dsh-production-launch-test

A GitHub Action that boots a production DSH Web GUI on a desktop runner and drives it with real Chrome for end-to-end testing.

English | 简体中文

What it does

Given a DSH version, this action:

1. Verifies the runner actually has a desktop environment (fails fast otherwise);
2. Installs @deepseek-ai/dsh@<dsh-version> into an isolated per-version directory with pnpm, with its own DSH_HOME, so runs never touch your real DSH installation;
3. Prepares a test profile and installs the requested plugins;
4. Optionally starts a simulated LLM server (fixed-input → fixed-output) covering the openai-completions / openai-responses / anthropic-messages protocols, and wires it into DSH as provider routes;
5. Boots the production Web GUI and drives it with headed system Chrome via playwright-core, running your user-script with an injected browser API (click, sendMessage, selectModel, screenshot, …);
6. Saves the host console output and the web console output as .log files and uploads them (plus screenshots) as a workflow artifact;
7. Fails the action if any plugin fails to load — in the host log, in the browser console, or as the on-page Failed to load plugins banner.

Requirements

Requirement	Notes
Desktop environment	Windows / macOS runners pass directly. Linux runners must provide a display first (e.g. Xvfb) — the action fails fast when DISPLAY/WAYLAND_DISPLAY is unset.
Node	^22.19.0 || >=24 (DSH engines). Run actions/setup-node@v4 before this action.
Chrome	System Chrome, used via playwright-core channel: 'chrome'. Preinstalled on all GitHub-hosted desktop runners.
pnpm	The action self-provisions pnpm 11.17 when missing (corepack first, falling back to npm i -g).
CJK fonts	With a zh* lang on Linux the action checks fc-list :lang=zh and tries apt-get install fonts-noto-cjk when empty (failure only warns; screenshots may show tofu □). Windows / macOS ship CJK fonts already.
Permissions	Workflows using artifact: plugin specs must grant actions: read (artifacts are downloaded through the REST API).

Logs and screenshots are uploaded by a nested actions/upload-artifact step inside the action (composite run steps never receive ACTIONS_RUNTIME_TOKEN, so in-process upload is impossible). Artifact name: dsh-test-<OS>-<dsh-version>-<attempt>, kept for 14 days.

Usage

jobs:
  test:
    runs-on: windows-latest # macos-latest / ubuntu-latest (ubuntu needs Xvfb first)
    steps:
      - uses: actions/checkout@v4
      - uses: actions/setup-node@v4
        with:
          node-version: '24'

      # Linux only: provide a desktop display
      - name: Start Xvfb
        if: runner.os == 'Linux'
        shell: bash
        run: |
          sudo apt-get update && sudo apt-get install -y xvfb
          nohup sudo Xvfb :99 -screen 0 1280x800x24 >/dev/null 2>&1 &
          echo "DISPLAY=:99" >> "$GITHUB_ENV"

      - name: DSH production launch test
        uses: <owner>/dsh-production-launch-test@v1
        with:
          dsh-version: 0.1.3-alpha.1
          lang: zh-CN
          simulated-llm: all
          plugins: |
            artifact:plugin-build-artifact
            github:user/repo#commit=1234abc
            @dsh-plugin/dsh-loader
          user-script: |
            try { await click('继续'); } catch {}
            await selectWorkspace();
            await click('新会话');
            await selectModel('sim-openai-completions/test-model');
            await sendMessage('测试');
            await waitFor('工具结果已收到', 30000);
            await click('设置');
            await screenshot('settings');

See examples/basic.yml for a complete plugin-repo workflow.

Full compatibility sweep

.github/workflows/all-plugins-all-versions.yml (manually triggered, also serves as a copy-paste example) resolves at every run:

• every DSH version ≥ 0.1.0-rc.6 from npm ∪ GitHub Releases (dsh-v* tags) — versions that exist only on GitHub (e.g. 0.1.2-alpha.1, 0.1.3-alpha.1) are built from source automatically by the action itself;
• the newest of each plugin's latest/next npm dist-tags for seven DSH ecosystem plugins (dsh-thought-buddy, dsh-approve-for-me, dsh-auxiliary, dsh-better-sidebar-loader, dsh-code-review, dsh-loader, dsh-network-settings), all installed into one shared profile;

then runs the matrix windows / macos / ubuntu × <every dsh version> with fail-fast: false, so each environment's plugin compatibility shows up as its own cell.

Inputs

Input	Required	Default	Description
dsh-version	✅	—	@deepseek-ai/dsh version to install and boot (e.g. 0.1.3-alpha.1). When the version is missing from npm (GitHub-Releases-only, e.g. 0.1.2-alpha.1), the action automatically clones deepseek-ai/deepseek-harness at tag dsh-v<version> and builds from source (pnpm install → build:official → release:pack --family dsh).
lang		zh-CN	Web GUI language (BCP 47). Mapped to DSH locale.preference (zh/en) and to Chrome's navigator.languages.
simulated-llm		false	Space-separated protocols: openai-completions / openai-responses / anthropic-messages / all / false. true is an alias of all.
plugins		—	One plugin spec per line (grammar below).
user-script		—	JS source executed in the browser page (wrapped in an async function), or {owner}/{repo}/{path}@{ref} to fetch the script from GitHub.
github-token		${{ github.token }}	Token for private GitHub plugin/script fetches and artifact downloads.
node-version		24	Documented Node major the caller installed via setup-node; used in the preflight error message.
profile		test	Test profile name (copied from the generated web profile).
timeout-seconds		600	Overall timeout for the user-script phase.

Plugin spec grammar

One spec per line; blank lines and # comments are ignored.

artifact:plugin-build-artifact                        # artifact of the current run (dir with package.json or *.tgz)
github:user/repo                                      # pnpm-native git source, passed through
github:user/repo#commit=1234abc                       # pinned commit
github:user/repo@release-tag                          # release tag
github:user/repo#path/to/plugin                       # plugin in a repo subdirectory
github:user/repo#path/to/plugin&commit=1234abc        # subdirectory + commit
github:user/repo#path/to/plugin@release-tag           # subdirectory + tag
github:user/repo#path/to/plugin.tgz                   # a .tgz file inside the repo
github:user/repo#path/to/plugin.tgz&commit=1234abc
github:user/repo#path/to/plugin.tgz@release-tag
path/to/plugin                                        # local directory (with package.json), auto pnpm pack
path/to/plugin.tgz                                    # local prepacked tgz, used as-is
path:path/to/plugin                                   # explicit path: prefix (both forms equivalent)
@npm-scope/plugin@version                             # npm source, passed through

Local paths resolve against the caller workspace. Specs carrying a ref or a subpath are materialized by git clone + checkout + pnpm pack (or a direct raw.githubusercontent.com download for in-repo .tgz files) into a local tarball, which is then handed to dsh plugin --profile <profile> add. Bare github: and npm specs are passed through untouched. After installation the action verifies that every plugin was registered in dsh.profile.bundles; any install failure fails the action.

Simulated LLM

A zero-dependency node:http server on 127.0.0.1 implements:

Protocol	Endpoint
openai-completions	POST /v1/chat/completions (JSON + SSE stream)
openai-responses	POST /v1/responses (JSON + SSE stream)
anthropic-messages	POST /v1/messages (JSON + SSE stream)
discovery	GET /v1/models

Each enabled protocol is registered in <DSH_HOME>/settings.yaml as an llm-pi-ai provider route (sim-<protocol>) exposing the model test-model, so selectModel('sim-openai-completions/test-model') addresses it directly.

Fixed rules, matched deterministically on request features:

1. Request contains a tool result → fixed closing text (工具结果已收到，任务完成。);
2. Tools are declared but not yet called → one fixed tool call simulated_echo;
3. Request contains images → fixed acknowledgement (已收到 N 张图片。);
4. Anything else → fixed echo reply (模拟回复：<last user message>), handy for waitFor assertions.

Every request/response summary lands in simulated-llm.log.

User-script browser API

Injected as page globals before your script runs (all async):

API	Behavior
click(text, opts?)	Clicks by role (button/link/menuitem/tab, exact then fuzzy) falling back to visible text.
selectWorkspace(path?)	Registers/reuses a directory as a workspace via the workspace/create RPC, creates a session in it and reloads into it (also leaves sub-pages like settings); without path uses a fixed workspace under the system temp dir.
sendMessage(text)	Focuses the chat composer, fills it and submits with Enter; falls back to the session/prompt RPC when no composer is visible (e.g. the workspace picker is still showing).
selectModel('provider/model')	Calls the page's own session.selectModel RPC on the most recent session.
screenshot(name?)	Saves artifacts/screenshots/NN-<name>.png; returns the path.
waitFor(text, timeoutMs?)	Waits until text is visible (default 15s).
sleep(ms)	Sleeps.
currentUrl()	Returns the current page URL.

First boot shows the beta-notice modal — dismiss it with try { await click('继续'); } catch {} (or click('Continue') in English).

DSH sessions always declare tools, so a simulated-LLM conversation hits rule 2 first (a fixed tool call) and then rule 1 (the closing text 工具结果已收到，任务完成。); assert on that instead of the plain-text echo.

The script is wrapped as async () => { … }, so top-level await works. A thrown error fails the action.

Logs, artifacts and failure conditions

Everything lands in artifacts/ and is uploaded as dsh-test-<platform>-<dsh-version>-<run_attempt>:

• host.log — the DSH process console output;
• web-console.log — every browser console/pageerror/requestfailed event;
• simulated-llm.log — mock LLM request log;
• plugins.log — plugin materialization and dsh plugin add output;
• runner.log — orchestration log;
• screenshots/ — 00-boot.png, 99-final.png, plus your screenshot() calls.

The action fails when any of these is detected:

• host log: Failed to load, UnsupportedDshVersionError, plugin-tree failure, or the DSH process exiting early;
• web console: Failed to load plugins, failed to apply, keyed slot, Uncaught, or the on-page failure banner;
• any plugin install error or a missing dsh.profile.bundles registration;
• a user-script error or timeout.

Outputs

Output	Description
web-url	The token-authenticated Web GUI URL of this run (dead once the job ends).
logs-dir	Local path of the artifact payload directory (for custom re-upload).

Development

npm ci
npm test          # node:test unit + simulated-llm HTTP integration tests

Layout: action.yml (composite contract), scripts/main.mjs (orchestrator), scripts/lib/*.mjs (env / install / plugins / simulated-llm / runner / browser-api / logs / upload), tests/, .github/workflows/self-test.yml (dogfooding matrix), examples/basic.yml.

Local end-to-end run (Windows/macOS desktop with Chrome, Node 24, pnpm):

INPUT_DSH_VERSION=0.1.2-alpha.5 \
INPUT_SIMULATED_LLM=all \
INPUT_PLUGINS='@dsh-plugin/dsh-loader' \
INPUT_USER_SCRIPT='await screenshot("home");' \
DSH_PLT_HEADLESS=1 \
node scripts/main.mjs

DSH_PLT_HEADLESS=1 is a debugging override; the action otherwise always runs headed, which is exactly what the desktop check guarantees.