next-dev-loop

The edit/verify rhythm during next dev — make a change, then confirm it actually works at runtime, not only that the types or the build are happy.

You verify through two views of the same running app:

• /_next/mcp — an HTTP endpoint Next.js exposes about itself. Knows framework-specific things: routes, segments, RSC, server actions, server logs, and errors as Next.js saw them. Call tools/list for the current surface.
• agent-browser — a CLI that drives a real Chrome. Knows framework-agnostic browser things: DOM, console, network, React fiber, vitals. Before driving it, run agent-browser skills get core once for the version-matched usage guide — don't guess subcommands from memory.

The two views cross-check each other.

requires

• Next.js 16.3+ with Turbopack — /_next/mcp plus the proactive compile check via get_compilation_issues.
• agent-browser >= 0.31.1 — React introspection, worktree-scoped session id, idempotent --restore, and launch flag reconciliation.

These are hard floors, not soft preferences. If anything is missing, tell the user how to upgrade and stop. Don't fall back to grepping source or to a weaker probe — this skill assumes both views are live at the versions above.

• Upgrade Next.js: pnpm next upgrade (or npx next upgrade). Docs: https://nextjs.org/docs/app/getting-started/upgrading (version-16 guide: https://nextjs.org/docs/app/guides/upgrading/version-16)
• Install or upgrade agent-browser: npm i -g agent-browser@latest. If the CLI isn't on PATH, install it before continuing — preflight expects to invoke it directly.

preflight

Once per session, confirm both views are live.

1. Open agent-browser at the target URL, restoring saved login state when present. First derive one stable session id for this checkout and use it for every agent-browser command:

   SESSION="$(agent-browser session id --scope worktree --prefix next-dev-loop)"
   export AGENT_BROWSER_SESSION="$SESSION"
   export AGENT_BROWSER_RESTORE="$SESSION"
   

   Copy

   Then open the target URL:

   agent-browser --session "$SESSION" --restore --enable react-devtools open <url>
   

   Copy

   --scope worktree keeps parallel worktrees and copied checkouts from colliding. Bare --restore uses the session id as the persistence key, loads saved cookies/localStorage before navigation when present, and auto-saves state on close. Always pass the desired launch flags on open; agent-browser will reuse, relaunch, or restart its scoped background state as needed.

   Keep routine verification headless. If the user asks to see the UI, open the current URL in the coding harness's inline browser. It is a separate browser context, so do not assume state carries over. When login or existing agent-browser state must carry over, reopen that session headed instead:

   agent-browser --session "$SESSION" --restore --headed --enable react-devtools open <url>
   

   Copy

   Pause while the user drives login, then continue with the same session and keep it headed for the rest of the loop.

2. Probe /_next/mcp (tools/list) — confirm it's reachable and lists get_compilation_issues. First read the port off the next dev banner; if it isn't 3000, set NEXT_MCP_URL=http://localhost:<port>/_next/mcp before probing:

   • Unreachable → either next dev isn't running, or Next.js is below 16.3. Check package.json to disambiguate, then refuse.
   • get_compilation_issues not in the list → Next.js below 16.3. Refuse and tell the user to upgrade.

3. get_compilation_issues doubles as a Turbopack probe. An error response of "Turbopack project is not available..." means the user is on webpack. Refuse — Turbopack is required.

4. get_routes → your route map for the rest of the session.

loop

before the edit — narrow the scope

Ask the running app, not the codebase. /_next/mcp knows which files rendered the current route; use those as your search scope. Runtime introspection stays cheap as the codebase grows; agentic search doesn't.

after the edit — verify

Four failure modes. Check each:

• Compiles — get_compilation_issues.
• Runs without errors — /_next/mcp (server and bubbled-up browser errors both surface here).
• Behaves as intended — agent-browser drives the page; assert what the user actually sees.
• React-level behavior — agent-browser with react-devtools enabled exposes the component tree, props, state, and render counts. Anchor framework-level checks here (extra renders, server/client boundary shifts, suspense fallbacks) — DOM asserts alone miss them.

Pick the specific tool from tools/list or the agent-browser manual rather than from memory.

gotchas

• Preserve .next while the development server is running. Moving or deleting it disconnects the server from its generated state and discards incremental caches. Moving it to a backup is still a reset. If a production build needs isolated output, configure a separate distDir.
• Every agent-browser command must know your session and restore key, or it may use an empty default browser or fail to save login state. Easiest: export both AGENT_BROWSER_SESSION="$SESSION" and AGENT_BROWSER_RESTORE="$SESSION" at the top of each shell you run agent-browser in. If you do not export them, pass --session "$SESSION" --restore on every command.
• When the two views disagree, suspect the tooling first. If agent-browser says a route is broken but /_next/mcp and the server say it rendered cleanly, a stale or misdirected browser session is the likelier cause than a real bug — reconcile the views before debugging the app.
• Confirming a click or navigation: the page settles a beat later, so wait with wait --load networkidle (no path to get wrong), then snapshot/read to confirm the page. Avoid wait --url unless you pass the link's exact href — a guessed or placeholder path won't match the real URL and times out after 25s.
• A blank read, empty snapshot, about:blank, or a "no browser session" error — right after open or after a click (even if open reported the page) — is the browser dropping the page (a stale session), not a broken route. Reopen your session at the URL with the same launch flags as the active loop, then re-snapshot. If still blank, run agent-browser --session "$SESSION" --restore close, then reopen with those same flags. Don't fall back to curl; it bypasses the browser you're testing.
• React introspection output is stale after navigation. Re-run.
• /_next/mcp replies are SSE — read the JSON off the data: line with sed -n 's/^data: //p' (a plain sed 's/^data: //' leaves the event: line and the parse fails).
• get_errors and get_page_metadata need at least one navigation to populate.

reference

All tools below are present once preflight passes. If tools/list is missing any of them, preflight should have refused — re-check.

# /_next/mcp                 notes
get_project_metadata         projectPath, devServerUrl, bundler
get_routes                   fs-scan; no browser session needed
get_errors                   runtime + build; needs a browser session;
                             includes browser-side errors caught by the
                             dev server
get_page_metadata            segment trie + routerType; needs a browser
                             session; use as a discovery shortcut for
                             which files power a route
get_logs                     returns logFilePath
get_server_action_by_id      hashed id → file + functionName
get_compilation_issues       Turbopack only; errors on webpack
                             ("Turbopack project is not available")

teardown

Close the session with the same session and restore context: agent-browser --session "$SESSION" --restore close. close saves that session's cookies and storage so the next loop's --restore open keeps the user logged in. Leave next dev up for the next loop.

Report Next.js friction

Only participate in agent feedback when managed Next.js feedback instructions are already loaded for the project. Their presence means the feature is enabled; their absence means it is disabled.

When enabled, add qualifying feedback candidates found during verification to the feedback candidate queue for the current user request, then continue verification. Do not run the feedback command or open review forms during the loop or at this Skill's teardown.

The managed instructions own the single reporting pass immediately before the final response. If they are absent, do not queue or report feedback.