Dev Server — Start & Monitor

Start the project's dev server and watch its output for errors. The user says "start dev" and you handle detection, startup, and monitoring — they only hear from you when something breaks.

Workflow

1. Detect the stack

Scan the working directory for project markers. Check in this order — first match wins:

Node.js framework detection — when package.json exists, check dependencies:

Node.js package manager — check in order:

1. bun.lock or bun.lockb → bunx / bun run
2. pnpm-lock.yaml → pnpm exec / pnpm run
3. yarn.lock → yarn / yarn run
4. Default → npx / npm run

Python environment — check in order:

1. .venv/ or venv/ exists → source .venv/bin/activate &&
2. poetry.lock → poetry run
3. Pipfile.lock → pipenv run
4. Default → direct command

If nothing matches, tell the user you couldn't detect the stack and ask what command to run.

If the detection picks a stack that seems wrong for the directory (e.g. package.json exists but it's just dev tooling in a Python project, or both manage.py and package.json present and unclear which is primary), show the user your guess and confirm before starting. Don't silently commit to a wrong stack.

2. Pre-flight checks

Before starting:

1. Port conflict — check if the default (or user-specified) port is taken:

   lsof -i :<port> -t 2>/dev/null
   

   If occupied, identify the occupying process (ps -p <pid>) and present two options to the user: kill PID N, or bind to port+1. Pick a default — don't block on the question if the occupying process is clearly another dev-server instance of the same project.

2. Dependencies installed?

   • Node: check node_modules/ exists
   • Python: check virtualenv exists or key packages importable
   • Ruby: check bundle check passes
   • Go/Rust/Java: no check needed (build tools handle it)

   If missing, ask the user if you should install first.

3. Start with Monitor

Use the Monitor tool with persistent: true so the server runs for the session.

The grep filter matters because a raw dev-server stdout stream is mostly routine request logs — thousands of lines of noise that would flood the conversation and hide real errors. The pattern below surfaces only startup failures, runtime exceptions, build errors, and warnings:

<dev-command> 2>&1 | grep --line-buffered -iE \
  '(error[:\[]| ERR[!_]|EADDRINUSE|EACCES|ENOENT|ECONNREFUSED|FATAL|panic|Segmentation|Module not found|Cannot find module|ModuleNotFoundError|ImportError|SyntaxError|TypeError|ReferenceError|NameError|AttributeError|KeyError|ValueError|RuntimeError|IndentationError|failed to|build failed|compile error|compilation error|WARN[:\[]|warning[:\[]|deprecated|port.*already|address already in use|unhandled|rejected|crash|killed|Traceback|Exception|FAILED|ActionView|ActiveRecord|LoadError|undefined method|NoMethodError|cannot find|not found|Permission denied|exit code [1-9]|exit status [1-9]|thread.*panic|cannot compile)'

Set the Monitor description to something specific: "Next.js dev :3000" or "Django dev :8000".

Port flag by stack (when user specifies a port):

• Next.js: -p PORT
• Vite/Astro/SvelteKit/Nuxt/Remix/Angular: --port PORT
• Django: 0.0.0.0:PORT
• Flask: --port PORT
• FastAPI/Uvicorn: --port PORT
• Rails: -p PORT
• Phoenix: via PORT=PORT env var
• Laravel: --port PORT
• Go/Rust: usually via env var or arg (check the code)

Announce once started and verified:

Started Next.js dev server (bun run dev) on :3000 — root URL returned 200, no error overlay. Monitoring for errors.

3.5. Verify it's actually serving

"No error in the Monitor" ≠ "the app works." After the dev command announces readiness, confirm the server is reachable before handing control back to the user:

1. HTTP probe — curl -sS -o /dev/null -w "%{http_code}\n" http://localhost:<port> with a 5s timeout. Expect 2xx/3xx. Non-2xx, connection refused, or timeout → investigate.
2. UI smoke check (frontend projects) — when the project is a frontend framework (Next.js, Vite, Remix, Astro, SvelteKit, Nuxt, Angular, Gatsby, Expo web), invoke the browser-use skill to load the root URL, take a screenshot, and check for a visible error overlay (Next.js red box, Vite overlay, React error boundary fallback). Report what you see.
3. Backend-only projects — if there's a known health/root route (/health, /api/health, /), curl it and show the status code. Otherwise one HTTP probe is enough.

If verification fails, surface it with the same classification as §4 (build/runtime/dependency/port) — don't just re-announce success.

4. React to errors

When a Monitor notification fires, surface everything the filter caught — do not silently drop items because they look minor. Claude's job is coverage; the user decides what to ignore.

1. Classify — build error, runtime exception, missing dependency, port conflict, type error, syntax error, warning, deprecation.
2. Context — quote the error, name the file and line if visible.
3. Fix or ask — for obvious issues (missing dep, typo, known pattern), suggest or apply the fix. For ambiguous ones, ask.
4. Severity shape — label each item (error / warning / info) so the user can filter; do not pre-filter on their behalf.

5. Stopping

When the user says "stop", "kill it", "shut down", or similar:

• TaskStop the Monitor
• Confirm stopped

Edge cases

• Monorepo: If package.json has workspaces or there's a turbo.json/nx.json, ask which package. Or check for a root dev script.
• Multiple servers: Frontend + backend? Start each Monitor in the same turn (parallel Bash calls) with distinct descriptions like "Rails API :3001" and "Vite web :5173". Don't serialize — they're independent.
• Docker Compose: If docker-compose.yml exists alongside a framework, mention both options.
• Custom commands: If the user says "run XYZ and watch it", skip detection — just run their command through the Monitor filter.
• Turbopack: If Next.js dev script includes --turbo, preserve it.

Session continuity

The Monitor is persistent: true and outlives context compaction. If the conversation has been compacted and the user asks about the server, restate: framework, port, PID (from startup), and how long it's been running. Don't assume the user remembers which dev server was started.

Source: alexandrbasis/claudops — distributed by TomeVault.