verify
GitHub在隔离环境中构建、启动并驱动 LLM Gateway 栈,用于验证 API、网关、仪表盘或截图变更。支持本地复现、端口偏移、种子数据截图及端到端工作流检查。
Trigger Scenarios
Install
npx skills add theopenco/llmgateway --skill verify -g -y
SKILL.md
Frontmatter
{
"name": "verify",
"description": "Build, launch, and drive the LLM Gateway stack in an isolated worktree environment to verify API, gateway, dashboard, playground, or screenshot changes. Use when locally reproducing a change, launching services on offset ports, taking seeded-data screenshots, or checking a workflow end to end."
}
Verify a change
Use the repository's current implementation and AGENTS.md as the source of
truth. Do not reuse ports, containers, credentials, or commands remembered from
another worktree.
Isolate the workspace
- Read Running an isolated stack per worktree in
AGENTS.md. - Reuse the worktree's exported
STACK_SUFFIX, database URLs, Redis ports, app ports, and service URLs only if they are complete. Otherwise choose an unused slot and export the full block fromAGENTS.md. - Confirm the selected ports are free. Never start or reset the default shared Docker stack from a Conductor worktree.
Start only the isolated services needed by the change:
pnpm build:core
docker compose up -d
pnpm wait-for-services
pnpm push-test
pnpm push-dev
pnpm seed
Use pnpm setup only when a full reset is required and STACK_SUFFIX is set;
it removes the selected stack's volumes.
Build and launch
Build affected apps through Turbo filters, including their workspace
dependencies. Run the full pnpm build before handoff when required by
AGENTS.md.
Launch the configured stack with pnpm dev. Use exported service URLs where
the repository defines them; otherwise construct the local origin from the
matching app port in AGENTS.md.
Drive and capture
- Use seeded accounts and identifiers from
packages/db/src/seed.ts; passwords equal their seeded email addresses. - Verify the changed behavior through the same surface a user exercises. Pin a
gateway provider and set
x-no-fallback: truewhen provider-specific behavior matters. - Use an available browser automation tool for UI verification. If a local
Playwright script is needed, run it from an app that declares
@playwright/test(apps/ui,apps/playground, orapps/code). - Capture only seeded local data. Resolve screenshot output paths from the tool being used instead of assuming a tool-specific directory.
Record the exact commands, URLs, and observed result. Do not describe a check as passing unless it ran successfully.
Clean up
Stop processes started for the check. Run docker compose down -v only when
this session started the worktree-specific STACK_SUFFIX stack. Never tear down
the default shared stack or a stack that was already running.
Version History
- 8fdff5b Current 2026-08-16 07:59


