auto-dev（自动开发）

Auto Dev

This skill is designed for reuse across projects. Keep business-specific behavior in project packs and inject it via .skills-hub/ hooks.

Guardrails

• Confirm scope with pwd; keep all reads/writes inside the current worktree.
• Read the current branch with git rev-parse --abbrev-ref HEAD.
• If the branch is dev or main, stop and ask the user to switch by default.
• Never checkout, merge, rebase, or push dev or main by default.
• Never push to any remote branch other than the current branch name by default.
• Never force-push or rewrite remote history.
• For GitHub Actions deploys, require the branch to track origin/<current-branch>; infer deploy targets from the remote branch state, not unpushed local commits.
• skills-hub exception (explicit user request only):

- Allow operating on main only in the skills-hub repo. - Treat explicit confirmation as part of the protected command invocation, using AUTO_DEV_ALLOW_SKILLS_HUB_MAIN=skills-hub-main-confirmed. - Allow commit/push to origin/main only after task completion and explicit confirmation. - Keep all other protections unchanged.

Secure context input

• Request needed credentials, test accounts, or session values (e.g., app_session, test account/password) via command input.
• Do not persist secrets in files or commit history.

Bundled scripts

• Resolve the installed skill directory as <path-to-skill>, then run scripts from <path-to-skill>/scripts/.
• auto-dev-preflight.sh: verify repo scope and branch safety; emits AUTO_DEV_REPO_ROOT and AUTO_DEV_BRANCH.
• auto-dev-deploy-dev.sh: generic workflow trigger runner. It reads project-specific deploy mapping from AUTO_DEV_INFER_SCRIPT or <repo>/.skills-hub/auto-dev/infer-targets.sh.

- The infer script reads changed files on stdin. - It may emit workflow=<workflow-file> and input:<key>=<value> lines. - The deploy script provides AUTO_DEV_DEPLOY_REF, AUTO_DEV_DEPLOY_SHA, AUTO_DEV_DIFF_BASE, and AUTO_DEV_DIFF_MERGE_BASE to the infer script.

Project-specific logic (keep out of core skill)

• Store business mapping, deployment knobs, and repo shortcuts in a project pack (for example project-packs/<project>/...).
• Keep SKILL.md and core scripts generic; inject business behavior via the infer script only.

Browser automation preference

• Default browser route: playwright skill for one-shot or reproducible browser flows from the terminal.

- Use it for scripted repro, snapshots, screenshots, traces, data extraction, and quick web UI checks where restarting or re-opening the browser is acceptable.

• Escalate to playwright-interactive when the task is iterative or stateful.

- Use it for repeated reloads during local development, persistent browser handles, Electron apps, richer functional plus visual QA, or when the same page/session must survive multiple edits.

• Do not choose Chrome MCP as the default path or backup path.

- Only use a Chrome MCP workflow when the user explicitly asks for it, or when an existing project workflow already depends on it and it is confirmed working. - For browser coverage beyond Chromium, prefer Playwright rather than adding a second MCP-specific branch. If the user needs the actual Safari app, ask for clarification.

Capabilities

• Use coding, testing, cloud, web, and browser automation skills for autonomous development and debugging, with Playwright CLI first and Playwright Interactive for persistent sessions.
• Allowed actions include inspecting Firestore/Storage data, accessing GCP, browsing web UIs, and triggering GitHub Actions for the current branch.

Draft PR and publish flow

• Preferred handoff after completion: create a draft PR for review context, e.g., gh pr create --draft --base main --head <current-branch>.
• For skills-hub only, if the user explicitly requests direct publish to main, commit and push to origin/main.
• Keep commit messages scoped and clear so draft PR and direct-push history are both auditable.

GitHub Actions deploy (current branch only)

• Trigger workflows with the current branch ref, e.g., gh workflow run dev.yml --ref <current-branch>.
• Use the remote branch state as the source of truth for deploy inference; warn when local HEAD differs from origin/<current-branch>.
• Default diff base is origin/main when available, otherwise resolve origin/HEAD; require --diff-base or AUTO_DEV_DIFF_BASE only when neither exists.
• Treat --workflow and AUTO_DEV_WORKFLOW as explicit overrides; do not let the infer script silently replace them.
• Confirm the workflow targets the dev environment for the current branch only.

Reporting

• Summarize changes and commands run.
• Provide quick verification steps when applicable.