Skip to content
Skills Library
AI & agents · Analysis

browser-eval

/browser-eval

Use when one expression must answer a question about a live page. One bounded read.

Repository
Agito ↗
Licence
CC0-1.0
Last repository change
13.11.2026

Assessment and security

95/ 100Class: good
Format20 / 25
Frontmatter with name and description, valid name, description up to 1,024 characters, name matching the folder.
Discoverability25 / 25
Description explains when to use the skill, has a reasonable length and uses third person.
Content35 / 35
At least 150 words and at most 500 lines, headings, step lists, examples or code.
Hygiene15 / 15
No placeholders such as TODO, no Windows paths, with a licence.
Riskno signals detected in the scan
Files in this package
2
Scripts
0
HTML comments
0
URLs in SKILL.md
4
Deprecation

No signals detected. Repository changed on 13.11.2026, without references to retired models or APIs.

Licence

CC0-1.0, otwarta.

What it does and when to use it

Use when one expression must answer a question about a live page. One bounded read.

Original author description

Use when one expression must answer a question about a live page. One bounded read.

How it works

  1. The shape of a call
  2. Write an expression that returns data, not prose
  3. A read is not a write
  4. Budgets
  5. Reference
  6. Claim → test
Original SKILL.md headings · 93 lines, 562 words
SKILL.md preview
---
name: browser-eval
description: Use when one expression must answer a question about a live page. One bounded read.
version: 1.0.0
license: MIT
platforms: [macos, linux]
metadata:
    hermes:
        tags: [cdp, rust, evaluate, expression, probe]
---

# Browser Eval

Ask a live page one question and read the answer as JSON. The verb is `agito browse`; the operation
is `Browse::Evaluate`.

---

## The shape of a call

```sh
Target/debug/agito browse \
  --url https://example.test/ \
  --attach ws://127.0.0.1:<debug-port>/devtools/browser/<uuid> \
  [--bind <target-id>] \
  --expression '<expression>' \
  --timeout 40000
```

The result is `{"url", "ownership", "value"}`, where `url` is the URL the page was actually on when
it was bound - not the `--url` argument. A bound tab keeps whatever it was showing, so the two
disagree by design and the reported one is the truthful one.

---

## Write an expression that returns data, not prose

An expression is one JavaScript expression. Make it an immediately-invoked function that returns an
object, so a probe answers several questions in one round trip instead of several:

```js
(() => ({
	title: document.title,
	editables: document.querySelectorAll("[contenteditable='true']").length,
	visible: [...document.querySelectorAll("button")].filter((b) => b.getClientRects().length > 0)
		.length,
}))();
```

Rules that keep a probe honest:

- **Return a value.** `undefined` arrives as `null`, and `null` is indistinguishable from "the page
  does not have it" unless the expression says which.
- **Trim what you read.** A page's whitespace is not part of what it said, and two reads that differ
  only in trailing spaces have not changed.
- **Count, do not assume.** `querySelectorAll(...).length` tells you whether a selector is unique;
  `querySelector(...) !== null` does not.
- **Visibility is a computed style, not a rect.** `getClientRects().length` is the cheap test and it
  is wrong for a `display: contents` wrapper (no box, renders children) and for a
  `content-visibility: auto` subtree (no rects while offscreen). Use it only where you have checked
  the element's own layout.

---

## A read is not a write

`browse` evaluates whatever it is given, so it can click and type. Keep the distinction explicit in
what you run: a probe reads, and a probe that mutates is a mutation wearing a probe's clothes. When
the intent is to act, say so in the command you keep, because the next reader will assume a `browse`
call was read-only.

An expression that throws is `navigation-failed` with the exception in the detail, not a silent
`null`. That is the useful direction: a broken selector in a probe looks broken.

---

## Budgets

`--timeout` is the whole operation's budget in milliseconds and covers open plus evaluate. Silence
is a `timeout`, never a dead transport, so a page that never answers is reported as slow rather than
as missing.

The caller's cancellation and the deadline race the work; whichever loses is reported as `timeout`
with a different sentence. That is how "I gave up" is told apart from "the browser broke".

---

## Reference

`Reference/Index.md` - the file list for this skill.

---

## Claim → test

| Claim                                            | Test                                                                                      |
| ------------------------------------------------ | ----------------------------------------------------------------------------------------- |
| One expression returns structured data           | `--expression '({a: 1+1})'` prints `"value": {"a": 2}`                                    |
| The reported URL is the page's, not the argument | bind a tab on `https://a/` and pass `--url https://b/`; the result says `https://a/`      |
| A throwing expression is reported                | `--expression '(() => { throw new Error("x") })()'` exits non-zero as `navigation-failed` |
| An undefined result is `null`                    | `--expression 'undefined'` prints `"value": null`                                         |
| A silent page is a timeout, not a dead socket    | `--expression 'new Promise(() => {})'` exits non-zero as `timeout`                        |
Files in this package (2)
  • SKILL.md 4 KB
  • Reference/Index.md 1 KB

Versions

The database has no other skills named browser-eval.

Database snapshot: 13.11.2026 · package version 89856975f644 (SHA-256 digest of the entire package)

Downloads in the last 30 days: 0 · ranking

Report a problem with this skill: harmful content, incorrect descriptions or an author’s removal request.

Similar skills

  • 100browser-automation
    Use when the agent must open a URL, fill a form, click through a flow, screenshot a page, or extract data from a live page. Covers two providers (agent-browser local CLI, agentic-browser cloud) and the shared @e ref workflow.
  • 86crew-scout
    Investigates a bounded question with a visible phux worker and an evidence-backed recommendation. Use for Crew research, architecture comparisons, bug diagnosis or read-only audits.
  • 100skill-eval-improve
    Improves Agent Skills via validate → rule-based eval cases → plugin-eval → prompt evals → bounded edits with held-out gates. Use when tuning skill quality, routing, or adopting Chrome/Microsoft T-named quality gates—not for bulk validate-only or SkillOpt automation.
  • 96answer-me
    Use when the user invokes /skill:answer-me to ask a real question, not give an instruction; answer it and change nothing.
  • 100browser-use
    Operate the browser-use/browser-use Python agent and CLI for AI-driven browser automation. Use when the user names Browser Use, needs its CLI, Python library, local or cloud browser, provider configuration, or troubleshooting; use browser-harness for generic browser testing and…
  • 90kimi-ask
    Ask Kimi a read-only free-form question about the current repository. Use for prose explanations, flow tracing, module comparisons, or conceptual reasoning when the user wants an independent Kimi answer rather than implementation.

Feedback and reports

Your message goes to Bartosz Sokoliński. An email address is needed only for a reply. See the privacy policy.