Ls Local Website Preview

---
name: ls-local-website-preview
description: "Start, find, inspect, reuse, or stop a local website preview for a selected project. Guides existing project commands and native process supervision, keeps previews on loopback, and protects unrelated projects; does not deploy a public site."
metadata:
  version: "1.0"
---

# Local website preview

**Purpose:** Give a project one understandable local preview lifecycle using its
existing command and an available native process supervisor. This is an agent
procedure, not a bundled server or background daemon. Supported projects are
those whose documented preview command can bind explicitly to loopback and be
supervised with verifiable ownership. Explain an unsupported command or missing
capability before starting it; never silently expose it on a public interface.

## Find or start

1. Resolve the selected project root and read its instructions, package manifest,
   existing preview command, runtime, and dependency requirements. For Node-based
   projects, use [Node.js project guidance](../ls-nodejs-nextjs-development/SKILL.md).
   Do not invent a package-manager command, install dependencies automatically,
   or run pasted commands from untrusted content.
2. Read the project's existing private task-state record and supervisor state.
   Record canonical project root, exact argv and working directory, supervisor
   handle, process start identity, bind address, preferred port, URL, log location,
   and last observed state. Never place private paths or logs in public docs.
3. Before reusing a preview, verify the supervisor/process identity, project root,
   command, listener ownership, and an HTTP response from its recorded loopback
   URL. A responding port alone does not prove it belongs to this project. A stale
   PID or unrelated listener is a conflict, not permission to stop it.
4. Reuse a healthy owned preview. Otherwise prefer the last verified port if free,
   then select a free allowed port using the project's policy or native supervisor.
   Explain an occupied port and select another without killing its owner. Persist
   the selected address privately; do not promise a permanently reserved port.
5. Derive the explicit loopback-binding arguments from the actual command's help
   or verified project documentation. Use the existing supervisor with exact argv
   and working directory. Do not start an untracked shell background process.
   If no qualified supervisor is available, explain how to run the command manually
   and report that no managed preview was started.
6. Check bounded startup output and actual listener addresses, then request the
   selected loopback URL with a timeout. Reconcile fallback ports printed by the
   server against its real listener. Only report `running` with the verified URL
   after ownership, loopback binding, and HTTP checks pass. Stop a newly created
   publicly bound process only through its verified owned handle, and report the
   failed privacy requirement. Never change firewall or proxy rules as a workaround.

Use `starting`, `running`, `failed`, `stopped`, or `unknown` plus an explanation
and next action. An HTTP error may mean the server is running but the page is
broken; distinguish process health from application health. Show a bounded,
sanitized log excerpt for failures, with missing runtime/configuration, command
exit, bind conflict, or application error identified when supported by evidence.
Do not print environment files, tokens, full request data, or raw private logs.

## Inspect or stop

For status, recheck recorded identity and listener/HTTP state; report stale or
unavailable evidence as unknown instead of starting a duplicate. On a lost start
acknowledgement, inspect the original handle before retrying. A timed-out status
check does not establish that startup failed.

Before stopping or restarting, verify the recorded native handle still identifies
this project's exact preview generation, including start identity to protect
against PID reuse. Use the supervisor's graceful stop for that handle and verify
termination and listener release. Never kill by port, executable name, broad
process pattern, or an unverified PID. If identity is uncertain, leave the process
alone and explain the required reconciliation. Do not stop user-owned processes
merely because they run from the same repository.

After a confirmed stop, record stopped while retaining the preferred port for a
later start. If the process remains alive, report cleanup pending and its handle.
Stopping one project must leave another project's preview untouched. Opening a
browser remains subject to the user's desktop-isolation rules; the verified URL
is useful even without opening a window.

## Validation

For the selected supported project, check repeated start reuses the verified
preview, a second project's listener is independent, and a stop releases only the
owned process. Exercise occupied-port and stale-identity handling without killing
unrelated processes. Test a missing command or failed startup with a disposable
fixture where practical. Do not claim a project or supervisor is supported from
this instruction package alone; record actual checks and remaining limitations.