Skip to content

SKILL.md

Latest commit

 

History

History
434 lines (325 loc) · 28.3 KB

SKILL.md

File metadata and controls

434 lines (325 loc) · 28.3 KB
name opencli-browser
description Use when an agent needs to drive a real Chrome window via opencli — inspect a page, fill forms, click through logged-in flows, or extract data ad-hoc. Covers the selector-first target contract, compound form fields, stale-ref handling, network capture, and the agent-native envelopes the CLI returns. Not for writing adapters — see opencli-adapter-author for that.
allowed-tools Bash(opencli:*), Read, Edit, Write

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.

Session lifecycle

  • opencli browser * commands require --session <name>. Use the same session name for a multi-step flow; use a different name to isolate parallel browser work.
  • Owned browser sessions keep a tab lease alive between calls. Release it with opencli browser --session <name> close or let the idle timeout expire.
  • opencli browser bind --session <name> binds the Chrome tab you already have open to that session. Use this for logged-in pages, SSO flows, or pages you manually positioned before handing control to the agent.
  • --window foreground|background (or OPENCLI_WINDOW=foreground|background) chooses whether OpenCLI creates/focuses a foreground browser window or uses a background browser window for owned sessions.

Bind Tab

opencli browser bind --session gmail
opencli browser --session gmail state
opencli browser --session gmail click "Search"
opencli browser --session gmail network
opencli browser unbind --session gmail

Binding never owns the user window and never closes the user tab. It fails closed if the tab is closed or becomes non-debuggable. Re-run bind --session <name> when you switch to a different real tab.

Navigation is allowed on bound sessions because the session now represents explicit agent ownership of that tab. Tab mutation (tab new, tab select, tab close) is still blocked for bound sessions. Use an owned session when you want OpenCLI to manage tab lifecycle.

opencli browser sessions returns idleMsRemaining: null for bound sessions. That means there is no OpenCLI idle-close timer; the binding lasts until unbind, tab close, window close, or daemon restart.

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

level meaning you should
exact Fingerprint agreed on tag + strong IDs with at most one soft drift Proceed.
stable Tag + strong IDs still agree, soft signals (aria-label, role, text) drifted Proceed, but if what you typed/clicked matters, re-check with get value or state.
reidentified Original ref was gone; a unique live element matched the fingerprint and was re-tagged with the old ref Double-check you hit the right element before chaining more writes.

Structured error codes

Branch on these, not on the human message:

code meaning
not_found Numeric ref is no longer in the DOM. Re-state.
stale_ref Ref exists but the element at that ref changed identity. Re-state.
invalid_selector CSS was rejected by querySelectorAll. Fix the selector.
selector_not_found CSS matches 0 elements. Try find with a looser selector.
selector_ambiguous CSS matches >1 and no --nth. Add --nth or narrow the selector.
selector_nth_out_of_range --nth beyond match count.
option_not_found select couldn't find an option matching that label/value. Error envelope includes available: string[] of the real option labels.
not_a_select select was called on a non-<select> element.

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

command purpose
browser state Snapshot: text tree with [N] refs, scroll hints, hidden-interactive hints, compounds (N): sidecar for date/select/file refs.
browser state --source ax Opt-in accessibility-tree snapshot. Use when custom controls, portals, or iframe contents are hard to identify in normal state. AX refs can recover stale React re-renders by role/name/nth and can route same-origin iframe refs. Cross-origin iframe refs are best-effort because Chrome may not expose attachable OOPIF targets to extensions.
browser state --compare-sources Metrics-only DOM vs AX comparison for deciding whether AX should become default. It prints counts and sizes, not page text, so it is safer to share for validation.
browser find --css <sel> [--limit N] [--text-max N] Run a CSS query and return one entry per match with {nth, ref, tag, role, text, attrs, visible, compound?}. Allocates refs for matches the prior snapshot didn't tag. Cheap alternative to state when you already know the selector.
browser find --role button --name Save Semantic locator query. Also supports --label, --text, and --testid. Use before raw CSS when a control has accessible labels.
browser frames List cross-origin iframe targets. Pass the index to --frame on eval.
browser screenshot [path] Viewport PNG. No path → base64 to stdout. Prefer state when you just need structure.
browser screenshot --annotate [path] Visual ref map. Refreshes DOM refs and overlays visible [N] labels so the screenshot maps back to browser click <ref> targets. Use for icon-only controls, visual layouts, charts, or when text state is ambiguous.

Get (read-only)

command returns
browser get title plain text
browser get url plain text
browser get text <target> [--nth N] {value, matches_n, match_level}
browser get value <target> [--nth N] {value, matches_n, match_level}
browser get attributes <target> [--nth N] {value: {attr: val, ...}, matches_n, match_level}
browser get text --role option --name Travel Semantic locator read without a prior state call. Same flags as browser find.
browser get html [--selector <css>] [--as html|json] [--depth N] [--children-max N] [--text-max N] [--max N] Raw HTML, or structured tree. JSON tree nodes have {tag, attrs, text, children[], compound?}. Truncation reported via truncated: {depth?, children_dropped?, text_truncated?}.

Interact

command notes
browser click <target> [--nth N] Returns {clicked, target, matches_n, match_level}.
browser click --role button --name Submit Semantic click. Write actions require a unique match; ambiguous locators return candidates instead of clicking the first match.
browser hover [target] [--role R --name N] [--nth N] Moves the mouse over an element. Use for hover menus/tooltips before taking state or clicking submenu items. Returns {hovered, target, matches_n, match_level}.
browser focus [target] [--role R --name N] [--nth N] Focuses an element without typing. Useful before keys or when a page reacts to focus/blur. Returns {focused, target, matches_n, match_level}.
browser dblclick [target] [--role R --name N] [--nth N] Double-clicks an element via native mouse events when available. Returns {dblclicked, target, matches_n, match_level}.
browser check [target] [--role R --name N] [--nth N] Ensures checkbox/radio/aria-checked control is checked. Returns {checked, changed, target, matches_n, match_level, kind}. Prefer this over blind click when target state matters.
browser uncheck [target] [--role R --name N] [--nth N] Ensures checkbox/aria-checked control is unchecked. Radio buttons cannot be unchecked directly; select another radio in the group instead.
browser upload [target] <file...> [--role R --name N] [--nth N] Attaches local file path(s) to an input[type=file] via CDP. With semantic flags, omit target and pass files as positionals. Returns {uploaded, files, file_names, target, matches_n, match_level, multiple?, accept?}.
browser drag [source] [target] [--from-role R --from-name N] [--to-role R --to-name N] [--from-nth N] [--to-nth N] Mouse-based drag from one resolved element center to another. Works for mouse-listener drag libraries; native HTML5 dataTransfer drops may need a site-specific fallback. Returns {dragged, source, target, source_matches_n, target_matches_n, ...}.
browser type [target] <text> [--role R --name N] [--nth N] Clicks first, then types. With semantic flags, omit target and pass text as the only positional. Returns {typed, text, target, matches_n, match_level, autocomplete}. autocomplete: true means a combobox/datalist popup appeared after typing — you almost always need keys Enter or a follow-up click on the suggestion to commit the value.
browser fill [target] <text> [--role R --name N] [--nth N] Exact replacement for input, textarea, and contenteditable targets. With semantic flags, omit target and pass text as the only positional. Returns {filled, verified, text, actual, matches_n, match_level}. Use this when you need raw text set and verified, not keyboard/autocomplete behavior. Pipeline form supports { fill: { ref, text, submit: true } }.
browser select [target] <option> [--role R --name N] [--nth N] Matches native <select> option by label first, then value. With semantic flags, omit target and pass option as the only positional. Use compound from find/state to see exactly what labels are available.
browser keys <key> Enter, Escape, Tab, Control+a, etc. Runs against the focused element.
browser scroll <direction> [--amount px] up / down. Default amount 500.

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 download [pattern] [--timeout ms]  # wait for a Chrome download whose filename/URL/mime contains pattern
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.

browser wait download requires Browser Bridge extension 1.0.8+ because it uses Chrome's downloads lifecycle API. Pass a narrow filename or URL substring such as receipt.pdf when possible; an empty pattern waits for the next/recent download in the timeout window. The command reports {downloaded, filename, url, state, elapsedMs} on success and a JSON error envelope on timeout/failure.

Extract

  • web read --url <url> — One-shot Markdown reader for arbitrary pages. It expands relevant same-origin iframes by default, so old iframe-shell sites work better than with a top-document-only scrape. Use --frames all-same-origin when completeness matters more than Markdown noise. For AJAX shell pages use opencli web read --url <url> --wait-for "<selector>" --wait-until networkidle --diagnose; diagnostics show frame URLs, empty containers, and API-like XHRs. If the value you need is table/API data, switch to browser network or a dedicated adapter instead of relying on Markdown.
  • 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.

Default output keeps JSON/XML/plain-text and JS-like API responses, then drops obvious static assets and telemetry by URL. If an expected endpoint is missing, run browser network --all once and check whether an unusual content type or URL filter hid it.

Tabs & session

command purpose
browser tab list JSON array of {index, page, url, title, active}. The page string is the tab identity you pass as <targetId> to tab select / tab close, or to --tab <targetId> on any subcommand. (--tab's placeholder is historical — the value is always page.)
browser tab new [url] Open a new tab. Prints the new page string.
browser tab select [targetId] Make a tab the default. All subcommands accept --tab <targetId> to target one without changing the default.
browser tab close [targetId] Close by page.
browser back History back on the active tab.
browser close Release the current owned browser session when done.
browser bind --session <name> Bind the current Chrome tab to a browser session.
browser unbind --session <name> Detach a bound session without closing the user tab/window.

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.

command rough cost when to use
state medium (bounded by internal budget) First call on any page, after every nav, when you need refs.
find --css <sel> small You already know the selector — one query, compact entries.
get title / get url tiny Sanity checks between steps.
get text/value/attributes tiny per call Verifying one specific field.
get html (raw) can be huge Avoid on unbounded pages. Always pair with --selector and a budget.
get html --as json --depth 3 --children-max 20 medium When you need to reason about structure, not a specific field.
screenshot large Only when the page is visual (CAPTCHA, charts). Prefer state.
extract medium per chunk Long-form reading. Loop via next_start_char.
network (default) small First look at APIs.
network --detail <key> varies Pull one body.
network --raw huge Only after --filter narrowed the candidate set.
eval "JSON.stringify(...)" controlled Targeted extraction when none of the above fit.

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

Pick from a long dropdown

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

Pick from a custom React dropdown

Use this for Radix, shadcn, Material UI, Mercury-style category fields, and other controls that are not native <select>.

opencli browser --session mercury state                          # find category trigger ref
# If the trigger/option is not clear, use AX:
opencli browser --session mercury state --source ax              # look for combobox/button/listbox/option names
opencli browser --session mercury click 7                        # click category trigger
opencli browser --session mercury state --source ax              # fresh refs after the portal/listbox opens
opencli browser --session mercury click 12                       # click option
opencli browser --session mercury get text 7                     # verify visible selected label

Do not use browser select on these widgets. browser select is only for native <select> elements. Custom dropdowns should be driven with state -> click trigger -> state -> click option -> verify.

Compare DOM vs AX observation

When deciding whether AX refs are better for a page, collect metrics without sharing page contents:

opencli browser --session compare state --compare-sources

Report sources.dom.refs, sources.ax.refs, frame_sections, approx_tokens, elapsed_ms, and any per-source error. Use this before arguing that AX should become the default on a site.

Scrape a list via network instead of DOM

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

Read a long article in chunks

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

Cross-origin iframe

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

browser state --source ax may omit cross-origin iframe contents or fail to route actions into them when Chrome does not expose an attachable OOPIF target to the extension. In that case use browser frames + browser eval --frame, a normal DOM state, or navigate/bind directly to the iframe URL when possible.

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

symptom fix
opencli doctor red: "Browser not connected" Start Chrome with --remote-debugging-port=9222, or install the extension from the Chrome Web Store.
attach failed: chrome-extension://... Disable 1Password / other CDP-hungry extensions temporarily.
selector_not_found right after state Page mutated. wait selector "..." then retry.
stale_ref across every command You are reusing refs from a prior page. Re-state.
click succeeds but nothing happens The element is probably a decorative wrapper stealing clicks from the real target. find --css "..." with a narrower selector and retry on the inner element.
type appears to finish but value is wrong Autocomplete, masked input, or React controlled re-render. Verify with get value. Add keys Enter or re-type.
Giant get html output Pass --selector + --as json --depth 3 --children-max 20 --text-max 200.
Network cache seems stale Bump --ttl down, or let it expire. The cache lives at ~/.opencli/cache/browser-network/.

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 --trace retain-on-failure evidence and filing a fix.