディングの手法「仕様駆動型開発（Specification-Driven Development、略してSDD）」を簡単に導入できるツールで、仕様から一貫性のあるコードを自動生成できます。

Spec-driven development (SDD) for AI coding assistants. - Fission-AI/OpenSpec

OpenSpec helps teams and coding agents create, refine, and manage living specifications.

OpenSpec has no native rejected-proposal state. Learn why, what the maintainers decided, and the decision.md convention that stops agents reopening settled investigations.

_First published on openspec-ui.dev._ OpenSpec gives an agent something to build from: a proposal, a design, a list of tasks, and the spec deltas the change will leave behind. What it does not give you is a way to see the agent while it works. Several tools draw OpenSpec changes as boards and lists, and they do it well. Read on 2026-09-20 in their own words, none of the three I compared starts an agent. OpenSpec Workbench does: it runs an agent CLI on a change, says what the run is doing while it does it, and lets a person stop it. This is a short tour of how, and of what it will not promise. ## A change becomes a card The Pipeline draws every active change in the order the changes declare, with what a live run last said and what can start alongside what. A card is where you act: it starts a run, answers a checkpoint, and asks a run to stop, with a reason. The same Pipeline is in the standalone web application and in the VS Code extension, and both read the same shared core. A run is a chain of stages: propose, review, apply, verify, archive, git. You choose the agent for each stage, and where the chain pauses for you. ## What a run says it is doing Every run writes a status record: whose it is, where it runs, what it last said it was doing and how long ago. The Pipeline draws it on the card, and `openspec-ui-cli status` prints it for every run of the repository, whichever host started it. It never says a run is stuck. A silent agent and a hung one look identical, and telling them apart is a person's judgement, so the tool leaves that call to you and gives you the last thing the run said and its age. Changes can run side by side. Each one can have its own git worktree, guarded by a lease, and `openspec-ui-cli ready` says which ready changes can start alongside which, and why the others cannot. ## Stop it where its work is sound A card's Stop asks for a reason. The request is signed with the machine's key. The run reads it at its next renewal and acts only if it is verified and fresh, and it stops where its work is sound. From a terminal: openspec-ui-cli stop <instanceId> --reason "wrong branch" `--after <task>` lets the run finish a named task first and stops at the next sound point after it. ## Your agent, per stage The picker drives the agent CLIs you already have: Claude CLI, GitHub Copilot CLI, Codex CLI, Gemini CLI, DeepSeek CLI, and a local model behind an OpenAI-compatible endpoint. The CLI-based ones also come in a variant that speaks the Agent Client Protocol, which carries structured progress instead of scraped text. The tool never handles an API key: each CLI keeps its own login. Here is what has actually been run. Claude CLI and Copilot CLI have been run against the real binaries, and so has DeepSeek CLI over ACP, on Node 22.18+ or 24.2+. Codex and Gemini never have, raw or over ACP: their adapters are written to the vendors' documented interfaces and tested against a mocked peer, and if either misbehaves for you, that is the likeliest reason and a report would be useful. Two ACP caveats are worth knowing before you choose one: the Claude adapter never asks for permission, and `copilot --acp` completed file writes and shell commands here without asking either. ## Checkpoints, ceilings and a record Where a person decides is configuration, and the runner enforces it: an autonomy level, checkpoints between stages, a review gate that only a change's own settings can relax, and tasks marked as human-only or delegated to a named agent. Each stage has its own spending cap in the unit its agent honours, each chain has a ceiling, and every run lands in an audit log. The project also states plainly which limits do not exist, in a document of their own. The Workbench ships 17 change templates, for things like a Vite migration, a production Dockerfile, an authentication middleware or a testing baseline, so a new change can start from a proposal that already has the right shape. ## Try it In VS Code, install it from the Marketplace: code --install-extension openspec-ui.openspec-ui-vscode Or run the standalone application from a clone of the repository: npm install npm run build --workspace @openspec-ui/server npm run start --workspace @openspec-ui/server -- <workspaceRoot> 4317 It prints a URL with a one-time token; open that exact URL. Everything runs on your machine, against your own agent CLIs. The repository, the packages and the extension are still called OpenSpec-UI; the product's name is OpenSpec Workbench. The code and the issues are at github.com/VeryComplexAndLongName/OpenSpec-UI, and I would like to hear from anyone who tries it, above all from people who run Codex or Gemini, which I have not been able to. ## Where each claim comes from * The Pipeline, the stages, the status record, `stop`, and the other viewers compared on 2026-09-20: the repository's README, sections "How this differs from the other OpenSpec viewers", "Agentic Harness" and "CI CLI". * Which agents have been run against a real binary, and the ACP caveats: the README's "Agent Selection" section. * Checkpoints, autonomy levels and per-stage caps: HARNESS.md. What does and does not cap a run: LIMITS.md. * The 17 templates: packages/core/src/templates, 17 files.

Spec-driven development (SDD) for AI coding assistants. - Fission-AI/OpenSpec

OpenSpec helps teams and coding agents create, refine, and manage living specifications.

_First published on openspec-ui.dev._ OpenSpec Workbench ships in two forms: a VS Code extension and a standalone web application that runs on your machine. They show the same OpenSpec and Git data and run the same workflows. The first architectural decision was how to stop them from becoming two different products, because two deliveries of "the same" behaviour diverge, and they do it quietly. The answer was to write the behaviour once and let each host be a thin adapter around it. This is what that looks like, what was turned down, and where "thin" turned out to cost something. ## The decision Five packages, one of them holding all of the behaviour: * `core` owns execution, OpenSpec and Git integration, security, persistence and the derived state of a change. * `webui` holds the React components, written so they do not know how they are reached. * `server` and `extension` are the two hosts, and `cli` is a third, non-interactive one for CI. The protocol began as five commands (`plan`, `implement`, `review`, `status` and `cancel`) and seven events (`started`, `stdout`, `stderr`, `progress`, `completed`, `failed` and `cancelled`), and it has grown since, with events such as `checkpoint` and `permissionRequest`. It is defined in `core` and nowhere else. A host serialises it and does not reimplement any of the execution behind it. Security is in the core too, from the start and not as a later addition. Every run is restricted to the workspace, commands and their arguments are allowlisted, executions are audited, and the contents of the repository are treated as data, never as instructions. ## What was rejected Three alternatives were written down and refused. **Running the standalone server inside the extension.** It would have made both hosts share one transport, but the extension host already runs Node and needs no HTTP for ordinary work. Dynamic ports, window collisions, authentication and process clean-up would have burdened the most common workflow. So the extension imports the core directly and talks to its webview over a message bridge, and the local server stays an optional mode. **Implementing execution separately in each host.** Streaming, cancellation, errors and security would have drifted apart. Both hosts adapt the same protocol instead. **Storing a change's state in`.openspec.yaml`.** That would have extended a format that belongs to OpenSpec. Draft, in progress, implemented and archived are inferred from where a change sits and from its `tasks.md`, by a heuristic in the core, so the tool never forks the format it depends on. The delivery model was also reviewed independently after the first proposal, and that review changed the recommendation for the extension's transport while keeping the shared core. ## What it bought A defect is fixed once, in the core, which depends on neither HTTP nor the VS Code API. The run status record that the Pipeline draws on a card is the same record `openspec-ui-cli status` prints, whichever host started the run. A new agent is one adapter in the core's registry, and both hosts see it. One choice in that registry is worth naming. The plain agent adapters treat a CLI's output as opaque text on purpose, so that a change in a CLI's output format between versions cannot break the event stream. The price is that they cannot see or gate a single action mid-run, which is what the adapters that speak the Agent Client Protocol were added for, beside the plain ones and not in place of them. ## Where "thin" was not free A thin adapter is still an adapter, and each host has to route each command. When an agent asked for permission in the middle of a chain, the chain had no way to receive the answer. The core routed `cancel` to the stage in flight but not `resolvePermission`. The fix was in the core, and both hosts gained the matching branch. The run did not fail. It waited. So the rule is narrower than "hosts contain no logic". Hosts contain no behaviour, but each of them still has to carry every command to the core, and a command that one host forgets to carry is invisible until somebody uses it. The project requires contract tests between the shared UI and the server before a change is archived, for that reason. ## Try it The code is at github.com/VeryComplexAndLongName/OpenSpec-UI, where the repository and packages keep the name OpenSpec-UI. A short tour of what it does is in Supervise agents on OpenSpec changes. ## Where each claim comes from * The layering, the protocol, the security model, the three rejected alternatives, the independent review and the consequences: the repository's ADR 0001. * The packages, the hosts and the direct import with a message bridge: the README's "Architecture at a Glance" and "Packages". * Plain adapters treating output as opaque text, and why ACP adapters were added beside them: ADR 0013. * The permission request that waited, and both hosts gaining the matching branch: the project's own write-up, What a run tells you: 0.40 to 0.44. * That the status record is the one `openspec-ui-cli status` prints: the README's "CI CLI" section.

Install OpenSpec, the change-centric spec-driven development CLI. Learn the explore-propose-apply-archive loop, brownfield setup, and where it breaks.

OpenSpec helps teams and coding agents create, refine, and manage living specifications.

OpenSpec brings spec driven development to AI coding. Stop chaotic AI code. Align specs before writing code.

Spec-driven development (SDD) for AI coding assistants. - Fission-AI/OpenSpec

AICoding 正從模型崇拜走向工程落地。Claude Code 與 OpenSpec 雙輪驅動，以上下文治理 + 規格化開發破解 AI 編碼意圖模糊、上下文漂移難題，打造可複用、可規模化的企業級 AI 研發範式。