Prove an OpenWork change works on the real product before calling it done. Start the product, drive the change the way a user or client would, capture screenshots, and show the proof fails on dev and passes on your branch. Use before opening a PR, when a reviewer asks "show me", or when a core journey is red.
インストールする前に、エージェントに与えられる指示の中身を確認できます。
A change is done when you have watched it work on the real product, not when it
compiles. The committed e2e suite is only the nine core journeys in
evals/journeys.ts; everything else is proved per PR and thrown away.
| The change is… | Prove it with |
|---|---|
| a server rule (permissions, isolation, validation, secrets) | an API test next to the code, e.g. ee/apps/den-api/test/api/*.api.test.ts (pnpm --filter @openwork-ee/den-api test:api) |
| a pure function or component logic | a colocated unit test |
| something a person or client does end to end | a proof: evals/proofs/<name>.proof.test.ts (write-a-spec) |
| one of the nine core journeys broke or changed | a step in that journey, evals/core/<journey>.e2e.test.ts |
pnpm world up dev-app-web --detach -- --browser # this checkout in headless Chrome; prints webUrl and cdp
pnpm world down dev-app-web # when done
Use a desktop window only for desktop-shell changes (see drive-desktop-cdp).
Den needs MySQL and Redis on localhost (pnpm dev:den:mysql, or a local MySQL
via OPENWORK_EVAL_MYSQL_URL=mysql://root@127.0.0.1:3306).
drive-desktop-cdp (new session, send, wait, screenshot).pnpm evals:e2e <name> --local; screenshots and the report land in
evals/results/test-runs/<latest>/.Drive the real user path. No internal setters, no test-only endpoints, no
reading app state; pnpm --dir evals lint:specs rejects those in proofs.
A proof that passes on dev proves nothing about your change. Run it on both:
git worktree add /tmp/ow-base origin/dev
cp evals/proofs/<name>.proof.test.ts /tmp/ow-base/evals/proofs/
(cd /tmp/ow-base && pnpm install --frozen-lockfile && pnpm evals:e2e <name> --local) # must fail
pnpm evals:e2e <name> --local # must pass
git worktree remove /tmp/ow-base
If it passes on dev, the assertion is looser than the change: tighten it.
Open the PR with open-a-pr. CI runs your proof and the nine core journeys on
the PR head and posts the evidence. Name the proof and its before → after in
the PR's Evidence section.
| Journey | What it proves | Run |
|---|---|---|
org-invite-join | an owner invites someone and they join with the right role | pnpm evals:e2e org-invite-join --local |
core-chat | a signed-in person sends a message and reads the reply | pnpm evals:e2e core-chat --local |
thread-continues-after-restart | a conversation, queued follow-up and draft survive a restart | desktop; pnpm evals:e2e thread-continues-after-restart --local |
agent-connection-journey | a connector call in chat shows what was called and what came back | pnpm evals:e2e agent-connection-journey --local |
agent-permissions | blocked commands stay blocked; allowed ones run | pnpm evals:e2e agent-permissions --local |
app-smoke | the packaged desktop app opens to a working screen | packaged smoke (apps/desktop/scripts/packaged-smoke.mjs) |
packaged-update-install | an installed app downloads and installs a newer build | macOS packaged smoke |
agent-signup-via-mcp | an MCP client connects, searches capabilities and runs one | pnpm evals:e2e agent-signup-via-mcp --local |
secure-keys-in-chat | keys stay out of the transcript, model, logs and other members' calls | pnpm evals:e2e secure-keys-in-chat --local |
まだレビューはありません。使ってみた感想をお寄せください。
概要と使いどころ
Add a new feature, put work behind a feature flag, roll something out (per organization, for everyone, cloud or self-hosted), revert or kill a feature, or make a big change to existing behavior such as a new engine. Use BEFORE writing the feature code, and when finishing or removing a rollout.
日本語の概要は準備中です。原文の説明を表示しています。
Local OpenWork Electron browser automation with CDP. Use when driving a local Electron dev app, browser_list, browser_snapshot, browser_eval, composer automation, or local UI smoke tests.
日本語の概要は準備中です。原文の説明を表示しています。
Flag customer, prospect, partner, or outside-person identities in the diff. Reported in the Warden security summary.
日本語の概要は準備中です。原文の説明を表示しています。
Connect the OpenWork MCP Gateway (https://api.openworklabs.com/mcp/agent) to Claude Code, Codex, Gemini CLI, Cursor, VS Code, Claude Desktop, or ChatGPT so the agent can use the user's OpenWork organization skills, plugins, and connections.
日本語の概要は準備中です。原文の説明を表示しています。
Screen a pull request from an outside contributor for hidden, obfuscated, or supply-chain-risky changes before any of its code runs.
日本語の概要は準備中です。原文の説明を表示しています。
Create an OpenCode plugin for OpenWork. Scaffolds the plugin file with the correct API shape, tool definitions, and hook registration. Use when the user asks to 'create a plugin', 'write a plugin', or 'make a plugin that does X'.
日本語の概要は準備中です。原文の説明を表示しています。