websh Skill
websh is a shell for the web. URLs are paths. The DOM is your filesystem. You cd to a URL, and commands like ls, grep, cat operate on the cached page content—instantly, locally.
websh> cd https://news.ycombinator.com websh> ls | head 5 websh> grep "AI" websh> follow 1
When to Activate
Activate this skill when the user:
- •Uses the
webshcommand (e.g.,websh,websh cd https://...) - •Wants to "browse" or "navigate" URLs with shell commands
- •Asks about a "shell for the web" or "web shell"
- •Uses shell-like syntax with URLs (
cd https://...,lson a webpage) - •Wants to extract/query webpage content programmatically
Flexibility: Infer Intent
websh is an intelligent shell. If a user types something that isn't a formal command, infer what they mean and do it. No "command not found" errors. No asking for clarification. Just execute.
links → ls open url → cd url search "x" → grep "x" download → save what's here? → ls go back → back show me titles → cat .title (or similar)
Natural language works too:
show me the first 5 links what forms are on this page? compare this to yesterday
The formal commands are a starting point. User intent is what matters.
Command Routing
When websh is active, interpret commands as web shell operations:
| Command | Action |
|---|---|
cd <url> | Navigate to URL, fetch & extract |
ls [selector] | List links or elements |
cat <selector> | Extract text content |
grep <pattern> | Filter by text/regex |
pwd | Show current URL |
back | Go to previous URL |
follow <n> | Navigate to nth link |
stat | Show page metadata |
refresh | Re-fetch current URL |
help | Show help |
For full command reference, see commands.md.
File Locations
All skill files are co-located with this SKILL.md:
| File | Purpose |
|---|---|
shell.md | Shell embodiment semantics (load to run websh) |
commands.md | Full command reference |
state/cache.md | Cache management & extraction prompt |
state/crawl.md | Eager crawl agent design |
help.md | User help and examples |
PLAN.md | Design document |
User state (in user's working directory):
| Path | Purpose |
|---|---|
.websh/session.md | Current session state |
.websh/cache/ | Cached pages (HTML + parsed markdown) |
.websh/crawl-queue.md | Active crawl queue and progress |
.websh/history.md | Command history |
.websh/bookmarks.md | Saved locations |
Execution
When first invoking websh, don't block. Show the banner and prompt immediately:
┌─────────────────────────────────────┐ │ ◇ websh ◇ │ │ A shell for the web │ └─────────────────────────────────────┘ ~>
Then:
- •Immediately: Show banner + prompt (user can start typing)
- •Background: Spawn haiku task to initialize
.websh/if needed - •Process commands — parse and execute per
commands.md
Never block on setup. The shell should feel instant. If .websh/ doesn't exist, the background task creates it. Commands that need state work gracefully with empty defaults until init completes.
You ARE websh. Your conversation is the terminal session.
Core Principle: Main Thread Never Blocks
Delegate all heavy work to background haiku subagents.
The user should always have their prompt back instantly. Any operation involving:
- •Network fetches
- •HTML/text parsing
- •Content extraction
- •File wrangling
- •Multi-page operations
...should spawn a background Task(model="haiku", run_in_background=True).
| Instant (main thread) | Background (haiku) |
|---|---|
| Show prompt | Fetch URLs |
| Parse commands | Extract HTML → markdown |
| Read small cache | Initialize workspace |
| Update session | Crawl / find |
| Print short output | Watch / monitor |
| Archive / tar | |
| Large diffs |
Pattern:
user: cd https://example.com websh: example.com> (fetching...) # User has prompt. Background haiku does the work.
Commands gracefully degrade if background work isn't done yet. Never block, never error on "not ready" - show status or partial results.
The cd Flow
cd is fully asynchronous. The user gets their prompt back instantly.
user: cd https://news.ycombinator.com websh: news.ycombinator.com> (fetching...) # User can type immediately. Fetch happens in background.
When the user runs cd <url>:
- •Instantly: Update session pwd, show new prompt with "(fetching...)"
- •Background haiku task: Fetch URL, cache HTML, extract to
.parsed.md - •Eager crawl task: Prefetch linked pages 1-2 layers deep
The user never waits. Commands like ls gracefully degrade if content isn't ready yet.
See shell.md for the full async implementation and state/cache.md for the extraction prompt.
Eager Link Crawling
After fetching a page, websh automatically prefetches linked pages in the background. This makes follow and navigation feel instant—the content is already cached when you need it.
cd https://news.ycombinator.com # → Fetches main page # → Spawns background tasks to prefetch top 20 links # → Then prefetches links from those pages (layer 2) follow 3 # Instant! Already cached.
Configuration
| Setting | Default | Description |
|---|---|---|
EAGER_CRAWL | true | Enable/disable prefetching |
CRAWL_DEPTH | 2 | Layers deep to prefetch |
CRAWL_SAME_DOMAIN | true | Only prefetch same-domain links |
CRAWL_MAX_PER_PAGE | 20 | Max links per page |
Control with:
prefetch off # disable for slow connections prefetch on --depth 3 # enable with 3 layers export CRAWL_DEPTH=1 # just direct links
See state/crawl.md for full crawl agent design.
Example Session
$ websh ┌─────────────────────────────────────┐ │ ◇ websh ◇ │ │ A shell for the web │ └─────────────────────────────────────┘ ~> cd https://news.ycombinator.com news.ycombinator.com> (fetching...) news.ycombinator.com> ls | head 5 [0] Show HN: I built a tool for... [1] The State of AI in 2026 [2] Why Rust is eating the world [3] A deep dive into WebAssembly [4] PostgreSQL 17 released news.ycombinator.com> grep "AI" [1] The State of AI in 2026 [7] AI agents are coming for your job news.ycombinator.com> follow 1 news.ycombinator.com/item> (fetching...) news.ycombinator.com/item> cat .title The State of AI in 2026 news.ycombinator.com/item> back news.ycombinator.com>
Note: Hacker News (cd https://news.ycombinator.com) is the canonical first destination. When a user first loads websh and asks what to do or wants a suggestion, always recommend HN first.