Opencli Browser

opencli-browser

The first reader of this CLI is an agent, not a human. Every subcommand returns a structured envelope that tells you exactly what matched, how confident the match is, and what to do if it didn't. Lean on those envelopes — do not guess.

This skill is for driving a live browser to accomplish an agent task. If you are building a reusable adapter under ~/.opencli/clis/<site>/ use opencli-adapter-author instead.

Prerequisites

opencli doctor

Until doctor is green, nothing else will work. Typical failures: Chrome not running, extension not installed, debug port blocked by 1Password / other extensions. The doctor output tells you which.

Window lifecycle

• opencli browser * commands already keep the automation session alive between calls. The window stays open until you run opencli browser close or the idle timeout expires.
• --focus (or OPENCLI_WINDOW_FOCUSED=1) opens the automation window in the foreground. Use it when you want to watch the page live.
• --live (or OPENCLI_LIVE=1) is mainly for browser-backed adapter commands such as opencli xiaohongshu note .... It keeps the adapter's automation window open after the command returns so you can inspect the final page state.

Mental model

1. Selector-first target contract. Every interaction command (click, type, select, get text/value/attributes) takes one <target>, which is either a numeric ref from state/find or a CSS selector. Use --nth <n> to disambiguate multiple CSS matches.
2. Every envelope reports matches_n and match_level. match_level is exact, stable, or reidentified — the CLI already rescued moderate DOM drift for you, but the level tells you how confident to be.
3. Compact output first, full payload on demand. state is a budget-aware snapshot; get html --as json supports --depth/--children-max/--text-max; network returns shape previews and you re-fetch a single body with --detail <key>. If you emit a giant payload you are burning context you did not need to burn.
4. Structured errors are machine-readable. On failure the CLI emits {error: {code, message, hint?, candidates?}}. Branch on code, not on message strings.

Critical rules

1. Always inspect before you act. Run state or find first. Never hard-code a ref or selector from memory across sessions — indices are per-snapshot.
2. Prefer numeric ref over CSS once you have it. Numeric refs survive mild DOM shifts because the CLI fingerprints each tagged element. A CSS selector written by hand will break the first time the site re-renders.
3. Read match_level after every write. exact = all good. stable = the element is the same but some soft attrs drifted — your action still applied. reidentified = the original ref was gone and the CLI found a unique replacement; double-check you hit the right element.
4. Use the compound field for form controls. Do not regex-guess a date format, do not state twice to get the full <select> options list. The compound envelope has the format string, full option list up to 50, options_total for overflow, and accept/multiple for <input type=file>.
5. Verify writes that matter. After type <target> <text>, run get value <target>. After select, run get value. Autocomplete widgets, React controlled inputs, and masked fields all silently eat characters. The CLI cannot detect this for you.
6. state → action → state after a page change. Navigations, form submits, and SPA route changes invalidate refs. Take a fresh snapshot. Do not reuse refs from before the transition.
7. Chain with &&. A chained sequence runs in one shell so refs acquired by the first command stay live for the second. Separate shell invocations lose the session context you just set up.
8. eval is read-only. Wrap the JS in an IIFE and return JSON. If you need to change the page, use the structured click / type / select / keys commands instead — they produce structured output and fingerprints, eval does not.
9. Prefer network to screen-scraping. If a page you care about fetches its data from a JSON API, the API is almost always more reliable than scraping the rendered DOM. Capture once, inspect the shape, then --detail <key> the body you need.

Target contract (<target> for click / type / select / get text|value|attributes)

<target> ::= <numeric-ref> | <css-selector>

• Numeric ref — the [N] index from state or find. Cheap, resilient to soft DOM drift.
• CSS selector — anything querySelectorAll accepts. Must be unambiguous on write ops, or pair with --nth <n>.

Envelope on success

{ "clicked": true, "target": "3", "matches_n": 1, "match_level": "exact" }

{ "value": "kalevin@example.com", "matches_n": 1, "match_level": "stable" }

match_level

Structured error codes

Branch on these, not on the human message:

Error envelope always includes error.code and error.message. Target errors (selector_not_found, selector_ambiguous, etc.) often add error.candidates: string[] with suggested selectors. option_not_found adds error.available: string[] instead.

Command reference

Inspect

Get (read-only)

Interact

Wait

browser wait selector "<css>" [--timeout ms]    # wait until the selector matches
browser wait text "<substring>" [--timeout ms]  # wait until the text appears
browser wait time <seconds>                     # hard sleep, last resort

Default timeout 10000 ms. SPA routes, login redirects, and lazy-loaded lists need wait before state/get.

Extract

• browser eval <js> [--frame N] — Run an expression in the page (or in a cross-origin frame via --frame). Wrap in an IIFE and return JSON. Read-only: no document.forms[0].submit(), no clicks, no navigations. If the result is a string, stdout is the raw string; otherwise it's JSON.
• browser extract [--selector <css>] [--chunk-size N] [--start N] — Markdown extraction of long-form content with a continuation cursor. Returns {url, title, selector, total_chars, chunk_size, start, end, next_start_char, content}. Loop on next_start_char until it is null. Auto-scopes to <main>/<article>/<body> if you don't pass --selector.

Network

browser network                        # shape preview + cache key list
browser network --detail <key>         # full body for one cached entry
browser network --filter "field1,field2"  # keep only entries whose body shape contains ALL fields as path segments
browser network --all                  # include static resources (usually noise)
browser network --raw                  # full bodies inline — large; use sparingly
browser network --ttl <ms>             # cache TTL (default 24h)

List entries look like {key, method, status, url, ct, size, shape, body_truncated?}. Detail envelope is {key, url, method, status, ct, size, shape, body, body_truncated?, body_full_size?, body_truncation_reason}. Cache lives in ~/.opencli/cache/browser-network/ so you can re-inspect without re-triggering the request.

Tabs & session

Compound form controls

Every date/time, select, and file input carries a compound field. Use it — do not regex attributes.

Date family

{
  "control": "date",
  "format": "YYYY-MM-DD",
  "current": "2026-04-21",
  "min": "2026-01-01",
  "max": "2026-12-31"
}

control is one of date | time | datetime-local | month | week. format is a concrete template string — type into the field using that exact format, or select by label if the site wraps the native input in a custom widget.

Select

{
  "control": "select",
  "multiple": false,
  "current": "United States",
  "options": [
    { "label": "United States", "value": "us", "selected": true },
    { "label": "Canada", "value": "ca" }
  ],
  "options_total": 137
}

options[] is capped at 50 entries. current is always correct even when the selected option is past the cap — it's computed by scanning every option, not from the truncated list. If options_total > options.length and you need an option that isn't in options[], call browser select <target> "<label>" directly — the CLI matches against the live DOM, not the truncated list.

File

{
  "control": "file",
  "multiple": true,
  "current": ["report.pdf", "cover.png"],
  "accept": "application/pdf,image/*"
}

Do not invent file paths. Upload is done via the normal click flow — respect accept when telling the user what to upload.

Where compounds show up

• browser find --css <sel> entries: inline on each match.
• browser get html --as json tree nodes: inline on matching nodes.
• browser state snapshot: in a compounds (N): sidecar keyed by numeric ref, so you can tell at a glance which [N] entries have rich metadata.

Cost guide

Think about payload size per call. Budgets exist for a reason.

Rule of thumb: one state per page transition, one find per follow-up query, one get/click/type per action. If your plan involves >10 calls per page you are probably scraping instead of interacting — consider extract or network.

Chaining rules

Good — one shell, live session:

opencli browser open "https://news.ycombinator.com" \
  && opencli browser state \
  && opencli browser click 3

Bad — each line is a fresh shell, refs from call 1 are already forgotten when call 2 runs. (Only a problem if you rely on shell-scoped state; browser refs themselves persist in-page, but interleaving unrelated shells invites races.) Prefer && when the steps are meant to be atomic.

Never chain a write and then an immediate state without a wait if the action causes a network round-trip — you will snapshot the pre-response DOM and make bad decisions off stale data.

Recipes

Fill a login form

opencli browser open "https://example.com/login"
opencli browser state                          # find [N] for email, password, submit
opencli browser type 4 "me@example.com"
opencli browser type 5 "hunter2"
opencli browser get value 4                    # verify (autocomplete can eat chars)
opencli browser click 6                        # submit
opencli browser wait selector "[data-testid=account-menu]" --timeout 15000
opencli browser state                          # fresh refs on the logged-in page

Pick from a long dropdown

opencli browser state                          # sidebar shows [12] <select name=country>
opencli browser find --css "select[name=country]"
# the compound.options_total is 137, but compound.current is "" — unselected.
opencli browser select 12 "Uruguay"
opencli browser get value 12                   # { value: "uy", match_level: "exact" }

Scrape a list via network instead of DOM

opencli browser open "https://news.ycombinator.com"
opencli browser network --filter "title,score"
# -> find the /topstories entry, note its key
opencli browser network --detail topstories-a1b2

Read a long article in chunks

opencli browser open "https://blog.example.com/long-post"
opencli browser extract --chunk-size 8000
# -> content + next_start_char: 8000
opencli browser extract --start 8000 --chunk-size 8000
# ...until next_start_char is null

Cross-origin iframe

opencli browser frames
# -> [{"index": 0, "url": "https://checkout.stripe.com/...", ...}]
opencli browser eval "(() => document.querySelector('input[name=cardnumber]')?.value)()" --frame 0

Pitfalls

• Do not submit forms via eval "document.forms[0].submit()" — modern sites intercept with JS handlers and silently drop the call. Either click the submit button via its ref, or (if you know the GET URL) just open it directly.
• Do not reuse refs across a page transition. wait for the new state, then re-state. Old refs will either 404 or (worse) reidentify onto a similarly-shaped element on the new page.
• match_level: reidentified is a warning, not an error. The action went through, but if you are chaining 5 more writes that all depend on that being the right element, verify with a get text or get value before continuing.
• Budget-aware commands silently cap. get html --as json with default budgets will return truncated: {...}. If your downstream logic needs the whole subtree, raise --depth / --children-max or tighten the selector.
• autocomplete: true on a type response is not an error. It means a suggestion popup is open and your value isn't committed yet. Typically keys Enter to accept the first suggestion, or click the one you want.
• network --filter is AND-semantics on path segments. --filter "title,score" keeps entries whose body shape contains both title and score as path segments, at any depth. It is not a regex.
• Screenshots are for humans, not for agents. Use state + find unless the page is genuinely visual (captcha, chart). Screenshots burn tokens and rarely add signal an agent can act on.

Troubleshooting

See also

• opencli-adapter-author — turning what you just figured out into a reusable ~/.opencli/clis/<site>/<command>.js.
• opencli-autofix — when an existing adapter breaks, this skill walks you through OPENCLI_DIAGNOSTIC and filing a fix.